Modelling CO2 Corrosion of Pipeline Steels
Modelling CO2 Corrosion of Pipeline Steels
Modelling CO2 Corrosion of Pipeline Steels
November, 2016
Abstract
Over the years, several attempts have been made by various research institutions and
petroleum companies to develop models for the prediction of CO2 corrosion in pipelines, in
order to better capture the underlying principles that cause it. Modelling CO2 corrosion is
important to the oil and gas and carbon capture and storage (CCS) industries, as it provides
the means by which the prevention of the financial costs from lost production, the
preservation of the environment as well as the health and safety of human lives can be
achieved.
In this thesis, existing models have been investigated and compared against newly derived
models in terms of their accuracy of prediction, by using an identical test dataset. A neural
network (NN) model was developed, in which a detailed sensitivity analysis was carried out
on Matlab training functions to determine their degree of suitability in CO2 corrosion
prediction. Results showed that the tansig transfer function was the most suitable and that a 2-
layer network was sufficient to obtain desirable R2-values of ~0.9 for both low and high
pressure CO2 corrosion data.
Also, a linear regression model was developed based on predictor variables: temperature (T),
CO2 partial pressure (𝑃𝐶𝑂2 ), fluid velocity (U) and pH, for both low and high pressure CO2
data. The respective R2-values obtained are 0.65 and 0.7. An R2-value of 0.8 can be achieved
for the low pressure CO2 data; however the derived regression equation is inelegant and
contains a combination of a large number of predictor terms.
From Monte Carlo analyses, the exponential and normal distributions were discovered to be
the best fits for the low and high pressure CO2 corrosion rate data, respectively. Further,
parametric sensitivity analyses revealed the pH and fluid velocity to be the least and most
significant variables for low pressure CO2, respectively, while the velocity and temperature
were the least and most significant variables for high pressure CO2 corrosion, respectively.
i
Dedication
I dedicate this thesis to my wonderful family – my mother, father and sister. I am very
grateful for all the words of advice, encouragement, support and prayers you have given me
over the years. Writing this thesis would indeed not have been possible if not for the help
you’ve all given me. Thank you so much for everything.
ii
Table of contents
Abstract ........................................................................................................................................ i
Dedication ...................................................................................................................................ii
List of Tables ...........................................................................................................................viii
List of Figures ............................................................................................................................. x
Nomenclature........................................................................................................................... xix
Greek letters ........................................................................................................................xxii
Chapter 1. Introduction ............................................................................................................... 1
1.1 Corrosion Costs .................................................................................................................... 2
1.2 Environmental Impact of CO2 Emissions............................................................................. 3
1.3 The Importance of CCS ........................................................................................................ 8
1.4 CO2 Corrosion in Petroleum and CCS pipelines .................................................................. 9
1.5 Aims and Objectives ........................................................................................................... 10
1.6 Thesis Structure and Content.............................................................................................. 13
Chapter 2. Literature Review on CO2 Corrosion...................................................................... 15
2.1 CO2 Corrosion - Introduction ............................................................................................. 16
2.2 An Outline of the CO2 Corrosion Process .......................................................................... 16
2.3 Types of CO2 Corrosion Damage ....................................................................................... 19
2.3.1 Pitting ........................................................................................................................... 19
2.3.2 Mesa-type Attack ......................................................................................................... 20
2.3.3 Flow-induced Localised Corrosion .............................................................................. 21
2.4 Major Factors Influencing CO2 Corrosion ......................................................................... 22
2.4.1 The Effect of Environmental Factors on CO2 Corrosion ............................................. 22
Effect of Solution Chemistry............................................................................................. 22
Effect of Temperature and CO2 Partial Pressure and Fluid Flowrate Velocities .............. 23
CO2 Corrosion Prediction – Corrosion Rate-Determining Parameters ............................. 24
Presence of H2S ................................................................................................................. 27
Determination of pH for a Mixed CO2/H2S-System ......................................................... 28
Conditions for H2S Corrosion Risks and H2S Corrosion Mechanism .............................. 29
Mechanics and Kinetics of Iron Sulphide Film/Scale Formation ..................................... 31
Species Transport through Boundary Layers .................................................................... 32
2.4.2 The Effect of Physical Factors on CO2 Corrosion ....................................................... 34
Effect of Crude oil ............................................................................................................. 34
Effect of Flow and Erosion ............................................................................................... 35
Effect of Organic Acid ...................................................................................................... 36
iii
2.4.3 The Effect of Metallurgical Factors on CO2 Corrosion .............................................. 36
Further Metallurgical Considerations - Alloying Elements.............................................. 37
The Influence of Chromium ............................................................................................. 38
2.5 Established CO2 Corrosion Models ................................................................................... 39
2.5.1 Hydrocor Model .......................................................................................................... 40
2.5.2 Corplus Model ............................................................................................................. 40
2.5.3 KSC Model .................................................................................................................. 41
2.5.4 Multicorp Model.......................................................................................................... 41
2.6 CO2 Corrosion Mechanism between Low and High Partial Pressures .............................. 42
2.6.1 Carbon Capture and Storage (CCS) ............................................................................ 42
2.6.2 Research Directions ..................................................................................................... 43
2.6.3 The Limitation of Henry’s law for High Pressure Conditions .................................... 44
2.6.4 Methodology for CO2 Corrosion Prediction at High Pressure Conditions ................. 45
2.7 Neural Networks ................................................................................................................ 47
2.7.1 Case Study on the Implementation of Neural Networks in Model Prediction ............ 49
2.8 Fuzzy Logic Systems ......................................................................................................... 51
2.8.1 Case Study on the Execution of the Fuzzy Logic System ........................................... 52
2.9 Conclusions ........................................................................................................................ 54
Chapter 3. CO2 Corrosion Models ........................................................................................... 57
3.1 Introduction – CO2 Corrosion Models ............................................................................... 58
3.2 The 1975 de Waard-Milliams Correlation ......................................................................... 59
3.3 The 1991 de Waard-Lotz-Milliams Correlation ............................................................... 61
3.3.1 Accounting for High-Temperature Protective Films................................................... 62
3.4 The 1995 de Waard-Lotz-Dugstad correlation ................................................................. 64
3.4.1 Parameter Study on the 1995 de Waard-Lotz-Dugstad Correlation – Investigating
Changes in Flow Velocity and CO2 Fugacity on CO2 Corrosion Rate ................................ 66
3.5 The NORSOK M-506 Model ............................................................................................ 68
3.5.1 Matlab Model Results based on NORSOK Correlations ............................................ 70
3.5.2 Parameter Study on the NORSOK M-506 Model – Investigating Changes in Water
Cut Percentage and CO2 Mole Percentage on CO2 Corrosion Rate ..................................... 71
3.6 The Nesic-Postlethwaite-Olsen Model (NPO Model) ....................................................... 72
3.6.1 Cathodic Reactions ...................................................................................................... 73
3.6.2 Anodic Reactions ........................................................................................................ 74
3.7 Introduction to Freecorp..................................................................................................... 76
iv
3.7.1 Organic Acid Corrosion ............................................................................................... 76
3.7.2 Hydrogen Sulphide Corrosion ..................................................................................... 78
3.7.3 CO2 Corrosion Rate Trends and Parametric Study on Velocity of Fluid Flow ........... 81
3.8 Conclusions ........................................................................................................................ 84
Chapter 4. Statistical Analysis .................................................................................................. 87
4.1 Introduction ........................................................................................................................ 88
4.2 Low Pressure CO2 Corrosion Data ..................................................................................... 88
4.2.1 Descriptive Statistics.................................................................................................... 88
4.2.2 Principal Component Analysis .................................................................................... 91
4.2.3 Multiple Linear Regression ......................................................................................... 94
Variable Interactions ......................................................................................................... 97
4.2.4 Nonlinear Regression ................................................................................................. 100
4.2.5 Response Surface Plots .............................................................................................. 102
4.3 High Pressure CO2 Corrosion Data .................................................................................. 104
4.3.1 Descriptive Statistics.................................................................................................. 104
4.3.2 Principal Component Analysis .................................................................................. 106
4.3.3 Multiple Linear Regression ....................................................................................... 108
Variable Interactions ....................................................................................................... 110
4.3.4 Nonlinear Regression ................................................................................................. 114
4.3.5 Response Surface Plots .............................................................................................. 115
4.4 Conclusions ...................................................................................................................... 117
Chapter 5. Neural Network Models ........................................................................................ 119
5.1 Introduction – Historical Background .............................................................................. 120
5.2 Definition of Neural Networks ......................................................................................... 122
5.3 The Simple Neuron and Neuron with Vector Input ......................................................... 123
5.4 The Transfer Function ...................................................................................................... 124
5.5 Neural Network Architectures .......................................................................................... 125
5.5.1 Single-layer Feedforward Networks .......................................................................... 125
5.5.2 Multilayered Feedforward Networks ......................................................................... 125
5.5.3 Recurrent Networks ................................................................................................... 126
5.6 Design of the NN Model for Predicting CO2 Corrosion .................................................. 127
5.6.1 NN Modelling of Low Pressure CO2 Corrosion Data ............................................... 132
Summary of Training and Testing data for NN Modelling ............................................. 132
Results ............................................................................................................................. 132
v
Discussion of the Final Neural Network Model Specifications ..................................... 136
5.6.2 NN Modelling of High Pressure CO2 Corrosion Rate Data ...................................... 138
Summary of Training and Testing data for NN Modelling ............................................ 138
Results............................................................................................................................. 141
Discussion of the Final NN Model Specifications ......................................................... 144
5.7 Conclusions ...................................................................................................................... 148
Chapter 6. Fuzzy Inference Systems ...................................................................................... 151
6.1 Introduction – Fuzzy Logic .............................................................................................. 152
6.2 Definition of Fuzzy Logic ................................................................................................ 153
6.3 Membership Functions ..................................................................................................... 154
6.4 Mamdani and Sugeno Fuzzy Inference Systems ............................................................. 156
6.5 Low Pressure CO2 Corrosion Prediction using a Fuzzy Inference System ..................... 157
6.5.1 Methodology ............................................................................................................. 157
6.6 High Pressure CO2 Corrosion Prediction using a Fuzzy Inference System ..................... 165
6.6.1 Methodology ............................................................................................................. 165
6.7 Conclusions ...................................................................................................................... 172
Chapter 7. Monte Carlo Simulation ....................................................................................... 175
7.1 Introduction ...................................................................................................................... 176
7.2 Probability Distributions .................................................................................................. 177
7.3 Types of Probability Distributions ................................................................................... 177
7.3.1 Discrete Probability Distributions ............................................................................. 177
7.3.2 Continuous Probability Distributions ........................................................................ 177
7.4 Monte Carlo Simulation of Low Pressure CO2 Corrosion Dataset .................................. 179
7.4.1 Experimental Dataset ................................................................................................ 179
7.4.2 Monte Carlo Dataset .................................................................................................. 182
Case Study I - Based on the Derived Regression Equation ............................................ 185
Case Study II - Based on the Derived Regression Equation .......................................... 187
Sensitivity Analysis ........................................................................................................ 190
7.5 Monte Carlo Simulation of High Pressure CO2 Corrosion Dataset ................................. 191
7.5.1 Experimental Dataset ................................................................................................ 191
7.5.2 Monte Carlo Dataset .................................................................................................. 193
Case Study I - Based on the Derived Regression Equation ............................................ 195
Case Study II - Based on the Derived Regression Equation .......................................... 197
Sensitivity Analysis ........................................................................................................ 200
7.6 Optimisation of Results.................................................................................................... 201
vi
7.7 Conclusions ...................................................................................................................... 203
Chapter 8. Extended Discussion ............................................................................................. 207
8.1 Introduction ...................................................................................................................... 208
8.2 Model Performances of the 1991, 1995 De Waard, Norsok and Freecorp ...................... 208
8.2.1 Low Pressure CO2 Corrosion Prediction Results....................................................... 208
The 1991 De Waard Model ............................................................................................. 208
The 1995 De Waard Model ............................................................................................. 209
Norsok Model.................................................................................................................. 210
Freecorp Model ............................................................................................................... 211
8.2.2 High Pressure CO2 Corrosion Prediction Results ...................................................... 212
The 1991 De Waard Model ............................................................................................. 212
The 1995 De Waard Model ............................................................................................. 213
Norsok Model.................................................................................................................. 214
8.3 Applicability and Limitations of the Various Derived Models ........................................ 215
8.4 Concluding Remarks on the Performances of each of the Models................................... 218
Chapter 9. Conclusions and Recommendations for Further Work ......................................... 221
9.1 Conclusions ...................................................................................................................... 222
9.2 Recommendations for Further Work ................................................................................ 225
Appendices ............................................................................................................................. 226
A1.1 Project Gantt Chart .................................................................................................... 226
A1.2 Microsoft Excel Corrosion Database - Sample .......................................................... 227
A4.1 Statistical Student t-table ........................................................................................... 228
A4.2 Three-dimensional Plot of Corrosion Rate against CO2 Partial Pressure and
Temperature for the Low Pressure CO2 Corrosion Dataset from Alternate Viewing Angles
............................................................................................................................................ 229
A6.1a Three-dimensional Mamdani FIS Plot of Corrosion Rate against Temperature and
CO2 partial pressure for the Low Pressure CO2 Dataset (Centroid Defuzzification) ......... 230
A6.1b Three-dimensional Mamdani FIS Plot of Corrosion Rate against Temperature and
CO2 partial pressure for the Low Pressure CO2 Dataset (Bisector Defuzzification) .......... 230
A6.2 Mamdani FIS Plot of Corrosion Rate against Temperature for the High Pressure CO2
Dataset ................................................................................................................................ 231
Developed Model Code ...................................................................................................... 232
Regression Model Code .................................................................................................. 232
Neural Network Model Code .......................................................................................... 258
Monte Carlo Simulation Model Code ............................................................................. 271
References .............................................................................................................................. 295
vii
List of tables
Table 1-1. Share of corrosion costs for the U.S. industry categories (Koch et al., 2002).......... 2
Table 1-2. RCP scenarios for the end of the 21st century (IPCC, 2014) .................................... 7
Table 1-3. Weight percentage chemical composition of St.52 carbon steel (Dugstad et al.,
1994b) ...................................................................................................................................... 10
Table 1-4. Weight percentage chemical composition of X65 carbon steel (Zhang et al., 2012a)
.................................................................................................................................................. 11
Table 2-1. Estimated corrosion rates of carbon-manganese steels for some fields in the
Norwegian sector of the North Sea (Dugstad et al., 1994a) .................................................... 25
Table 3-1. Summary of variables and their respective magnitudes required to obtain results in
figure 3-7 .................................................................................................................................. 70
Table 3-2. Scaling temperature maxima for the various models at 𝑓𝐶𝑂2 =0.3MPa ................... 84
Table 4-1. Descriptive statistics of the CO2 corrosion sample data ......................................... 89
Table 4-2. Test data variables and their corresponding ranges ................................................ 91
Table 4-3. Principal component analysis ................................................................................. 91
Table 4-4. Regression model terms and statistics .................................................................... 94
Table 4-5. Descriptive statistics of the CO2 corrosion sample data ....................................... 104
Table 4-6. Test data variables and their corresponding ranges .............................................. 106
Table 4-7. Principal component analysis ............................................................................... 106
Table 4-8. Regression model terms and statistics .................................................................. 109
Table 4-9 Low pressure CO2 corrosion model performance ................................................. 118
Table 4-10 High pressure CO2 corrosion model performance ............................................... 118
Table 5-1. Definitions of the transfer functions used (Beale et al., 2014) ............................. 124
Table 5-2. Definitions of the training functions used (Beale et al., 2014; Sharma and
Venugopalan, 2014) ............................................................................................................... 131
Table 5-4. Summary of the results obtained for the tested neuron configurations ................ 136
Table 5-5. Summary of neural network properties ................................................................ 137
Table 5-6. Summary of the results obtained for the tested neuron configurations ................ 145
Table 5-7. Summary of neural network properties ................................................................ 146
Table 5-8. Model performance for each dataset .................................................................... 149
Table 6-1. Summary of FIS processes and methods .............................................................. 157
Table 6-2. Membership function descriptor and corresponding corrosion rate ranges ......... 159
Table 6-3. Some rules for the development of the Mamdani and Sugeno FIS ...................... 160
Table 6-4. Variables used in the study and their corresponding ranges ................................ 165
viii
Table 6-5. Descriptors assigned to each of the input and output variable membership functions
................................................................................................................................................ 166
Table 6-6. Rules for the development of Mamdani and Sugeno FIS ..................................... 166
Table 6-7. Magnitudes of input variables and respective output variable results .................. 167
Table 6-8. Model performance for each dataset ..................................................................... 173
Table 7-1. Probability density functions for the normal and Weibull distributions (Harrison,
2010), (Johnson et al., 1994) .................................................................................................. 178
Table 7-2. Test statistic summary for normal and Weibull Probability density functions ..... 180
Table 7-3. Probability density functions for the lognormal and exponential distributions
(Harrison, 2010; Johnson et al., 1994) ................................................................................... 181
Table 7-4. Statistical parameters for each distribution ........................................................... 182
Table 7-5. Summary statistics for the use of the Weibull distribution as a direct fit for the low
pressure CO2 corrosion rate .................................................................................................... 184
Table 7-6 Summary statistics for the exponential and lognormal probability distributions .. 184
Table 7-7. Test statistic summary for normal and Weibull Probability density functions ..... 185
Table 7-8. Summary statistics for Case Study I ..................................................................... 187
Table 7-9. Best-fit distributions for each input parameter ..................................................... 187
Table 7-10. Test statistic summary for normal and Weibull probability density functions ... 188
Table 7-11. Summary statistics for Case Study II .................................................................. 190
Table 7-12. Test statistic summary for normal and Weibull probability density functions ... 192
Table 7-13. Statistical parameters for each distribution ......................................................... 193
Table 7-14 Summary statistics for the normal and logistic probability distributions ............ 194
Table 7-15. Test statistic summary for normal and Weibull probability density functions ... 195
Table 7-16. Summary statistics for Case Study I ................................................................... 197
Table 7-17. Best-fit distributions for each input parameter.................................................... 197
Table 7-18. Test statistic summary for normal and Weibull probability density functions ... 198
Table 7-19. Summary statistics for Case Study II .................................................................. 200
Table 7-20. Monte Carlo simulation results based on the derived regression equations ....... 205
Table 7-21. The least and most influential predictor variables on CO2 corrosion ................. 206
Table 8-1. R2-values for the derived models .......................................................................... 218
Table 8-2. R2-values for the established models .................................................................... 219
ix
List of figures
Figure 1-1. The production and manufacturing corrosion costs share ....................................... 3
(Koch et al., 2002)...................................................................................................................... 3
Figure 1-2. Share of each greenhouse gas in total global emissions (IPCC, 2014) ................... 4
Figure 1-3. Keeling curve. ......................................................................................................... 5
(NOAA, 2015) ........................................................................................................................... 5
Figure 1-4. World CO2 emissions by sector (IEA, 2010) .......................................................... 6
Figure 1-5. Projected Keeling curve based on climate modelling ............................................. 8
(Meinshausen et al., 2011) ......................................................................................................... 8
Figure 1-6. World total primary energy supply (IEA, 2012) ..................................................... 8
Figure 2-1. Diagram showing CO2 corrosion including the effect of O2 ................................. 17
(Song et al., 2005) .................................................................................................................... 17
Figure 2-2 Pitting corrosion of a steel pipe (Newman, 2010) ................................................. 19
Figure 2-3 Pitting corrosion of a steel pipe (Newman, 2010) ................................................. 20
Figure 2-4. Mesa-type corrosion in a CO2-rich system (Kermani and Morshed, 2003) .......... 20
Figure 2-5. Flow-induced localised pipeline corrosion ........................................................... 21
Figure 2-6. Major factors affecting CO2 corrosion .................................................................. 22
Figure 2-7. Plot illustrations indicating the effect of the primary parameters on CO2 corrosion
of carbon steel (Dugstad et al., 1994a)..................................................................................... 24
Figure 2-8. Morphologies for protective and nonprotective scale formation (Crolet et al.,
1996, Kermani and Morshed, 2003) ........................................................................................ 27
Figure 2-9. Schematic of H2S corrosion – corrosion rate of steel in H2S solutions is said to be
under mass-transfer control. (Sun and Nesic, 2007) ................................................................ 32
Figure 2-10. Plot showing comparative trend of elemental contributions to corrosion-
resistance (Kermani et al., 2001), (Kermani et al., 2003) ........................................................ 38
Figure 2-11. Plot showing the category of chromium content of steel and corresponding
corrosion rate (Kermani et al., 2001; Kermani et al., 2003) .................................................... 39
Figure 2-12. Phase diagram for pure CO2 – the captured gas can be transported strictly on
either side of the dashed line without cross over. Dense-phase region represents the most
efficient means of pipeline transport. (IEA GHG, 2010) ......................................................... 42
Figure 2-13. Phase envelopes of CO2 with a combination of impurities ................................. 43
(Downie et al., 2007)................................................................................................................ 43
Figure 2-14. Solubility of CO2 in H2O (Spycher et al., 2003) ................................................. 44
Figure 2-15. Flow loop of input to output via continuous adjustment of weights ................... 47
x
(Demuth et al., 2009) ................................................................................................................ 47
Figure 2-16. Basic layered structure of a three-layer artificial neural network ........................ 48
(Al-Fattah and Startzman, 2003) .............................................................................................. 48
Figure 2-17. The purelin and log-sigmoid transfer functions (Beale et al., 2014) ................... 48
Figure 2-18. Plot showing the effect of sulphate ion on pitting potential of steel in 3% NaCl
solution. (Cottis et al., 1999) .................................................................................................... 50
Figure 2-19. Steps in fuzzy inference process .......................................................................... 52
Figure 2-20. 3-D Surface contour plots showing the relationships between input variables
(CO2 mole % with temperature and gas production rate with temperature) and the output
variable, CO2 corrosion rate (Hajizadeh, 2007) ....................................................................... 53
Figure 3-1. Plot of Corrosion Rate against Temperature for varying carbon dioxide partial
pressures,𝑃𝐶𝑂2 = 0.01, 0.03, 0.1 and 0.3MPa (without inhibition from carbonate scale) ....... 60
Figure 3-2. Typical plot of corrosion rate against temperature from experimental CO2
corrosion data (Dugstad et al., 1994b; Halvorsen et al., 1999; Sun and Nesic, 2004) ............. 60
Figure 3-3. Plot of corrosion rate against temperature for varying CO2 fugacities, 𝑓𝐶𝑂2 =
0.01, 0.03, 0.1and 0.3MPa ....................................................................................................... 63
Figure 3-4. Plot of corrosion rate against temperature for varying CO2 fugacities 𝑓𝐶𝑂2 =
0.01, 0.03, 0.1and 0.3MPa ....................................................................................................... 65
Figure 3-5. Plot of Corrosion rate against temperature (Conditions are: T=10-100oC,
fCO2 =1MPa, pH=5.0 and for varying flowrate velocity) .......................................................... 67
Figure 3-6. Plot of corrosion rate against temperature (Conditions are: T=10-100oC,
U=6.0m/s, pH=5.0 and for varying carbon dioxide fugacities) ................................................ 68
Figure 3-7. NORSOK M-506-generated and Matlab 2012a-generated plots of corrosion rate
against temperature for varying CO2 fugacities: 0.03MPa (0.3bar), 0.1MPa (1bar) and
0.3MPa (3 bar). Conditions are: pH=4, CO2 mole =100%, Shear stress=2Pa ......................... 69
Figure 3-8. Plot of carbon dioxide corrosion against temperature ........................................... 71
(For varying water cut percentages) ......................................................................................... 71
Figure 3-9. Plot of carbon dioxide corrosion against temperature ........................................... 72
(For varying CO2 mole percentages) ........................................................................................ 72
Figure 3-10. Nesic-Postlethwaite-Olsen (NPO) model output for T=100oC – Tafel plot of
potential against log10 (current density).................................................................................... 75
Figure 3-11. Plot of corrosion rate against temperature for the NPO model for ...................... 75
CO2 fugacities: 0.01MPa (0.1bar), 0.03MPa (0.3bar), 0.1MPa (1bar) and 0.3MPa (3 bar).
Other conditions where applicable are: pH=4, CO2 mole =100%, Shear stress=2Pa .............. 75
xi
Figure 3-12. Scatterplot of Freecorp model readings against experimental readings for two
datasets: Crolet et al, (1999) and Guo et al, (2005) ................................................................. 78
Figure 3-13. Species concentration profile - H2S concentration as a function of the distance
from steel surface ..................................................................................................................... 79
Figure 3-14. Scatterplot of Freecorp model readings against experimental readings for two
datasets: Lyle and Schutt (1998) and Nesic et al. (2008)......................................................... 81
Figure 3-15. Plot of corrosion rate against temperature for varying CO2 partial pressures for a
fluid velocity, U=0.1m/s .......................................................................................................... 83
Figure 3-16. Plot of corrosion rate against temperature for varying fluid velocities ............... 83
Figure 3-17. Plot of CO2 corrosion rate against temperature for the various models for
𝑓𝐶𝑂2 =0.3MPa (3bar) ................................................................................................................. 85
Figure 4-1. Boxplot showing the overall range (box and whiskers), inter-quartile range (white
space in blue edged box), median, skewness and outliers (red plus symbols) in each of the
variables in the dataset ............................................................................................................. 90
Figure 4-2. Boxplot showing the range of values, inter-quartile range, median and skewness
of the 𝑃𝐶𝑂2 and pH variables (Outliers are red plus symbols).................................................. 90
Figure 4-3. Pareto chart showing the spread of the data (variance) among each of the principal
components .............................................................................................................................. 92
Figure 4-4. Two-dimensional biplot showing the orientation (magnitude and direction) of the
variables in the principal component axes ............................................................................... 92
Figure 4-5. Scree plot showing decreasing eigenvalues (variances) for all the principal
components .............................................................................................................................. 93
Figure 4-6. Three-dimensional plot of corrosion rate against temperature and CO2 partial
pressure 𝑃𝐶𝑂2 ) showing regression model through the points.................................................. 95
Figure 4-7. Multiple linear regression model corrosion rate against experimental corrosion
rate plot and error bar chart for selected test data – error is the difference between
experimental and model corrosion rate .................................................................................... 96
Figure 4-8. Variable interactions between: .............................................................................. 97
A. 𝑃𝐶𝑂2 and Temperature B. Velocity and Temperature ......................................................... 97
Figure 4-9. Variable interactions between: .............................................................................. 98
A. pH and Temperature B. Velocity and 𝑃𝐶𝑂2 ........................................................................ 98
Figure 4-10. Variable interactions between: ............................................................................ 99
A. pH and 𝑃𝐶𝑂2 ......................................................................................................................... 99
B. pH and Velocity................................................................................................................... 99
xii
Figure 4-11. Hougen-Watson nonlinear model corrosion rate against experimental corrosion
rate plot and error bar chart for selected test data – Error is the difference between
experimental and model corrosion rate .................................................................................. 101
Figure 4-12. Leverage-CO2 corrosion data plot and residuals-CO2 corrosion data plot ........ 102
Figure 4-13. Response surface modelling plots – CO2 corrosion rate can be investigated by
changing the magnitude of each predictor variable ................................................................ 103
Figure 4-14. Response surface model corrosion rate against experimental corrosion rate plot
and error bar chart for selected test data – error is the difference between experimental and
model corrosion rate ............................................................................................................... 103
Figure 4-15. Boxplot showing the overall range (box and whiskers), inter-quartile range
(white space in blue-edged box), median, skewness and outliers (red plus symbols) in each of
the variables in the dataset ...................................................................................................... 105
Figure 4-16. Boxplot showing the range of values, inter-quartile range, median, skewness and
outliers of the pH variable ...................................................................................................... 105
Figure 4-17. Pareto chart showing the spread of the data (variance) among each of the
Principal Components ............................................................................................................ 107
Figure 4-18. Two-dimensional showing the orientation of the variables in the principal
component axes ...................................................................................................................... 107
Figure 4-19. Scree plot showing the decreasing magnitudes of eigenvalues for all the principal
components ............................................................................................................................. 108
Figure 4-20. Three-dimensional plot of corrosion rate against temperature and CO2 partial
pressure (𝑃𝐶𝑂2 ) showing regression model through the points ............................................... 109
Figure 4-21. Multiple linear regression model corrosion rate against experimental corrosion
rate plot and error bar chart for selected test data................................................................... 110
Figure 4-22. Variable interactions between:........................................................................... 111
A. PCO2 and Temperature B. Velocity and Temperature........................................................ 111
Figure 4-23. Variable interactions between:........................................................................... 112
A. pH and temperature............................................................................................................ 112
B. Velocity and 𝑃𝐶𝑂2 ............................................................................................................... 112
Figure 4-24. Variable interactions between:........................................................................... 113
A. pH and 𝑃𝐶𝑂2 ....................................................................................................................... 113
B. pH and velocity .................................................................................................................. 113
Figure 4-25. Hougen-Watson nonlinear model corrosion rate against experimental corrosion
rate plot and error bar chart for selected test data................................................................... 114
xiii
Figure 4-26. Leverage-CO2 corrosion data plot and residuals-CO2 corrosion data plot........ 115
Figure 4-27. Response surface modelling plots – CO2 corrosion rate can be investigated by
changing the magnitude of each predictor variable ............................................................... 116
Figure 4-28. Response surface model corrosion rate against experimental corrosion rate plot
and error bar chart for selected test data ................................................................................ 116
Figure 5-1. An artificial model of a neuron showing the transformation of input signals, xi, to
an output signal, yi (Veelenturf, 1995) ................................................................................... 120
Figure 5-2. Stages in machine learning (Haykin, 1999) ........................................................ 122
Figure 5-3. Flow loop of input to output via continuous adjustment of weights ................... 123
(Demuth et al., 2009) ............................................................................................................. 123
Figure 5-4. A simple neuron (Beale et al., 2014) ................................................................... 123
Figure 5-5. A Single-layer feedforward network ................................................................... 125
Figure 5-6. A multilayered feedforward network (Al-Fattah and Startzman, 2003) ............. 125
Figure 5-7. A recurrent network with no self-feedback loop (Haykin, 1999) ....................... 126
Figure 5-8. Flowchart showing the sequence of steps in the development of the neural
network model........................................................................................................................ 127
Figure 5-9. Block diagram of the neural network model ....................................................... 128
Figure 5-10. Arbitrary plot of predicted corrosion rate against experimental corrosion rate 129
Figure 5-11. Combined performances of each training function defines the degree of
suitability of the transfer function in determining the final NN model parameters ............... 131
Figure 5-12. Line plots of the sum total of absolute error against number of neurons for the
logsig transfer Function ......................................................................................................... 132
Figure 5-13. Line plots of R2-value against number of neurons for the logsig transfer function
................................................................................................................................................ 133
Figure 5-14. Line plots of R2-value against number of neurons for the tansig transfer function
................................................................................................................................................ 134
Figure 5-15. Multiple bar charts showing the Mean R2-coefficients for each training function
for both logsig and tansig transfer functions .......................................................................... 134
Figure 5-16. Multiple bar charts showing the absolute error means for each training function
for both logsig and tansig transfer functions .......................................................................... 135
Figure 5-17. Line plots of sum total of absolute error against number of neurons for the tansig
transfer function ..................................................................................................................... 135
Figure 5-18. NN model corrosion rate against experimental corrosion rate .......................... 137
Figure 5-19. Error bar chart ................................................................................................... 138
xiv
Figure 5-20. Bar chart showing the distribution of data-points from each source ................. 139
Figure 5-21. Bar chart showing the distribution of data-points for the grouped experimental
corrosion rates ........................................................................................................................ 140
Figure 5-22. Bar chart showing the distribution of data-points for the recorded temperatures
................................................................................................................................................ 140
Figure 5-23. Corrosion rate against temperature plot for the training dataset ........................ 141
Figure 5-24. Line plots of R2-values of training functions against number of neurons for the
logsig transfer function ........................................................................................................... 142
Figure 5-25. Line plots of R2-values of training functions against number of neurons for the
tansig transfer function ........................................................................................................... 142
Figure 5-26. Bar charts showing the means of the R2-values of each training function for both
logsig and Tansig transfer functions ....................................................................................... 143
Figure 5-27. Bar charts showing the means of absolute corrosion rate error of each training
function for Both logsig and tansig transfer functions ........................................................... 144
Figure 5-28. Line plots of sum of absolute corrosion rate error against neuron configuration
for the traincgf and trainlm functions ..................................................................................... 145
Figure 5 -29. NN Model corrosion rate against experimental corrosion rate plot.................. 147
Figure 5-30. Error bar chart – Error is the difference between experimental and model
corrosion rate .......................................................................................................................... 147
Figure 5-31. Corrosion rate against temperature plot for the testing dataset with NN model
predictions .............................................................................................................................. 148
Figure 6-1. Close connection between classical logic and classical set and their relationships
with predicates of logic........................................................................................................... 152
Figure 6-2. Membership function of a fuzzy set .................................................................... 153
Figure 6-3. Triangular and trapezoidal membership functions .............................................. 154
(Fuzzy Logic Toolbox User’s Guide, 2015) ........................................................................... 154
Figure 6-4. Linear piece-wise membership function (Ukil, 2010) ......................................... 155
Figure 6-5.Generalised form of a continuous membership function ...................................... 155
(Fuzzy Logic Toolbox User’s Guide, 2015) ........................................................................... 155
Figure 6-6. Fuzzy inference diagram showing all parts of the fuzzy inference process – from
fuzzification through to defuzzification (Fuzzy Logic Toolbox User’s Guide, 2015) ........... 158
Figure 6-7. pH membership function plots............................................................................. 159
Figure 6-8. Three-dimensional plot of corrosion rate against PCO2 and temperature – result of
Mamdani FIS .......................................................................................................................... 160
xv
Figure 6-9. Mamdani FIS plot of corrosion rate against temperature .................................... 161
Figure 6-10. Mamdani FIS plot of corrosion rate against CO2 partial pressure .................... 162
Figure 6-11. Three-dimensional plot of corrosion rate against flowrate velocity and
temperature – result of Mamdani FIS .................................................................................... 163
Figure 6-12. Mamdani FIS plot of corrosion rate against flowrate velocity.......................... 163
Figure 6-13. Mamdani FIS model corrosion rate against experimental corrosion rate plot and
error bar chart for selected test data – error is the difference between experimental and model
corrosion rate.......................................................................................................................... 164
Figure 6-14. A generic fuzzy inference system showing input variables to the left (T, 𝑃𝐶𝑂2 ,
pH, fluid flowrate) and the output variable to the right (Corrosion Rate) ............................. 165
Figure 6-15. Rule Viewer for all seven rules showing mini graphical representations of each
of the input variables and their effect on the output variable................................................. 167
Figure 6-16.Three-dimensional plot of corrosion rate against 𝑃𝐶𝑂2 and temperature – result of
Mamdani FIS.......................................................................................................................... 168
Figure 6-17.Three-dimensional plot of corrosion rate against 𝑃𝐶𝑂2 and temperature – result of
Sugeno FIS ............................................................................................................................. 168
Figure 6-18. Mamdani FIS plot of Corrosion Rate against CO2 partial pressure .................. 169
Figure 6-19. Sugeno FIS plot of Corrosion Rate against CO2 partial pressure ..................... 170
Figure 6-20. Mamdani FIS Model Corrosion Rate against Experimental Corrosion Rate plot
and Error Bar chart for selected test data – Error is the difference between Experimental and
Model Corrosion Rate ............................................................................................................ 171
Figure 7-1. Normal probability plot and Weibull probability plot ........................................ 179
Figure 7-2. Normal probability density and Weibull probability density plots for experimental
CO2 corrosion rate dataset...................................................................................................... 181
Figure 7-3. Number of runs used in Monte Carlo simulations .............................................. 183
Figure 7-4. Monte Carlo simulation using statistical parameters of the experimental dataset
................................................................................................................................................ 183
Figure 7-5. Probability density function plot for Monte Carlo simulation of low pressure CO2
corrosion rate.......................................................................................................................... 186
Figure 7-6. Cumulative density function plot for Monte Carlo simulation of low pressure CO2
corrosion rate.......................................................................................................................... 186
Figure 7-7. Probability density function plot for Monte Carlo simulation of low pressure CO2
corrosion rate.......................................................................................................................... 188
xvi
Figure 7-8. Cumulative density function plot for Monte Carlo simulation of low pressure CO2
corrosion rate .......................................................................................................................... 189
Figure 7-9. Sensitivity analysis – effect of each parameter on corrosion rate........................ 190
Figure 7-10. Normal probability plot and Weibull probability plot ....................................... 192
Figure 7-11. Normal probability density and Weibull probability density plots for
experimental CO2 corrosion rate dataset ................................................................................ 193
Figure 7-12. Monte Carlo simulated probability density function plot showing normal and
Logistic curve fits ................................................................................................................... 194
Figure 7-13. Probability density function plot for Monte Carlo simulation of high pressure
CO2 corrosion rate .................................................................................................................. 196
Figure 7-14. Cumulative density function plot for Monte Carlo simulation of high pressure
CO2 corrosion rate .................................................................................................................. 196
Figure 7-15. Probability density function plot for Monte Carlo simulation of high pressure
CO2 corrosion rate .................................................................................................................. 198
Figure 7-16. Cumulative density function plot for Monte Carlo simulation of high pressure
CO2 corrosion rate .................................................................................................................. 199
Figure 7-17. Sensitivity analysis – effect of each parameter on corrosion rate...................... 201
Figure 7-18. Probability plot showing the Weibull distribution as the best-fit for CO2 partial
pressure variable ..................................................................................................................... 203
Figure 8-1. The 1991 De Waard model corrosion rate against experimental corrosion rate plot
and error bar chart for selected test data – error is the difference between experimental and
model corrosion rate ............................................................................................................... 209
Figure 8-2. The 1995 De Waard model corrosion rate against experimental corrosion rate plot
and error bar chart for selected test data ................................................................................. 210
Figure 8-3. Norsok model corrosion rate against experimental corrosion rate plot and error bar
chart for selected test data ...................................................................................................... 211
Figure 8-4. Freecorp model corrosion rate against experimental corrosion rate plot and error
bar chart for selected test data – error is the difference between experimental and model
corrosion rate .......................................................................................................................... 212
Figure 8-5. The 1991 De Waard model corrosion rate against experimental corrosion rate plot
and error bar chart for selected test data – error is the difference between experimental and
model Corrosion rate .............................................................................................................. 213
Figure 8-6. The 1995 De Waard model corrosion rate against experimental corrosion rate plot
and error bar chart for selected test data ................................................................................. 214
xvii
Figure 8-7. Norsok model corrosion rate against experimental corrosion rate plot and error bar
chart for selected test data ...................................................................................................... 215
xviii
Nomenclature
[𝑎, 𝑏] a given range in which the random variable X may fall within
𝐵𝐻2 𝑆 = 15500J/mol
Ecorr corrosion potential at the point where the sum of the individual current
densities passes through zero
F Faraday constant
F(x) integral of f(x) and is the cumulative density function (cdf) of X over
the range [a, b]
𝑖(𝐻 +) 𝑜𝑟 𝑖𝛼(𝐻 + ) , 𝑖𝑜(𝐻 +) instantaneous activation controlled cathodic current and initial
exchange current density for hydrogen ion respectively (A/m2)
iFe, io(Fe) instantaneous anodic current term for the dissolution of the metal (Fe)
and initial exchange current density of the metal respectively (A/m2)
H0 null hypothesis
km, kr mass transfer rate constant and activation reaction kinetics constant of
CO2 corrosion reaction – applicable to De Waard and NPO correlations
𝑘𝑚,𝐻2 𝑆 mass transfer coefficient for H2S in the hydrodynamic boundary layer,
xx
K1 dissociation constant for the hydration of aqueous H2S
𝐾𝐶𝑂2, 𝐾𝐻2 𝑂 equilibrium constants for the solubility of H2O in CO2 and CO2 in H2O
respectively
n all possible values of the random variable X; the sum of the product of
the input scalar (p) and weight (w) added to bias (b); any given number
T temperature (oC)
V volume (m3)
xi, yi , (xi, yi) experimental corrosion rate, model predicted corrosion rate and
abscissa-ordinate coordinate pair respectively
Greek letters
𝛽0 , 𝛽1, 𝛽2 , … , 𝛽𝑛 constants or predictor variable coefficients for regression equation
η overpotential (volts, V)
𝜇 mean for the normal probability distribution; scale parameter for the
log-normal distribution
xxiii
xxiv
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Chapter 1. Introduction
1
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The effect of carbon dioxide (CO2) corrosion on pipelines is of great relevance to the
petroleum as well as the Carbon Capture and Storage (CCS) industries. CO2 corrosion is
responsible for lost production as it brings about the gradual degradation of pipe internals
with time.
The U.S. refinery capacity was approximately one-fifths of the total world refinery capacity as
of 1991 (Campbell, 1991), and is ~27% of the global refinery capacity as of 2014. In 2002,
the production and manufacturing industry accounted for approximately 13% (or $18 billion)
of the corrosion costs in industry categories (Koch et al., 2002). The share for each of the
industry categories is shown in Table 1-1.
Table 1-1. Share of corrosion costs for the U.S. industry categories (Koch et al., 2002)
Industry Category Percentage Share (%)
Infrastructure 16
Utilities* 35
Transportation 21
Production and Manufacturing 13
Government 15
* Utilities includes corrosion costs from the natural gas distribution system (~10% of Utilities)
2
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
In addition, the natural gas distribution corrosion costs account for 10% of the Utilities
industry category, as seen from Table 1-1. Hence, the cumulative share for the petroleum
industry was in excess of 40% of the U.S. total corrosion costs as of 2002. As of 2014, the
cost of corrosion in the production and manufacturing sector was $34.4 billion, almost double
the reported 2002 amount, with the oil and gas industry accounting for more than half.
As of 2011, the annual cost of corrosion rose to $2.2 trillion, amounting to over 3% of the
world‘s gross domestic product (GDP) (Hays, 2011). The cost of general corrosion is said to
be between 3 to 5% of an industrialised nation’s GDP (Schmitt et al., 2009). Also,
approximately 60% of all oil and gas field failures are related to CO2 corrosion (Nyborg,
2005).
3
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
A recent report by the IPCC also suggests that most of the observed increase in global average
temperatures since the mid-20th century is most likely due to the observed increase in
anthropogenic greenhouse gas concentrations (IPCC, 2014).
CO2 is evidently the most significant greenhouse gas given that annual emissions of the gas
have risen by almost 80% between 1970 and 2010, from 21 to 49 gigatonnes, and as of 2010,
represented approximately 77% of the total global greenhouse gas emissions (IPCC, 2014).
The pie chart in Figure 1-2 shows the share of each greenhouse gas in global emissions.
F-gases Nitrous
2% Oxide (N2O)
6%
Figure 1-2. Share of each greenhouse gas in total global emissions (IPCC, 2014)
On this note, atmospheric CO2 concentration has elevated beyond its pre-industrial revolution
magnitude, 280ppm, to the value of 400ppm, as of 2013 (IPCC, 2014; Dlugokencky, 2014).
This observational evidence is depicted by the rise in atmospheric CO2 concentration as
shown in the Keeling curve of Figure 1-3. From Figure 1-3, it is seen that atmospheric CO2
concentration increased steadily on an annual basis from 1960 to date. This rise in greenhouse
gas levels ultimately results in projected changes in climate such as increasing temperatures,
changes in precipitation, sea-level rise and increased frequency and intensity of some extreme
climatic events leading to increased climate variability.
4
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Globally, these environmental changes have a profound impact on biodiversity such that the
ecosystems are unable to adopt or induce self-regulatory mechanisms to nullify the changes,
thus mitigation processes are required. CO2 is being emitted anthropogenically at a faster rate
than the natural processes of the carbon cycle are able to remove it. One major reason for the
earths’ ecosystems inability to counteract the effect of these changes is land use –
deforestation. The impact of deforestation is two-fold; firstly, it reduces the number of trees,
which are the earth’s natural carbon sinks and secondly, when trees are felled, they are often
burned, releasing their carbon contents into the atmosphere.
Mitigation is the anthropogenic intervention to reduce net greenhouse gas emissions that
would lessen the pressure on natural ecosystems as well as human systems from the effects of
climate change. Furthermore, mitigation involves the reduction of greenhouse gases by the
reduction of fossil fuel use or by the increase in the natural carbon uptake rate by ecosystems.
An examination of CO2 emissions by sector indicates that electricity and heat generation is
the largest sector (41%) responsible for CO2 emissions as shown in Figure 1-4.
5
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Residential
6%
Other
10%
Electricity and
Heat (Power
generation) Industry
41% 20%
Transport
22%
__________________________________________________________________________________
*Other refers to emissions from commercial services, agriculture, fishing and energy industries other than
electricity and heat generation
__________________________________________________________________________________________
Further evidence of the impact of CO2 emissions can be inferred from climate modelling. In
2014, the IPCC proposed four Representative Concentration Pathway (RCP) scenarios. These
are: RCP 2.6, RCP4.5, RCP6.0 and RCP8.5. The RCPs are consistent with a wide range of
possible changes in predicted anthropogenic GHG emissions (IPCC, 2014).The names of the
RCPs come from their respective radiative forcing values from the end of the 21st century
with reference to pre-industrial revolution values: 2.6, 4.5, 6.0 and 8.5W/m2 (IPCC, 2014).
Radiative forcing of a gas is defined as the difference between the incident solar radiation on
the earth’s atmosphere and the outgoing infrared radiation caused by the increased
concentration of the gas (IPCC, 2014). CO2 has the highest contribution to radiative forcing
and accounts for more than 80% of the radiative forcing from GHGs for the climate modelling
pathway scenarios due to its high atmospheric concentration, warming effect, long residence
time in the atmosphere and its global spatial distribution (Clarke et al., 2007). Since 1750, its
radiative forcing has increased by 1.88W/m2 or ~65% of the increased forcing by all long-
lived greenhouse gases (LLGHG) (Dlugokencky et al, 2014). Table 1-2 is a summary of the
RCP scenarios.
6
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Table 1-2. RCP scenarios for the end of the 21st century (IPCC, 2014)
Future Definitions Global Mean Surface Global Mean Sea
Scenario Warming (oC) level rise (m)
Of all these scenarios, RCP2.6 represents the best possible future pathway for emissions
however; it represents a near-ideal situation. In terms of the current global emission trends,
based on socio-economic factors, RCP 4.5 and RCP6.0 represent the more likely future
pathways while RCP8.5 is an extreme. The effect each of these scenarios has on the CO2
atmospheric concentrations is shown in Figure 1-5.
It can be observed from Figure 1-5 that the atmospheric CO2 concentration peaks for RCP2.6,
4.5 and 6.0 do not coincide with the stated emissions peaks listed in Table 1-2. For instance,
the emissions peak for RCP2.6 is between 2010 and 2020, yet the atmospheric concentration
of the gas does not peak until approximately 2040. The atmospheric CO2 concentration peaks
occur much later due to the long-life or long residence time of CO2.
7
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Worldwide, the power generation sector is heavily reliant on coal which is the most carbon
intensive of fuels thus amplifying its share in global emissions as shown in the pie chart of
Figure 1-4.
50%
Coal/Peat
45%
40%
Percentage Share
35% Oil
30%
25% Natural Gas
20%
15% Nuclear
10%
5%
0% Biofuels and Waste
1973 2010 2035 (Renewables)
Year Hydro
8
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
It is worth noting that from Figure 1-6, fossil fuel (oil, gas and coal) combustion accounts for
approximately 80% of the total world CO2 emissions. Also shown in Figure 1-6, is the
projected world energy consumption for the year 2035, and this shows very little change in
the usage of coal. This implies that while some countries will inevitably reduce their reliance
on coal, other countries may increase their consumption of it – in general, some countries will
depend on it more than others.
In order to reduce greenhouse gas emissions however, coal-fired plants will have to be
deployed in conjunction with CCS technologies. This is a low-cost option (World Energy
Council, 2007; IPCC, 2014). In a context where a gas-fired system is deployed instead of the
coal-fired alternative, CCS technologies should be used alongside even though, generally
speaking, the former is more efficient and environmentally cleaner than the latter (IEA, 2012).
Corrosion of pipelines under low pressure CO2 conditions has been covered extensively with
readily accessible corrosion data made available (Dugstad et al., 1994b; Halvorsen et al.,
1999; Sun and Nesic, 2004). CO2 corrosion under low pressure conditions is characteristic of
the oil and gas industry and comes about when CO2 in produced gases dissolve in water, thus
forming carbonic acid (Garverick, 1994). For gas condensate wells producing formation
water, produced gas coupled with acid gases, (H2S, CO2), if liquid water is allowed to
condense on the tubing, CO2 corrosion may occur (Garverick, 1994). CO2 corrosion has also
become an increasingly common phenomenon in enhanced oil recovery applications, where
CO2 and water mixtures are injected directly into the depleted well to facilitate extraction
(Garverick, 1994; Satter et al., 2007).
Corrosion of pipelines under high pressures comes about by the transportation of the gas in
the dense liquid-phase from its point of capture (Power plant) to a storage site (abandoned
coal mine or oilfield) (Downie et al., 2007). Elevated pressures in this context exceed 7.4MPa
(74bar), hence corrosion risks are possible. Corrosion risks are potentially more severe when
impurities such as traces of water droplets, H2S, SOx and NOx gases are present (Downie et
al., 2007). These gases are present in CO2 streams depending on the fuel source of the power-
generating emitter: fuel or flue gas (Granite and O’Brien, 2005).
9
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Fuel gases are combustible gases which are direct products of the partial combustion of a fuel
during gasification. They usually contain higher amounts of CO than flue gases (Zevenhoven
and Kilpinen, 2004). Flue gases are the final gaseous products of complete combustion and
typically contain higher amounts of CO2 than fuel gases (Zevenhoven and Kilpinen, 2004).
The costs associated with pipeline CO2 corrosion coupled with the unique challenges of
modelling corrosion risks at elevated CO2 pressures, in order to reduce global emissions of
the gas signifies the importance of a research undertaking in this field.
1. The writing-up of a literature review in order to be fully aware of the latest research
activities being carried out in the field of CO2 corrosion. Preparation of the literature
review would aid the preparation of a project Gantt chart (see Appendix A1.1),
planning of project activities and to direct the focus of the research.
2. The gathering of CO2 corrosion data from sources in research literature. Data is to be
collected from multiple experimental sources. In these sources, corrosion tests were
carried out on samples of ferritic-pearlitic carbon steel for low pressure data. One such
chemical composition for the low pressure CO2 dataset is presented in Table 1-3
(Dugstad et al., 1994b):
Table 1-3. Weight percentage chemical composition of St.52 carbon steel (Dugstad et al.,
1994b)
C Si Mn S P Cr Ni V Mo Cu
0.180 0.340 1.500 0.017 0.023 0.080 0.030 0.003 0.030 0.100
10
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
was the X65 pipeline carbon steel with a ferritic-pearlitic microstructure (Zhang et al.,
2012a). Its chemical composition is shown in Table 1-4:
Table 1-4. Weight percentage chemical composition of X65 carbon steel (Zhang et al.,
2012a)
C Si Mn P S N Cr Mo Ni Cu
0.100 0.310 1.480 0.350 0.005 0.007 - - - -
4. The analysis of the stored data. This involves the plotting of profiles for corrosion
rates against temperature or with other variables such as the CO2 partial pressure, flow
velocity, and pH, depending on what is being modelled at that instance.
5. The writing, debugging and execution of Matlab model codes for the respective
established CO2 corrosion models obtained from research literature. All model codes
are written in Matlab 2012a. These models include the 1975, 1991 and 1995 De
Waard as well as the Norsok and Nesic-Postlethwaite-Olsen (NPO) models. The
limitations of each of the models are then fully discussed.
7. The division of the low and high pressure CO2 datasets into model-developing and
testing datasets.
11
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
9. The development of neural network (NN) models for both low and high pressure CO2
corrosion data based on the designated model-developing datasets. Given that the use
of NNs as prediction tools for low pressure CO2 data has been established (Nesic and
Vrhovac, 1999), but not for high pressure; it would be useful to develop a NN model
for predicting high pressure CO2 corrosion as this would constitute a novel approach.
10. The development of a fuzzy inference system (FIS) model capable of predicting
corrosivity for the low and high pressure CO2 corrosion datasets.
11. The application of Monte Carlo simulations to develop a parametric sensitivity study
on the influence of the predictor variables (temperature, CO2 partial pressure, flow
velocity and pH) on CO2 corrosion rates; for both low and high pressure CO2
corrosion datasets.CO2 corrosion research literature has very limited content when it
comes to the use of statistical tools and Monte Carlo modelling. Some aspects of
statistical modelling such as the use of PCA as well as the Monte Carlo-simulated
sensitivity analysis constitute novel approaches in this field.
12. The established 1991 and 1995 De Waard, Norsok and Freecorp models are tested
with the same test datasets used with the derived models. The model results based on
the application of the R2-coefficient for both low and high pressure CO2 datasets are
compiled together and discussed.
12
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
This Chapter has so far, dealt with the monetary costs of CO2 corrosion and its impact on
world climate. There was also a discussion of the recently adopted climate modelling
scenarios applicable from current timelines until the end of the 21st century. The thesis aims
and objectives have been fully discussed, showing a progression from a review of the
literature to data collection, storage and analysis. Specific modelling techniques considered to
be novel in approach with respect to CO2 corrosion research literature have also been
highlighted.
The following chapters have generally been organised the same way as the ordered list of the
thesis aims and objectives. Chapter 2 is the literature review for CO2 corrosion and this is
dealt with in great detail. Topics covered include: CO2 corrosion mechanism, the factors
affecting CO2 corrosion, the different types, an introductory account of CO2 corrosion models
as well as the methodology for modelling CO2 corrosion from low to high CO2 partial
pressures. Also discussed are introductory reviews of the use of NNs and fuzzy logic as tools
suitable for CO2 corrosion prediction.
Chapter 3 deals with the established CO2 corrosion correlations and models. These include the
aforementioned De Waard, Norsok and Freecorp models. The NPO model is also discussed as
it is the basis for the development of the Freecorp model. These models are compared against
each other by applying identical parametric conditions for the temperature, CO2 partial
pressure, flow velocity and pH and assessing their predictions. Other topics covered include
organic acid and H2S corrosion.
Chapter 4 covers the statistical analysis of the low and high pressure CO2 datasets.
Explanations on the full descriptive statistics, principal component analysis and regression
relationships are outlined. Also, response surface model plots for the CO2 corrosion datasets
are discussed. Again, the R2-coefficient is used to report model performances for the derived
regression equations, which are subsequently used in Chapter 7.
13
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Matlab 2012a training and transfer functions. A process of elimination is carried out to
determine which combination of transfer and training functions is best-suited to CO2
corrosion modelling. Also, the determination of the final NN model size is discussed and this
is proven to be a direct consequence of the size of the datasets used. The low and high
pressure CO2 datasets were each split into model-developing and testing sets. The R2-value is
then used to report model performance following testing with test set.
Chapter 6 covers fuzzy modelling. An account is given of how fuzzy logic originated. The
various membership functions are discussed. The methodology for the development of a fuzzy
logic model based on the creation and applicability of rules (‘if-then’) is explained. The R2-
statistic is used as an assessment of model performance when tested with a test set.
The seventh chapter covers Monte Carlo simulation. Definitions of relevant concepts such as
the random variable and the different types of probability distributions are stated. A brief
basis is outlined on how to determine best-fit distributions for a given dataset. The low and
high pressure CO2 datasets are analysed in order to determine their best-fit theoretical
distributions. The pre-determined regression equations are then used as case study scenarios
for establishing a parametric sensitivity analysis for CO2 corrosion rates in terms of the given
predictor variables.
Chapter 8 is an extended discussion of the performance results obtained for the established
models outlined in Chapter 3. An identical test dataset to the ones used for the assessment of
performances of the derived models in Chapters 5 to 8, is used as the basis for evaluating the
accuracy of model predictions. Performance results of the derived models in Chapters 4 to 8
are discussed in an expansive manner. The advantages, limitations and their applicability are
explained in detail.
The Conclusions and Recommendations for Further work are outlined in Chapter 9.
Generally, the conclusion is discussed in the same logical manner as the thesis aims and
objectives. Furthermore, those aspects of modelling that have are beyond the scope of this
research are stated as recommendations for further work.
14
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
15
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Some established models developed for the prediction of CO2 corrosivity are also outlined.
Modelling techniques for the accurate prediction of high pressure CO2 corrosion in pipelines
such as neural networks and fuzzy logic are also discussed as viable research directives in the
latter sections of this chapter.
Carbon dioxide corrosion was first documented to occur as early as the 1940’s in the U.S. oil
and gas industry (Perez, 2013) and has since been followed by several studies on corrosion
rate prediction (Crolet and Bonis, 1991; Song et al., 2005; Hernandez et al., 2006). CO2
corrosion, also termed sweet corrosion involves the dissolution of gaseous carbon dioxide into
an aqueous phase where, via a hydration reaction with water forms carbonic acid, H2CO3, the
specie which induces corrosive attack on the surface of the metal (Kermani and Morshed,
2003).
Dry carbon dioxide is itself non-corrosive at oilfield temperature and pressure conditions
(T=0-100oC, P<7.4MPa) however when dissolved in an aqueous solution, it initiates a series
of electrochemical reactions culminating in the aqueous phase coming in contact with the
metal surface thus inducing the dissolution of the latter.
In general, gaseous carbon dioxide exerts a partial pressure on water resulting in its
dissolution. The aqueous carbon dioxide then undergoes hydration thus forming carbonic acid
which does not fully dissociate, unlike strong mineral acids, which completely dissociate in
aqueous media (Kermani and Morshed, 2003).
16
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 2-1 is a summary of reactions occurring on the surface of the metal. The horizontal
short-dashed line represents the aqueous boundary situated in close proximity, at a distance,
y=δs, to the steel surface. The steel surface is represented by a distance, y=0. CO2 present in
the oil and gas stream exerts a partial pressure, CO2 (𝑃𝐶𝑂2 ), leading to its dissolution in the
aqueous phase, CO2 (aq). Its reaction with water leads to the formation of carbonic acid
(H2CO3). Carbonic acid eventually undergoes dissociation into the bicarbonate ion (HCO3-)
and hydrogen ion. The bicarbonate ion undergoes further dissociation into the carbonate ion
and hydrogen ion. Meanwhile, steel undergoes oxidation into ferrous ions (Fe2+). The
eventual reaction between ferrous ions and the carbonate ions results in the formation of
ferrous carbonate (FeCO3).
Also, a possibility in the corroding system is the presence of O2 (Song et al., 2005). This
exerts a partial pressure, O2 (𝑃𝑂2 ), leading to its dissolution in the aqueous phase O2 (aq).
Aqueous O2 then gets oxidised in the presence of water, forming hydroxide ions (OH-). The
consequent reaction between the aforementioned ferrous ions and hydroxide ions results in
the formation of ferrous hydroxide, Fe(OH)2.
It is often the case that the corrosion of iron is generally summarised as follows (Crolet et al.,
1999):
The CO2 corrosion mechanism is a series of electrochemical reactions that involve the anodic
dissolution of iron and the cathodic evolution of hydrogen. (Nesic and Vrhovac, 1999)
17
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The anodic dissolution of iron leads to the formation of various films such as FeCO3
(siderite), Fe3O4 (magnetite) and Fe3C (cementite) which can be protective or non-protective
depending on film-formation conditions and the type of film formed.
For the cathodic reactions, it has been assumed that the presence of CO2 increases the rate of
hydrogen evolution reaction. For strong acids that are known to dissociate fully, hydrogen
ions will be evolved readily in solution however since the CO2 corrosion mechanism is mass
transfer limited therefore hydrogen ion evolution cannot exceed the rate at which it is
transported to the metal surface from bulk solution. (Nesic and Vrhovac, 1999)
In solutions of pH>4, the mass transfer controlled current is small but the presence of the
weak acid, H2CO3, which is the hydrated form of CO2, enables hydrogen evolution at a much
higher rate. Hence at any given pH, the presence of CO2 leads to a much higher corrosion rate
than would be found in a solution of a strong acid. Carbonic acid may be reduced at the
cathode leading to the evolution of hydrogen gas. In any case, both reactions are said to
procced independently of each other (Gray et al., 1989). Another cathodic reaction is the
reduction of water (Tanupabrungsun et al., 2013).
Also it has been suggested that at high pH conditions in CO2-rich solutions, that the reduction
of the bicarbonate ion becomes important:
2HCO− − 2−
3 (aq) + 2e → H2 (g) + 2CO3 (aq) (2.8)
18
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
2.3.1 Pitting
This is a form of localised corrosion that usually occurs in low-velocity fluid flow
environments around dew point temperatures in a gas-producing well (Kermani and Morshed,
2003). It results in the appearance of holes in the metal surface, as seen in Figures 2-2 A and
2-2B.
A B
Figure 2-2A is a close-up view of pipe steel that has undergone pitting corrosion and Figure
2-2B shows the pit size on the steel surface in relation to the rest of the pipe. Pitting corrosion
is difficult to predict and may give the appearance of small holes on the steel surface and is
known to develop under the surface of the metal triggering mechanical failures by fatigue, as
seen in Figure 2-3 (Newman, 2010).
19
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Pitting susceptibility increases with increasing temperature and CO2 partial pressure. It is also
reported that nearly all metals and alloys regardless of type are likely to undergo pitting
corrosion provided the conditions are right. While some researchers have pointed out that
chloride ions are not a necessary component for propagating pitting corrosion (Videm and
Dugstad, 1989) and conversely, the addition of lead (Pb) inhibited pitting corrosion through
deposition at local anodes (Schmitt and Feinen, 1983; Schmitt and Engels, 1988); there is
generally no applicable rule for the prediction of this type of corrosion since its propagation
has been put down to its dependency on various factors.
This is a type of localised corrosion that takes place in low to medium fluid-flow
environments as shown in Figure 2-4, where the protective film forms, however it is unstable
and unable to withstand the intensity of the operating regime. Low to medium velocity fluid
flow is still capable of washing away any protective films formed on the metal surface.
Figure 2-4. Mesa-type corrosion in a CO2-rich system (Kermani and Morshed, 2003)
20
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Although a study performed by Ikeda and others (Ikeda et al., 1984) attributed the initiation of
mesa-attack to the competitive film formation reactions between Fe3O4 and ferrous carbonate
(FeCO3), the presence of magnetite scale has not been detected in actual field conditions
(Crolet, 1994; Crolet, 2002). It was concluded that the initiation and propagation of mesa-
attack corrosion is due to inadequate protection offered by ferrous carbonate film formation
on the metal surface (Crolet, 1994; Crolet, 2002). However there is still no definitive
understanding of the way in which mesa-attack occurs and the precise nature of prevailing
conditions required for its propagation is uncertain hence further systematic studies are
necessary to prevent future occurrence in the field.
This form of corrosion usually starts from pits/troughs that had previously been sites of
localised mesa attack (Kermani and Morshed, 2003). This form of corrosion is solely
dependent on high fluid flowrates.
High velocity fluid flow proceeds to generate turbulent eddies around mesa-attack sites,
sweeping away any films formed as well as scale growths (Schmitt and Feinen, 1983),
21
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
(Schmitt and Engels, 1988). Once the scales are damaged or destroyed, the flow velocity then
prevents the re-formation of protective film thereby leaving fluids to corrosively attack the
exposed metal surface.
A number of key factors play a role in influencing the extent of CO2 corrosion. These
parameters are broadly classified into the following categories as shown in Figure 2-6 below:
CO2 Corrosion
22
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
[𝐴+ ][𝐵 − ]
𝑆𝑎𝑡 = (2.10)
𝐾𝑠𝑝
𝐴+ + 𝐵 − ⇌ 𝐴𝐵 (2.11)
Where:
mol
𝐴+ and 𝐵 − are ionic species in ′ ′ , that form an insoluble salt (𝐴𝐵)
l
mol 2
𝐾𝑠𝑝 is the solubility limit in ′ ( ) ′
l
For CO2 corrosion, A+ and B- are Fe2+ and CO32-, respectively. Thus:
High supersaturation of Fe2+ and CO32- ions leads to the precipitation of FeCO3 on the metal
surface as a film layer, once the ferrous carbonate solubility limit (Ksp) is exceeded, which
leads to a consequent reduction in corrosion rate by providing an extended diffusion length
between the metal surface and corrosive medium, as seen in the Corrosion rate vs. pH plot of
Figure 2-7B (Kermani and Morshed 2003). Low Fe2+ concentrations in solution has the effect
of reducing corrosion rates but at a much slower rate because the rate of formation of FeCO3
scale is much slower, as seen in the Corrosion Rate vs. pH plot of Figure 2-7B.
23
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The following illustrations depict the general trend for some environmental factors on CO2
corrosion.
A B
C D
Figure 2-7. Plot illustrations indicating the effect of the primary parameters on CO2
corrosion of carbon steel (Dugstad et al., 1994a)
Figure 2-7A shows a linear relationship between CO2 partial pressure to an exponent of 0.7
and corrosion rate (Dugstad et al., 1994a). An increase in CO2 partial pressures results in
incremental rates of corrosion. Figure 2-7B shows a near-linear decrease in corrosion rate
with increasing pH. Corrosion rate however decreases more slowly for low dissolved ferrous
ion concentrations than for higher ferrous ion concentrations, due to supersaturation occurring
more readily for the latter than the former thus forming a protective scale sooner on the steel
surface. Figure 2-7C shows an increase in temperature resulting in a corresponding increase in
corrosion rates albeit a peak occurs at the 60-80oC range due to protective magnetite scale
formation, which subsequently leads to a fall in corrosivity. Figure 2-7D shows an initial
increase in fluid flow velocity leads to an increase in corrosion rates, after which increasing
fluid flow velocity leads to a levelling-off of corrosivity. For very high fluid velocities
exceeding 7m/s, there is the possibility of a further increase in corrosion rate, particularly for
elevated temperatures (Eriksrud and Sφntvedt, 1983), due to the mechanical removal of
surface film (Dugstad et al., 1994a).
24
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
For both processes, CO2 corrosion is induced by the dissolution of the gas in an aqueous
medium coupled with the initiation of carbonic acid equilibria; and is dependent on the CO2
mole fraction or percentage. In general, CO2 corrosion will increase in magnitude, the greater
its mole fraction within the oil and gas stream, likewise this rule is applicable to CCS
processes albeit the higher CO2 mole fractions and presence of acid gases potentially implies
greater corrosion rates are to be expected for the latter. Other factors intrinsic to petroleum
fluids such as operating pressures, temperatures and pH also have an influence on CO2
corrosion rates as shown in Table 2-1.
Table 2-1. Estimated corrosion rates of carbon-manganese steels for some fields in the
Norwegian sector of the North Sea (Dugstad et al., 1994a)
Field Type CO2 Pressure Max. Calculated Estimated
Operator content (bar) temperature pH corr. rate
(mol %) (oC) (mm/year)
Tommeliten Oil/Gas 3 235 90 6.1 11 (90oC)
Statoil 15 (67oC)
Lille-Frigg Gas 2.4 440 80 Sat.* 13 (80oC)
Elf Aquitaine 15 (68oC)
TOGI Gas 0.2 100 55 Sat.* 5.1
Norsk Hydro
Sleipner Gas/oil 9 150 90 6.8 12 (90oC)
West Statoil 19 (56oC)
Corrosion rates are calculated by using the 1991 De Waard correlation (De Waard et al., 1991)
* The pH at FeCO3 saturation has been used in the corrosion rate calculation
25
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
From Table 2-1, it is observed that though CO2 mole content is the primary determinant of the
magnitude of predicted corrosion rates, pressures, temperatures and pH also contribute hence
these are utilised in CO2 corrosion prediction modelling. The 1991 De Waard model is
described in greater detail in Section 3.3. The high CO2 mole content (9%) of the Sleipner
West field implies that it has the greatest predicted CO2 corrosion rates (12 and 19mm/year)
of all the fields.
It must be noted that while these parameters are the main causes of corrosion, small variations
in film formation layers and corrosion product layers (of the aqueous phase) also play a part.
For instance, in a scenario when all primary parameters are kept constant: temperature, partial
pressure, fluid velocity and pH, while the amount of corrosion product in the water phase is
varied, relatively high corrosion rates can be expected. (Dugstad et al., 1994a). The reason for
these high corrosion rates is attributed to changes in the properties of the thin layer of
corrosion products and scale which accumulate on the surface of the metal. The morphology
and composition of the film layer determine the type of corrosive attack – worst-case, low
corrosion under protective film or localised corrosion (mesa/pitting) (Dugstad et al., 1994a). It
is also said that the film layers can interact with the transport of corrosion inhibitors to the
surface of the metal thereby controlling inhibitor performance (Dugstad et al., 1994a).
In ‘worst-case’ (nonprotective film) corrosion, the presence of ferrous ions (Fe2+) ions in
solution leads to the formation of iron carbide film, Fe3C. This film deposits on the metal
surface, un-corroded and due to its naturally porous nature, offers no protection to the metal
surface underneath as seen in Figure 2-8, for the nonprotective Fe3C or Fe3C-FeCO3 films
(Dugstad et al., 1994a).
In most cases, it enhances CO2 corrosion as a result of it having a greater overpotential than
iron. It sets up a galvanic contact between itself and iron and this accelerates cathodic
reactions that lead directly to the anodic dissolution of the metal in the presence of << 1ppm
Fe2+ (Dugstad, 1998).
It may also lead to increased anodic dissolution of the metal by enhancing local acidification
around the metal surface – since cathodic reactions occur preferentially at iron carbide sites,
the aqueous phases in these regions become more alkaline as they are separated from the
metal. The net effect is a change in the water composition where cathodic regions become
more alkaline and anodic ones become more acidic leading to increased dissolution rates of
the metal by internal localised acidification. (Crolet, 1994; Dugstad, 1998).
26
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Nonprotective
Fe3C + FeCO3
Fe3C
Fe3C
Metal Metal
Protective
Fe3C
Fe3C + FeCO3
Fe3C + FeCO3
Metal Metal
Figure 2-8. Morphologies for protective and nonprotective scale formation (Crolet et al.,
1996, Kermani and Morshed, 2003)
It is worth noting that the FeCO3 and Fe3C films are the most commonly observed surface
films on carbon and low-alloy steels in CO2-containing environments (Kermani and Morshed,
2003). This film is protective when the FeCO3 phase covers the metal surface, effectively
sealing it completely as seen for the protective film in Figure 2-8, or if it is integrated within
the Fe3C phase.
Presence of H2S
CO2/H2S corrosion is widely encountered in the oil and gas industry. Hydrogen sulphide
corrosion, often termed as sour corrosion, occurs in fields whose oil and gas streams have
some amount of hydrogen sulphide satisfying the following condition: 𝑃𝐶𝑂2 /𝑃𝐻2 𝑆 < 200 (Yap
27
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
and Srinivasian, 2010). This is the condition for corrosive effect as stated in Table 2-2. It is
suggested that the simultaneous presence of both acid gases leads to a very aggressive
environment that can cause severe corrosion of steels but ultimately, the extent of corrosivity
is dependent on the aforementioned relative partial pressures of each gas with respect to each
other (Yap and Srinivasian, 2010).
In general, the higher the partial pressure of H2S, the lower the pH of the system since it is an
acid gas. When the system temperature, H2S and CO2 partial pressures are constant, the pH
increases with time, as the hydrogen ion concentration falls in the aqueous phase.
The presence of CO2 and H2S in bulk solution greatly influences the solution pH however it
should be noted that while pH is an important parameter for the estimation of corrosion rates,
the in-situ pH is of particular importance. The pH reported from results of water analyses is
not in-situ as the water analyses are carried out after samples have been exposed to the
atmosphere. This pH does not accurately describe the system (Yap and Srinivasian, 2010). In-
situ pH is said to be determined from ionic modelling of the solution or can be calculated
using the Brönsted concept as stated below (Yap and Srinivasian, 2010):
𝑊ℎ𝑒𝑟𝑒:
By the Brönsted concept, a Brönsted acid is a proton donor while a Brönsted base is a proton
acceptor (Naiman, 1948).
− 2− 2− 2−
H + = [HCO− 2− − 2− −
3 ] + 2[CO3 ] + [HS ] + 2[S ] + [OH ] − CHCO3 − 2CHS − CCO3 − 2CS (2.15)
The species in square brackets represent equilibrium concentrations which can be determined
by dissociation reactions for both CO2 and H2S. In this thesis, only the dissociation of H2S is
28
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
shown. The following set of reactions summarise the dissolution of H2S into the aqueous
medium and its subsequent dissociation:
𝐾𝐻
Dissolution of H2S in aqueous phase: H2 S (g) → H2 S (aq) (2.16)
𝐾1
Hydration of aqueous H2S: H2 S (aq) + H2 O(l) → HS − (aq) + H3 O+ (aq) (2.17)
K2
Dissociation of bisulphide ion: HS − (aq) → S 2− (aq) + H3 O+ (aq) (2.18)
The equilibrium constant terms, KH, K1 and K2 are then written and simplified as follows:
[H2 S]
𝐾𝐻 = (2.19)
pH2 S
Making [𝐻2 𝑆] the subject of the equation, thus [𝐻2 𝑆] = 𝐾𝐻 𝑝𝐻2 𝑆 (2.20)
[HS − ][H + ]
𝐾1 = (2.21)
[H2 S]
K1 [H2 S]
Making [HS − ] the subject of the equation, thus [HS − ] = [H+ ]
(2.22)
[S 2− ][H + ]
𝐾2 = (2.23)
[HS − ]
𝐾2 [HS− ]
Making [𝑆 2− ] the subject of the equation, thus [S 2− ] = [H+ ]
(2.24)
𝐾1 𝐾𝐻 p[H2 S]
Substituting eqn. (2.16) in eqn. (2.18), we have ⟹ [HS − ] = [H+ ]
(2.25)
𝐾1 𝐾2 𝐾𝐻 p[H2 S]
Substituting eqn. (2.25) in eqn. (2.24), we have ⟹ [S 2− ] = [H+ ]2
(2.26)
Once all the equilibrium constants are determined in this way, several measured parameters
such as temperature, pressure, CO2 and H2S mole fraction, ionic strength, Henry’s law
constants and equilibrium constants in conjunction with an accurate water analysis are
required to solve Equation. (2.11) (Yap and Srinivasian, 2010).
29
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Table 2-2.In a similar way to the CO2 corrosion mechanism, H2S dissolves in the aqueous
bulk solution to exhibit corrosive effects. Gaseous hydrogen sulphide exerts a partial pressure
thus enabling it to dissolve in solution to form bisulphide ion (HS-). The bisulphide ion then
undergoes further dissociation to form the sulphide ion (S2-), as shown as Equations. (2.16),
(2.17) and (2.18). These equations constitute the cathodic reactions.
30
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
A mechanistic model was developed based on results from the experiments. From the
research findings, it was discovered that iron sulphide scale in the form of mackinawite is
predominant over other forms of iron sulphide scale particularly at the initial stages of scale
formation. Also, H2S is observed to react with the metal surface directly by a ‘solid-state’
reaction which differs from the supersaturation-precipitation theory mode by which iron
carbonate scale is formed in CO2-rich environments. By ‘solid-state’ reaction, the implication
is that there is still evidence that iron sulphide precipitates out of solution following
supersaturation but in addition, a direct reaction between H2S and the metal surface is the
prevalent mode of combination. Furthermore, there are several scales formed in the H2S
environment in contrast to a CO2-only environment. Some of these scales are: amorphous
ferrous sulphide, mackinawite ((Fe, Ni)1+xS [where x=0 to 0.11]), cubic ferrous sulphide,
smythite (Fe3+xS4 [where x=0 to 0.3] or Fe3S4), greigite (Fe2+Fe3+2S4 or Fe3S4), pyrrhotite
(Fe1-xS), troilite (FeS) and pyrite (FeS2) (Sun and Nesic, 2007; James and Fleischer, 1966).
These are all various forms of iron sulphide scale but it is worth noting that the actual number
of these sulphides in existence is still a source of debate between mineralogists and
thermodynamicists (Smith and Joosten, 2006). These scales are covered in more detail in
Section 3.7.2 In addition, evidence from a preceding journal article: (Sun et al., 2006),
supports the fact that the supersaturation-precipitation theory does not always hold for H2S
corrosion, in that, corrosion rates always exceed precipitation rates and the explanation is that
iron sulphide scale is formed mainly by ferrous ions released from the metal surface by
corrosion and not by ferrous ions present in bulk solution.
In terms of the kinetics of iron sulphide film formation, it is concluded that H2S corrosion
rates generally increase with increasing concentrations of H2S however corrosion rates
decrease with increasing reaction time. The reason for the decrease of corrosion rate with time
is due to the formation of stable mackinawite scales whose thickness increases over time on
the metal surface (Sun et al., 2006).
31
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Both H2S and H+ species will travel to the metal surface from the bulk solution via convective
diffusion (turbulent eddies) and then by molecular diffusion.
Figure 2-9. Schematic of H2S corrosion – corrosion rate of steel in H2S solutions is said
to be under mass-transfer control. (Sun and Nesic, 2007)
32
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
A number of assumptions were made in the development of the mathematical model. These
assumptions are discussed here. Firstly, there is always a very thin but dense mackinawite
film of about << 1μm on the metal surface and it acts as solid state diffusion barrier for the
sulphide species involved in corrosion. Secondly, film growth is cyclical going through
growth, cracking and delamination stages. Also, the film (outer scale) continuously grows in
thickness over time and lastly, the outer scale is layered, very porous and is loosely attached
hence is prone to peeling and spalling (Sun and Nesic, 2007).
These assumptions coupled with the findings from experimentation have led to the conclusion
that using the concept of fluxes and the fact that the corrosion rate of steel in H2S solutions is
mass-transfer limited; at steady-state (equilibrium), all fluxes are equal.
𝛿0.5 1
𝐵𝐻 𝑆 𝑐𝑏,𝐻2 𝑆 − 𝐶𝑅𝐻2 𝑆 (𝐷 + )
− 2
𝑅𝑇 𝐻2 𝑆 𝜀𝜓 𝑘𝑚,𝐻2 𝑆
𝐶𝑅𝐻2 𝑆 = 𝐴𝐻2 𝑆 𝑒 𝑘 𝑙𝑛 (2.28)
𝑐𝑠,𝐻2 𝑆
Where:
𝑚𝑜𝑙
𝐴𝐻2 𝑆 , 𝐵𝐻2 𝑆 are the Arrhenius constants, 𝐴𝐻2 𝑆 = 1.30 × 10−4 𝑚2 𝑠 and 𝐵𝐻2 𝑆 = 15500 J/mol
𝑐𝑠,𝐻2 𝑆 is the concentration of H2 S on the steel surface and is set to 1.00 × 10−7 in mol/m3
𝛿𝑜𝑠 is the thickness of the mackinawite scale ⟹ 𝛿𝑜𝑠 = 𝑚𝑜𝑠 ⁄(𝜌𝐹𝑒𝑆 𝐴𝑜𝑠 ) in m
𝐷𝐻2 𝑆 is the diffusion coefficient for dissolved H2 S in water, 𝐷𝐻2 𝑆 = 2.00 × 10−9 , in m2 /s
𝑘𝑚,𝐻2 𝑆 is the mass transfer coefficient for H2 S in the hydrodynamic boundary layer, 𝑘𝑚,𝐻2 𝑆 =
1.00 × 10−4 in nearly stagnant condition, in m/s
33
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
This equation is non-linear with respect to 𝐶𝑅𝐻2 𝑆 but can be solved by the application of
appropriate numerical techniques.
These are also very important factors that include water-wetting, the presence of wax and wax
deposition, corrosion film characteristics, the effect of crude oil and erosion.
While experiments have been carried out to determine the magnitudes of corrosion of
associated steel pipelines in brine environments, it is not the same as carrying out these
experiments in the presence of the particular crude oil and brine. Differences are likely to be
present in measurements recorded for the two instances and these have often led to gross
errors when using corrosion test results to predict corrosion scenarios in the field.
Crude oil composition is therefore said to have an effect on CO2 corrosion. Experimental
research provided supporting evidence of this, which also showed that more sophisticated
analyses of the data presented in the study were required such as the use of artificial neural
networks (ANNs) (Hernandez et al., 2006). The article suggests the fact that though
interfacial tension and fluid dynamics play an important role in CO2 corrosion, the role of the
given crude oil is also significant, in that the composition of the latter affects interfacial
tension. (Hernandez et al., 2006). It is also suggested that wettability is strongly affected by
the presence of surface active compounds and these are believed to be polar molecules
containing oxygen (O), nitrogen (N) and sulphur (S) molecules.
The heavier the fraction of the crude, the greater the number of oxygen, nitrogen and sulphur
molecules because such crudes contain asphaltene and resins. In other words, the polar
molecules, O, N and S, in these crudes are as a result of the presence of polarisable
compounds, asphaltenes and resins (Hernandez et al., 2006). Polarisable compounds and polar
molecules thus exert a synergistic effect which changes the wettability of crude oil by
reducing the interfacial tension between oil and water therefore causing the system to be more
oil-wet. The net effect is a decreased dissolution rate of the metal at its surface because
34
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
corrosion of the metal proceeds at high rates when the metal surface is water-wet as opposed
to when it is oil-wet.
The prevailing flow regimes in a segment of pipe will depend on the fluid velocity through
the pipe, pipe orientation and geometry, pipe inclination and length as well as temperature and
pressure conditions. Once any of the aforementioned variables undergo a slight change along
the length of the pipe segment, then this change induces a change in flow regime.
Typical flow regimes found in oil and gas production facilities include stratified, wavy
stratified, rolling wave, plug flow and annular (Bondos et al., 2007). The significance of
knowledge of the flow regime, particularly of mixed phase fluids, is the fact that the type of
wetting occurring in any given scenario is easily determined. The key factors here are
oil/water ratio, emulsion tendency/stability and water cut percentages (Hernandez et al.,
2006). When water cut percentages are greater than 30%, water then becomes the continuous
phase and there are higher corrosion risks (Kermani and Morshed, 2003).
Erosion occurs when solid particles that may range in size such as sand accompany fluid flow
thus creating an abrasive effect on the metal surface (Giourntas et al., 2015). Erosion-
corrosion is a form of tribo-corrosion material loss mechanism caused by flowing fluid (in the
presence of solid particles) damaging both the surface layers such as the passive
film/corrosion products and the base metal (Hu et al., 2011). It involves electrochemical
corrosion processes and mechanical wear (Hodgkiess et al., 1999). This form of corrosion is
gaining greater attention not only because of its destructive nature but also this regime of
degradation has a greater likelihood of occurring particularly for old oil wells known to
produce high levels of sand in produced fluids.
Experimentation is often very useful when erosion occurs as a result of particulate fluid flow.
Currently, there are no industry guidelines that assess and control erosional corrosion caused
35
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
by particulate fluid flow however the commonly cited equation formulated by the American
Petroleum Institute (API): API RP-14E is used as a resource tool for evaluating pure gas-in-
liquid-erosion and is stated as follows (API, 1981):
𝑉𝑒 = 𝐶⁄ (2.29)
√𝜌𝑚
Where:
This equation is empirical and it is more often the case that any judgements made using the
equation in predicting corrosion risks are by the operator’s personal experience or discretion.
Molybdenum (Mo) has also been found to improve the corrosion-resistant characteristics of
steel alloys and while the addition of copper (Cu) also gives the alloys similar beneficiary
traits; it tends to reduce the effectiveness of corrosion inhibitors (Gulbrandsen and Nyborg,
2000). In general, the addition of small quantities of copper, nickel, chromium and possibly
molybdenum together have been found to improve corrosion-resistance of carbon steels
(Videm and Dugstad, 1989).
I. Lowering carbon (C) and adding carbide-forming alloying elements to maximise the
effect of any subsequent addition of chromium (Cr) and Molybdenum (Mo) by
ensuring that they remain in solid solution.
II. Achieving the desired properties by micro-alloying additions and mechanical and heat
treatments.
In microalloying, steel compositions are designed with low carbon content containing
microalloying elements such as titanium (Ti), niobium (Nb) and vanadium – stronger carbide-
forming metals. The reason for micro-alloying is that the aforementioned metals will combine
with carbon (in the given steel) preferentially thus leaving chromium and nickel uncombined
in the ferrite thereby enhancing corrosion resistance of the alloy.
Additionally, while the removal of carbon from these low-alloy steels is likely to reduce their
overall strength, the addition of silicon and nickel is used to restore the strength caused by the
removal of carbon. Extensive metallurgical studies have led to confirmed results in terms of
the degree of corrosion-resistance of each alloying element. A comparative trend is shown in
Figure 2-10.
37
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Corrosion Rate (arbitrary)
The following table categorises the effect of chromium in terms of addition percentages.
The following figure shows a schematic of CO2 corrosion performance for each of the
chromium-content categories discussed in Table 2-3.
38
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 2-11. Plot showing the category of chromium content of steel and corresponding
corrosion rate (Kermani et al., 2001; Kermani et al., 2003)
While optimum levels for the addition of chromium is still undetermined, a level between 2 to
3% Cr was considered essential to achieve good corrosion performance. It is also concluded
that Vanadium-microalloyed steel containing Cr, Si, Mo and Cu is the most promising
combination in terms of producing alloys with the desired features of strength and corrosion
resistance (Kermani et al., 2001; Grau, 2000).
Many research institutions and petroleum companies have developed CO2 corrosion models
over the years. Some of the oil companies that have contributed to the understanding of CO2
corrosion through the development of pipeline corrosion models include: Shell, Statoil, Total
and BP (Nyborg, 2010). In addition some research institutes have developed pipeline CO2
corrosion models and these are: the Institute for Energy Technology (IFE), OLI Systems and
Ohio University (Nyborg, 2010).
The models discussed in this Section include the HydroCor, Corpus, KSC and Multicorp
models. Other models such as the De Waard, Norsok M-506 and Freecorp Models are
discussed in detail in Chapter 3.
39
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
It should be noted that a separate CO2 corrosion model is used to generate a corrosion rate
profile for pipelines with multi-phase flow. Due to this fact, pH calculation is not included in
the Hydrocor model it therefore does not account for iron and bicarbonate concentrations.
It also has a separate model for computing corrosion rates. In general the Corpus model relies
on huge amounts of data to work out corrosion rates. The full CO2 corrosion description in
this case comprises results from the CO2 potential corrosivity index and the corrosion rate
values. It must be noted that the model does not account for any instances of oil-wetting.
Model operation is based on detailed water chemistry analyses including the effects of CO2,
organic acids and calcium.
It includes a facility for correcting calcium carbonate (CaCO3) concentration. Where user-
specified CaCO3 concentrations are likely to lead to CaCO3 super-saturation, the user is
warned to check the calcium carbonate concentration to reduce its magnitude. If this message
is ignored by the user, the program then proceeds to correct this by calculation of a pH of
lower magnitude (Nyborg, 2010).
40
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The model simulates chemical reactions in the aqueous phase, electrochemical reactions on
the metal surface and diffusion of species to and from the bulk solution, as well as diffusion
through porous iron carbonate films. Corrosion rate calculations are carried out to include the
risk of mesa attack. It includes a relatively strong effect of protective films, which is sensitive
to pH and temperature. For this reason, there is a tendency for it to predict low corrosion rates
for high temperatures and high pH (Nyborg, 2010).
Advancements to the original KSC model came by way of the development of a multiphase
flow model with precipitation of iron carbonate films and effects of oil wetting. The model
results were then further verified against laboratory and field data (Nesic et al., 2005; Nyborg
2010).
The model also includes the effects of organic acids such as acetic acid and H2S, including
iron sulphide film precipitation. It is based on detailed mechanistic modelling of the kinetics
of chemical and electrochemical reactions of species in the bulk solution and on the metal
surface respectively and the transient transport of species from the bulk solution to the metal
surface. The model is well suited for facilitating an understanding of the various CO2
corrosion mechanisms through the inclusion of corrosion inhibitor performance and the
effects of multi-phase flow in addition to the other facets mentioned.
41
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
2.6 CO2 Corrosion Mechanism between Low and High Partial Pressures
In the CCS industries, carbon dioxide is captured and transported at elevated pressures in
pipes for the purpose of controlling emissions of the gas in order to reduce its greenhouse gas
effects in the atmosphere. The gas is transported at pressures above those characteristic of oil
and gas pipelines; usually exceeding 20 bar, and as a single-component stream, to suitable
storage sites at secure locations for geologically significant timescales (Downie et al., 2007).
CO2 can be transported on the liquid side or on the vapour side of the vapour/liquid line
running between the critical point (74bar, 31oC) and the triple point (5bar, -56oC) while not
allowed to cross the dashed line as shown in Figure 2-12 (IEA GHG, 2010). It is often
transported as a dense-phase liquid at elevated pressures for convenience and efficiency
(Downie et al., 2007). However assessing the corrosion risks for pipelines at these elevated
pressures is challenging and potentially unpredictable due to the non-ideal dissolution of CO2
in the liquid phase coupled with the presence of impurities such as SOx and NOx gases with
traces of water.
Figure 2-12. Phase diagram for pure CO2 – the captured gas can be transported strictly
on either side of the dashed line without cross over. Dense-phase region represents the
most efficient means of pipeline transport. (IEA GHG, 2010)
It should be noted that prior to transport, the captured gas has to be elevated to pressures
above its critical point however the presence of impurities in the stream influences the critical
42
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
temperature and pressure as well as the width of the phase envelope that consequently dictates
the region of two-phase flow as shown in Figure 2-13.
Increasing percentage content of impurities in the CO2 stream is known to reduce the
supercritical phase of the gas which impacts on the operating region of the pipeline thus the
pipeline may have to be operated at higher pressures in order to maintain single-phase flow of
the gas stream (supercritical/dense phase), for instance. The presence of impurities also
affects the density and viscosity of the stream and the extent of their influence are dependent
on the type, quantity and combination of impurities present (Downie et al., 2007). A well-
documented implication of the presence of impurities is its effect on the temperature drop and
consequent hydrate formation (Downie et al., 2007).
departure of real gases from ideal gas behaviour. The acidity of the aqueous solution is thus
known by converting the solubility values to solution concentrations.
It is also possible to determine the acidity of the aqueous media for mixed CO2-H2S, CO2-SOx
and CO2-NOx, by application of the same principles stated for the CO2-H2O system.
I II
Pressure (bar)
Figure 2-14. Solubility of CO2 in H2O (Spycher et al., 2003)
There is therefore a need to take into account the non-ideality of the gas phase as it dissolves
in the aqueous phase. Due to the fact that co-solubility of carbon dioxide in water and water in
carbon dioxide are the means by which CO2 corrosion occurs, solubility of these species is
thus investigated in terms of solution thermodynamics as is briefly outlined in the following
Section. The methodology for accurate supercritical CO2 corrosion modelling is discussed by
(Choi and Nesic, 2009) as well as (Mohamed et al., 2011).
44
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Of all these reactions, the dissolution of CO2 is briefly discussed because it was derived from
basic thermodynamic relationships. The dissolution of carbon dioxide from its gaseous form
to its aqueous form is summarised by the following reaction:
𝐾𝐻
CO2 (g) ⇔ CO2 (aq) (2.30)
KH is the Henry’s law constant and its unit is mol/ (litre x bar)
Where:
Accounting for non-ideality of the gas phase, the following equations are derived from the
definition of fugacity and applying necessary corrections for the pressure:
𝑉𝑖 (𝑃−𝑃 𝑜 )
The correction for K is given by: 𝐾(𝑇,𝑃) = 𝐾(𝑇,𝑃) 𝑒𝑥𝑝 { } (2.34), and serves as a
𝑅𝑇
Where:
𝑃, 𝑃0 and 𝑃𝑡𝑜𝑡 are the applied pressure, atmospheric pressure (0.101325 MPa) and the
𝑇 = Absolute Temperature in K
These equations were then used to solve for the fugacity coefficients so that the volumes of
the compressed CO2 gas and H2O are then calculated using an equation of state. In this case,
the Redlich-Kwong equation was used and yielded cubic equations which were then solved
using the Newton-Raphson numerical algorithm (Mohamed et al., 2011). The solubility
constant, KH was then computed. The Redlich-Kwong EOS is stated as follows (Redlich and
Kwong, 1949):
𝑅𝑇 𝑎
𝑃= − (2.35)
𝑉 − 𝑏 𝑉(𝑉 + 𝑏)𝑇𝑘0.5
Where:
𝑉 = Volume (in m3 )
46
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Artificial neural networks are information-processing units that are quite similar in terms of
mode of operation to biological nervous systems such as the brain and it’s highly complex
network of neurons – artificial neural networks (ANNs) are inspired by their biological
counterparts (Haykin, 1999; Krogh, 2008). Essentially, neural networks are models which can
be ‘trained’ to forecast, by developing a correlation between a known set of input variables
and output problem descriptors. They obtain a correlation from the use of transfer functions
thereby assigning ‘weighted scalars’ to input data which is subsequently matched to a given
output. Matching is achieved by adjusting and re-adjusting the weighted scalars accordingly
in a continuous loop, thereby reducing the error between the target and output as seen in
Figure 2-15. In so doing, the neural network learns and trains such that it utilises its
‘experience’ in predicting the outcome for a different set of input data.
Figure 2-15. Flow loop of input to output via continuous adjustment of weights
Often neural networks (NNs) are designed in the form of neural architectures. A typical
layered structure of a NN is shown in Figure 2-16. A NN may contain as many layers as
necessary depending on the task to be carried out. However, if a NN has been trained with a
large number of layers on a small dataset, then there is the risk of over-fitting. When a NN is
over-trained in this manner, it loses its ability to adapt and fit a wider range of datasets and as
such is said to lose its ability to generalise (Radonja and Stankovic, 2002).
47
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
As earlier stated, the transfer function determines the correlation that maps input data to
output. For this reason, the type of transfer functions to be used should be based on the
characteristics of the given input and output data, such as the range and magnitude of each
data-value within the dataset (Vogl et al., 1988). The purelin and log-sigmoid transfer
functions are shown in Figure 2-17, on the left and right respectively.
+1 +1
n n
O O
-1 -1
Figure 2-17. The purelin and log-sigmoid transfer functions (Beale et al., 2014)
The purelin transfer function is often used in the last layer of a multilayer NN as function
approximator while the logsig transfer function is used in the hidden layer of a multilayer NN
because it is differentiable (Beale et al., 2014). Training algorithms are a set of instruction
code that controls the optimisation method for a given NN. They therefore control how the
48
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
weights of a NN are set (Beale et al., 2014). The definitions of some training algorithms are
shown in Table 2-4.
Table 2-4. Definitions of the training functions used (Beale et al., 2014)
The neural network employed was a simple 4-node network with a single hidden layer.
Training of this network involved making use of two-thirds of the experimental data available
while the remaining one-third of the data was used for validation purposes to signal the
termination of training. This resulted in a network with a total of 75 weights and plots were
then produced to observe the effect of each ion on pitting corrosion potential. Plots were
produced in order to compare experimental results by (Man and Gabe, 1981), with those of
the neural network where experimental results were measured using the same material and
test method as was used to produce the training data (Cottis et al., 1999). One such plot is
shown in Figure 2-18.
49
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 2-18. Plot showing the effect of sulphate ion on pitting potential of steel in 3%
NaCl solution. (Cottis et al., 1999)
Overall, the training dataset was small therefore there was the need to design a small neural
network such that it relies on generalisation as opposed to memorisation. The neural network
performs reasonably well in situations where elements of extrapolation are required when
modelling problems with small data sets.
Furthermore, based on observations from these studies it can be said that, the performance of
neural networks is largely dependent on the quality of training data however given that
corrosion data is largely inconsistent and is known to contain a substantial amount of ‘noise’,
results are not likely to be meaningful as are likely to be models of the noise rather than that
of the average behaviour (Cottis et al., 1999).
Noise in corrosion data may be as a result of the use of Monte Carlo techniques to fill gaps in
summarisation plots, the use of faulty equipment, poor conduction of experiments and a
failure to report significant variables while carrying out corrosion experiments (Cottis et al.,
1999). Monte Carlo techniques, in this instance, are used to improve upon the quality of the
reported data and the appearance of summarisation plots before the application of neural
networks for corrosion prediction. In another study in which sulphuric acid corrosion of steel
was investigated, the question of validation was raised (Helliwell et al., 1996). Validation had
been carried out to a limited extent. It is also apparent that the degree of confidence that could
be placed in prediction was questionable.
50
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
In conclusion, there is a need to state the confidence that can be placed in the prediction of a
given neural network and it is also necessary to predict both model variability and mean
behaviour though it will require copious amounts of training data sets.
Fuzzy logic systems provide a language with syntax and semantics that can translate
qualitative knowledge into numerical reasoning. Fuzzy linguistic variables include low, high,
good and so on and when these are combined with fuzzy set operators such as and/or, the
process is termed approximate reasoning (Hajizadeh, 2006).
Approximate reasoning is a term used when referring to fuzzy logic due to the fact that it does
not deal in absolutes. In fact, it is based on binary logic, but it differs from it. Where binary
logic assigns 0 or 1, for a false or true statement, fuzzy logic instead assigns an approximate
value in the range: 0-1.
Essentially, fuzzy systems process input data converting it to output data according to the
steps shown in Figure 2-19. Fuzzification involves assigning linguistic variables to input data,
often applying ‘if-then’ rules as conditional statements. The ‘if-then’ rule is discussed more in
Section 2.8.1. The implication phase assigns a membership function to the data. This is the
mapping of the data into the range: 0-1, also known as categorising the input data into degrees
of membership (Zadeh, 1965), as seen in Set X of Figure 2-19. It is worth noting that
membership functions may take several forms: triangular, trapezoidal and continuous among
others. These are discussed further in Section 6.3.
The membership function selected for a given input dataset depends upon the distribution of
the data within the dataset. Aggregation is the process of putting all fuzzy sets together in an
aggregate fuzzy set and defuzzification translates the fuzzy information back to readable form
as an output, as seen in Figure 2-19.
There are two main types of fuzzy inference systems (FIS), and these are the Mamdani and
Takagi-Sugeno systems. These will be discussed further in the next Section however the main
difference between these two systems is in the defuzzification step. The Mamdani re-
transforms fuzzy sets to output using a defuzzification technique – directly striping away
fuzzified data to reveal the output.
51
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The Takagi-Sugeno system on the other hand, makes use of a weighted average technique to
convert fuzzified data into output (Kaufmann and Gupta, 1985; Fuzzy Logic Toolbox User’s
Guide, 2015). In general, fuzzy logic systems excel particularly during instances where gaps
in numerical banks of data exist and where in such cases, exact measured process variables
are considered too imprecise to justify the use of numbers. This is often quite common in
several engineering problems (Hajizadeh, 2006).
A simple set of if-then rules was used in the execution of a fuzzy set in which the input
variables are classed as members or non-members given that a fuzzy set is defined as a
collection of ordered pairs of the form A={x, μ(x)}, where A describes the relationship
between an unknown quantity, x, and a membership function, μ(x). It should be noted that in
general, fuzzy sets can be divided into crisp boundary and non-crisp boundary fuzzy sets but
for each type, quantities are classed as members or non-members.
In terms of the execution of the if-then rules of a fuzzy system, the closer an argument is to
the ‘if’ part, the greater the influence the ‘then’ part has on the mapping of input variables.
52
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
For instance, for the statement: ‘If temperature is low, then corrosion rate is low’; for small
input temperature values, the corrosion rates will definitely be assigned small magnitudes.
The fuzzy system then sums up the ‘then’ parts and implements a defuzzification system
which yields an output. The Takagi-Sugeno (TS) inference system was used for this research
study in which 50 different corrosion rate data points are used for training however a choice
can be made between its use and the use of the Mamdani inference system. The main
difference between these inference systems is that the former involves the use of only
constant output membership functions however the TS system can be applied to mimic the
commonly-used Mamdani by implementing a hyperbolic transfer function (Hajizadeh, 2007;
Fuzzy Logic Toolbox User’s Guide, 2015).
Furthermore, three-dimensional (3-D) surface contour plots were produced for the given user-
input parameters and the trends were found to be in general agreement with research
literature. Two such plots are shown in Figure 2-20. In addition, the 3-D plots can be used in
CO2 corrosion rate prediction and can also serve as an informative tool which lends itself to
the understanding of specific relationships between model input variables and the output,
corrosion rate (Hajizadeh, 2007).
Figure 2-20. 3-D Surface contour plots showing the relationships between input
variables (CO2 mole % with temperature and gas production rate with temperature)
and the output variable, CO2 corrosion rate (Hajizadeh, 2007)
53
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
2.9 Conclusions
Since the discovery of CO2 corrosion in gas wells located in Texas, USA, in the 1940’s,
several research undertakings on this subject have been carried out. CO2 in the dry gas form is
non-corrosive at oilfield and CCS conditions. It however becomes corrosive when it dissolves
in the aqueous phase, forming carbonic acid. It is particularly more corrosive when localised
mesa-type attacks take place. Mesa-type attacks occur in low fluid-flow environments, in
which Fe3C scale is somewhat prevalent. It is also very dependent on the composition of fluid
flow. Fe3C scales provide inadequate protection, thus expose the metal surface underneath to
severe corrosive attack.
The major factors influencing CO2 corrosion were also discussed and these were broadly
categorised into: environmental, physical and metallurgical factors. The environmental factors
were also discussed with respect to actual oil and gas producing fields in the North Sea; data
contained in Table 2-1. Explanations were provided for the quoted corrosion rate estimates
given in Table 2-1 in terms of the corrosion rate-determining factors, i.e. the environmental
factors. An increase in the magnitudes of temperature and CO2 partial pressures were
observed to lead to elevated corrosion risks for pipelines. For pH, the reverse was the case,
low pH implies high corrosivity while for fluid flow; very high velocities appear to cause
corrosion rates to plateau, as seen in Figure 2-7D.
The condition for corrosivity with regards to a H2S-containing environment is: 𝑃𝐶𝑂2 /pH2S <
200. Also, H2S corrosion is fundamentally different from CO2 corrosion, in that it is not
electrochemical whereas the latter is. H2S corrosion proceeds via direct combination of the
gas with the metal in a ‘solid state corrosion reaction’ in which species transport from the
bulk solution through boundary layers to the metal surface dictates the rate of corrosion.
For accurate modelling of CO2 corrosion at high pressures, the application of appropriate
equations of state is required. Some research directives were stated, one of which was the
determination of the pH for a mixed CO2-H2O system.
The use of NNs for the prediction of low pressure CO2 corrosion rates is well-documented
however in this thesis; it is also used for the prediction of CO2 corrosion rates under high
pressure conditions. Some reference was made to the training of neural networks for small
datasets in order to retain their ability to generalise. High pressure CO2 corrosion datasets in
open literature are very small in size therefore the need to adapt the NNs for this particular
54
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
purpose is crucial. NN modelling of high pressure CO2 is discussed in the chapter on neural
network models.
For fuzzy inference systems (FIS), the process in which input data is converted to output data
was discussed. FIS employ approximate reasoning techniques to problem-solving through the
use of human language descriptors and assigning values between: 0 to 1 to represent these
descriptors. The way in which the assignments are carried out is governed by the use of the
membership function. It is imperative to understand the data being modelled in order to be
able to select an appropriate membership function. The advantage this modelling technique
has over the others is its ability to directly visualise the various relationships between input
data variables in both two-dimensional and three-dimensional space.
55
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
56
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
57
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
A large number of models have been developed by oil companies for the purpose of
predicting pipeline CO2 corrosion (Nyborg, 2002; Nyborg, 2010). In Section 2.5 of the
previous chapter, some established CO2 corrosion models were discussed such as Shell’s
HydroCor, Total’s Corplus, IFE’s KSC model and Ohio University’s Multicorp models.
These models differ from each other because of the different philosophies behind their
development. One way to classify these models is by conservatism. The concept of
conservatism is explained by the scale formation effect. For those models described as
conservative or known to over-predict corrosion rates, these models are said to take little or
no account of the effect of the corrosion film (passive layer) on the overall CO2 corrosion
mechanism. It is known that, particularly at high temperatures, scale formation layers are
more stable and compact hence the net effect on the surface of the metal is that corrosion is
reduced.
The less conservative models often assume the influence of film and scale formation hence
generally predict lower corrosion rates than conservative models. It is also said that the less
conservative models also assume oil-wetting tendencies as opposed to the mostly water-
wetting tendencies of conservative models (Nyborg, 2002; Nyborg, 2010). The aqueous phase
is the continuous phase in contact with the metal surface for water-wetting systems hence
corrosion risks are greater than for oil-wetting systems whose continuous phase is oil.
In general, these two factors – the effect of corrosion products/film layer and oil/water
wettability are the main distinguishing features in the modes of operation of the several
models developed for CO2 corrosion prediction. In this chapter, the corrosion models
discussed fall into one of two categories – empirical or mechanistic. The former implies
models developed with the use of raw experimental or field data to extract underlying
relationships between measured variables while the latter involves the development of a
prediction system based on the mechanisms that underpin the physical phenomena being
investigated. The 1975, 1991, 1995 De Waard and Norsok models are empirical models
(Nyborg, 2010) while the Nesic-Postlethwaite-Olsen (NPO) model is purely mechanistic
(Nesic et al., 1996).
Also, in Section 2.4.1, as part of the environmental factors affecting CO2 corrosion, the
presence of H2S was discussed. The scenarios for the severity of corrosion risks in the
58
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
presence of H2S as well as the mechanics and kinetics of formation of iron sulphide scale
were also outlined. In this chapter, H2S corrosion is discussed in terms of the use of Ohio
University’s freely accessible Freecorp software (Ohio University, 2008) to predict
corrosivity for H2S corrosion data in research literature (Lyle and Schutt, 1998; Nesic et al,
2008).
In the mechanics and kinetics of iron sulphide scale formation of Section 2.4.1, some H 2S
scale types were listed, one of which was mackinawite. In Section 3.7.2, with the aid of the
aforementioned Freecorp software, the H2S solution concentration is expressed in terms of the
distance of corroding species (H3O+ and H2S) from the steel surface. Mackinawite is said to
constitute the inner film formed on the steel surface. The Freecorp model is also used to
predict corrosivity for CO2 corrosion data in research literature (Crolet et al, 1999; Guo et al,
2005) as it incorporates the electrochemical philosophy of the NPO model.
1710
log10 (𝑉𝑐𝑜𝑟𝑟 ) = 5.8 − + 0.67 log10 (𝑃𝐶𝑂2 ) (3.1)
𝑇𝑘
Where:
𝑇𝑘 = Temperature (K)
𝑃𝐶𝑂2 = carbon dioxide partial pressure (MPa); 𝑃𝐶𝑂2 = 𝑚𝑜𝑙% 𝐶𝑂2 × 𝑡𝑜𝑡𝑎𝑙 𝑝𝑟𝑒𝑠𝑠𝑢𝑟𝑒(𝑃)
The correlation expresses corrosion rates in units of mm/year in terms of temperature, T and
carbon dioxide partial pressure, 𝑃𝐶𝑂2 . A typical plot for the correlation is shown in Figure 3-1.
59
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 3-1. Plot of corrosion rate against temperature for varying carbon dioxide partial
pressures, 𝑷𝑪𝑶𝟐 = 𝟎. 𝟎𝟏, 𝟎. 𝟎𝟑, 𝟎. 𝟏 𝐚𝐧𝐝 𝟎. 𝟑𝐌𝐏𝐚 (without inhibition from carbonate
scale)
From Figure 3-1, it is observed that corrosion rates increase as a power function – power of
10, as the temperature increases. Herein lies the limitation of the model, in that typically plots
of corrosion data from laboratory results as well as field data show a distinct peak at
approximately 60-90oC followed by a decrease in magnitude of corrosion rate with a further
increase in temperature beyond 90oC, as shown in Figure 3-2.
Figure 3-2. Typical plot of corrosion rate against temperature from experimental CO2
corrosion data (Dugstad et al., 1994b; Halvorsen et al., 1999; Sun and Nesic, 2004)
60
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
However for the plot in Figure 3-1, there is no peak for this range of temperatures hence the
reason a subsequent correlation was formulated in order to account for the film formation
effect.
The other limitation of this corrletion is the use of carbon dioxide partial pressures instead of
fugacity. At higher CO2 pressures, such as for pressures that exceed 20bar, for instance, a
more appropriate and accurate approach will involve the use of the concept of fugacity as
opposed to partial pressure (de Waard et al., 1991).
The 1991 correlation is basically the same as the 1975 equation in terms of statement and
variables with the exception of the fugacity term, fCO2, which replaces the previous carbon
dioxide partial pressure term, 𝑃𝐶𝑂2 . The equation is stated as follows (de Waard et al., 1991):
1710
log10 (𝑉𝑐𝑜𝑟𝑟 ) = 5.8 − + 0.67log10 (𝑓𝐶𝑂2 ) (3.2)
𝑇𝑘
Where:
𝑇𝑘 = Temperature (K)
Application of this model involves the use of correction factors which are associated with a
physical or chemical effect that causes a slight change in the corrosion prediction result (de
Waard et al., 1991). The physical or chemical effect includes parameters affecting CO2
corrosion such as pH and scale formation, which are not stated directly in Equation 3.2 as
variables. Therefore, separate corrrection factors for each of these parameters are determined
Generally, for any physical or chemical effect being corrected for, the correction factor, which
is usually less than unity (<1), is multiplied by the result obtained from the Equation 3.2 and
this often yields corrosion rates of decreased magnitudes as compared to the original predicted
results from Equation 3.2, which are considered to be over-conservative (de Waard et al.,
1991). Specifically, the correction for pH will tend to yield correction factor results that
exceed unity (>1) for media that are under-saturated with corrosion products, Fe2CO3 or
Fe3O4. The reason being that for under-saturated systems, there is a steady anodic dissolution
61
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
rate of iron on its surface and this proceeds until the system is sufficiently saturated after
which the rate of dissolution slows down. Thus corrosion rates are higher for under-saturated
systems because film formation has not yet proceeded to the point where it inhibits further
corrosion at the metal surface therefore any pH correction factors obtained will exceed unity.
These FeCO3 films are formed and adhere onto the metal surface and reduce the surface area
of the metal exposed to corrosive attack thereby reducing corrosion rates and resulting in
temperature maximum (peak) for corrosion rates. The temperature corresponding to the
maximum corrosion rate is the scaling temperature and is given by Equation 3.3. At this
temperature, ferrous ion concentration (Fe2+) formed at the metal surface and local pH are
such that a protective film is formed. (de Waard et al., 1991)
2400
𝑇𝑠𝑐𝑎𝑙𝑒 = (3.3)
6.7 + 0.6log10 (𝑓𝐶𝑂2 )
Where:
While film formation can occur at any temperature, its protection of the metal surface is
greater at higher temperatures. At lower temperatures, for instance, temperatures less than
60oC, the corrosion product film has a smudge-like appearance and is easily removed by
flowing liquids while at higher temperatures, the texture is different, usually coarser and is
less easily washed away (de Waard et al., 1991). It is also worth noting that flowrate has an
effect on scaling temperature; a higher flowrate will result in a higher scaling temperature.
Also, a higher bulk pH will tend to lower this temperature.
62
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 3-3 shows the plot of corrosion rate against temperature for varying CO2 fugacities.
Figure 3-3. Plot of corrosion rate against temperature for varying CO2 fugacities,
𝒇𝑪𝑶𝟐 = 𝟎. 𝟎𝟏, 𝟎. 𝟎𝟑, 𝟎. 𝟏𝐚𝐧𝐝 𝟎. 𝟑𝐌𝐏𝐚
For the ascending part of the curve, Equation 3.2 is applied, however for the descending part
of the curve, Equation 3.5 is applied. Equation 3.5 is derived as follows:
Multiplying the right hand side of Equation (3.2) by the left hand side of Equation (3.4),
yields the result for calculating the new corrosion rate, Equation (3.5):
1710 2400
log10 (𝑉𝑐𝑜𝑟𝑟 ) = {5.8 − + 0.67 log10(𝑓𝐶𝑂2 )} × { − 0.6log10 (𝑓𝐶𝑂2 ) − 6.7} (3.5)
𝑇𝑘 𝑇𝑘
The scaling temperature is evaluated from Equation 3.3 using the given CO2 fugacity, to
determine where the ascending part of the curve ends and where the descending part begins. It
is observed from Figure 3-3, that the scaling temperatures for each of the CO2 fugacities is
different, the higher the CO2 fugacity, the smaller the magnitude of its corresponding scaling
temperature and this is due to the fact that at lower temperatures, any scale formation is
63
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
washed away. This therefore results in greater corrosion rates for high CO2 fugacities as the
rate at which scale is being washed away is also high. These equations were implemented in
Matlab and verified against results published in research literature (Woollam and Hernandez,
2006).
[𝐶𝑂2 ]
𝐶𝐶𝑂𝑅𝑅 = (3.6)
1 1
+
𝑘𝑟 𝑘𝑚
Where:
𝑘𝑟 and 𝑘𝑚 are rate constants associated with reaction kinetics of the corrosion reaction,
the charge transfer reaction and the mass transfer of dissolved 𝐶𝑂2 from
Equation 3.6 can be expressed as Equation 3.7, shown below. Equation 3.7 is the overall
resistance model.
1 1
𝐶𝐶𝑂𝑅𝑅 = + (3.7)
𝐶𝑅 𝐶𝑀𝑇
Where:
𝐶𝑅 is the highest possible reaction rate, i. e. when mass transfer is infinitely fast
64
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Hence CR, also known as the contribution from activation reaction kinetics is associated with
‘worst-case’ corrosion and takes the form of the previously established 1991 de Waard-Lotz-
Milliams correlation:
1119
log10 (𝐶𝑅 ) = 4.93 − + 0.58 log10 (𝑓𝐶𝑂2 ) (3.8)
𝑇𝑘
𝑈 0.8
Also, CMT is given by: 𝐶𝑀𝑇 = 2.45 𝑑0.2 (3.9)
Where:
The pH term stated in the correlation in research literature, is neglected here as this
correlation was used for condensed water whose pH is determined by the solubility and
dissociation of CO2 (Woollam and Hernandez, 2006). The plot of corrosion rate against
temperature in Figure 3-4 shows the shape and trend for each CO2 fugacity value.
Figure 3-4. Plot of corrosion rate against temperature for varying CO2 fugacities
𝒇𝑪𝑶𝟐 = 𝟎. 𝟎𝟏, 𝟎. 𝟎𝟑, 𝟎. 𝟏𝐚𝐧𝐝 𝟎. 𝟑𝐌𝐏𝐚
65
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
It is observed from Figure 3-4, that corrosion rates increase as CO2 fugacity increases from
0.01 to 0.3MPa (0.1-3bar). The reason being that greater CO2 fugacities translate into greater
concentrations of the gas in bulk solution and consequently greater carbonic acid
concentrations. This in turn leads to greater corrosion rates hence greater dissolution rates of
the metal at the anode.
Also, it is observed that the scaling temperatures increase as CO2 fugacity decreases from 3
bar to 0.1 bar. This is due to the fact that at higher temperatures, more stable film layers are
formed that are considerably harder to wash away (de Waard et al., 1991). At lower
temperatures, scale takes up a smudge-like texture and does not offer protection to the metal
surface hence a greater proportion of the metal surface is exposed to corrosive attack leading
to greater corrosion rates. Scaling is more likely to occur at lower temperatures for greater
CO2 fugacities because corrosion rates are higher, therefore the bulk solution is likely to get
super-saturated more readily before the effect of high temperatures set in – solubility
decreases with increasing temperature.
The plot in Figure 3-5 shows the corrosion profile with temperature for 𝑓𝐶𝑂2 =1MPa (10bar),
T=10-100oC, pH=5.0 and varying flow velocities. It is observed that corrosion rates increase
with increasing temperatures regardless of the fluid flowrate velocity until the scaling
temperature is reached (~63oC) after which corrosion rates decrease gradually. It is generally
known that corrosion specie solubilty decreases with increasing temperature and upon
reaching the scaling temperature of 63oC, supersaturation takes place and as temperatures
proceed to increase further, there is said to be suffcient scale build-up on the metal surface
such that the effective area of the metal exposed to corrosive action is reduced hence
corrosion rates fall even though it follows from Arrhenius theory, that increasing temperatures
lead to an increase in reaction rates.
66
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 3-5. Plot of corrosion rate against temperature (Conditions are: T=10-100oC,
𝒇𝑪𝑶𝟐 =1MPa, pH=5.0 and for varying flowrate velocity)
For the plot of corrosion rate against temperature for T=10-100oC, U=6.0m/s, pH=5.0 and
varying carbon dioxide partial pressures, Figure 3-6, there is an increase in corrosion rates
with increasing temperatures until the scaling temperature is reached after which there is a
drop in the magnitudes of corrosion rates. Carbon dioxide partial pressure to the power of 0.7
is said to have a direct proportional relationship with corrosion rate (Dugstad et al., 1994a), as
seen in Figure 2-7A, Section 2.4.1 – there is an increasing trend for each curve in that the
greater the CO2 partial pressure, the greater the corresponding corrosion rate and this is a
result of greater concentrations of carbonic acid forming. Given that carbonic acid readily
dissociates into the bicarbonate ion and hydrogen ion, it implies that a greater concentration of
carbonic acid will result in a greater concentration of hydrogen ions in solution thereby
making the solution more acidic thus enhancing CO2 corrosion.
67
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 3-6. Plot of corrosion rate against temperature (Conditions are: T=10-100oC,
U=6.0m/s, pH=5.0 and for varying carbon dioxide fugacities)
This is an empirical model developed by the Norwegian oil companies, Statoil, Norsk Hydro
and Saga petroleum. It is originally based on laboratory data that was previously used to
calibrate the de Waard model, however in addition it is capable of estimating corrosion rates
at temperatures of up to 150oC (Nyborg, 2010). The model is said to yield ‘worst-case’ CO2
corrosion rate results since the flow loop experiments upon which it is based assume low
ferrous (Fe2+) ion concentration in the aqueous phase (Olsen et al., 2005). The model equation
can be summarised as follows Woollam and Hernandez (2006), Norsok Standard M-506,
(2005):
0.62
𝑆 0.146+0.0324log(𝑓𝐶𝑂2 )
𝐶𝑁𝑂𝑅 = 𝐾𝑡 𝑓𝐶𝑂 ( ) 𝑓(𝑝𝐻)𝑡 (3.10)
2
19
Where:
68
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The model takes account of the formation of protective films at higher temperatures as seen in
the plot in Figure 3-7. This Figure was obtained from the open-source Norsok M-506,
accessed by a Microsoft Excel spreadsheet (coloured lines with markers). The NORSOK
source code was then written, implemented and plotted separately using Matlab 2012a
software (coloured lines without markers). These two plots were then superimposed together.
Corrosion rates are observed to increase with increasing temperature after which there are
peaks at a temperature of about 78oC followed by a dip in magnitudes. The peaks in corrosion
rate are due to the effect of the presence of protective film layers on the surface of the metal
which have the dual action of reducing the number of sites on the metal surface available to
corrosive attack as well as forming a compact shield of ferrous carbonate on the metal surface
that is neither easily swept away by fluid flow nor does it offer channels through its structure
for the possibility of localised attack.
69
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
One limitation of the model is that it does not account for oil-wetting and its use is not
applicable to systems in which pH stabilisation is used for corrosion control (Nyborg, 2010).
Figure 3-7 shows the Matlab 2012a model results (shown as blue, gree and cyan lines)
overlaid on the Norsok M-506-generated results (shown as magenta circles, black crosses and
red stars). The code was written based on identical parameters as indicated in published
research literarture to validate Matlab 2012a-generated results. (Woollam and Hernandez,
2006). Other conditions required to obtain the plot shown in Figure 3-7 are tabulated in Table
3-1 as follows:
Table 3-1. Summary of variables and their respective magnitudes required to obtain
results in figure 3-7
Parameter/unit Magnitude
pH 4.0
Pipe Diameter (mm) 100
Water cut (%) 20
3
*Liquid volumetric flowrate (m /s) 0.0307
CO2 mole (%) 100
*Superficial liquid velocity (m/s) 3.9063
* These parameters have been assumed. Superficial liquid velocity is calculated from an assumed value of liquid
m Volumetric flowrate (m3 /s)
volumetric flowrate using the formula: Velocity, 𝑈 ( ) = . 𝑇𝑘 /Tstd was assumed to be
s Area (m2 )
approximately equal to unity given that the instantaneous temperature of the gas in Kelvin, 𝑇𝑘 , is almost
identical to the temperature of the gas at standard conditions, Tstd.
All other parameters such as pipe roughness, oil specific gravity, density and viscosity of
water, density and viscosity of oil as well as compressibility have been assigned identical
values as the default Norsok M-506 corrosion rate model. There are slight differences in the
superimposed plots and these are likely due to the assumed figures used in modelling.
70
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
3.5.2 Parameter Study on the NORSOK M-506 Model – Investigating Changes in Water
Cut Percentage and CO2 Mole Percentage on CO2 Corrosion Rate
Upon observation of the plot in Figure 3-8, corrosion rates increase with increasing
temperature as expected due to the reasoning behind the Arrhenius theory (Section 3.7.3);
there is a temperature dependence on the rate constant hence the chemical reaction, CO2
corrosion in this context is dependent on temperature change.
In general, the greater the water cut percentage in the fluid, the greater the rates of corrosion.
This is explained by the relative ease of formation of carbonic acid when the water cut
percentages are greater and the fact that the pressure or shear stress exerted on the metal
surface or pipe walls is greater for conditions of high water cut. Fluid flow with a great water
cut percentage, say 70%, implies that a bigger force is exerted per unit area of the pipe
internal wall than the same fluid flow with 30% water cut, for instance. For the plot shown in
Figure 3-9, corrosion rates increase with increasing CO2 mole percentages due to the fact that
there is an increased likelihood for the formation of carbonic acid in solution of a greater
71
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
concentration hence a CO2 mole percentage of 90% would induce greater corrosion rates than
one of 70%, for instance.
This is a mechanistic model that combines the equations for the cathodic reactions which
include the reduction of H+ and the reduction of H2CO3 against a single current equation for
the anodic dissolution of Fe. Hence, when the sum of cathodic current equations is said to be
equal to the anodic current equation, therefore the rate of corrosion can be determined (Nesic
et al., 1996; Woollam and Hernandez, 2006):
𝑖𝐻 + , 𝑖𝐻2 𝐶𝑂3 are the cathodic current terms for hydrogen ion and carbonic acid respectively
𝑖𝐹𝑒 is the anodic current term for the dissolution of the metal (Fe)
72
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
1 1 1
= + 𝑑 (3.12)
𝑖(𝐻 +) 𝑖𝛼(𝐻 + ) 𝑖𝑙𝑖𝑚(𝐻 +)
Where:
𝑑
𝑖𝑙𝑖𝑚(𝐻 + ) is the limiting current
𝑑
The term, 𝑖𝑙𝑖𝑚(𝐻 + ) , is due to mass transfer kinetics playing an important role in the transport
𝑑 +
𝑖𝑙𝑖𝑚(𝐻 + ) = 𝑘𝑚 𝐹[𝐻 ]𝑏 (3.13)
Where:
𝐹 = Faraday constant
The activation controlled current is given by the following expression (Woollam and
Hernandez, 2006):
𝜂
− ⁄𝑏
𝑖𝛼(𝐻 +) = 𝑖𝑜(𝐻 + ) . 10 𝑐 (3.14)
The cathodic current equation for carbonic acid is expressed in a similar manner as that of
hydrogen ions, H+.
73
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
𝜂
⁄𝑏
𝐻𝑒𝑛𝑐𝑒, 𝑖(𝐹𝑒) = 𝑖𝑜(𝐹𝑒) . 10 𝑎 (3.15)
𝜂 is overpotential (Volts, V)
These equations are used to derive the plot shown in Figure 3-10. This is a Tafel plot of
potential against the log of current density and the corrosion rates are read off from the point
of intersection of Ecorr=-513mV and the anodic dissolution line (A1). The current density is
then read off from the horizontal axis as the value of icorr. This current is then converted into
units of corrosion rate in mm/year by the following expression in order to obtain the plot
shown in Figure 3-11.
1.155A 1mm
Conversion of Current density to Corrosion rate: = (3.16)
m2 year
The plot shown in Figure 3-10 was produced for a temperature of 100oC and a CO2 fugacity
of 0.3 MPa (3bar). The temperature was then changed accordingly in the model to determine
the corresponding corrosion rates. The CO2 fugacity was then changed to 0.1MPa (1 bar),
0.03MPa (0.3 bar) and 0.01MPa (0.1 bar) and the same procedure was repeated to obtain the
plot shown in Figure 3-11.
74
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 3-10. Nesic-Postlethwaite-Olsen (NPO) model output for T=100oC – Tafel plot of
potential against log10 (current density)
Figure 3-11. Plot of corrosion rate against temperature for the NPO model for CO2
fugacities: 0.01MPa (0.1bar), 0.03MPa (0.3bar), 0.1MPa (1bar) and 0.3MPa (3 bar).
Other conditions where applicable are: pH=4, CO2 mole =100%, Shear stress=2Pa
75
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
From Figure 3-11, it is observed that corrosion rates generally increase with increasing
temperature which is as expected following the explanation given by the Arrhenius theory for
the effect of increasing temperature on increasing the rate of chemical reactions. It is also
observed that the corrosion rates are greater for CO2 fugacities of higher magnitudes in the
following order: 𝑓𝐶𝑂2 (at 0.3MPa) 𝑓𝐶𝑂2 (at 0.1MPa) > 𝑓𝐶𝑂2 (at 0.03MPa) > 𝑓𝐶𝑂2 (at 0.01MPa).
The reason for this is that higher magnitudes of CO2 fugacity translate into greater
concentrations of the gas in solution and consequently higher concentrations of carbonic acid
hence enhanced corrosion risks and higher anodic dissolution rates of the metal.
In terms of Freecorp version 1.0’s ability to accurately estimate corrosion, it has been
designed with a mechanistic background at its core (Ohio University, 2008). The model is
capable of indicating the relative contributions to corrosion as well as the dominant corrosion
mechanism based on the corrosive species input by the user. In addition, polarisation sweeps
and polarisation curves for each individual electrochemical reaction are displayed.
76
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Acetic acid is of a major concern in top-of-line corrosion due to its ability to provide an
additional source of hydrogen ions brought about by its weak dissociative property. It
partially dissociates into a hydrogen ion and an acetate ion as shown below:
HAc + H2 O ⇔ H3 O+ + Ac − (3.17)
Apart from being a reservoir for hydrogen ions, it has also been discovered experimentally
that the undissociated CH3COOH specie can be directly reduced after adsorbing onto the
metal surface (Nesic et al., 2009).
Thus, the presence of free CH3COOH becomes problematic, in that, particularly at low pH
ranges, the equilibrium of the dissociation reaction shifts to the left. It is also said that the
reduction of free acetic acid is strongly affected by the velocity of fluid flow indicating that it
is a mass transfer-controlled process and this implies that corrosion rate is dependent upon the
acetic acid concentration and the kinetics of transport of the species from the bulk solution to
the metal surface (Nesic et al., 2009).
Figure 3-12 shows a scatterplot of Freecorp model results for acetic acid corrosion against
experimental results. The two datasets, Crolet et al., 1999 and Guo et al., 2005, are the results
of the investigation of CO2-CH3COOH corrosion on X65 carbon steel and N80 carbon steel
respectively. Corrosion rates were determined by varying the concentrations of acetic acid for
both studies. Experiments were carried out at temperatures of 22oC for Crolet et al., 1999 and
50oC for Guo et al., 2005. Although the CO2 concentration in the former was constant, 1bar
(0.1MPa), there were two magnitudes for CO2 concentration in the latter case: 1 (0.1MPa) and
10bar (1MPa).
77
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 3-12. Scatterplot of Freecorp model readings against experimental readings for
two datasets: Crolet et al, (1999) and Guo et al, (2005)
The model over-predicts all data points of the Guo et al. (2005) study while the y=x line splits
the Crolet et al. (1999) dataset roughly into two, particularly for high HAc concentrations.
High HAc concentrations correspond to the experimental readings of the greatest magnitude
(≈5mm/year) and as experimental readings increase in magnitude, disproportionately high
readings are observed for model predictions. The Freecorp model predictions for the Guo et
al., 2005 dataset for 1MPa (shown in red crosses) greatly exceed the corresponding
experimental readings even though higher than normal corrosion rates are expected as a result
of the CO2 partial pressure being equal to 1MPa (10bar).
H2S corrosion proceeds via solid state reaction – the initial and final states of iron exist in the
solid state and unlike CO2 and HAc corrosion, it is not electrochemical. There is no
significant separation of the oxidation and reduction reactions at the steel surface. Hence, no
current flows between the cathode and anode (Nesic et al., 2009).
Fe + H2 S ⇔ FeS + H2 (3.19)
78
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Freecorp version 1.0 models H2S as a chemical reaction showing the species concentration
profile for H2S concentration from the metal surface across the liquid boundary layer to the
bulk solution as shown in Figure 3-13. The following concentration profile is representative
of the following conditions: T=20oC, Ptotal=0.1MPa (1bar), 𝑃𝐶𝑂2 =0.01MPa (0.1bar),
H2S(g)=40ppm and pH=5.
In general, H2S corrosion is limited to the kinetics of diffusion of corroding species (H3O+ and
H2S) from the bulk solution across the boundary layer and onto the metal surface. A
concentration gradient is set up across the boundary layer whereby actively corroding species
at the metal surface induce a drive resulting in the influx of more corroding species from the
bulk solution to the metal surface.
79
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
As corrosion proceeds over time, mackinawite scales thicken and become more protective of
the metal surface beneath and in some instances, may lead to the formation of pyrrhotite,
another less soluble form of ferrous sulphide scale. In cases where, the H2S concentrations are
high, pyrite and elemental sulphur form on the metal surface (Nesic et al., 2009).
The reason for there being different forms of iron sulphide is attributed to the fact that they
vary in terms of thermodynamic stability. Troilite and pyrrhotite are more stable than
mackinawite, for instance, combined with the fact that the equilibrium ferrous ion
concentration and pH are greater on the metal surface than in the bulk solution (Benning et al,
2000; Criaud et al, 1989).
Figure 3-14 shows a scatterplot for two sets of data: Lyle and Schutt (1998) and Nesic et al
(2008). While overall, the model estimates corrosion fairly accurately, for both datasets, it
tends to under-predict experimental readings of larger magnitudes. For instance, the model
predicts less accurately for the last two data points (shown in green crosses) from the Lyle and
Schutt (1998) study, which investigated H2S and pitting corrosion for a constant temperature
of 15.6oC, at H2S pressures of the range: 0-0.013MPa (0-0.13bar), CO2 partial pressures (0-
0.069MPa), for a pH range of 3.85 to 4.75 and for a duration of 14 days. The less accurate
model predictions correspond to the data points with low pH values (3.8-4.0) as well as with
the high H2S gas phase pressures (0.0069-0.013MPa) while the varying CO2 pressures do not
appear to have any significant impact on trends. Hence, the model struggles with corrosion
estimations for high H2S pressures and very low pH ranges.
A similar explanation can be used to describe model predictions on the Nesic et al., 2008
study, as well. Here, H2S corrosion was investigated at a constant temperature (20oC), pH (5)
and CO2 partial pressure (0.01MPa), with varying H2S concentrations (0-180ppm) and for a
period of 24 hours. The main difference between the two studies is that the pH was constant
for the Nesic et al (2008) study while it was varied for the Lyle and Schutt study. While the
CO2 partial pressure varies for the former and is constant in the latter, CO2 does not influence
the trends in these studies.
80
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 3-14. Scatterplot of Freecorp model readings against experimental readings for
two datasets: Lyle and Schutt (1998) and Nesic et al. (2008)
3.7.3 CO2 Corrosion Rate Trends and Parametric Study on Velocity of Fluid Flow
Corrosion rates increase with increasing temperature. This is explained by the kinetic theory
in that greater temperatures provide larger amounts of energy to colliding molecules in a
given reaction hence the frequency of collisions and the likelihood of the formation of a
product increases. In terms of CO2 corrosion, this implies that an increase in temperature
results in increased reaction rates, thus an increased tendency to form corrosion products such
as the evolution of hydrogen gas, the dissolution of iron to ferrous ions as well as the
formation of carbonate scales. This reaction is summarised as follows (Nesic and Vrhovac,
1999):
The Arrhenius theory also follows from the Kinetic molecular theory and further emphasises
this point (Tan and Chan, 2011). The Arrhenius equation is stated as follows:
𝐸
− 𝐴
𝑘 = 𝐴𝑒 𝑅𝑇𝑘
(3.21)
81
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Where:
By the Arrhenius theory, reacting molecules will have to acquire a minimum threshold energy
(activation energy, EA) in order to form products and temperature provides this energy (Tan
and Chan, 2011). Thus, from the definition of the Arrhenius equation, the natural logarithm of
the reaction rate constant, loge k, is proportional to the inverse of the absolute temperature of
the reaction, implying that the rate constant for any given reaction increases with increasing
temperature (Tan and Chan, 2011).
It is observed from Figure 3-15, that corrosion rates generally increase with temperature and
this is in accordance with the kinetic theory and the Arrhenius equation since it accelerates all
processes involved in corrosion – species transport, electrochemical and chemical reaction
rates.
Corrosion rates increase quite markedly as temperatures increase without accounting for the
effect of scale formation which results in peaks usually between 60-80oC depending on flow
conditions and water chemistry as well as the formation of Fe3O4 at higher temperatures. It is
also worth noting that even when ferrous ion concentrations are included in the simulation,
[Fe2+] = 50-120ppm, the result is simply a decrease in the overall magnitude of the corrosion
rates with no distinctive peak whatsoever (de Waard et al., 1991).
It is also observed that from Figure 3-16, the greater fluid velocities result in greater
magnitudes of corrosion rate (Corr Rate), thus Corr Rate at U=1.0m/s > Corr Rate at
U=0.5m/s > Corr Rate at U=0.1m/s. The reason for this is due to the fact that iron carbonate
scales deposited on the surface of the metal have an inhibitory effect to corrosion as they
build up and thicken however with greater fluid velocities, there is a greater tendency for
these scales to be eroded and washed away hence exposing a greater area of the metal surface
to further corrosive attack.
82
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 3-15. Plot of corrosion rate against temperature for varying CO2 partial
pressures for a fluid velocity, U=0.1m/s
Figure 3-16. Plot of corrosion rate against temperature for varying fluid velocities
83
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
3.8 Conclusions
The comparison of the different models discussed was based on identical values for the
underlying physical and chemical conditions that affect CO2 corrosion – the parameters
common to all these models were the temperature, CO2 fugacity (partial pressure), flow
velocity and pH.
The 1975 De Waard correlation is the most basic function. CO2 corrosion increased with
increasing temperatures, as do all of the other models. However it does not account for mid-
temperature scaling and this is its main limitation. The 1991 and 1995 De Waard correlations
account for mid-temperature scaling but the former does not explicitly contain a velocity term
in its function whereas the latter incorporated a velocity term as part of its resistance model.
The 1995 De Waard model also made use of a pH correction as part of the activation reaction
kinetics term.
The Norsok model also exhibited similar CO2 corrosion rate-temperature profile plots as the
De Waard models whilst accounting for shear stress in pipe flow. An interesting feature was
the scaling temperatures and how they varied across the different models. Table 3-2 shows the
scaling temperatures for each of these models.
Table 3-2. Scaling temperature maxima for the various models at 𝒇𝑪𝑶𝟐 =0.3MPa
Model Peak temperature, Tscale (oC)
1975 De Waard-Milliams Continuous Increase
1991 De Waard-Lotz-Milliams 71
1995 De Waard-Lotz-Dugstad 75
Norsok 78
NPO Continuous Increase
For the 1991 and 1995 De Waard correlations, as the CO2 fugacity decreased from 0.3MPa to
0.01MPa, the scaling temperatures increased consequently, which is depicted in Figures 3-3
and 3-4 respectively. Film formation is not included in both the 1975 De Waard and NPO
models hence there are no peaks in their corrosion rate-temperature profiles, as indicated in
Table 3-2. Film formation is included in the Norsok model correlations, and so a peak is
depicted in its corrosion rate-temperature profile. However, this peak remains fairly constant
regardless of changes in the magnitude of CO2 fugacity, as seen in Figure 3-7. While for the
84
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
1991 and 1995 De Waard models, changes in the magnitude of CO2 fugacity consequently
leads to changes in the magnitude of the corresponding scaling temperatures as seen in
Figures 3-3 and 3-4, respectively. A comparison of all the investigated corrosion models is
shown in Figure 3-17.
Figure 3-17. Plot of CO2 corrosion rate against temperature for the various models for
𝒇𝑪𝑶𝟐 =0.3MPa (3bar)
From the plot, all the model predictions appear to be in close proximity to each other for the
temperature range of 20-80oC but after 80oC, their behaviours are markedly different. While
the 1991 De Waard, 1995 De Waard and Norsok model predictions decrease once their
respective scaling temperatures are exceeded, the 1975 De Waard and NPO models continue
to increase due to the fact that as previously mentioned, their correlations do not account for
film formation. Therefore, their accuracy in prediction is limited to the low to mid-
temperature range only.
In addition, the Norsok model’s predicted results are consistently higher than those of the De
Waard’s correlations for most temperatures, due to the fact that the applied shear stress of
2Pa, as seen in Figure 3-7, is in the lower end of its applicable range of 1-150Pa. The NPO
model is mechanistic and electrochemical in its basis and as such has a tendency for over-
85
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
prediction, particularly for temperatures exceeding 80oC. However, it must be said that while
the empirical models such as the De Waard correlations and the Norsok model may appear to
predict more accurately than the mechanistic NPO, drawing conclusions from results
predicted outside their range of applicability is questionable. Mechanistic models are more
likely, in general, to better model a wider range of parameters than their empirical
counterparts. Overall, based on the corrosion rate vs temperature plot trends, the 1991, 1995
and Norsok models are the more realistic ones, with porous Fe2O3 porous films forming
before peak temperatures and Fe3O4 passive films forming beyond peak temperatures.
For the Freecorp version 1 model, in predicting CO2 corrosion, the model does not account
for Fe3C, FeCO3 and Fe3O4 scaling and the implication of this is that corrosion rates
continuously increase with increasing temperatures. Despite a ferrous ion concentration,
[Fe2+], provision in the model, corrosion rates still increase with increasing temperatures even
though [Fe2+] in excess of 80ppm should lead to supersaturation that will eventually cause the
build-up of scale (de Waard et al., 1991), culminating in a peak-like trend as depicted by the
Norsok model, for instance. The model is also limited to uniform corrosion hence it is
incapable of estimating pitting or any form of localised corrosion. Also, the range of CO2
prediction is from 1Pa (0.01mbar) to 1MPa (10bar).
In HAc corrosion prediction, model predictions are fairly accurate provided HAc
concentrations do not greatly exceed 0.01M (600ppm), for which it grossly over-predicts.
Also, the model makes no provision for TLC.
In H2S corrosion prediction, the model seems to perform fairly well although as with HAc,
the greater the concentrations of H2S, the less accurate are the model predictions. The model’s
maximum limit is a H2S partial pressure of 1MPa (10 bar). The model does not account for
the precipitation of ferrous sulphide hence is incapable of identifying the likeliest form of
ferrous sulphide scale for a given flow condition given that there is a high tendency for the
formation of a mixture of different scales on the steel surface. The composition of the scale
will affect the species concentration profile from the metal surface across boundary layer to
the bulk solution because the scale acts as a barrier to the transport of corroding species.
These models were assessed in terms of their accuracy of predicted corrosion rates using a
separate test dataset. The results are discussed in Section 8.2.
86
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
87
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
4.1 Introduction
In the previous chapter, established CO2 corrosion models in research literature such as the De
Waard-Milliams correlations, Norsok and NPO models were discussed. Though these models
are satisfactory for their intended purpose – low pressure CO2 corrosivity prediction, their
limitations are their unsuitability for high pressure CO2 corrosion prediction (Mohamed et al.,
2011) and when more complex effects are present such as the growth of protective scale
(Nesic et al., 2008).
In this chapter, low and high pressure CO2 corrosion datasets are analysed using various
statistical techniques. The statistical techniques applied include descriptive statistics, principal
component analysis (PCA), variable interactions, regression and response surface modelling.
PCA is a mathematical procedure which transforms potentially correlated data into an
orthogonal system of linearly uncorrelated principal components (Suryanarayana and Mistry,
2016). PCA is carried out in such a way that the first principal component accounts for much
of the variability within the dataset. The datasets comprise the primary environmental factors:
temperature (T), CO2 partial pressure (𝑃𝐶𝑂2 ), flowrate velocity (U), pH and corrosion rate
(Corr Rate). A smaller test dataset was then separately selected from related sources and used
as unseen data in the prediction of CO2 corrosion rates. Assessments of the performance of
each model were then made by plotting model vs experimental scatter diagrams and
determining the difference (error) in magnitudes between each of the plotted points. In
addition, the R2-value and 95% confidence intervals were also used in the evaluation of model
accuracy and reliability.
88
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
A summary of the descriptive statistics for each variable is displayed in Table 4-1.
The boxplots in Figure 4-1 shows the distribution of each of these variables. The boxplots for
temperature and pH data have approximately symmetrical data distributions, implying that the
data within each of these variables is distributed evenly around the measures of central
tendency (mean, median and mode). In general, the input variables: T, 𝑃𝐶𝑂2 , pH and the
output variable, Corr Rate, exhibit positive skewness, implying that their means are not
substiantially greater than their respective medians and modes. For pH, the skewness is 0.20
therefore, the mean, median and mode are approximately equal in magnitude. Velocity
exhibits a negative skew, hence its mode is greater than its respective median and mean
values. In the boxplots of Figure 4-1 and Figure 4-2, the closer the horizontal centre-lines of
each variable to the bottom of the box is, the greater the degree of positive skewness and the
closer the horizontal centre-line is towards the top of the box, the greater the degree of
negative skewness. The red plus symbols above the corrosion rate boxplot indicate that the
outliers for this dataset are in the 60-70mm/year range. There are outliers for the 𝑃𝐶𝑂2 variable
as shown in Figure 4-2, however the range is not as significant as that of the corrosion rate
variable.
89
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 4-1. Boxplot showing the overall range (box and whiskers), inter-quartile range
(white space in blue edged box), median, skewness and outliers (red plus symbols) in
each of the variables in the dataset
Magnitude of Data (from Range)
Figure 4-2. Boxplot showing the range of values, inter-quartile range, median and
skewness of the 𝑷𝑪𝑶𝟐 and pH variables (Outliers are red plus symbols)
90
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
A summary of the test dataset used in this study is shown in Table 4-2. The number of data-
points in this dataset is 15. This dataset is used for validation of the model and is not included
in the development of the final model.
This dataset is a randomly selected group of CO2 corrosion data (Dugstad et al., 1994b;
Nordsveen et al., 2003). It is used as unseen data in each of the developed models in order to
assess their performances in CO2 corrosion prediction.
91
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
It is seen that from the cumulative figures, 96.2% of the variation in this dataset, is explained
by the first 3 principal components, and so the fourth is not necessary to describe the variation
in this particular dataset. This further implies that three variables are sufficient in describing
CO2 corrosion rate. The Pareto chart shown in Figure 4-3 illustrates this point further.
Figure 4-3. Pareto chart showing the spread of the data (variance) among each of the
principal components
Pictorially, the biplot shown in Figure 4-4, below, illustrates the relative influence of each
variable in two-dimensional (2-D) space.
Variable Data-points
Figure 4-4. Two-dimensional biplot showing the orientation (magnitude and direction)
of the variables in the principal component axes
92
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The 2-D biplot shows that CO2 partial pressure, temperature and velocity are three most
dominant variables in terms of variation within the dataset and as such are likely to have a
greater influence in terms of the prediction of CO2 corrosion rates, due to their respective
lengths and directions in the first and second component axes. The CO2 partial pressure and
temperature variables point in the positive direction of the first component (PC1) axis while
the velocity and pH variables do not. The velocity is considered more significant than the pH
in terms of its effect on the variance because its angular orientation to the positive direction of
the PC1 axis is comparatively smaller.
The scree plot shown in Figure 4-5 is essentially a line graph plot of the eigenvalues of all the
principal components joined together. The slope of the line joining PC1 to PC2 is quite steep
while the following line joining PC2 to PC3 is less steep, while the slope of the line joining
PC3 to PC4 increases again. This serves to indicate that three principal components are
required to describe this dataset.
Figure 4-5, reinforces the fact that three principal components are responsible for the greatest
variation in the dataset, and as indicated by the biplot in Figure 4-4, the principal variables are
Temperature, 𝑃𝐶𝑂2 and Velocity. This deduction is used as a basis in the determination of a
suitable multiple linear regression model shown in the following section.
Figure 4-5. Scree plot showing decreasing eigenvalues (variances) for all the principal
components
93
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
𝑦𝑖 = 𝛽0 + 𝛽1 𝑥1 + 𝛽2 𝑥2 + ⋯ + 𝛽𝑛 𝑥𝑛 𝑓𝑜𝑟 𝑖 = 1, 2, … , 𝑛 (4.1)
Where:
Given that the principal variables of CO2 corrosion rate for this sample data as determined by
principal component analysis are temperature and 𝑃𝐶𝑂2 , the multiple linear regression model
shown in Figure 4-6 comprises these variables. The following table shows the estimated
coefficients for the simple multiple linear regression model.
The R-squared and adjusted R-squared values are 0.507 and 0.480 respectively for N=60
(number of data-points) and 56 error degrees of freedom. The t-stat statistic and p-value from
Table 4-4 indicate the level of importance of the given term. The p-values of all the regression
model terms with the exception of the intercept are less than the critical p-value of 0.05, thus
implying that these terms are significant. The implication for the intercept term is that there is
no significant difference in the regression model should this term be included or excluded
completely.
94
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
For the t-stat values in Table 4-4, the absolute t-stat values for each regression model term is
considered. The absolute t-stat values for each of the regression model terms with the
exception of the intercept, exceeds the critical t-statistic of 2.004 (see Statistical t-tables in
Appendix A4.1), for 56 degrees of freedom and the 0.025 significance level. Therefore, the
intercept is the only term that is not significant in the model – the same conclusion as that of
the p-value statistic.
Based on the estimate values in Table 4-4, the regression equation is given by:
The polygonal plane through the points in the three-dimensional scatter plot in Figure 4-6, is a
graphical representation of the corrosion rate regression model derived above. Alternate views
for this plot are shown in Appendix A4.2
Corrosion Rate (mm/year)
70
60
50
40
30
20
10 2.5
0 2
20 1.5
40
1
60
80 0.5
100 0
Temperature (oC)
CO2 Partial Pressure, 𝑷𝑪𝑶𝟐 (MPa)
Figure 4-6. Three-dimensional plot of corrosion rate against temperature and CO2
partial pressure (𝑷𝑪𝑶𝟐 ) showing regression model through the points
95
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Another linear regression model was derived with seven terms given by the following
relationship.
There is a good positive correlation between the model and experimental data. The data-
points are also in close proximity to the 95% confidence intervals.
Figure 4-7. Multiple linear regression model corrosion rate against experimental
corrosion rate plot and error bar chart for selected test data – error is the difference
between experimental and model corrosion rate
96
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Variable Interactions
Interaction prediction plots between single input variables, temperature, 𝑃𝐶𝑂2 , U and pH, are
simulated to examine their effect on corrosion rate. Adjusted corrosion rate responses are a
function of the x-axis variable, shown in each of these plots, with the other variables fixed at
constant values – low, medium and high magnitudes. These plots make it possible to examine
whether the effect of one variable depends on the value of the other variable, and if so, to
what extent.
Figure 4-8A shows the prediction interaction between temperature and 𝑃𝐶𝑂2 . It is observed
that at T=20oC, the corrosion rate is low, barely exceeding 5mm/year while corrosion rates are
greater for T=55oC than for T=90oC, at lower CO2 partial pressures. The corrosion rate for
T=90oC is greater than that of T=55oC when 𝑃𝐶𝑂2 exceeds 1MPa. It is generally the case that
the higher temperatures induce greater corrosion rates because the rates of chemical reactions
increase with increasing temperatures (Tan and Chan, 2011).
A B
97
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 4-8 B shows the prediction interaction plot for temperature and velocity. The T=20oC
line (shown in green) results in the least rate of corrosion. In general, corrosion rates increase
from 0 to 11mm/year as the velocity increases from 0 to 13m/s. Initially, the T=55oC line
(shown in red) exceeds the T=90oC (shown in blue) in terms of the magnitude of corrosion
rate, i.e. 16>14mm/year, however as velocity increases to a value of 12m/s, the two lines
intersect and then the T=90oC line exceeds the T=55oC line at U=13m/s.
A B
B
Figure 4-9B shows the prediction interaction plot for 𝑃𝐶𝑂2 and velocity. It is observed that
higher 𝑃𝐶𝑂2 values result in greater corrosion rates, as greater acid concentrations within the
fluid medium in contact with the metal surface become more prevalent. Initially,
98
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
𝑃𝐶𝑂2 =0.04MPa yields the greater rate of corrosion, i.e. corrosion rate at 𝑃𝐶𝑂2 =0.04MPa
>corrosion rate at 𝑃𝐶𝑂2 =1.07MPa>corrosion rate at 𝑃𝐶𝑂2 =2.1MPa, for U=0m/s. However, at
U=2m/s, the three lines intersect and then the order of adjusted corrosion rates is reversed:
corrosion rate at 𝑃𝐶𝑂2 =2.1MPa>corrosion rate at 𝑃𝐶𝑂2 =1.07MPa>corrosion rate at
𝑃𝐶𝑂2 =0.04MPa.
A B
Figure 4-10A shows the prediction-interaction plot for 𝑃𝐶𝑂2 and pH. Each of the lines in the
plot has negative gradients. Corrosion rates commence at approximately 32, 28 and
24mm/year for 𝑃𝐶𝑂2 =2.1, 1.07 and 0.04MPa respectively and decrease steadily until they
reach 17, 11 and 4mm/year.
Figure 4-10B shows the prediction-interaction plot for velocity and pH. Corrosion rates fall
for each of the line plots with decreasing pH, thus negative slopes. The magnitudes of
adjusted corrosion rates are in the following order: corrosion rate at U=13m/s> corrosion rate
99
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
𝑦 = 𝑓(𝑥) + 𝜀𝑟 (4.4)
Where:
𝑓 is the function that transforms predictor variables (input)to response variables (output)
𝑥 is the predictor variable or input variable (Temperature, PCO2 , Velocity and pH)
In particular, the Hougen-Watson equation has proven to be a very useful nonlinear equation.
The Hougen-Watson equation is a derived equation which expresses the rate of a chemical
reaction in terms of the concentration of reacting species (Carberry, 2001; Staelens et al.,
2002). This relation is as follows:
𝑥3
𝛽1 𝑥2 −⁄𝛽
5
𝑦𝑖 = (4.5)
1 + 𝛽2 𝑥1 + 𝛽3 𝑥2 + 𝛽4 𝑥3
Where:
100
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
In this study, the predictor variables replace the concentrations of reacting species. From the
formulation, it is clear that the corrosion rate can be expressed in terms of three variables. The
R-squared values for this equation using the study dataset with a combination of temperature,
𝑃𝐶𝑂2 and velocity, as predictor variables is 0.7; while the combination of temperature, 𝑃𝐶𝑂2
and pH as predictor variables yields an R-squared value of 0.45.
The magnitudes of the R-squared values provide supporting evidence to the deduction from
principal component analysis, which is that temperature, 𝑃𝐶𝑂2 and velocity are the principal
variables for this dataset.
The plots in Figure 4-11 show the performance of the model. The R2-value is moderately high
with majority of the data-points falling within the 95% confidence interval. This indicates that
the model is reasonably accurate and reliable.
101
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The plots in Figure 4-12 show how well the model fits the original data. The leverage plot
indicates that a few outliers are present (points above the dashed line). There is a possibility
that these high-leverage datapoints have greater random errors associated with their
experimental measurements than the rest of datapoints in this particular dataset. Nevertheless,
majority of the datapoints lie well below the dashed line (Chatterjee and Hadi, 1986).
The residuals plot shows that there is an even spread of the datapoints about the zero mark
(dashed line). This implies that the model is being sufficiently explained by the three predictor
variables.
Figure 4-12. Leverage-CO2 corrosion data plot and residuals-CO2 corrosion data plot
The response surface plots also show trends for each variable, for instance, greater
magnitudes of all the variables except pH lead to an increase in CO2 corrosion rate (shown by
102
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
the centre straight line shown in each plot slice). CO2 corrosion rate outputs are predicted
using 95% confidence limits. The performance of the response surface model is summarised
in the plots of Figure 4-14. The response surface tool produced a low value for the R2-statistic
and this is because of a combination of the variation within the dataset as well as the presence
of outliers.
Figure 4-14. Response surface model corrosion rate against experimental corrosion rate
plot and error bar chart for selected test data – error is the difference between
experimental and model corrosion rate
103
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
From the boxplot shown in Figure 4-15, temperature has the widest range of values (24-
150oC) while pH has the smallest range of values (3.1-6). All of the variables excluding the
corrosion rate are said to be positively skewed, implying that their means are greater than
their respective medians and modes but not by substantial amounts as the skewness
104
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
magnitudes (Table 4-5), are only slightly greater than zero. The pH distribution, shown in
Figure 4-16, is the only distribution with outliers (indicated by the red plus symbol). The
boxplot of the pH variable indicates that the outlier is that of pH=6. This is the single greatest
value for pH in the dataset. The relatively long top whisker of the pH boxplot also indicates it
is highly positively skewed. The pH skewness of 1.15, from Table 4-1 confirms this.
Magnitude of Data (from Range)
Figure 4-15. Boxplot showing the overall range (box and whiskers), inter-quartile range
(white space in blue-edged box), median, skewness and outliers (red plus symbols) in
each of the variables in the dataset
Magnitude of Data (from Range)
Figure 4-16. Boxplot showing the range of values, inter-quartile range, median,
skewness and outliers of the pH variable
105
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
A summary of the test dataset is shown in Table 4-6. The size of this dataset is small as the
actual dataset for the study also has a small size (N=16). This dataset is used for validation of
the model and is not included in the development of the final model.
The Pareto chart shown in Figure 4-17 shows the relative significance of the principal
components to the variance of the data.
106
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 4-17. Pareto Chart showing the spread of the data (variance) among each of the
principal components
The relative influence of each of the variables to the first and second principal components is
depicted in the two-dimensional biplot in Figure 4-18.
Variable Data-points
Figure 4-18. Two-dimensional Showing the orientation of the variables in the principal
component axes
107
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The 2-D biplot (Figure 4-18) shows that the CO2 partial pressure (𝑃𝐶𝑂2 ) variable is the most
dominant variable in terms of accounting for the variation within the dataset followed by the
temperature, velocity and pH respectively. This is because the direction of these three
variables shows they are within close proximity to the Principal component 1 (PC1) axis. The
length of each variable indicates their relative contribution to each principal component.
The scree plot shown in Figure 4-19 is a line graph of the eigenvalues from Table 4-7. The
slopes change from PC1 to PC2, PC3 and PC4. However, the most significant change in slope
occurs between PC1 and PC2, afterwards, the slope from PC2 to PC4 is uniform. This
indicates that PC1 and PC2 are sufficient to describe the variation within the dataset.
Figure 4-19. Scree plot showing the decreasing magnitudes of eigenvalues for all the
principal components
Temperature (oC)
CO2 Partial Pressure, 𝑃𝐶𝑂2 (MPa)
Figure 4-20. Three-dimensional plot of corrosion rate against temperature and CO2
partial pressure (𝑷𝑪𝑶𝟐 ) showing regression model through the points
Table 4-8 shows the model terms for a suitably derived regression equation.
For N=16 observations (data-points) and 9 degrees of freedom at the 95% confidence limit,
the p-values of all the regression terms with the exception of the intercept and velocity terms,
is less than the critical p=0.05 value. This indicates all these terms are significant to the model
except the intercept, velocity and Tx𝑃𝐶𝑂2 xUxpH terms. For the t-stat values, the absolute,
109
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
non-negative figure is considered. All t-stat values for the terms are greater than 2.262 with
the exception of the intercept, velocity and the Tx𝑃𝐶𝑂2 xUxpH terms. The value: 2.262 in this
case is the critical t-statistic for 9 degrees of freedom at the 0.025 significance level (see
Statistical t-tables in Appendix A4.1). Therefore the t-stat analysis arrives at the same
conclusion as the p-value. Hence the intercept, velocity and Tx𝑃𝐶𝑂2 xUxpH terms can be
neglected altogether. This equation is given by:
Figure 4-21. Multiple linear regression model corrosion rate against experimental
corrosion rate plot and error bar chart for selected test data
Variable Interactions
The interaction between input variables is discussed by the use of variable interaction plots.
By varying the x-axis variable and keeping a second variable constant at three magnitudes:
low, medium and high, the changes in adjusted corrosion rate can be investigated.
110
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
In Figures 4-22A and 4-22B, the intearctions between temperature and CO2 partial pressure
and velocity are shown as straight lines with negative gradients. These generally indicate that
an increase in the x-variable leads to a decrease in the adjusted corrosion rate. For the
tempertaure interaction with CO2 partial pressure, Figure 4-22A, the higher the CO2 partial
pressures, the smaller the magnitude of the corrosion rate. This might be due to an increased
effect of the thicker scale formation at higher pressures forming a protective barrier on the
metal surface against corrosion (Zhang et al., 2012b). The changes in temperature from the
low to medium and high values indicate that corrosion rates peak at the medium temperature
(T=87oC), starting off and ending with small corrosion rates for low and high temperatures
respectively. This is true for high pressure CO2 corrosion (Zhang et al., 2012b).
A B
For Figure 4-22B, higher flow velocities have a decreased effect on corrosion rates but only
very slightly and this may be because for high CO2 pressures, thicker scales are formed and
flow velocities of 0 to 4m/s in magnitude are insufficient to induce the washing away of these
111
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
scales from the metal surface. For Figure 4-23A, the interaction between temperature and pH
results in straight lines with negative gradients.
A B
An increase in pH leads to a drop in the corrosion rate, which is because the higher the pH,
the lesser its influence on CO2 corrosion as the acidity of the aqueous medium decreases. The
T=24oC line is not visible on this plot because its effect on corrosion rate is very small. Again,
the mid-temperature value of 87oC is responsible for the greatest corrosion rates. For Figure
4- 23B, increasing flow velocities leads to an overall slight decrease in the adjusted corrosion
rates and this may be attributed to the tendency for thicker scale formation for high pressure
CO2 conditions. It is also observed that as the CO2 partial pressures increase from 3.5-
23.3MPa, the adjusted corrosion rates fall. It is worth noting that at high CO2 pressure
conditions, the formation of dense, more compact and thicker iron carbonate scale is favoured
(Zhang et al., 2012b). Invariably, the greater the CO2 partial pressures, the thicker the scales
formed thus exerting an inhibitory effect on corrosion of the metal beneath.
112
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
For both Figures 4-24A and 4-24B, the adjusted corrosion rates decrease with increasing pH.
The negative gradients of the lines are expected due to decreased effect that the aqueous
medium has on corrosivity at high pH conditions. For Figure 4-24A, increasing CO2 partial
pressure has an inhibitory effect on the corrosion rate due to the formation of thicker scales at
P=13.4 and P=23.3MPa. For Figure 4-24B, greater corrosion rates are observed at U=0m/s
than at U=4m/s.
A B
In general, the greater the flow velocity, the greater the corresponding corrosion rate due to
the washing away of protective layers formed by scale on the metal surface. However, when
low fluid velocities are causing high corrosion rates, there is the likelihood that some degree
of localised corrosion is taking place as opposed to the more common uniform corrosion,
which is more characteristic of high fluid flow velocities.
113
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The plot in Figure 4-25 shows the performance of the model. In general, this model is a poor
fit for the data. The R2-value was very low though nonlinear modelling is tricky especially
when there is a shortage of data-points in the given high pressure CO2 corrosion dataset.
The leverage plot in Figure 4-26 indicates that there only two notable outliers, which is in
partial agreement with the descriptive statistics of the dataset that suggests there are no
outliers. The residulas plot of Figure 4-26, indicates that the there is an evem spread of the
data about the zero dotted line, implying that the model adequately represents the data even
114
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
though its performance with the test set is poor as seen by the low R2-coefficient as well as the
overall distribution of the data-points shown in Figure 4-25.
Figure 4-26. Leverage-CO2 corrosion data plot and residuals-CO2 corrosion data plot
The performance of the model is shown in Figure 4-28. It is seen that while this model
permitted the use of all variables, the R2-value was extremely low. This model struggled to fit
the data due to an insufficient size of the dataset. For instance, modelling involved the use of
numerical techniques, and in this case a definitive solution was not reached because
convergence could not be attained.
115
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 4-27. Response surface modelling plots – CO2 corrosion rate can be investigated
by changing the magnitude of each predictor variable
Figure 4-28. Response surface model corrosion rate against experimental corrosion rate
plot and error bar chart for selected test data
116
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
4.4 Conclusions
Descriptive statistics of the low pressure CO2 dataset revealed that the CO2 partial pressure
and corrosion rate variables contain outliers and these variables also exhibit positive
skewness, implying their means are greater than their medians. Additionally, the corrosion
rate variable outliers are in the 60-70mm/year range. These outliers are data-points which lie
outside of the overall pattern of the distribution. For the high pressure CO 2 dataset, the pH
was the only variable distribution containing outliers. The presence of the outliers also had an
effect on its skewness. Again, the pH was the only variable exhibiting notable positive
skewness, with its mean only slightly greater than its median.
The results from principal component analysis on the low pressure CO2 dataset revealed that
two components were responsible for 76% of the variation within the dataset. Subsequent
biplot analysis revealed that the respective variables in terms of contribution to the variation
within the dataset are CO2 partial pressure, temperature, flow velocity and pH, in that order.
For the high pressure CO2 dataset, two components accounted for 81.5% of the variation
within the dataset implying that a biplot analysis in two components is sufficient to find the
principal variables. Analysis form the 2-D biplot, revealed that in terms of contribution to the
variance, the variables are in exactly the same order as that of the low pressure CO2 case. The
major difference between these datasets is that the pH has a more significant influence on the
variance of the high pressure CO2 dataset.
The multiple linear regression, nonlinear regression and response surface techniques were
developed based on the bulk of the low pressure corrosion dataset and the accuracy of these
models was carried out by assessing their prediction performances using a smaller, identical
and randomly selected test dataset. This was repeated for the high pressure CO2 corrosion
dataset.
The statistical models generally have good correlations with experimental results for low
pressure CO2 corrosion. The Hougen-Watson nonlinear model has the highest R2-value of
0.67; test data-points have close proximity to the 95% confidence intervals. The multiple
linear regression model while also achieving a similarly high R2-value as the nonlinear
regression model, it involved the use of square-terms for the temperature variable and
interaction terms comprising all variables. However, the main limitation of the Hougen-
Watson nonlinear regression model is its inability to permit the use of all predictor variables.
117
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The response surface model performed relatively poorly in comparison to the other statistical
models in terms of the magnitude of the R2-squared coefficient. In addition, unlike the
Hougen-Watson nonlinear model, it permits the use of all four predictor variables. It is also
possible to interactively adjust any given predictor variable via graphical plot panes and
observe the subsequent change in CO2 corrosion rate instantaneously. Table 4-9 is a summary
of the statistical model performances.
For high pressure CO2 corrosion, the multiple linear regression model had a moderately high
R2-value as seen in Table 4-10, with data-points lying in close proximity to the 95%
confidence limits. There is however a slight tendency for over-prediction, particularly for low
to medium magnitudes of corrosion rates. This may be due to a greater degree of variation of
these points causing the model to struggle to adequately fit the data. Diagnostic plots such as
the leverage and residuals revealed that there are very few outliers and that the nonlinear
model satisfactorily represented the data however the model performance stated in Table 4-10
suggests otherwise. The response surface model performed very poorly in terms of the
accuracy of prediction.
In general, the statistical models struggled considerably to fit the data in terms of accuracy of
prediction. This can be attributed to the small dataset for high pressure CO2 corrosion. It is
worth noting that for both nonlinear regression and surface response surface models,
numerical techniques are applied and for both cases definitive solutions were not reached due
to a lack of convergence during numerical computation.
118
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
119
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Thus,
Where:
𝑤𝑗 = weight
𝑥𝑗 = input signal
𝑦𝑗 = output signal
Figure 5-1 is an illustration of such a neuron yielding an output signal, yi, from the various
input signals, xi, and their corresponding weights, wi.
x1
w1
x2 w2
w3 y
x3
w4
x4
Figure 5-1. An artificial model of a neuron showing the transformation of input signals,
xi, to an output signal, yi (Veelenturf, 1995)
However, since no training was available, all the neural network (NN) parameters had to be
designed but by the late 1950’s, the idea of the perceptron was conceived by Rosenblatt and
others (Ukil, 2010). Perceptrons are a class of NNs that were primarily designed for pattern
recognition (Veelenturf, 1995). The key difference between Rosenblatt’s neurons and those
120
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
of McCulloch is the introduction of the learning rule to the former. While these perceptrons
were very useful, their limitations are well-documented, one of which is their ability to solve
only a limited class of problems (Haykin, 1999). Another noteworthy contribution during this
period was made by von Neumann. Von Neumann made it known that there was a tendency
for NNs to have redundant neuronal connections and these yielded unreliable contributions,
which in turn affected overall performance (Haykin, 1999).
In the 1980’s however, two key developments were made to the advancement of NNs. The
first was the use of statistical mechanics to explain the operation of a recurrent network by
John Hopfield, which was capable of being used as an associative memory. A recurrent
network is an NN configuration with at least one feedback loop. The second key development
was the formulation and use of the backpropagation algorithm for training multilayer
perceptron networks, proposed by David Rumelhart and James McClelland. At present,
backpropagation remains the most popular learning algorithm for the training of multilayer
perceptrons (Ukil, 2010).
In chapter 3, the popular and well-established De Waard Milliam, Norsok and NPO corrosion
models were discussed. In chapter 4, a derived statistical model was obtained as a means to
address some of the shortcomings of the aforementioned established models. In this chapter,
neural network modelling of collated low and high pressure CO2 corrosion datasets will be
carried out by dividing these sets into training and test sets.
The training set comprising variables T, 𝑃𝐶𝑂2 , U, pH and Corr Rate, was used to develop the
neural network solely and is summarised as descriptive statistics in Tables 4-1 and 4-5
(Chapter 4, Sections 4.2.1 and 4.3.1), for low and high pressure CO2 respectively. Figures 4-1
and 4-15 are corresponding boxplots for the summary statistics in Tables 4-1 and 4-5
respectively. The test set is summarised in Tables 4-2 and 4-6 for low and high pressure CO2
corrosion respectively. The test set is strictly used for the purpose of prediction as unseen
data and for the computation of the R2-coefficient.
121
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The stages in machine learning are explained with respect to artificial intelligence (AI). The
goal of artificial intelligence is the development of algorithms that require machines to
perform cognitive tasks and solve problems in a manner similar to human reasoning (Haykin,
1999).
The first stage of machine learning for an AI system is to set up an environment that supplies
the given input data to the learning element. The learning element then uses the data to make
improvements in its existing knowledge base. The performance element makes use of the
knowledge base in order for the AI system to perform its task. Input data from the
environment often contains flaws; however the AI system’s knowledge base is unaware
initially and assesses itself by receiving feedback from the performance element. The
feedback loop mechanism allows the AI system to evaluate its hypotheses and to revise them
if necessary (Haykin, 1999). The error-minimisation concept in NNs applies the same logic,
in which the goal is to reduce the error between a specified target and an output in a
continuous loop as shown in Figure 5-3 (Demuth et al., 2009).
122
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 5-3. Flow loop of input to output via continuous adjustment of weights
123
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
There are three distinct functional operations for the computation of output data from input
data. Firstly, the input scalar, p, is multiplied by the weight, w. Secondly, the product of the
weight and input scalar, p, is added to a bias and is the net input (n). The bias is similar to the
weight except that it has a constant value of 1. Thirdly, the net input is passed through the
transfer function, f. The operation of ‘f’ produces the scalar output, a. For a neuron with
vector input, p1, p2, …, pj with weights w1, w2, …, wj, the net input, n is the sum of the
products of p1, p2, …, pα and w1, w2, …, wj, added to bias, b. This is shown as follows:
𝑛 = 𝑤1 𝑝1 + 𝑤2 𝑝2 + ⋯ + 𝑤𝑗 𝑝𝑗 + 𝑏 (5.1)
⟹ 𝑎 = 𝑓(𝑤𝑝 + 𝑏) (5.2)
Table 5-1. Definitions of the transfer functions used (Beale et al., 2014)
Transfer Algorithm Definition
Function
Purelin(n) Linear transfer function
𝑎1 = 𝑝𝑢𝑟𝑒𝑙𝑖𝑛(𝑛) = 𝑛 (5.3)
Logsig(n) Log-Sigmoid transfer
1
𝑎2 = 𝑙𝑜𝑔𝑠𝑖𝑔(𝑛) = (5.4) function. Based on the
(1 + 𝑒 −𝑛 )
logistic sigmoid function:
1
𝑆(𝑥) = (1+𝑒 −𝑥 ) (5.5)
124
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
125
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The hidden layers in such architectures enable the NN to extract higher-order statistics and it
acquires a global perspective which is particularly useful when the input layer is large
(Churchland and Sejnowski, 1992).
126
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 5-8. Flowchart showing the sequence of steps in the development of the neural
network model
There are two sets of input to the NN model: the variable training dataset and the corrosion
rate training dataset. Firstly, the NN is trained with the variable dataset which consists of the
parameters which influence CO2 corrosion rate, i.e., temperature, CO2 partial pressure (𝑃𝐶𝑂2 ),
pH and flowrate velocity. Secondly, the NN is trained with the corresponding corrosion rate
data. These corrosion rate data are direct laboratory measurements from corrosion rate
experiments reported in the literature: (Dugstad et al., 1994b; Choi and Nesic, 2009; Hesjevik
et al., 2003; Zhang et al., 2012a; Cui et al., 2006).
The NN model, net_01, is shown in Figure 5-8 as a process step in the flowchart sequence
because the model is developed and defined by the two sets of input preceding it. The model
127
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
is then used for corrosion rate prediction. Figure 5-9 shows a block diagram view of the input
and output datasets as well as the neural network, net_01.
Input Output
The variable test dataset is contained within the second input set and is used for testing. It
contains parameters identical to those in the variable training dataset. The model reads this
data and based on learned knowledge from the first input, yields output. These output results
are then compared against actual corrosion rates from the original data in order to compute
the correlation coefficient (R2-value). The method in developing the final NN model is based
on the use of the R2-value, and is shown in Figure 5-8 as the decision step (Draper and Smith,
1998), (Abyaneh, 2014). The condition, R2≥0.3, was chosen in the initial testing phase in
order to characterise the selected training functions to be used: a low-level screening process
to rule out undesirable training functions. For fine-tuning of the final NN model, however
R2≥0.7 was used and then later increased to R2≥0.8, for the final model result. R2-coefficients:
R2≥0.7 and R2≥0.8 were selected as further fine-tuning (pruning) criteria. The actual corrosion
rates are a direct consequence of the parametric conditions contained in the variable test
dataset, measured from experimental corrosion rate plots. The errors are calculated by
subtracting the actual (experimental) CO2 corrosion rates from model predicted CO2
corrosion rates. Absolute values of these errors are then evaluated and then results are
summed up to obtain the ‘Sum Total of Absolute Error’. Equations for these expressions are
shown as follows:
128
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
2
2
∑(𝑥𝑖− 𝑥̅ )(𝑦𝑖 − 𝑦̅)
𝑅 =( ) (5.12)
√∑(𝑥𝑖 − 𝑥̅ )2 (𝑦𝑖 − 𝑦̅)2
Where xi and yi are the ith observation and model predicted values respectively. 𝑥̅ and 𝑦̅ are
the mean values of xi and yi and n is the total number of data-points: (Draper and Smith,
1998; Abyaneh, 2014). The arbitrary x-y plot of predicted against experimental corrosion rate
in Figure 5-10, illustrates how the equations for the error expressions are obtained.
Predicted Corrosion Rate (mm/year)
xi
(xi, yi)
yi
Figure 5-10. Arbitrary plot of predicted corrosion rate against experimental corrosion
rate
129
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
In order to determine the final model parameters such as size of the network, the number of
layers (network configuration) and the type of training function(s) to use, simulation runs
were performed using Matlab 2012a version on a Windows 7, 64-bit system with 6 gigabytes
(GB) of random access memory (RAM).
Given that in any given neural network model, the transfer function converts input data to a
given output, trial runs to determine the final NN model, for the purpose of this study, are
broadly based on the type of transfer function (Vogl et al., 1998). Hence simulation tests were
carried out on two main groups – the logsig and tansig transfer function groups. The reason
for their selection as the head of their respective groups is that their respective algorithms
allow them to accept data of any magnitude (negative to positive infinity) whilst returning an
output in the range 0 to 1 for the former and -1 to 1 for the latter, thus making them the most
suitable to use as a starting transfer function in the network (Beale et al., 2014). Training
algorithms are a set of instruction code that governs how the NN will be optimised such as
the setting of weights and biases, for instance (Beale et al., 2014). There are three types of
training algorithms and these are: the gradient descent, conjugate gradient and quasi-Newton
algorithms (Sharma and Venugopalan, 2014).
The gradient descent algorithms evaluate function gradients at the initial guess value and take
steps proportional to the negative direction of the gradient in order to find a local minimum.
In conjugate gradient methods, the local minimum is located by searching along conjugate
directions. This implies that a faster convergence is achieved than with the gradient-descent
method. The quasi-Newton algorithms usually give better and faster optimisation than
conjugate methods however a greater amount of memory is required (Sharma and
Venugopalan, 2014).
Training functions are network functions that dictate a universal algorithm which sets weights
and biases in any given network to optimise performance. Matlab training functions can be
classified based on the type of training algorithm they use – all Matlab training functions thus
fall into one of the three given training algorithms (Sharma and Venugopalan, 2014). For this
study, two training functions were selected from each of the training algorithm groups to
make a total of six functions. The definitions of each of these functions are shown in Table 5-
2.
130
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Table 5-2. Definitions of the training functions used (Beale et al., 2014; Sharma and
Venugopalan, 2014)
Training Training Algorithm (Definition) Updates network weights
Function Group according to:
Traingdm Gradient Descent Gradient descent with momentum
Trainrp Gradient Descent Resilient backpropagation (Rprop)
Trainscg Conjugate Gradient Scaled conjugate gradient method
Traincgf Conjugate Gradient Conjugate gradient backpropagation with
Fletcher-Reeves updates
Trainbfg Quasi-Newton Broyden-Fletcher-Goldfarb-Shanno
(BFGS) quasi-Newton method
Trainlm Quasi-Newton Levenberg-Marquardt optimisation
The overall performance of each Matlab transfer function – logsig and tansig for this study, is
an aggregate of the individual performances of the training functions listed in Table 5-2.
Figure 5-11 is a pictorial representation of how the degree of suitability of the transfer
functions for determining the final NN model parameters is dependent on the combined
performances of the individual training functions. The testing phase of the NN model
development involved the use of two transfer functions, each with several training algorithms
as shown, in Figure 5-10.
Figure 5-11. Combined performances of each training function defines the degree of
suitability of the transfer function in determining the final NN model parameters
131
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Data from Dugstad et al. (1994b), Nordsveen et al. (2003) was used in developing the NN
model. The information displayed in Table 4-1, Section 4.2.1 is a detailed summary of the
collated data and the descriptive statistics. In total, there are 75 data-points and these were
divided into training and testing sets with 60 and 15 data-points for each set respectively. A
summary of the test set variables are shown in Table 4-2, Section 4.2.1.
Results
Figure 5-12 shows the plot of the sum total of the absolute error against number of neurons
for the logsig transfer function.
Sum Total of Absolute
Error (mm/year)
Figure 5-12. Line plots of the sum total of absolute error against number of neurons for
the logsig transfer function
The sum total of absolute errors decreases with increasing number of neurons for half of the
training functions tested, namely: traingdm, trainbfg and trainlm. Of all the training
algorithms tested, it is observed that the trainscg and trainlm functions have the lowest
overall absolute errors. This implies that predicted corrosion rates were consistently closer in
magnitude to the actual corrosion rates from the original data source (Dugstad et al., 1994b),
(Norsdveen et al., 2003) for the given range of number of neurons. The maximum number of
132
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
neurons tested was set at 20 to avoid overtraining. The sum total of absolute errors for
trainscg is approximately 100mm/year, which is equivalent to 100/15 ≈ 6.67mm/year (for 15
data-points).
From Figure 5-13, it is observed that for half of the training functions (traincgf, trainbfg and
trainlm), the magnitude of the coefficient of determination (R2-coefficient) increased from 5
neurons to 20 neurons. Again, the trainlm function outperformed all others in terms of the
magnitude of the R2-value.
Figure 5-13. Line plots of R2-value against number of neurons for the logsig transfer
function
The R2-value line plots for the tansig function are shown in Figure 5-14. The magnitudes of
the R2-coefficients increase with increasing number of neurons in a similar trend to that seen
in Figure 5-13. The trainlm function has the highest R2-values indicating greater precision in
prediction. It is also noted that while the R2-coefficients for each training function fluctuates
from 5 to 20 neurons for both logsig and tansig transfer functions, the only exception is that
of the consistently high-levelled trainlm.
133
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 5-14. Line plots of R2-value against number of neurons for the tansig transfer
function
The bar charts for the mean R2-coefficients for each training function from the logsig and
tansig transfer function tests shown in Figure 5-15, confirm that the trainlm functions
consistently performs better than all the others.
Figure 5-15. Multiple bar charts showing the mean R2-coefficients for each training
function for both logsig and tansig transfer functions
134
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
It is also noted that the trainlm, being a quasi-Newton algorithm, is generally considered to be
better than the other training functions at optimisation as they converge faster in performing
nonlinear computations (Sharma and Venugopalan, 2014). Furthermore, there is additional
evidence of better optimisation capability when observing the bar charts of Figure 5-16. It has
the least average absolute corrosion rate error, indicating that it is the most suitable training
function for the NN model due to its greater accuracy and precision.
Corrosion Rate Error (mm/year)
Mean of Absolute
Training Functions
Figure 5-16. Multiple bar charts showing the absolute error means for each training
function for both logsig and tansig transfer functions
Sum Total of Absolute Error (mm/year)
Figure 5-17. Line plots of sum total of absolute error against number of neurons for the
tansig transfer function
135
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 5-17 shows the plot of the sum total of absolute errors against the number of neurons
for the tansig transfer function. It is observed that for the training functions, namely,
traingdm, trainscg, trainbfg and trainlm, there is a decrease in the sum total of absolute error
with increasing number of neurons. The reason for this being that the NN model has a greater
capacity to learn about the inherent patterns in the dataset more easily than when there are too
few neurons (Alsmadi et al., 2009). Again, the sum total of absolute errors for trainlm is
approximately 95mm/year, which is equivalent to 95/15 ≈ 6.33mm/year (for 15 data-points).
The number of layers to be used was determined by testing four different neuron
configurations in multiples of 5 as shown in Table 5-4. Essentially, the total number of
neurons was distributed in an organised manner within layers.
Table 5-4. Summary of the results obtained for the tested neuron configurations
Neuron Configuration Total Number of Neurons R2-value
[5 5] 10 0.71
[5 5 5] 15 0.66
[5 10] 15 0.72
*[10 5] 15 0.91
* The layers are arranged with 10 neurons in the first layer and 5 neurons in the second layer
From the R2-coefficients of Table 5-4, it is seen that the [10 5] neuron configuration is the
final choice for developing the NN model. It produces the most accurate CO2 corrosion
predictions. A summary of the NN model properties is presented in Table 5-5.
136
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The plot in Figure 5-18 shows the performance of the 15-neuron, [10 5] configuration, 2-
layer neural network model. The magnitude of the R2-value and the closeness of all data-
points to the 95% confidence intervals in Figure 5-18 indicate the model is accurate. The
number of points outside of the upper boundary of the 95% confidence interval suggests
some degree of over-prediction is occurring which may be due to variation within the dataset
at low-mid temperatures.
137
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 5-19 shows the error bar chart for the NN model. This plot highlights that there is
over-prediction occurring for small corrosion rates.
Data from multiple sources was used in developing the NN model: (Choi and Nesic, 2009;
Cui et al., 2006; Hesjevik et al., 2003; Zhang et al., 2012a). The information displayed in
Table 4-5, Section 4.3.1 is a detailed summary of the collated data and the descriptive
statistics. In total, there are 22 data-points and these were divided into training and testing
sets with 16 and 6 data-points for each set respectively. The bar chart of Figure 5-20 shows
the distribution of the given data-points from each of the sources.
138
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Number of data-points
Data Sources
Figure 5-20. Bar chart showing the distribution of data-points from each source
Overall the number of data-points from the Zhang et al (2012a) study exceeds those of the
other sources (63% share) due to the fact that the experimental corrosion rate results for this
source were carried out for the widest range of temperatures (50-130oC) and pressures (9.5-
23.3MPa).
For the other sources, corrosion rate tests were carried out by maintaining a constant
temperature whilst varying pressures or maintaining a constant pressure while varying
temperatures as is the case with the Choi and Nesic (2009) and Cui et al (2006) studies
respectively. For the study by Hesjevik et al (2003), tests were focused on measurement of
corrosion rates for temperatures less than 30oC. Experimental corrosion rate measurements
were grouped into classes of 0-4, 4-8,...,16-20mm/year. Figure 5-21 shows the distribution of
corrosion rates in these classes.
139
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Number of data-points
Figure 5-21. Bar chart showing the distribution of data-points for the grouped
experimental corrosion rates
From Figure 5-21, the number of data-points for the mid-corrosion rate magnitude (8-
12mm/year) is greater than those for end-point corrosion rate groups (0-4 and 16-
20mm/year).
Number of data-points
Temperature (oC)
Figure 5-22. Bar chart showing the distribution of data-points for the recorded
temperatures
140
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The bar chart in Figure 5-22 shows the distribution of data-points for the recorded
experimental temperatures. There are more data-points in the mid-temperatures (50 and 60oC)
than for end-point temperatures (24 and 150oC).
Figure 5-23 shows the corrosion rate-temperature profile for the training dataset. A
polynomial curve fit through the points depicts the classic peak observed for CO2 corrosion
rate as a function of temperature (De Waard and Lotz, 1993). It is noted that the statistical
range of corrosion rates in the mid-temperatures (50-80oC) is ~11mm/year, highlighting that
the greatest variation in the magnitudes of corrosion occurs in these temperatures.
Corrosion Rate (mm/year)
Temperature (oC)
Figure 5-23. Corrosion rate against temperature plot for the training dataset
Results
The variation of the magnitudes of R2-coefficients of the training functions with number of
neurons for the logsig transfer function is shown in Figure 5-24. There is a general decrease
in the magnitude of the R2-coefficients with increasing number of neurons for all training
functions except for the Trainscg. This fall in the magnitude of the correlation coefficient
with increasing number of neurons indicates that model performance is decreasing despite the
increasing network size. This may be due to the presence of a greater number of redundant
nodes or synaptic weights in the NN or that training is possibly diverging, resulting in
predicted values being very different from expected results (Haykin, 1999). Other reasons
141
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
may be that the training dataset is small or that the inherent characteristic of the data is such
that modelling with an increasing number of neurons beyond 5, results in a fall of the R2-
value.A similar trend was observed in an environmental research study (Abyaneh, 2014).
R2-Coefficient
Number of Neurons
Figure 5-24. Line plots of R2-values of training functions against number of neurons for
the logsig transfer function
For the tansig transfer function, Figure 5-25 is the variation of the magnitudes of R2-
coefficients against number of neurons.
2
R -Coefficient
Number of Neurons
Figure 5-25. Line plots of R2-values of training functions against number of neurons for
the tansig transfer function
142
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
There is a general dip in the magnitudes of the R2-coefficients as the number of neurons
increases from 5 to 20. Again, given that the training dataset is small, there is greater
certainty that as the number of neurons increases in the NN, the number of excess weights
also increases and there is a tendency for their presence to reduce model accuracy (Haykin,
1999).The only exceptions are the quasi-Newton algorithms, trainbfg and trainlm. The quasi-
Newton algorithms are generally known to give better optimisation results than the others,
particularly for network sizes that are not large Beale et al., (2014), Sharma and
Venugopalan, (2014).
Figure 5-26 shows the performances of each training function for both logsig and tansig
transfer functions in terms of the average R2-value.
R2-Coefficient
Training Functions
Figure 5-26. Bar charts showing the means of the R2-values of each training function for
both logsig and tansig transfer functions
The trainlm function again outperforms all the others. The traincgf and trainbfg rank second
and third respectively on the R2-value performance test. It is also worth noting that the
training functions mostly perform better with the tansig transfer function than the alternative
logsig transfer function, the only exception being the traingdm function. This may be due to
the effect of the unique combination of the input data, the logsig transfer function and the
algorithm in the traingdm training function culminating in a good performance level. Overall,
a steady, high magnitude of R2-coefficient is maintained by the trainlm function with an
143
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
increasing number of neurons. This thus shows its suitability for use in developing the final
NN model.
The training functions performances were also assessed in terms of the sum of their absolute
errors. Figure 5-27 is the bar chart variation of the sum of absolute errors of the training
functions against number of neurons.
Sum of Absolute Error (mm/year)
Training Functions
Figure 5-27. Bar charts showing the means of absolute corrosion rate error of each
training function for both logsig and tansig transfer functions
It is seen that the traingdm has the least favourable performance in terms of the mean
absolute corrosion rate errors obtained while the trainrp and trainbfg have the smallest errors
for both logsig and tansig transfer functions as a combination. On a singular basis, the trainlm
has the least error when the tansig function is in use, followed closely by the traincgf
function. In summary, the best training functions appear to be the trainlm and traincgf. It
must be borne in mind that for all modelling tests conducted and discussed so far, the
condition applied is that R2≥0.3.
144
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
results as shown in Figures 5-24 and 5-25. In particular, the NN performances were relatively
high for neuron numbers of 5 to 15. The number of layers to be used was determined by
testing four (4) different neuron configurations in multiples of 5 as shown in Table 5-6.
Table 5-6. Summary of the results obtained for the tested neuron configurations
Figure 5-28 shows the comparative line plots for both traincgf and trainlm functions.
Sum of Absolute Error (mm/year)
Neuron Configuration
Figure 5-28. Line plots of sum of absolute corrosion rate error against neuron
configuration for the traincgf and trainlm functions
From Figure 5-28, the trainlm function consistently has the least absolute corrosion rate
errors implying that it is a better training function for developing the final NN model. It is
also noted that NNs with one hidden layer often encounter difficulties with approximation
and model-fitting due to global interaction between neurons. On the other hand, NNs with
two hidden layers have no such issues, as the first layer extracts local features within the data
while the second layer extracts the global features (Abyaneh, 2014).
145
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Also, from Table 5-7, the [5 5] neuron layer configuration offers the highest R2-value. This
value coincides with the smallest corrosion rate error, as seen in Figure 5-28. This corrosion
rate error is equivalent to 10.61/6 ≈ 1.77mm/year per plotted data-point on the NN model
corrosion rate against experimental corrosion rate plot shown in Figure 5-29.
The plots in Figure 5-29 and Figure 5-30 show the performance of the 10-neuron, [5 5]
configuration, 2-layer neural network model. The magnitude of the R2-value and the
closeness of all data-points to the 95% confidence intervals in Figure 5-29 indicate the model
is highly accurate. Figure 5-31 shows the corrosion rate-temperature profile for the test
dataset with corresponding NN model predictions. The best fit line is a polynomial function
with an R2-coefficient of 0.55. A higher R2-value can be attained however this results in a
distortion of the overall shape and is badly conditioned. There is an increase in corrosion
rates as temperatures increase from 20oC to 80oC owing to temperatures accelerating the
chemical and electrochemical corrosion reactions. The precipitation rate though is said to
increase as temperatures increase hence protective layers form on the metal surface leading to
a reduction in corrosion rate: Johnson and Tomson, (1991), Yin et al., (2009).
NN predictions are close to the test data results shown in Figure 5-31, appearing mostly as
pairs of data-points. There is a slight degree of inaccuracy in model predictions for the mid-
temperature region of 50-80oC and this is attributed to greater variation in the training and
test set corrosion rates.
146
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 5 -29. NN Model corrosion rate against experimental corrosion rate plot
Corrosion Rate Error (mm/year)
Test Data
Figure 5-30. Error bar chart – Error is the difference between experimental and model
corrosion rate
147
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Temperature (oC)
Figure 5-31. Corrosion rate against temperature plot for the testing dataset with NN
model predictions
5.7 Conclusions
For both low and high pressure CO2 corrosion data, the tansig transfer function consistently
outperformed the logsig transfer function for most training functions with the exception of
the traingdm and trainrp training functions, in terms of the absolute corrosion rate errors. As
these two algorithms are gradient descent functions, their under-performance relative to the
conjugate gradient and quasi-Newton algorithm function alternatives is to do with the
technique of finding and locating local minima while carrying out nonlinear optimisation.
Also, in terms of the magnitude of the R2-coefficent, the tansig transfer function offered
better results for all training functions except the traingdm. Again, the quasi-Newton
algorithm functions, trainbfg and trainlm are known to offer better optimisation results.
The R-squared coefficient for the final neural network model is 0.91 for both low and high
pressure CO2 datasets. The models provide good fits for the corrosion test set. For the low
pressure CO2 corrosion NN model, all test points are in close vicinity to the 95% confidence
bounds indicating a high degree of accuracy however for test experimental results of lower
magnitude; there is a slight tendency for over-prediction. For the high pressure CO2 corrosion
148
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
NN model, all test points are within the bounds of the 95% confidence limits indicating a
high degree of accuracy. The error test bars indicate that predicted results are within
3mm/year of the CO2 corrosion test set.
As is usually the case with CO2 corrosion, the mid-temperature experimental values of ~60oC
induce the greatest corrosion rates due to the Arrhenius theory - the fact that higher
temperatures speed up corrosion rates by providing greater amounts of kinetic energy to
reacting species on a molecular level. However the limiting factor preventing a continuous
induced elevation of CO2 corrosion rates beyond levels of the mid-temperature values to
higher levels of the very high end-point temperatures (110-150oC), is the consequent
formation of iron carbonate (FeCO3) and magnetite. Though the ferrous ion concentration
([Fe2+]) increases in solution with higher temperatures, solubility decreases, resulting in scale
formation which coats the metal surface and brings about a decrease in corrosion rates.
This phenomenon reflects on both training and testing datasets with mid-point temperatures
coinciding with very high corrosion rates. Also, in these datasets, and in particular, the high
pressure CO2 dataset, 50% of the data-points lie around the 50-80oC range, which represents
a range of ~11mm/year in the training dataset – a wide variation.
NNs tend to predict less accurately when trained with data of considerable variation. This is
depicted in the model vs experimental plot where the high corrosion rate points are on
opposite bounds of the 95% confidence limits while other data-points are not as widely-
spaced apart on the plot. A summary of the NN model performances for low and high
pressure is presented in Table 5-8.
149
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
150
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
151
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The model-developing data was summarised as descriptive statistics, given in Tables 4-1 and
4-5 (Chapter 4, Sections 4.2.1 and 4.3.1) for the low and high pressure CO2 corrosion datasets
respectively. Figures 4-1 and 4-15 are box plots of the data summarised in Tables 4-1and 4-5,
for low and high pressure CO2 corrosion respectively. Also, the test set was summarised in
Tables 4-2 and 4-6 for low and high pressure CO2 data respectively, for prediction and the
subsequent calculation of the R2-coefficient.
Classical logic is based on the assumption of bivalence; that there are two truth-values (true
and false) and that the value of any given logical formula is uniquely defined by the truth-
values of its components: bivalence and truth functionality respectively (Behlolavek and Klir,
2011). Various many-value logics employ the use of truth functionality while abandoning the
bivalence concept (Behlolavek and Klir, 2011).
Classic logic and predicates of classic logic are linked closely with classic sets. A predicate is
a system that contains formulae and quantifiable variables. A logic operation on a predicate
produces a unique solution – an operation on a classical set.
Figure 6-1. Close connection between classical logic and classical set and their
relationships with predicates of logic
However with the abandoning of bivalence at the root core of many-value logics, the
connection between logic and sets was severed. Historically, fuzzy logic was created by Lotfi
Zadeh in the 1960s (Ukil, 2010). Lotfi Zadeh renewed the connection between many-value
logic and classical sets in his 1965 publication on fuzzy sets (Zadeh, 1965). A subsequent
152
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
publication was made ten years after the first seminal paper was released, in which the term
fuzzy logic, was used (Zadeh, 1975).
Fuzzy logic systems are an extension of multi-valued logic with the aim of developing
approximate reasoning that is closer in spirit to human thinking and natural language than
traditional logic systems: (Hajizadeh, 2006; Ukil, 2010; Fuzzy Logic Toolbox User’s Guide,
2015).
For any set X, the membership function on X, 𝜇(X), is a function from X to the real unit
interval [0,1] (Zadeh, 1965). Membership functions will be discussed further in the next
section. The broad definition of fuzzy logic is almost synonymous with set theory, to which
fuzzy sets belong. It is a branch of set theory that deals with the impreciseness of certain
phenomena such as classes of objects with unsharp boundaries in which membership is a
matter of degree. It is thus related to both logic as well as set theory (Behlolavek and Klir,
2011).
153
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The membership degree 𝜇A(X) quantifies the grade of membership of the element x to the
fuzzy set A. The simplest membership functions consist of straight lines and the most basic is
the triangular membership function as shown in Figure 6-3A. This function is essentially a
collection of three points forming a triangle. The other straight-line function, the trapezoidal
membership function, which has a flat top and two sloped sides, is shown in Figure 6-3B .
A B
154
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
High (𝜇=1.0)
Low (𝜇=0.0)
In general, the more complex membership functions such as the Gaussian, generalised bell,
sigmoidal, polynomial and the Z, S and Pi curves exhibit a continuous form as shown in
Figure 6-5. However, it must be noted that of all the continuous membership functional
forms, particularly the sigmoidal and bell-shaped functions have relatively little practical use
in fuzzy control (Ukil, 2010).
High (𝜇=0.90)
Low (𝜇=0.30)
155
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The Mamdani FIS was proposed by Ebrahim Mamdani in 1975 following Lotfi Zadeh’s 1973
publication, as an attempt to control a steam engine and boiler configuration by the
implementation of a set of linguistic control rules obtained from experienced, well-trained
human operators (Mamdani and Assilian, 1975; Mamdani, 1976; Mamdani, 1977).
In general, both Mamdani and Sugeno fuzzy inference systems involve the transformation or
fuzzification of input variables from a given data set into fuzzy sets, categorising the data into
degrees of membership (by membership function) and the consequent re-transformation of
the fuzzy sets into desirable output (Fuzzy Logic Toolbox User’s Guide, 2015). Where the
main difference lies between these two methodolgies, is in the re-transformation of fuzzified
data into desirable output. While for the Mamdani FIS , a defuzzification technique is used,
for the Sugeno FIS, a weighted-average technique is implemented to compute the output
(Fuzzy Logic Toolbox User’s Guide, 2015). Also, crisp functions are used as the
consequences for rules in the Sugeno FIS while the Mamdani FIS is not characterised by this
approach in defining rules (Ukil, 2010). Both methodologies can be used to effectively model
any system however the Mamdani FIS is more versatile as it does not matter what form the
desired output takes whereas the Sugeno FIS is only capable of rendering a constant or linear
output.
156
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
6.5 Low Pressure CO2 Corrosion Prediction using a Fuzzy Inference System
6.5.1 Methodology
The Mamdani fuzzy inference system (FIS) was used in the development of the Mamdani
model based on the given low pressure CO2 datasets: (Dugstad et al., 1994b; Nordsveen et
al., 2003). The model’s performance was then tested using the test sample dataset. The
Mamdani FIS processes and its corresponding methods are listed in Table 6-1. These FIS
properties produce the most desirable output for CO2 corrosion modelling in Matlab.
Fuzzy operators: ‘And’ and ‘Or’, as listed in Table 6-1, are then applied to the fuzzified data,
thus converting the data into an antecedent. Once the fuzzy rule weights are set, implication
involves taking a single number from the antecedent as input and forming a fuzzy set as
output, the consequent. The consequent is represented by a membership function, which
weights appropriately the linguistic features that are attributed to it. Implication is
implemented for each fuzzy rule. The ‘Min’ or minimum implication method tends to
truncate the output fuzzy set as opposed to the ‘Prod’ or product implication method which
scales the output fuzzy set. Also, the centroid defuzzification method was chosen over the
bisector method for its suitablity (See Appendix A6.1 a and A6.1b).
157
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
medium
medium
high
high high
Figure 6-6. Fuzzy inference diagram showing all parts of the fuzzy inference process –
from fuzzification through to defuzzification (Fuzzy Logic Toolbox User’s Guide, 2015)
Aggregation is the process by which the fuzzy sets that represent the output of each rule are
combined to form an aggregate fuzzy set. The input to aggregation, is the set of fuzzy sets for
each rule, defined by the implication process. The aggregate output is such that there is one
fuzzy set to each input variable.
Defuzzification follows aggregation, thus receiving the input of aggregation, a single fuzzy
set and then returns a single number, that best represents it. There are five different options
for Mamdani FIS defuzzification – centroid, bisector, largest of maximum (lom), smallest of
maximum (som) and middle of maximum (mom). The centroid was selected because it
returns the centre of the area under the curve, thus producing the most desirable corroiosn
rate plot shape. It is important to note that while the fuzzification to defuzzification steps
shown in Figure 6-6 are slightly different to the approach used by Mamdani, the overall idea
is identical (Matlab Fuzzy Logic User’s Guide, 2015). Also, the Mamdani FIS is based on the
original idea by Zadeh (Zadeh, 1973). Each of the input variables was assigned three
158
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
membership functions – Low, Medium, and High. For instance, the pH variable is shown in
Figure 6-7.
Magnitude of membership functions
Table 6-2. Membership function descriptor and corresponding corrosion rate ranges
Membership Function Descriptor Corrosion Rate Range (mm/year)
Very Low 0-10
Low 10-20
Low-Medium 20-30
Medium 30-40
Medium—High 40-50
High 50-60
Very High 60-70
In order to develop the FIS for the dataset, fourteen rules are defined. The following Table
shows some of the rules of the FIS based on close observation of variable trends within the
dataset.
159
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Table 6-3. Some rules for the development of the Mamdani and Sugeno FIS
Rule Temperature 𝑃𝐶𝑂2 Velocity pH Corrosion Weight
Number Rate
1 Medium High High Low Very High 1
2 Medium Low High Medium Low-Medium 1
3 *Not Medium *Not High Low Low Very Low 1
4 Medium Low High Low Medium 0.1
5 High High High Low Medium-High 0.1
6 High High Medium Low High 0.01
7 Low Low High Medium Low 0.001
___________________________________________________________________________
* Implies that these descriptors are negations
The rule viewer is set once these rules are established in the FIS. Three-dimensional (3-D)
plots of each of the variables in the dataset permit observation of any two input variables with
CO2 corrosion rate. One such plot is shown in Figure 6-8.
Corrosion Rate (mm/year)
Figure 6-8. Three-dimensional plot of corrosion rate against 𝑷𝑪𝑶𝟐 and temperature –
result of Mamdani FIS
The 3-D plot in Figure 6-8 shows a dome-like shape for the variation of CO2 corrosion rate
with temperature and CO2 partial pressure. There is a general increase of corrosion rate with
increasing temperature and 𝑃𝐶𝑂2 , however while there is a distinct peak for the change in
160
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
corrosion rate with temperature as shown in the two-dimensional (2-D) plot in Figure 6-9, the
corrosion rate appears to level off for very high values of 𝑃𝐶𝑂2 ., as shown in Figure 6-10.
Corrosion Rate (mm/year)
Temperature (oC)
Figure 6-9. Mamdani FIS plot of corrosion rate against temperature
The corrosion rate plot against temperature shows a peak at a temperature of 55 oC. The
corrosion rate value at the peak is 57mm/year and as temperature increases further to 90oC,
corrosion rate decreases steadily. Thus the plot in Figure 6-9 follows the typical corrosion
rate against temperature plots described in research literature, for scaled corrosion (De Waard
and Lotz, 1993; Zhang et al., 2012b). For this particular data, the greatest magnitudes for
corrosion rate are at temperatures of 60oC, while there are lower corrosion rates at 90oC. The
main shortcoming of this plot is that corrosion rates are rather conservative and it takes into
account, the very high corrosion rates within the dataset. This plot does not represent the
mean corrosion rates, particularly at 60oC and at 90oC.
For the variation of corrosion rate with 𝑃𝐶𝑂2 , Figure 6-10, corrosion rates increase very
steadily from 15mm/year, for CO2 partial pressures in the range, 0-0.7MPa, to 17mm/year,
after which it increases steeply to 60mm/year for 𝑃𝐶𝑂2 values of about 1.2MPa. Beyond this
point, corrosion rates plateau and this depicts the dataset accurately. Theoretically, corrosion
rates are expected to increase with increasing 𝑃𝐶𝑂2 . In fact corrosion rates increase with
increasing 𝑃𝐶𝑂2 to an exponent of 0.7, as seen in Figure 2-7A, Section 2.4.1 (Dugstad et al.,
1994b). The reason corrosion rate does not carry on increasing throughout the range of 𝑃𝐶𝑂2
161
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
values is due to the fact that film formation on the metal surface limits CO2 corrosion rates,
by reducing the surface area of the metal exposed to the acidic medium.
Corrosion Rate (mm/year)
Figure 6-10. Mamdani FIS plot of corrosion rate against CO2 partial pressure
Although, for this particular study, the water is un-buffered, implying a more corrosive
medium, the protection provided by the surface film is sufficient to reduce corrosion rates at
very high 𝑃𝐶𝑂2 values. In un-buffered water, the film is said to comprise ferrous carbides and
alloying elements from the steel (Dugstad et al., 1994b).
The 3-D plot of Figure 6-11, shows that there are corrosion rate peak values for both
temperature and velocity due to the presence of the central crest of the dome. The 2-D plot of
corrosion rate against flowrate velocity (Figure 6-12) shows that the peak occurs at (7m/s,
57mm/year). An increase in flowrate velocities results in a fairly steep elevation of corrosion
rates. This is attributed to the erosional effect of fluid flow on surface film growth and
thickness. The greater the velocity of fluid flow, the greater the tendency for any surface film
to be washed away, exposing a greater area of the metal surface to the corrosive medium and
increasing corrosion rates (Dugstad et al., 1994b).
162
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Corrosion Rate (mm/year)
Figure 6-11. Three-dimensional plot of corrosion rate against flowrate velocity and
temperature – result of Mamdani FIS
However, corrosion rates reach a peak at (7m/s, 57mm/year) and level-off at (13m/s,
35mm/year). The higher fluid velocities wash away the surface film and at the same
preventing the medium from further inducing the dissolution of iron to ferrous ions in
solution at a sufficiently fast rate.
Corrosion Rate (mm/year)
Figure 6-12. Mamdani FIS plot of corrosion rate against flowrate velocity
163
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The overall effect is that corrosion rates slow down and then decreases as the fluid velocity
reaches the 13m/s mark. In general, it is difficult to model CO2 corrosion rate based on
established fluid mechanical theories of mass transfer and shear stress for this particular study
because of the effect that flow velocities have on the composition and the thickness of the
surface film layer (Dugstad et al., 1994b).
Figure 6-13 shows the prediction results from the Mamdani FIS model using the test dataset.
The R2-value is relatively low, with about half of the data-points lying outside of the 95%
confidence interval. The implication is that the model is not the most accurate of the all the
discussed CO2 prediction models and in terms of reliability, a greater number of high-
magnitude corrosion rates fall within the boundaries of the confidence interval than the low-
magnitude corrosion rates. Hence it is more accurate and reliable in predicting high-
magnitude corrosion rates and much less accurate in predicting low-magnitude corrosion
rates.
Figure 6-13. Mamdani FIS model corrosion rate against experimental corrosion rate
plot and error bar chart for selected test data – error is the difference between
experimental and model corrosion rate
164
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
6.6 High Pressure CO2 Corrosion Prediction using a Fuzzy Inference System
The data used for the fuzzy inference analysis in this study was obtained from the following
research publications: Choi and Nesic (2009), Choi et al (2010), Hesjevik et al (2010) and
Zhang et al (2012a).
6.6.1 Methodology
Mamdani and Sugeno fuzzy inference systems were implemented usng the Matlab fuzzy
toolbox. The variables used for the analyses are listed in Table 6-4, below:
Table 6-4. Variables used in the study and their corresponding ranges
Variable Range
o
Temperature ( C) 0-150
𝑃𝐶𝑂2 (MPa) 3.5-23.3
pH 3.1-6
Fluid flowrate (m/s) 0-4
Corrosion Rate (mm/year) 0-15
The data used in the analyses was gathered from the aforementioned sources and sorted as
listed in Table 6-4. A diagramatic representation of the FIS is shown in Figure 6-14, below.
Temperature
𝑷𝑪𝑶𝟐
Corrosion Rate
CO2 Fuzzy Inference System
pH
Fluid flowrate
Figure 6-14. A generic fuzzy inference system showing input variables to the left (T,
𝑷𝑪𝑶𝟐 , pH, fluid flowrate) and the output variable to the right (Corrosion Rate)
165
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Three membership functions are assigned to each input variable and five membership
functions are assigned to the output variable. Descriptors are used in naming the membership
functions. The input variable membership functions are: low, medium and high while the
output variable membership functions are: very low, low, medium, high and very high as
shown in Table 6-5. The membership function type used for both input and output variables
is the Gausssian function. This function type was chosen for its practicality and also, for the
fact in general, by the central limit theorem, the arithmetic means of samples from randomly
different data sets are likely to be normally distributed (Rice, 2007).
Table 6-5. Descriptors assigned to each of the input and output variable membership
functions
Input variable membership function Output variable membership function
Low Very low
Medium Low
High Medium
High
Very High
In order to obtain an output from an FIS, a number of rules and the respective weight
contribution for each rule has to be developed. The following table shows the seven self-
developed rules used by this particular FIS.
Table 6-6. Rules for the development of Mamdani and Sugeno FIS
Rule Temperature 𝑃𝐶𝑂2 pH Fluid Corrosion Weight
Number flowrate Rate
1 Medium High Low High Very High 0.4
2 High Medium High Low Low 0.1
3 Low Low High Low Very low 0.2
4 Medium High Low High High 0.1
5 High Medium Medium Medium Medium 0.1
6 High * Low Low Very low 0.05
7 * High Low Low Very low 0.05
__________________________________________________________________________________
* Implies that these variables are absent in the given rule
166
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Upon creation of the rules, the rule viewer is set as shown in Figure 6-15. The rule viewer
provides the facility to change the magntiude of the input variables while showing an
instantaneous change in the output (high pressure CO2 corrosion rate).
6
1
7
1
1
1
1
Figure 6-15. Rule viewer for all seven rules showing mini graphical representations of
each of the input variables and their effect on the output variable
The following Table shows the magnitudes of the variables used in producing the Mandani
and Sugeno FIS corrosion plots shown in Figures 6-16 and 6-17, respectively.
Table 6-7. Magnitudes of input variables and respective output variable results
Input Variable Output Variable
Temperature = 78.4oC Mamdani FIS corrosion rate = 7.5mm/year
𝑃𝐶𝑂2 = 15.5MPa
pH = 4.44 Sugeno FIS corrosion rate = 5.43mm/year
167
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Three-dimensional plots of corrosion rate against 𝑃𝐶𝑂2 and temperature are shown for both
Mandani and Sugeno systems in Figures 6-16 and 6-17.
Corrosion Rate (mm/year)
168
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Table 6-7 is a summary of results from the rule viewer for Figure 6-15. The Mamdani FIS
plot shown in Figure 6-16, illustrates the variation of corrosion rate with 𝑃𝐶𝑂2 and
temperature in three-dimensional space. It can be seen that for the temperature variable,
corrosion rates increase steadily and peaks at approximately 80oC and 12 mm/year at the
centre of the dome in accordance with the Van’t Hoff, Arrhenius and Kinetic theories (Tan
and Chan, 2011). Corrosion rates then decrease steeply as temperatures drop (See Appendix
A6.2).
For CO2 partial pressures, the two-dimensional Mamdani plot, shown in Figure 6-18, is
derived from the three-dimensional plot in Figure 6-16. It is observed that the corrosion rate
increases steadily with increasing 𝑃𝐶𝑂2 up till pressures of approximately 12MPa after which
there is a slight dip before it reaches the point, (~23MPa, 13.5mm/year): this is the highest
corrosion rate.
Corrosion Rate (mm/year)
𝑷𝑪𝑶𝟐 (MPa)
Figure 6-18. Mamdani FIS plot of corrosion rate against CO2 partial pressure
Higher corrosion rates are possible for high pressure CO2 corrosion than for low pressure
CO2 corrosion (Zhang et al., 2012b). However, there is a tendency for scale formation to limit
corrosion rates as scales are found to be thicker. However for this plot, the velocity is high
169
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
(4m/s) and the pH is low (3.5), hence corrosion rates continue to increase with increasing
CO2 partial pressures.
The Sugeno FIS plot shown in Figure 6-17, illustrates the variation of corrosion rate with
𝑃𝐶𝑂2 and temperature in three dimensions. The plot trend for the temperature axis of Figure 6-
17, shows an initial rise and subsequent plateau in corrosion rates hence exhibiting an
identical pattern to the front part of Mamdani plot dome. The variation of corrosion rates with
𝑃𝐶𝑂2 is only slightly different to the observed trend exhibited by the Mamdani plot. Corrosion
rates increase with increasing CO2 and reaches a peak at ≈13.8MPa and plateaus for
increasing 𝑃𝐶𝑂2 values as seen in Figure 6-19.
Corrosion Rate (mm/year)
𝑷𝑪𝑶𝟐 (MPa)
Figure 6-19. Sugeno FIS plot of corrosion rate against CO2 partial pressure
A peak is reached in CO2 corrosion rates because super-saturation of the aqueous medium at
the metal surface would ensue. Super-saturation in this context implies that the concentration
of ferrous ions (Fe2+) and carbonate ions (CO32-) have exceeded their solubility limits in the
aqueous medium or fluid hence film growth occurs on the metal surface. For high pressure
CO2 corrosion, thick scale formation is favoured (Zhang et al., 2012b) and this serves to limit
corrosion rates, even with increasing CO2 partial presssures. Thick scale covers the metal
170
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
surface and reduces the rate of iron dissoultion, moreover the aqueous medium is super-
saturated thus cannot accommodate additional Fe2+ ions which ultimately results in limiting
the magnitude of CO2 corrosion rate. Figure 6-20 shows the model vs experimental plot and
error bar chart for the Mamdani FIS. The model has a tendency for under-prediction. The
95% confidence bounds indicate that the reliability in prediction is not very high but this is to
be expected since the size of the dataset is small.
Figure 6-20. Mamdani FIS model corrosion rate against experimental corrosion rate
plot and error bar chart for selected test data – Error is the difference between
experimental and model corrosion rate
171
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
6.7 Conclusions
Classical logic comprises two parts: bivalence and truth functionality. Bivalence is the
concept of having two truth value (true or false) outcomes while truth functionality is the use
of a truth-value to represent the solution of a given logical formula. The link between
classical logic and classical sets was established through the application of bivalence.
By abandoning bivalence, these two systems were disconnected from each other only to be
re-connected once more by Lotfi Zadeh in the 1960’s. The development of fuzzy inference
systems has led to intelligent systems which use human thinking and natural language to
transform input data to output data.
For modelling low and high pressure CO2 corrosion, each of the input variables, namely:
temperature CO2 partial pressure, flow velocity, pH as well as the output variable, corrosion
rate were simulated using the Gaussian membership function or bell functions. The reason
that these were chosen is that they are continuous membership functions hence they depict
smoother and more rounded edges during profile prediction than the straight-lined or sharp-
edged results of the triangular, trapezoidal or linear piece-wise functions.
Additionally, suitable descriptors such as very low, low, medium, high and very high were
used with corresponding ranges such as 0-10mm/year to categorise corrosion rates into
different classes in order to aid corrosion simulation. Also, the rules were developed
separately for the low and high pressure CO2 datasets so that the inherent traits within each
dataset is fully captured during simulation.
Fuzzy systems in general, map input data to output data via a procedural process: input data
→ fuzzification → implementation →aggregation → defuzzification → output data. For the
fuzzification stage, which is the assigning of linguistic descriptors to the initial input data,
input is converted to the antecedent. The ‘Prod’ and ‘Max’ options were the selected fuzzy
operators for ‘And’ and ‘Or’ respectively because the alternative fuzzy operator options, the
‘Min’ and ‘Max’ produced simulated corrosion rate magnitudes that did not match-up
against expected results and so were unsuitable.
For the implication stage, which involves taking a single number from the antecedent as input
and forming a consequent by representing it with a membership function, the ‘Min’ option
was selected for its suitability. The ‘Prod’ option tended to scale the simulated results by a
small factor. For the aggregation stage, which is the combining of all fuzzy sets into a single
172
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
aggregate fuzzy set, ‘Max’ was chosen due to the apparent likeness of simulated results to the
experimentally measured data of the given datasets.
For the defuzzification stage, which is the re-transformation of fuzzified data into output data,
the centroid was selected out of a possible five alternatives because it returns the centre of the
area under the curve thus producing the most desirable results, particularly when simulated
trends for the various input variables are to be assessed against expected profiles.
These FIS processes were applied to both the low and high pressure CO 2 datasets following
testing and trial runs of the various other alternatives. The Mamdani system was the FIS
method deployed for both datasets. The Sugeno method was also applied to the high pressure
CO2 dataset however the R2-value was considerably low.
In terms of performance, the Mamdani FIS model attained a higher level of accuracy for the
high pressure CO2 dataset than for the low pressure CO2 dataset as shown in Table 6-8.
This may be due to the fact that though the size of the dataset for the high pressure CO 2 case
is considerably smaller when compared to that of the low pressure CO2 dataset, there is also
much less variation within it. There is also a substantially less scatter of data-points in the
model vs experimental plots for the high pressure CO2 dataset when compared to that of the
low pressure CO2 one.
Overall, the FIS model is not the most accurate prediction system there is for CO2 corrosion,
but it excels at showing realistic relationships between the various variables and the corrosion
rate. The simulated corrosion rate against temperature, 𝑃𝐶𝑂2 and flow velocity plots depict the
expected trend that is characteristic of experimental profiles.
173
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
174
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
175
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
7.1 Introduction
Monte Carlo simulation is a computational technique for determining the sensitivity of a
complex system by varying parameters within statistical constraints (Statistics and Machine
Learning Toolbox User’s Guide, 2015). It can also be defined as a numerical method of
solving mathematical problems by random sampling (Sobol, 1994). Historically, the birth of
this technique is widely accepted to be in 1949 and its main originators are: Stanislav Ulam
and John von Neumann (Sobol, 1994). Interestingly however, the theoretical foundation
behind the technique which involves the use of randomness in a determinative manner can be
traced back to the 18th century (Harrison, 2010).
Sections 4.2.1 and 4.2.3 covered the descriptive statistics and regression analysis for the low
pressure CO2 dataset respectively, while Sections 4.3.1 and 4.3.3 of the previous chapter
covered the descriptive statistics and regression analysis for the high pressure CO2 dataset
respectively. In the previous chapter, a fuzzy model was derived with some limitations
particularly in the prediction accuracy for low pressure CO2 data.
In this chapter, the focus is a mathematical model that involves the prediction of CO2
corrosion rate from the following input variables: temperature, CO2 partial pressure (𝑃𝐶𝑂2 ),
pH and flow velocity. In order to carry out Monte Carlo simulations, appropriate probability
distributions for the given datasets must be assumed or obtained in conjunction with a suitable
best-fit regression equation, which in this thesis, was derived from the chapter on statistical
analysis.
For each of the low and high pressure CO2 corrosion datasets, the multiple linear regression
equations were deemed to be most suitable for further analysis due to their moderately high
R2-coefficient values: ~0.7 each, when these equations were assessed with separate test
datasets, in Sections 4.2.3 and 4.3.3. Probability distribution plots and probability plots were
used to measure the goodness of fit of each corrosion dataset from one of several probability
distributions – normal, Rayleigh, exponential, lognormal and Weibull probability
distributions. Tests for normality were first carried out on the datasets such as the Anderson-
Darling and Kolmogorov-Smirnov tests followed by other probability distribution tests.
Sensitivity analysis plots were also used to determine the relative significance of each of the
input variables on CO2 corrosion rate.
176
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
∑ 𝑃𝑟(𝑋 = 𝑛) = 1 (7.1)
𝑛
Where:
177
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Hence, the probability of X falling within a given range [a, b] is given by (Sobol, 1994),
(Harrison, 2010):
Where:
𝐹(𝑥) is the integral of 𝑓(𝑥)and is the cumulative density function (cdf) of 𝑋 over
Given that all the variables discussed in this thesis (temperature, 𝑃𝐶𝑂2 , flow velocity, pH, and
corrosion rate) are continuous variables, Monte Carlo simulations were carried out using
continuous probability distributions. The main probability distributions used as starting
references were the normal and Weibull distributions. The probability density functions for
these distributions are shown in Table 7-1.
Table 7-1. Probability density functions for the normal and Weibull distributions
(Harrison, 2010), (Johnson et al., 1994)
Probability Distribution Probability Density Function
Normal (𝑥−𝜇)2
1 −
𝑓(𝑥; 𝜇, 𝜎) = 𝜎√2𝜋 𝑒 2𝜎2 𝑓𝑜𝑟 𝑥 ≥ 0 (7.3)
The 3-parameter Weibull pdf becomes the 2-parameter Weibull pdf when the location
parameter, 𝜃 = 0. The 2-parameter Weibull pdf is therefore a special case of the 3-parameter
Weibull pdf.
178
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
It is observed from Figure 7-1, that the Weibull probability plot is a better fit for the dataset
than the normal probability plot and as such will be used in modelling the Monte Carlo
simulation. Table 7-2 shows a summary of statistical tests carried out on the experimental
CO2 corrosion rate data that confirms the Weibull distribution as a better fit than the Normal
distribution.
179
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Table 7-2. Test statistic summary for normal and Weibull Probability density functions
Test Statistic Normal Distribution Weibull Distribution
Anderson-Darling (AD) 3.586; p-value<0.005 0.326; p-value>0.5
Kolmogorov-Smirnov (KS) Rejects Null hypothesis (H0) -
It can be concluded from the results that a distribution is a better fit provided that the
magnitude of the AD-test statistic is small and the p-value is greater than 0.5 at the 5%
significance level (α=0.05). The KS-test for normality rejects the null hypothesis at the 5%
confidence level. Given that the null hypothesis in this instance is the statement that there are
no significant differences between the dataset and the distribution fit, the rejection of the null
implies that there are indeed significant differences between the dataset and the distribution
fit. This provides further proof that the Weibull pdf is better at modelling experimental CO2
corrosion rates.
Probability density plots of the normal and Weibull distributions for the given dataset are
shown in Figure 7-2. It is also observed from Figure 7-2, that the histogram fit for the Weibull
probability density function, is a better representation of the dataset than the histogram fit for
the normal pdf. The characteristic feature of the density plot is that the sum of the areas of all
the histogram bars is equal to unity. This implies that the area of a histogram bar for a given
corrosion rate range represents the probability of obtaining that given corrosion rate range.
The exponential and lognormal probability distributions are also close fit functions for the
data. The exponential distribution is directly related and belongs to the Weibull family of
distributions while the lognormal distribution is known to be a good fit for natural processes,
chemical reactions as well as for materials undergoing failure due to reactions such as stress
and corrosion (Lee and Wang, 2003; Gronhölm and Annila, 2007; Baboian et al., 2005). In
general, both the Weibull and lognormal distributions are used extensively in reliability and
survival analyses (Lee and Wang, 2003). Both are good fits for data that is positively skewed
(Fahidy, 2005). The corrosion rate variable from the low pressure CO2 dataset is positively
skewed as discussed in Section 4.2.1 and seen from the boxplot in Figure 4-1. The lognormal
and exponential pdfs are shown in Table 7-3.
180
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Table 7-3. Probability density functions for the lognormal and exponential distributions
(Harrison, 2010; Johnson et al., 1994)
Probability Distribution Probability Density Function
2-parameter Lognormal (𝑙𝑛𝑥−𝜇)2
1 {− }
𝑓(𝑥; 𝜇, 𝜎) = 𝑥𝜎√2𝜋 𝑒 2𝜎2 𝑓𝑜𝑟 𝑥 ≥ 0 (7.6)
The 3-parameter lognormal pdf reduces to the 2-parameter lognormal pdf, when the threshold
parameter: 𝛾 = 0.
Figure 7-2. Normal probability density and Weibull probability density plots for
experimental CO2 corrosion rate dataset
181
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Simulations were carried out for two groups. Case Study I involved the modelling of each
input parameter as a uniform probability distribution function in conjunction with a
previously-derived regression equation. Case study II involved the determination of the best-
fit probability distribution function for each parameter and the subsequent application of these
distribution functions in a pre-derived regression equation.
For every test simulation, 100,000 (1x105) Monte Carlo simulation runs were executed,
resulting in a randomly generated dataset of 100,000 CO2 corrosion rates. In general,
increasing the simulation runs decreases the size of the class intervals (bins) of the
histograms, leading to smoother representative probability density function plots however this
comes as a cost to computation. Also, there is a risk of the appearance of very thin comb-like
lines in the pdf plot if the simulatoin runs are increased indefinitely. The number of runs
chosen is shown in Figure 7-3. As simulation runs increased by an order of magnitude,
likewise time elapsed increased. This however meant that at one million (1x106) runs, the
time elapsed for this section of code increased to 0.1 seconds and subsequently, a whole
second for ten million (1x107) runs. Therefore, the point before the angled deviation on the
plot was chosen as the ideal for simulation runs. The difference between 0.1s and 1s is small
182
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
however in a context where optimisation is required (discussed in Section 7.6), a bit more
time may be required to run through all the lines of code. A Monte Carlo simulation was
performed based on the statistical parameters of the experimental dataset. Figure 7-4 shows
the result of this process. The simulated data was modelled specifically on the shape and scale
factors of the Weibull distribution, shown in Table 7-4. This was done because the probability
plot and the two statistical tests, whose results are shown in Table 7-2, revealed the Weibull to
be the better fit.
Figure 7-4. Monte Carlo simulation using statistical parameters of the experimental
dataset
183
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The Weibull model fit statistics were then compared against those for the experimental data.
Table 7-5 is a summary of the results.
Table 7-5. Summary statistics for the use of the Weibull distribution as a direct fit for
the low pressure CO2 corrosion rate
Statistical Experimental Weibull Monte Percentage Error
Parameter Corrosion Rate Carlo Model (%)
(mm/year) (mm/year)
Mean 17.03 17.02 0.00*
Standard Deviation 17.02 16.78 1.41*
____________________________________________________________________________________________________
____________________________________________________________________________________________________
Two other theoretical distributions had also been found to be good fits to the corrosion rate
experimental data as seen in Figure 7-2. However, where the Monte Carlo simulation runs of
the exponential model improved on the statistics of the Weibull, the variance for the
lognormal model was quite high.
Table 7-6 Summary statistics for the exponential and lognormal probability
distributions
Statistical Experimental Exponential Monte Lognormal Monte
Parameter Corrosion Rate Carlo (mm/year) Carlo (mm/year)
(mm/year)
Mean 17.03 17.03 (0.00)* 19.39 (13.86)*
Standard Deviation 17.02 17.03 (0.06)* 32.76 (92.50)*
____________________________________________________________________________________________________
____________________________________________________________________________________________________
From the results of Table 7-5 and 7-6, it is concluded that the exponential distribution is a
better distribution than the Weibull for representing low pressure CO2 corrosion data.
184
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Where:
The form of the equation shown in (7.1) was used for each of the other variables, replacing T
with 𝑃𝐶𝑂2 , pH and velocity accordingly.Given that the number of simulated corrosion rates is
very large (N=100,000), a test sample subset (N=100) was randomly selected for the purpose
of carrying out the AD and KS statistical tests. The results of these tests are shown in Table 7-
7. The Mersenne twister was chosen as the algorithm for the generation of pseudo-random
numbers. It was implemented in Matlab with a seed of zero due to its high number generation
speed in comparison to other modern generators, its long period length and has also passed
several stringent statistical tests for randomness (Matsumoto and Nishimura, 1998).
Table 7-7. Test statistic summary for normal and Weibull probability density functions
Test Statistic Normal Distribution Weibull Distribution
Anderson-Darling (AD) 0.427; p-value=0.307 0.232; p-value>0.5
Kolmogorov-Smirnov (KS) Rejects Null hypothesis (H0) -
The AD-statistic of the test sample for both the normal and Weibull distributions is low in
magnitude and as such the dataset can be adequately modelled by both distributions. The KS-
test for normality rejects the null hypothesis. Given that the null hypothesis is the assumption
that no significant differences exist between the sample test and a typical normal distribution,
rejection of the null therefore implies significant differences exist.
185
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 7-5 shows the probability density function plot for the simulated corrosion rates. It is
observed that both normal and Weibull probability curves fit the data adequately particularly
for very low and very high corrosion rates (tailends of the dataset) however both curves seem
slightly off for corrosion rates in the 20-35mm/year range. This imprecise prediction can also
be observed on the cumulative density function (cdf) plot for the 20-35mm/year range, shown
in Figure 7-6. In general, the Weibull distribution is seemingly a closer fit to the corrosion
rate data than the normal hence is used to model this dataset.
0.035
Weibull
Normal
0.03
0.025
Density
0.02
Density
0.015
0.01
0.005
0
0 10 20 30 40 50 60 70 80
Corrosion Rate (mm/year)
Figure 7-5. Probability density function plot for Monte Carlo simulation of low pressure
CO2 corrosion rate
0.9
0.8
Cumulative probability
0.7
Cumulative probability
0.6
0.5
0.4
0.3
0.2
Corr_Rate_01_monte data
0.1 Weibull
Normal
0
0 10 20 30 40 50 60 70 80
Corrosion Rate (mm/year)
Figure 7-6. Cumulative density function plot for Monte Carlo simulation of low pressure
CO2 corrosion rate
186
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Table 7-8 shows the summary statistics for Case study I and how they compare against those
of the experimental corrosion dataset. It is observed that while there is a sizeable error in the
estimated mean, the standard deviation estimate is closer to the true value. A second case
study is carried out while assuming best-fit distributions for each predictor variable and the
results of the study will be compared against that of Case Study I. This is described in the
next Section.
____________________________________________________________________________________________________
The best-fit distributions listed in Table 7-9 were used as equivalent representations for the
corresponding predictor variables in the derived expression obtained from multiple linear
187
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
regression on the low pressure CO2 corrosion dataset (Equation. 7.3), and is used to perform
Monte Carlo simulation. Table 7-10 is a summary of the test statistics.
Table 7-10. Test statistic summary for normal and Weibull probability density functions
Test Statistic Normal Distribution Weibull Distribution
Anderson-Darling (AD) 0.605; p-value=0.113 0.579; p-value=0.114
Kolmogorov-Smirnov (KS) 0.0766; p-value=0.5881 0.0749; p-value=0.6209
The AD-test value for the Weibull distribution is smaller in magnitude than the corresponding
test-value for the normal distribution, 0.579<0.605. Also, the p-value of the AD-test for the
Weibull distribution is greater than the corresponding value for the normality test at the 5%
significance level. This proves the Weibull distribution is marginally better suited to fit the
data. The values from the KS-test also support the results of the AD-test.
0.045
Weibull
0.04 Normal
0.035
0.03
0.025
Density
0.02
0.015
0.01
0.005
0
0 10 20 30 40 50 60 70 80
Corrosion Rate (mm/year)
Figure 7-7. Probability density function plot for Monte Carlo simulation of low pressure
CO2 corrosion rate
188
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The plot in Figure 7-7 is the result of the Monte Carlo simulation using the best-fit Weibull
statistical parameters for shape and scale. This plot is approximately normal in appearance
despite the pH being the only predictor variable to be modelled as normally-distributed. This
observation is explained by the Central limit theorem, which states that with sufficiently large
or a near-infinite number of draws, the arithmetic mean of independent random variables,
each with their distinct averages and variances converges to the normal distribution (Rice,
2007).
Figure 7-8 is the cumulative density function plot of the distribution in Figure 7-7. It is seen
that both Weibull and normal distributions closely match the simulated data, such that it is
difficult to tell which distribution is more suitable by mere inspection. This emphasises the
significance of the statistical test carried out in Table 7-10.
0.9
0.8
0.7
Cumulative probability
0.6
0.5
0.4
0.3
0.2
Corr_Rate_01_monte data
0.1 Weibull
Normal
0
0 10 20 30 40 50 60 70 80
Corrosion Rate (mm/year)
Figure 7-8. Cumulative density function plot for Monte Carlo simulation of low pressure
CO2 corrosion rate
Thus, comparing the results from Case Study I, shown in Table 7-8, with the results of Case
Study II, shown in Table 7-11, the drop of approximately 20% in the mean percentage error
189
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
suggests that the results of Case Study II are marginally better than Case Study I, even though
the standard deviation increases by ~13%.
____________________________________________________________________________________________________
Sensitivity Analysis
Case Study II was used as the basis for the Monte Carlo sensitivity analysis. Figure 7-9 shows
the sensitivity plot for corrosion to each of the predictor variables for the low pressure CO 2
dataset.
Temperature
𝐏𝐂𝐎𝟐
Corrosion Rate (mm/year)
Velocity
pH
away of protective film formation on the metal surface. Higher velocities are likely to trigger
greater corrosion rates in carbon steel pipes by more readily exposing the metal surface. This
is so because scales are thinner, less dense and less compact than for high CO2 pressure
conditions (Zhang et al., 2012b). The effect of temperature and CO2 partial pressure can be
seen as the second most-influential predictor variables. Their effect here can be explained
from the principal component analysis (PCA) results from Section 7.2.2, as they are the
greatest contributors to the variation within this dataset. The pH is the least contributor to the
variation within the dataset from PCA analysis, which also can be seen to have little impact
on corrosion rates even with a 50% change in the magnitude of its standard deviation. The
velocity variable is the only variable that does not follow the PCA order in terms of influence
and statistically, this may be due to the imperfect mean and standard deviation percentage
errors for Case Study II with respect to the experimental dataset. The best-fit distributions are
the best possible matches to the predictor variable datasets but cannot completely describe
them due to the presence of random errors in measurements or noise (Cottis et al., 1999).
191
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Weibull
0.95
Normal
0.9
0.75
Probability
0.5
0.25
0.1
0.05
0 2 4 6 8 10 12 14 16 18 20
Corrosion Rate (mm/year)
For the KS-test statistic, again, the smaller the magnitude of the test result, the better, while
the p-value is greater for the normal than for the Weibull distribution suggesting that the
probability of obtaining a good fit is slightly better for the Normal than the Weibull.
Table 7-12. Test statistic summary for normal and Weibull probability density functions
Test Statistic Normal Distribution Weibull Distribution
Anderson-Darling (AD) 0.314; p-value=0.513 0.492; p-value=0.211
Kolmogorov-Smirnov (KS) 0.164; p-value=0.815 0.168; p-value=0.776
Figure 7-11 shows the probability density plot for the high pressure CO2 corrosion dataset. By
inspection, both normal and Weibull curves appear to fit the data accurately. However, the
AD and KS tests show the data more likely comes from a normal distribution.
192
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
0.12
Weibull
Normal
0.1
0.08
Density
0.06
0.04
0.02
0
0 5 10 15 20 25 30
Corrosion Rate (mm/year)
Figure 7-11. Normal probability density and Weibull probability density plots for
experimental CO2 corrosion rate dataset
Table 7-13 shows comparative statistics for the Weibull and normal distributions.
193
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 7-12. Monte Carlo simulated probability density function plot showing normal
and logistic curve fits
Both the normal and logistic curves are good fits for the data. However, the logistic curve
hangs slightly over the centrally-located histograms at the peak of the pdf. Since the data
comes from a normal distribution, the normal curve is a good fit. The following Table shows
the summary statistics for both Normal and Logistic distributions. The normal Monte Carlo
outperforms the alternative Logistic distribution in terms of the calculated statistical
parameters.
Table 7-14 Summary statistics for the normal and logistic probability distributions
Statistical Experimental Normal Monte Logistic Monte
Parameter Corrosion Rate Carlo (mm/year) Carlo (mm/year)
(mm/year)
Mean 10.78 10.79 (0.09)* 10.79 (0.09)*
Standard Deviation 4.75 4.76 (0.21)* 4.94 (4.00)*
____________________________________________________________________________________________________
____________________________________________________________________________________________________
194
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Table 7-15. Test statistic summary for normal and Weibull probability density functions
Test Statistic Normal Distribution Weibull Distribution
Anderson-Darling (AD) 0.624; p-value=0.101 0.618; p-value=0.075
Kolmogorov-Smirnov (KS) 0.064; p-value=0.813 0.064; p-value=0.814
The results of the given test statistics imply that both distributions closely match the data. The
magnitude of the AD-test result is slightly smaller for the Weibull than for the normal
distribution however the p-value at the 5% significance level is higher for the Weibull than for
the normal distribution. Hence both distributions are as good as each other. The result of the
KS-test statistic implies given that both distributions have the same KS-test value, the Weibull
is slightly better because of the higher p-value.
For the probablity density function plot of Figure 7-13, both distributions also show how
closely they match-up against the given data. The cdf plot of Figure 7-14 also shows that by
inspection, the better fit might be difficult to deduce. The Weibull curve fits the data less
accurately for corrosion rates in the range: 8 to16mm/year as well as for corrosion rates
greater than 20mm/year. The Normal distribution matches up quite closely with the data
except for corrosion rates in the range: 0-8mm/year. Despite the close fit of the normal curve,
by visual inspection, the AD-test statistic is sufficiently conclusive.
195
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
0.07
Weibull
Normal
0.06
0.05
0.04
Density
0.03
0.02
0.01
0
0 5 10 15 20 25 30
Corrosion Rate (mm/year)
Figure 7-13. Probability density function plot for Monte Carlo simulation of high
pressure CO2 corrosion rate
0.9
0.8
0.7
Cumulative probability
0.6
0.5
0.4
0.3
0.2
Corr_Rate_01_monte data
0.1 Weibull
Normal
0
0 5 10 15 20 25 30
Corrosion Rate (mm/year)
Figure 7-14. Cumulative density function plot for Monte Carlo simulation of high
pressure CO2 corrosion rate
196
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Table 7-16 shows summary statistics for Case Study I. The mean of the Monte Carlo dataset
is accurate with a percentage error of 0.56% however the standard deviation is not as accurate.
Optimisation of the Monte Carlo dataset would be beneficial. This is discussed in Section 7.6.
____________________________________________________________________________________________________
Table 7-18 shows the summary statistics for the Normal and Weibull probability density
functions.
197
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Table 7-18. Test statistic summary for normal and Weibull probability density functions
Test Statistic Normal Distribution Weibull Distribution
Anderson-Darling (AD) 0.694; p-value=0.068 0.801; p-value>0.5
Kolmogorov-Smirnov (KS) 0.066; p-value=0.784 0.067; p-value=0.773
From the Table, the AD and KS test statistic values have smaller magnitudes for the normal
distribution than for the Weibull distribution. This implies that the data comes from a normal
distribution. However, the p-value for the AD test is greater for the normal distribution than
for the Weibull distribution. For the KS test, the p-value at the 5% significance level is 0.784,
which is greater than the corresponding value for the Weibull distribution. This leads to the
conclusion that the data is normally-distributed. Figure 7-13 shows the Monte Carlo
probability density function plots for both normal and Weibull distributions.
0.07
Weibull
Normal
0.06
0.05
0.04
Density
0.03
0.02
0.01
0
0 5 10 15 20 25 30
Corrosion Rate (mm/year)
Figure 7-15. Probability density function plot for Monte Carlo simulation of high
pressure CO2 corrosion rate
198
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
It is observed from the pdf plot, that while the Weibull curve originates at point (0, 0), it does
not match-up closely against the data for very low and very high-magnitude corrosion rates.
The normal curve on the other hand, performs better at fitting the data, particularly for the
high corrosion rates of 15-30mm/year. Figure 7-16 shows the Monte Carlo cdf plot for the
data with normal and Weibull cumulative distributions.
0.9
0.8
0.7
Cumulative probability
0.6
0.5
0.4
0.3
0.2
Corr_Rate_01_monte data
0.1 Weibull
Normal
0
0 5 10 15 20 25 30
Corrosion Rate (mm/year)
Figure 7-16. Cumulative density function plot for Monte Carlo simulation of high
pressure CO2 corrosion rate
While the shapes for both curves roughly match-up against the data, the Weibull curves
deviates from the data for corrosion rates of the range: 7-17mm/year and for the tail-points of
the data. The normal curve generally approximates the data better than the Weibull curve with
evidently smaller deviations. Table 7-19 is the summary statistics for Case Study II.
199
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
____________________________________________________________________________________________________
The results of this Case Study were compared with those of Case Study I, presented in Table
7-16. It is apparent that for both statistical parameters, Case Study I outperforms Case Study
II. The predictor variables in Case Study I were modelled using Uniform distributions and not
their best-fit distributions, the latter was applied for Case Study II, yet the Monte Carlo
simulation results are better for the first Case Study than the second. This may be because the
experimental dataset is quite small in size therefore what may appear to be the best-fit
distribution may not necessarily be the case. The small size of the dataset makes it somewhat
flexible to alternatives when Monte Carlo simulations are carried out.
Sensitivity Analysis
Case Study I was used to evaluate the sensitivity of corrosion rates to the various predictor
variables of the high pressure CO2 corrosion dataset. The Monte Carlo sensitivity plot is
shown in Figure 7-17.
From the plot, it is observed that the corrosion rate was most sensitive to changes in
temperature, CO2 partial pressure, pH and flow velocity, in that order. The reason this is so is
partly explained by the results from PCA of the high pressure CO2 dataset in Section 4.3.2.
From a statistical perspective, the temperature and CO2 partial pressure induce the greatest
influence on the variance for this dataset. In addition, the flow velocity and pH are less
significant in terms of their contribution to the variance.
200
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Temperature
𝐏𝐂𝐎𝟐
Corrosion Rate (mm/year)
Velocity
pH
For the low pressure CO2 corrosion dataset, sensitivity analysis revealed corrosion rates to be
most sensitive to changes in the velocity. For this reason, optimisation was performed on the
velocity variable. The other three variables (temperature, 𝑃𝐶𝑂2 and pH) were modelled using
201
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
their best-fit distributions while carrying out Monte Carlo simulations. The best-fit
distributions were used given that the summary statistics (Case Study II) produced marginally
better results than when modelling was carried out assuming uniformly distributed variables
(Case Study I). The mean and standard deviation of the Monte Carlo simulated dataset were
then equated with the corresponding statistics for the experimental dataset. An initial estimate
for the velocity was chosen after which an iterative scheme was implemented in Matlab 2012a
to determine the actual flow velocity for the optimisation process. During optimisation, the
solutions for the flow velocity were: 5.47m/s. and 6.96m/s. Even though there is more than
one solution for the optimised velocity, a conclusion can be drawn, which is the velocity
needs to be greater than 5.47m/s for the simulated data to more accurately model the
experimental dataset. The solution is more likely to be closer to 6.96m/s, given that the
recorded mean for the velocity at low pressure stated in Section 4.2.1, Table 4-1 is: 7.20m/s.
Though, the best-fit distributions were used in optimisation for each of the variables, these
distributions were only the best-match to the variable datasets and not their ideal
representations. It is highly unlikely that a given theoretical distribution will completely
represent every given data-point in any given experimental dataset due to the presence of
random errors or noise (Cottis et al., 1999). For instance, Figure 7-18 shows the CO2 partial
pressure dataset and its best-fit distribution. It can be seen that though the Weibull is a good-
fit, it does not accurately model the tail-points of the CO2 partial pressure data, i.e., the very
low and the very high data-points.
Similarly for the high pressure CO2 dataset, optimisation is carried out with respect to the
variable that the corrosion rate is most sensitive to, which is temperature. Again, for high
pressures, the percentage errors in Case Study I suggests adequate precision in modelling,
hence the other variables are modelled using the uniform distribution. There is one solution
result when the means and standard deviations of the Monte Carlo and experimental corrosion
rates are identical. This solution is at 60.36oC. Though this value is well below the mean of
temperatures for the high pressure corrosion dataset stated in Section 4.3.1, Table 4-5:
(80.25oC), it must be noted that the assumption during optimisation is that all other variables
are uniformly distributed.
In general, optimising the Monte Carlo data proved to be unusual because for the high
pressure CO2 dataset in particular, the size of the experimental dataset is quite small hence
there is the possibility of having several fitting solutions or no solutions at all.
202
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 7-18. Probability plot showing the Weibull distribution as the best-fit for CO2
partial pressure variable
7.7 Conclusions
Monte Carlo simulation is a technique involving the use of repeated sampling of a random
discrete or continuous variable. The experimental datasets for low and high pressure CO2
corrosion are of a continuous nature, in that the respective variables can be measured at
different points and contain decimal numbers and not simply whole number integers as is
applicable to discretised variables. An attempt was made to find a suitable range for the
number of simulations to be used when carrying out Monte Carlo simulations. A range of
between a hundred thousand to one million (1 × 105 − 1 × 106 ) simulations seemed sufficient
for modelling.
The experimental datasets were first tested for normality using probability plots and then were
subsequently fitted against several other theoretical distributions. The low CO2 pressure
corrosion dataset was found to come from the Weibull family of distributions. The Weibull,
lognormal and exponential distributions were all good fits judging by the low percentage
errors of the estimated distributions statistical parameters with respect to those of the
203
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
experimental dataset. The statistical reason for this is due to the positive skewness of the
corrosion rate predictor variable, which was shown as a boxplot in the descriptive statistics of
Section 4.2.1. The importance of these theoretical distributions to reliability and survival
analysis is well-known. The lognormal distribution, for instance, is a good fit for materials
undergoing failure due to stress or corrosion.
For the high pressure CO2 dataset, the normal distribution was found to be a closer fit than the
Weibull distribution. This may partly be due to its small size. Unlike the low pressure CO 2
dataset, the corrosion rate response variable is approximately symmetrical and not positively
skewed as shown in the boxplot in Section 4.3.1. It therefore does not lend itself to being
modelled accurately by the Weibull family of distributions. The regression equations derived
from Sections 4.2.3 and 4.3.3 were used in the Case Studies for the evaluation of the
corrosion rate sensitivity plots, for the low and high pressure CO2 corrosion datasets
respectively.
For the low pressure CO2 corrosion dataset, Case Study II was used to derive the corrosion
rate sensitivity plot. This Case Study involved the use of best-fit distributions, which
produced better statistics than the alternative Case Study I, for which Uniform distributions
were assumed for the predictor variables: temperature CO2 partial pressure, flow velocity and
pH. Table 7-19 shows these results. Corrosion rates were proven to be most sensitive to the
flow velocity, temperature, CO2 partial pressure and pH, in that order. Flow velocities tend to
have a great impact on low CO2 corrosion because according to experimental findings, the
ferrous carbonate (FeCO3) scale formed is less-dense, less-compact and comparatively thinner
than the scale formed under high pressure CO2 conditions. The flow velocities under low
pressure CO2 conditions will have to be closely monitored to avoid a scenario where
moderately high to very high flowrates sweep away protective scale on the metal surface, thus
exposing the metal to further corrosivity.
For the high pressure CO2 corrosion dataset, Case Study I, in which all variables were
assumed to follow the Uniform distribution, outperformed Case Study II, in which the best-fit
distributions were used to represent the predictor variables. This may be due to the small
experimental dataset used as the basis for carrying out Monte Carlo simulations. Table 7-20
shows summary statistics for low and high pressure CO2 corrosion datasets.
204
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Table 7-20. Monte Carlo simulation results based on the derived regression equations
Corrosion rates for the high pressure CO2 dataset were found to be most sensitive to
temperature, followed by the CO2 partial pressure, pH and flow velocity, in that order. The
temperature and CO2 partial pressures are influential variables regardless of the dataset. This
is partly explained statistically by the PCA results of Sections 4.2.3 and 4.3.3. These variables
have the greatest contributions to the variances of both datasets.
In terms of the corrosion mechanism, temperature affects corrosion rates because it speeds up
the underlying transport and electrochemical processes. It is generally known that the rate
constants of most reactions increase with an increase in temperature and corrosion is not an
exception. This implies that where possible, temperatures will have to be kept low, to control
corrosivity. A 5 or 10% increase in the temperature is roughly equivalent to a 2.5mm/year
increment in corrosion rates. This is obtained from the sensitivity analysis plot in Section
7.5.2, Figure 7-17. The flow velocity has the least effect on corrosion due to the fact that the
thicker scales under high pressure conditions diminish its effect. Table 7- 21 shows the least
and most influential variables to corrosion rates for both low and high pressure CO2 corrosion
datasets.
205
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Table 7-21. The least and most influential predictor variables on CO2 corrosion
As with all attempts at modelling, there are usually limitations. The limitations of the
sensitivity analysis results are that they are based on derived regression equations whose R2-
values are moderately high, with adequate reliability but are not perfectly accurate models.
The use of best-fit distributions as equivalent representations in Monte Carlo simulations can
be somewhat simplistic, given that these theoretical distributions are not always representative
of the experimental datasets and are capable of adding noise into simulations.
Also, the high pressure dataset is small in size therefore simulations based on the use of this
dataset may not fully capture the intrinsic patterns as thoroughly as possible, such as when a
sufficiently large dataset is used.
206
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
207
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
8.1 Introduction
In Chapter 3, various CO2 corrosion models were discussed including the empirical 1991,
1995 De Waard and Norsok correlations (Nyborg, 2002; Nyborg, 2010) as well as the
mechanistic Freecorp model. Also, in Chapters 4 to 7, statistical, neural network, fuzzy, and
Monte Carlo models were derived respectively.
In this Chapter, the models discussed in Chapter 3 are assessed using an identical
experimental dataset used in determining the performances of the derived models of Chapters
4 to 7. Model performances were evaluated using the R2-statistic in conjunction with 95%
confidence intervals. In addition, the applicability and limitations of the derived models from
Chapters 4 to 7 are discussed.
8.2 Model Performances of the 1991, 1995 De Waard, Norsok and Freecorp
The test dataset was used to evaluate the performance of the models. The predicted outputs
were plotted against their equivalent experimental corrosion results in the same manner as
described for the previously discussed models.
Also, all of the data-points lie outside of the 95% confidence interval. This is due to the fact
that for this model, only temperature and 𝑃𝐶𝑂2 variables are required with the other two
variables, velocity and pH not represented. The model is therefore very limited hence the
model-experiment plot indicates that the parameters are inadequately describing the data.
208
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 8-1. The 1991 De Waard model corrosion rate against experimental corrosion
rate plot and error bar chart for selected test data – error is the difference between
experimental and model corrosion rate
209
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 8-2. The 1995 De Waard model corrosion rate against experimental corrosion
rate plot and error bar chart for selected test data
Norsok Model
Figure 8-3 shows the correlation of the Norsok model results with experimental results. It is
seen that the Norsok model has a very strong correlation with experimental results, as
indicated by the very high R2-value even though there is evidence of over-prediction – the
data-points are consistently above the ‘experiment=model line’.
210
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 8-3. Norsok model corrosion rate against experimental corrosion rate plot and
error bar chart for selected test data
Freecorp Model
Figure 8-4 shows the correlation of the Freecorp model results with experimental results. It is
seen that the Freecorp model has a rather weak correlation with experimental results, as
indicated by the very low R2-value. The data-points appear to be divided into two sub-groups,
where the low-magnitude corrosion rates are close together, under-predicted and the high-
magnitude corrosion rates that are more scattered, yet evenly dispersed around the
‘experiment=model line’.
Also, a significant number of data-points lie outside of the wide-range 95% confidence
interval. This information coupled with the fact that the R2-value is low implies that the model
is not sufficiently accurate and reliable. The wide confidence bounds also indicate there is a
greater degree of uncertainty in prediction results in comparison to other models.
211
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 8-4. Freecorp model corrosion rate against experimental corrosion rate plot and
error bar chart for selected test data – error is the difference between experimental and
model corrosion rate
Also, the model over-predicted all test data-points indicating a high degree of conservatism.
While the predictions are mostly within the 95% confidence bounds, the boundary limits are
very wide indicating a high degree of uncertainty in predictions.
212
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 8-5. The 1991 De Waard model corrosion rate against experimental corrosion
rate plot and error bar chart for selected test data – error is the difference between
experimental and model Corrosion rate
The model vs experimental results, shown in Figure 8-6, is characterised by a very wide
confidence interval. This implies a high degree of imprecision in prediction, even though most
data-points lie within the confidence bounds. In addition, all test points with the exception of
one is under-predicted.
213
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 8-6. The 1995 De Waard model corrosion rate against experimental corrosion
rate plot and error bar chart for selected test data
Norsok Model
Figure 8-7 shows the correlation of Norsok model results with the test dataset. The R2-
coefficient for the model predictions of high pressure CO2 corrosion is low. The model was
developed for predicting low pressure CO2 corrosion thus struggles for this test dataset. The
model under-predicts test set data hence it is not a conservative model for high pressure CO2
data unlike the very conservative De Waard models. Though all predicted points lie within the
95% confidence bounds, it is not a reliable model given the wide range for the confidence
intervals.
The Freecorp model was not discussed because for the high pressure dataset, it was very
conservative leading to considerable over-predictions.
214
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Figure 8-7. Norsok model corrosion rate against experimental corrosion rate plot and
error bar chart for selected test data
The derived NN models for both low and high pressure CO2 datasets were highly accurate, in
terms of their computed performances using the R2-statistic. In addition, the use of the tansig
transfer function in conjunction with the trainlm training function produced the best results.
This is because the trainlm being a quasi-Newton algorithm is known for offering excellent
optimisation results. The size of the networks obtained were also a reflection of the size of the
datasets being used for modelling. The low pressure dataset was bigger in size than the high
pressure dataset and a network configuration of [10 5] was sufficient for obtaining accurate
predictions. For the high pressure dataset, a network configuration of [5 5] produced desirable
results. In general, NN models are very useful; however their performances are dependent on
215
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
the selection of the training data used. It is imperative that the training data adequately
captures the trends within the overall data or is at least representative of the entire dataset
being modelled.
The advantage of the NN models is that despite the small size of the high pressure CO2
corrosion dataset, with the appropriate data division for training and testing coupled with
pruning, high model accuracy was attained. It is capable of predicting corrosion rates to
within 3mm/year. They are somewhat robust and are adaptable to handle datasets of
different sizes.
For the derived Mamdani FIS model, the obtained R2-coefficients varied markedly for the low
and high pressure CO2 corrosion datasets, as seen in the Conclusion Section of Chapter 6,
Table 6-7. The reason for this is because the low pressure dataset is bigger and contains a
greater degree of variation within it while the high pressure dataset is small and has much less
variation within it. In general, both derived models are not accurate. There is a more even
spread of test data-points about the experiment=model line for the low pressure CO2 corrosion
dataset, as seen in Section 6.5.1, Figure 6-12, than for the high pressure CO2 corrosion
dataset, as seen in Section 6.6.1, Figure 6-19.
Fuzzy models have been described as useful models that excel at filling banks or gaps of data
which is either unknown or otherwise unavailable. They also excel at producing accurate
profiles for the given predictor variables in a two or three-dimensional format. The plots of
simulated corrosion rate against each of the predictor variables (temperature, CO2 partial
pressure, flow velocity and pH) showed realistic depictions of what is expected in an
experimental setting. They have also been described as models that improve substantially in
terms of performance when combined with NN models, forming neuro-fuzzy models
(Hajizadeh, 2007). This model aspect is beyond the scope of this thesis.
For the described statistical models, the multiple linear regression equations were moderately
accurate for both low and high pressure CO2 corrosion datasets. They also produced good
scatters for test data-points around the experiment-model line on the Model vs experimental
plots, though with a slight degree of over-prediction for both low (Section 4.2.3, Figure 4-7)
and high pressure CO2 corrosion datasets (Section 4.3.3, Figure 4-21) respectively.
For the nonlinear regression equations, the performances were markedly different for the
datasets. While the R2-coefficient obtained for the low pressure dataset was moderately high
216
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
(R2=0.67), that of the high pressure dataset was very low (R2=0.10). The reason for this is
because nonlinear regression modelling involves the application of numerical techniques and
while the dataset is sufficiently large for the low pressure dataset, the high pressure dataset is
quite small. Therefore, there was a lack of convergence during numerical computation. This
manifests itself in the model vs experiment line, shown as wide 95% confidence bounds, as
seen in Section 4.3.4, Figure 4-25. This implies that the results are imprecise and unreliable.
The multiple linear regression models are versatile and lend themselves for further statistical
analysis such as Monte Carlo simulations.
The statistical models have a drawback, which is the size of the dataset being modelled. They
generally require sizeable amounts of data in comparison to NN models in order to attain high
performance levels in terms of accuracy. The advantage they have over the NN models is that
relationships between predictor variables are inferred. A given response variable can be
expressed explicitly in terms of its predictor variables, when parametric modelling is carried
out, as was done in this thesis, unlike the ‘black-box’ approach of the NN models.
Monte Carlo simulations use random sampling and statistical modelling to estimate statistical
parameters or mathematical functions of interest. This technique is therefore an extension of
statistical analysis. The advantage of this technique is that it offers a solution for modelling
data-points as a ‘collective’. For instance, low pressure corrosion rates were found to come
from a Weibull family of distributions which includes the exponential distribution as a special
case. Though, the lognormal distribution was also found to fit the data, after a 100,000 Monte
Carlo simulations, the best-fit distribution was the exponential; with mean and standard
deviation percentage errors given as: 0.00 and 0.06%.
The Weibull, exponential and lognormal distributions are extensively applied in reliability
and survival analysis as they accurately model naturally-occurring phenomena such as
corrosion and wear accurately. For the high pressure CO2 dataset, it is not as clear-cut which
distribution is a best-fit. However, the logistic, normal and Weibull distributions were good
solutions. After carrying out Monte Carlo simulations, the normal proved to be the best fit,
owing to the central limit theorem (Rice, 2007). The best-fit percentage errors were within
0.09 and 0.21% of the mean and standard deviation respectively.
While these inferences are very useful, they still do not provide insight as to how the various
predictor variables affect the response variable: corrosion rate. The derived regression
equations enabled the evaluation of the corrosion rate sensitivity with respect to each of the
217
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
predictor variables. The flow velocity and pH were found to be the most and least influential
predictor variables respectively on corrosion rate, for the low pressure CO2 corrosion dataset.
Also, the temperature and flow velocity were found to be most and least influential predictor
variables respectively on corrosion rate for the high pressure CO2 corrosion dataset.
Overall, it was challenging to model the low pressure data due to a high degree of variation.
The presence of outliers in this dataset causes skewness which the models struggled to cope
with thereby increasing the errors in prediction and lowering their accuracy. For the high
pressure dataset, variation within the dataset did not prove problematic in modelling as the
dataset was small however because of the few data-points, the nonlinear regression model, for
instance was unable to converge to a solution. This is because the model applies numerical
techniques to compute the solution which relies on the use of a large dataset to be more
effective. Table 8-1 is summary of the results obtained for the derived models.
The 1991 De Waard model with an R-squared coefficient of 0.54 showed a positive
correlation between model and experimental results however it generally under-predicted the
test dataset. This inadequate representation of the data is apparently due to the correlation
only having two predictor variables: 𝑃𝐶𝑂2 and temperature and as such misses out completely
218
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
on the effects of the other two missing variables on the data. The 1995 De Waard model
performed better than each of all the other models. In addition to attaining a high R2-value, all
the test data-points are in very close proximity to the 95% confidence bounds.
The Freecorp model has a low R2-coefficient, even for the low pressure CO2 test dataset
owing to its lack of scaling modelling in its predictions. As a result, it has a great tendency to
over-predict. Predictions can be partially split into two groups, as seen in Figure 8-4: test
data-points less than 25mm/year and test data-points greater than 25mm/year. Model
performances are summarised in Table 8-1.
For high pressures, all the established models performed poorly, as indicated by the low R2-
values in Table 8-2. This is the case because all these models were developed with the sole
purpose of predicting corrosion under low pressure CO2 conditions, characteristic of oil and
gas pipelines. The wide 95% confidence bounds for all models indicate that they are
imprecise and model predictions are unreliable.
*The Freecorp model R2-value was not computed for the high pressure CO2 test dataset because it was excessively
conservative which led to massive over-predictions
219
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
220
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
221
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
9.1 Conclusions
The effect of carbon dioxide (CO2) corrosion on pipelines is of great relevance to the
petroleum as well as the CCS industries. CO2 corrosion is responsible for lost production as it
brings about the gradual degradation of pipe internals with time. CO2 corrosion was
discovered in gas wells in Texas in the 1940’s and ever since there have been multiple
research undertakings on the topic.
The cost of corrosion cannot be overstated - general corrosion costs are said to be between 3
to 5% of an industrialised nation’s gross domestic product. The cost of CO2 corrosion in the
U.S. oil and gas industry as of 2014 was said to be ~$18 billion. Due to the negative impact of
current global emissions of the gas on the World climate, geo-sequestration efforts are
currently being explored by the CCS industries which involve the use of abandoned oilfields
and coal mines for the storage of the gas for geologically significant timescales.
CO2 as a dry gas is non-corrosive at oilfield and CCS conditions, however when it becomes
dissolved in water, it forms carbonic acid, which is corrosive. CO2 corrosion potentially poses
a greater risk when it manifests in the localised mesa-form as opposed to uniform corrosion.
In mesa-type corrosion, the medium velocity fluid flow coupled with the ‘right’ fluid
composition gives rise to localised corrosive attack on the metal surface. Also the iron carbide
(Fe3C) scale is prevalent in this form of corrosion. Unlike the ferrous carbonate (FeCO3)
scale, Fe3C provides less of a protective barrier to corrosion for the metal surface beneath.
These forms of corrosion risk coupled with the aforementioned costs necessitate the
importance of understanding the core concepts underpinning CO2 corrosion hence the need
for reliable corrosion rate prediction models. Various models have been developed over the
years for predicting corrosion rates in pipelines. Some of these include the 1975, 1991 and
1995 De Waard, Norsok and Freecorp models. These models were developed for application
in the oil and gas setting, known for their characteristically low CO2 partial pressures,
typically less than 2MPa (20bar). While their prediction performances are good for low CO2
partial pressure conditions, their performances for high CO2 pressure environments have been
shown to be poor.
Separate low and high pressure CO2 datasets obtained from research literature were used to
evaluate model performance in terms of their accuracy in prediction. The statistical measure
used is the R2-coefficient, with a minimum value of 0 and maximum value of 1, for a positive
correlation between the test dataset and model predictions.
222
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Various models were derived based on neural networks (NN), fuzzy, statistical and Monte
Carlo simulations. These will be discussed in the order in which they are listed. Neural
networks are inspired by the operation of the human brain and its neurons in the nervous
system. For modelling CO2 corrosion, performance-related tests were carried out on Matlab
2012a transfer and training functions.
The derived NN models for the low and high pressure CO2 datasets produced very good
results with high R2-values of ~0.9 and the narrow 95% confidence bounds indicated model
precision and a great certainty in predictions. All predictions were within 3mm/year of the
experimental data. In general, NN models are particularly useful when the size of the dataset
is small, which is currently the case for high pressure CO2 corrosion. They are adaptable and
are capable of producing excellent results for very small datasets however their drawback is in
their need for a good user-selected training set. Without a training set that captures the
variation sufficiently enough, results may not be forthcoming.
Fuzzy Inference Systems (FIS) or fuzzy logic systems are based on binary systems, but where
binary systems deal with extremes (0 or 1); fuzzy logic systems operate in the ‘fuzzy’ region:
0-1. In terms of performance, the Mamdani FIS attained a higher R2-value for the high
pressure CO2 test set than for the low pressure CO2 test set. The reason the model struggled in
model prediction, particularly for the low pressure CO2 dataset might be due to a greater
degree of variation within this dataset than with the comparatively smaller high pressure test
set. The models do however excel at modelling the corrosion rate vs predictor variable
profiles realistically
The statistical models were derived for both low and high pressure CO2 corrosion datasets.
Firstly, principal component analysis (PCA) was performed on both datasets. The PCA results
reveal the relative contribution of each of the predictor variables to the response variable. The
analysis revealed that the temperature and CO2 partial pressure had the greatest share in terms
of the contribution to the variance of both low and high pressure datasets. The multiple linear
regression equations were obtained with moderately high R2-values. For the high pressure
CO2 dataset, there was a tendency for over-prediction indicating some degree of
conservatism.
The Monte Carlo model performances were evaluated by estimating the means and standard
deviations of the test dataset as well as for the derived models, and computing the errors
between these and expressing them as a percentage error.
223
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
The low and high pressure datasets were analysed to find their best-fit theoretical
distributions. The low pressure CO2 dataset was found to match-up well against the Weibull
family of distribution. This is because the corrosion rates in the low pressure CO2 dataset
exhibit positive skewness. The outliers within this dataset are in the 60-70mm/year range. The
lognormal, Weibull and exponential distributions are good models for naturally-occurring
phenomena such as metals undergoing wear and corrosion. They are also good fits for
positively skewed data. The high pressure dataset was found to follow the normal distribution
although this is not as clear-cut as the low pressure case. This is attributed to the small size of
the dataset. The regression equations were applied in the derivation of the corrosion rate
sensitivity. Since the regression equations contain predictor variable terms, Monte Carlo
simulations were used to determine the variables that the corrosion rate is most sensitive to.
The sensitivity analysis for the low pressure CO2 data revealed that corrosion rates are most
sensitive to flow velocity, closely followed by the temperature and CO2 partial pressure and
least sensitive to the pH. The result implied that the flowrate for the low CO2 pressure dataset
has to be closely monitored as for instance, an initial 10% increment in the standard deviation
of the flow velocity results in a corresponding increase in corrosion rates by 5mm/year.
For high pressures, it was observed that the temperature is the most influential variable on
corrosion rate followed by the CO2 partial pressure, pH and flow velocity, in that order.
Interestingly, results from PCA revealed the temperature and CO2 partial pressure to be the
predictor variables that had the greatest contribution to the variance of the dataset. This is true
for the both low and high pressure datasets. For both datasets, corrosion rates are consistently
sensitive to these variables.
Also, as previously stated, temperatures increase corrosion rates as explained by the Kinetic
theory, and so is significant. At high pressure conditions, thicker, more compact and denser
FeCO3 scales are formed on the metal surface which offers greater resistance to the sweeping
effect of fluid flowrates. Under low pressure CO2 conditions, FeCO3 scales are thinner and
less dense therefore are more vulnerable to fluid flow effects. These can be easily washed off
thus exposing the metal surface beneath to further corrosivity.
The limitations of the Monte Carlo sensitivity analyses are that they are based on fairly
accurate regression equations however they are not flawless and do not completely capture the
variability in the datasets. Also, in some cases, the distributions selected as best-fit were the
best-match distributions, which do not necessarily mean that they describe every data-point
224
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
within the predictor variable dataset accurately. For some predictor variables, the best-fit
distributions fit the central set of data-points well but perform relatively poorly at the tail-
points (very low and very high-magnitude data). In these instances, the assumptions for best-
fit distributions introduce more uncertainties to further modelling.
In this thesis, a Mamdani FIS model was derived for the purpose of CO2 corrosion prediction
for both low and high partial pressures of CO2. However, a Sugeno FIS model was not
developed for low and high CO2 partial pressures. Also, due to the fact that the derived
models had limited accuracy, which is somewhat a feature of these models when used as
stand-alone tools, a neuro-fuzzy model can be developed. This model would combine the
good characteristics of both the NN and fuzzy models, which can potentially result in a more
robust tool for corrosion prediction.
225
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
_________________________________________________________________________________________________________________________________
Appendices
226
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
_________________________________________________________________________________________________________________________________
Temperature 𝑃𝐶𝑂2 Velocity pH Corrosion Data Source Temperature 𝑃𝐶𝑂2 Velocity pH Corrosion
(oC) (MPa) (m/s) Rate (oC) (MPa) (m/s) Rate
(mm/year) (mm/year)
20 0.05 0.1 4.1 0.7 Dugstad et 20 0.05 0.1 4.1 0.7
al., 1994
20 0.13 0.1 3.8 0.9 Halvorsen et 20 0.13 0.1 3.8 0.9
al., 1999
20 0.17 0.1 3.7 0.5 Nordsveen 20 0.17 0.1 3.7 0.5
et al., 2003
20 0.28 0.1 3.7 1.9 20 0.28 0.1 3.7 1.9
20 0.38 0.1 3.7 2.5 20 0.38 0.1 3.7 2.5
20 0.05 13 4.1 3.5 Zhang et al., 20 0.05 13 4.1 3.5
2012a
20 0.05 8.5 4.1 4 Hesjevik et 20 0.05 8.5 4.1 4
al., 2010
20 0.28 13 3.7 13 20 0.28 13 3.7 13
20 0.28 8.5 3.7 10 20 0.28 8.5 3.7 10
20 0.28 3.1 3.7 9 20 0.28 3.1 3.7 9
20 0.13 13 3.8 7 20 0.13 13 3.8 7
20 0.13 8.5 3.8 8 20 0.13 8.5 3.8 8
20 0.17 3.1 3.7 4.8 20 0.17 3.1 3.7 4.8
20 0.17 8.5 3.7 5.4 20 0.17 8.5 3.7 5.4
20 0.17 13 3.7 8 20 0.17 13 3.7 8
40 0.2 3.1 3.8 1 20 0.205 1.13 4.8 2.5
40 0.2 8.5 3.8 6 20 0.225 1.13 4.7 3
40 0.04 8.5 4.15 5 20 0.305 2.25 4.6 6
227
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
228
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
A4.2 Three-dimensional Plot of Corrosion Rate against CO2 Partial Pressure and
Temperature for the Low Pressure CO2 Corrosion Dataset from Alternate Viewing Angles
229
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
A6.1a Three-dimensional Mamdani FIS Plot of Corrosion Rate against Temperature and
CO2 partial pressure for the Low Pressure CO2 Dataset (Centroid Defuzzification)
Corrosion Rate (mm/year)
A6.1b Three-dimensional Mamdani FIS Plot of Corrosion Rate against Temperature and
CO2 partial pressure for the Low Pressure CO2 Dataset (Bisector Defuzzification)
Corrosion Rate (mm/year)
230
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
A6.2 Mamdani FIS Plot of Corrosion Rate against Temperature for the High Pressure CO2
Dataset
Corrosion Rate (mm/year)
Temperature (oC)
231
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Developed Model Code
%Script to analyse Low Pressure CO2 corrosion data from Dugstad 1994 journal article using
multiple linear
%regression and non-linear regression
%The data in columns are low pressure CO_2 corrosion data from Dugstad 1994
%First column => Temperature (^oC)
%Second column => CO2 Pressure (MPa)
%Third column column => Velocity/flow (m/s)
%Fourth column => pH
%Fifth column => CO2 Corrosion rate (mm/year)
Dugstad_1994_data_01=...
[20.0000 0.0500 0.1000 4.1000 0.7000
20.0000 0.1300 0.1000 3.8000 0.9000
20.0000 0.1700 0.1000 3.7000 0.5000
20.0000 0.2800 0.1000 3.7000 1.9000
20.0000 0.3800 0.1000 3.7000 2.5000
20.0000 0.0500 13.0000 4.1000 3.5000
20.0000 0.0500 8.5000 4.1000 4.0000
20.0000 0.2800 13.0000 3.7000 13.0000
20.0000 0.2800 8.5000 3.7000 10.0000
20.0000 0.2800 3.1000 3.7000 9.0000
20.0000 0.1300 13.0000 3.8000 7.0000
20.0000 0.1300 8.5000 3.8000 8.0000
20.0000 0.1700 3.1000 3.7000 4.8000
20.0000 0.1700 8.5000 3.7000 5.4000
20.0000 0.1700 13.0000 3.7000 8.0000
40.0000 0.2000 3.1000 3.8000 1.0000
40.0000 0.2000 8.5000 3.8000 6.0000
40.0000 0.2000 13.0000 3.8000 18.0000
40.0000 0.0400 3.1000 4.1500 4.0000
40.0000 0.0400 8.5000 4.1500 5.0000
40.0000 0.0400 13.0000 4.1500 6.5000
40.0000 0.0530 3.1000 4.1000 4.5000
40.0000 0.0530 8.5000 4.1000 6.0000
40.0000 0.0530 13.0000 4.1000 8.5000
40.0000 0.0700 8.5000 4.0500 2.5000
40.0000 0.0700 13.0000 4.0500 9.0000
40.0000 0.7000 3.1000 3.6000 4.6000
40.0000 0.7000 8.5000 3.6000 17.0000
40.0000 0.7000 13.0000 3.6000 23.0000
40.0000 1.2000 8.5000 3.4000 15.0000
60.0000 0.2000 0.1000 3.8000 3.0000
60.0000 0.2100 0.1000 3.8000 5.0000
232
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
60.0000 0.2200 0.1000 3.8000 6.8000
60.0000 0.2200 3.1000 3.8000 16.0000
60.0000 0.2200 8.5000 3.8000 23.0000
60.0000 0.2200 13.0000 3.8000 32.0000
60.0000 0.2200 13.0000 3.8000 44.0000
60.0000 1.8000 3.1000 3.4000 33.0000
60.0000 1.8000 8.5000 3.4000 60.0000
60.0000 1.8000 13.0000 3.4000 70.0000
60.0000 0.0700 3.1000 4.1500 8.8000
60.0000 0.0700 13.0000 4.1500 20.0000
60.0000 0.4000 3.1000 3.7000 30.0000
60.0000 0.4000 8.5000 3.7000 34.0000
60.0000 0.4000 13.0000 3.7000 50.0000
90.0000 0.1600 3.1000 4.1000 1.9000
90.0000 0.1600 8.5000 4.1000 8.5000
90.0000 0.1600 13.0000 4.1000 19.0000
90.0000 0.3700 3.1000 3.8000 16.0000
90.0000 0.3700 3.1000 3.8000 17.0000
90.0000 0.3700 8.5000 3.8000 21.0000
90.0000 0.3700 13.0000 3.8000 32.0000
90.0000 1.4000 3.1000 3.6000 6.8000
90.0000 1.4000 3.1000 3.6000 15.0000
90.0000 1.4000 8.5000 3.6000 30.0000
90.0000 1.4000 13.0000 3.6000 41.0000
90.0000 1.4000 8.5000 3.6000 50.0000
90.0000 2.1000 3.1000 3.5000 15.0000
90.0000 2.1000 3.1000 3.5000 38.0000
90.0000 2.1000 13.0000 3.5000 65.0000];
%'Temperature ' 'CO_2 Partial Pressure (MPa)' 'Flowrate(m/s) ' 'pH
' 'Corrosion Rate (mm/year) '};
X_01=Dugstad_1994_data_01(:,1:3);
x_01=Dugstad_1994_data_01(:,1);
%x_01, x_02, x_03 and x_04 are column matrices comprising temp, pressure, velocity and
%pH values respectively
x_02=Dugstad_1994_data_01(:,2);
x_03=Dugstad_1994_data_01(:,3);
x_04=Dugstad_1994_data_01(:,4);
y_01=Dugstad_1994_data_01(:,5);
%y_01 represents a column matrix comprising the Corrosion rate values
b_1=0.7;
b_2=1.9;
b_3=2.4;
b_4=2.1;
b_5=1.92;
%rate_01=Sc_data_01_Fuzzy_data(:,3);
%nlintool(x_01,y_01,@model_fun_ScCO2_01)
%nlintool(X_01,y_01,@hougen,beta_01,0.01,xname,yname)
%Using the Hougen non-linear system to analyse this problem
%X_01 is a matrix comprising three columns: x_01, x_02 and x_03
%X_001=lsqcurvefit(@hougen,[1],X_01,y_01);
%betahat_01=nlinfit(X_01,y_01,@hougen,beta_02);
%[betahat_01,resid,J]=nlinfit(X_01,y_01,@hougen,beta_02);
%betaci=nlparci(betahat_01,resid,J);
%--using rstool--%
X_02=Dugstad_1994_data_01(:,1:4);
X_04=[x_01 x_02 x_04];
%X_02 is a matrix containing all the input varaiables for Temperature,
%P_CO2, Velocity and pH and will be used with the rstool
%beta_star=randn(nVars,1);
mdl_10=NonLinearModel.fit(X_01,y_01,@hougen,beta_02);
plotSlice(mdl_10)
subplot(1,2,1)
theta_01=plotDiagnostics(mdl_10,'leverage');
set(theta_01,'LineWidth',2.5,'MarkerSize',8.0)
set(gca,'FontName','Times New Roman')
title('')
xlabel('Corrosion Rate (mm/year)','FontWeight','Bold')
ylabel('Leverage','FontWeight','Bold')
234
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
subplot(1,2,2)
theta_02=plotResiduals(mdl_10,'fitted');
set(theta_02,'LineWidth',2.5,'MarkerSize',8.0)
set(gca,'FontName','Times New Roman')
title('')
xlabel('Corrosion Rate (mm/year)','FontWeight','Bold')
ylabel('Residuals','FontWeight','Bold')
display(mdl_10)
figure()
subplot(1,2,1)
plotDiagnostics(mdl_11,'leverage')
subplot(1,2,2)
plotResiduals(mdl_11,'fitted')
%mdl_11=NonLinearModel.fit(X_02,y_01,@hougen,beta_02);
display(mdl_11)
%These next two lines for carrying out the regression analysis without errors%
ones_01=ones(60,1);
X_03=[X_02 ones_01];
rstool(X_02,y_01,'linear');%,xname_01,yname)
%For all four variables
235
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
%--Variable Interaction plots--%
[b,bint,r,rint,stats] = regress(y_01,X_03);
%mdl_01=stepwise(X_02,y_01);
mdl_02=LinearModel.fit(X_02,y_01);
display (mdl_02)
%mdl_03=LinearModel.fit(X_02,y_01,'interactions');
%display (mdl_03)
%plotEffects(mdl_03);
d_01=dataset(x_01,x_02,x_03,x_04,y_01);
subplot(1,2,2)
kappa_02=plotInteraction(mdl_04,'x_01','x_03','predictions');
set(gca,'FontName','Times New Roman')
set(kappa_02,'LineWidth',2.5)
axis([3.3 Inf 0 Inf]);
xlabel('Velocity (m/s)','FontWeight','Bold')
ylabel('Adjusted Corrosion Rate (mm/year)','FontWeight','Bold')
legend('\it T\rm=20^oC','\it T\rm=55^oC','\it T\rm=90^oC',2)
title('')
subplot(1,2,2)
kappa_04=plotInteraction(mdl_04,'x_02','x_03','predictions');
set(gca,'FontName','Times New Roman')
set(kappa_04,'LineWidth',2.5)
xlabel('Velocity (m/s)','FontWeight','Bold')
ylabel('Adjusted Corrosion Rate (mm/year)','FontWeight','Bold')
legend('\it P\rm=0.04MPa','\it P\rm=1.07MPa','\it P\rm=2.1MPa',2)
title('')
figure('Name','Interaction Plots of CO_2 Partial Pressure with pH and Velocity with pH')
subplot(1,2,1)
kappa_05=plotInteraction(mdl_04,'x_02','x_04','predictions');
set(gca,'FontName','Times New Roman')
set(kappa_05,'LineWidth',2.5)
axis([3.3 Inf 0 Inf]);
xlabel('pH','FontWeight','Bold')
ylabel('Adjusted Corrosion Rate (mm/year)','FontWeight','Bold')
legend('\it P\rm=0.04MPa','\it P\rm=1.07MPa','\it P\rm=2.1MPa',3)
title('')
subplot(1,2,2)
kappa_06=plotInteraction(mdl_04,'x_03','x_04','predictions');
set(gca,'FontName','Times New Roman');
set(kappa_06,'LineWidth',2.5)
axis([3.3 Inf 0 Inf]);
xlabel('pH','FontWeight','Bold')
ylabel('Adjusted Corrosion Rate (mm/year)','FontWeight','Bold')
legend('\it U\rm=0.1m/s','\it U\rm=6.55m/s','\it U\rm=13m/s',3)
title('')
Temperature=x_01;
PCO2=x_02;
Velocity=x_03;
pH=x_04;
237
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Corrosion_Rate=y_01;
d_02=dataset(Temperature,PCO2,Velocity,pH,Corrosion_Rate);
mdl_05=LinearModel.fit(d_02,'Corrosion_Rate ~ Temperature + PCO2 + Velocity + pH +
Temperature*PCO2*Velocity*pH + Temperature^2 + PCO2^2 + Velocity^2');
mdl_06=LinearModel.fit(d_02,'Corrosion_Rate ~ Temperature + PCO2 +
Temperature:PCO2');
mdl_07=LinearModel.fit(d_02,'Corrosion_Rate ~ PCO2 + Temperature:PCO2 + PCO2:pH +
Temperature^2 + PCO2^2 + Temperature:PCO2:pH - 1 - Temperature');
mdl_08=LinearModel.fit(d_02,'Corrosion_Rate ~ Temperature + PCO2 + Velocity + pH +
Temperature:PCO2:Velocity:pH + Temperature^2');
%mdl_08 is the same as mdl_04 except that the x and y's in mdl_4 represent
%temperature and co. in mdl_08
figure()
subplot(2,3,1)
plotInteraction(mdl_05,'Temperature','PCO2','effects')
%xlabel('CO_2 Partial Pressure, P_C_O_2 (MPa)')
%ylabel('Adjusted Corrosion Rate (mm/year)')
%legend('Green=>T=20^oC','Red=>T=55^oC','Blue=>T=90^oC',2)
title('Interaction of Temperature and CO_2 Partial Pressure')
xlabel('Effect on Corrosion Rate')
subplot(2,3,2)
plotInteraction(mdl_05,'Temperature','Velocity','effects')
%axis([3.3 Inf 0 Inf]);
%xlabel('Velocity (m/s)')
%ylabel('Adjusted Corrosion Rate (mm/year)')
%legend('Green=>T=20^oC','Red=>T=55^oC','Blue=>T=90^oC',2)
title('Interaction of Temperature and Velocity')
xlabel('Effect on Corrosion Rate')
subplot(2,3,3)
plotInteraction(mdl_05,'Temperature','pH','effects')
%axis([3.3 Inf 0 Inf]);
%xlabel('pH')
%ylabel('Adjusted Corrosion Rate (mm/year)')
%legend('Green=>T=20^oC','Red=>T=55^oC','Blue=>T=90^oC',2)
title('Interaction of Temperature and pH')
xlabel('Effect on Corrosion Rate')
subplot(2,3,4)
plotInteraction(mdl_05,'PCO2','Velocity','effects')
%xlabel('Velocity (m/s)')
%ylabel('Adjusted Corrosion Rate (mm/year)')
%legend('Green=>P=0.04MPa','Red=>P=1.07MPa','Blue=>P=2.1MPa',2)
title('Interaction of CO_2 Partial Pressure and Velocity')
xlabel('Effect on Corrosion Rate')
subplot(2,3,5)
plotInteraction(mdl_05,'PCO2','pH','effects')
238
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
%axis([3.3 Inf 0 Inf]);
%xlabel('pH')
%ylabel('Adjusted Corrosion Rate (mm/year)')
%legend('Green=>P=0.04MPa','Red=>P=1.07MPa','Blue=>P=2.1MPa',3)
title('Interaction of CO_2 Partial Pressure and pH')
xlabel('Effect on Corrosion Rate')
subplot(2,3,6)
plotInteraction(mdl_05,'Velocity','pH','effects')
%axis([3.3 Inf 0 Inf]);
%xlabel('pH')
%ylabel('Adjusted Corrosion Rate (mm/year)')
%legend('Green=>V=0.1m/s','Red=>P=6.55m/s','Blue=>P=13m/s',3)
title('Interaction of Velocity and pH')
xlabel('Effect on Corrosion Rate')
display(mdl_04);
display(mdl_05);
display(mdl_06);
display(mdl_07);
display(mdl_08);
figure('Name','Multiple Lin Regression Case Order Plot of Levereage and Plot of Residuals
against Fitted Values')
subplot(1,2,1)
zeta_01=plotDiagnostics(mdl_08);
set(zeta_01,'LineWidth',2.5,'MarkerSize',8.0)
set(gca,'FontName','Times New Roman')
%xlim([0 80])
xlabel('Corrosion Rate (mm/year)','FontWeight','Bold')
ylabel('Leverage','FontWeight','Bold')
title('')
%Leverage Plot -> Identifies outliers within the data
subplot(1,2,2)
zeta_02=plotResiduals(mdl_08,'fitted');
set(zeta_02,'LineWidth',2.5,'MarkerSize',8.0)
set(gca,'FontName','Times New Roman')
%Plot of Residuals -> plot should be an even random scatter
xlim([0 80])
xlabel('Corrosion Rate (mm/year)','FontWeight','Bold')
ylabel('Residuals','FontWeight','Bold')
title('')
%figure()
%plotEffects(mdl_04);
tbl_02=anova(mdl_06,'summary');
display(tbl_02)
%Corr_Rate_01=11.787-(4.0357.*Temperature)-(1214.3.*PCO2)+(0.90548.*Velocity)-
(4.5541.*pH)+(75.02.*Temperature.*PCO2)+(0.18747.*Temperature.*Velocity)...
% -(31.663.*PCO2.*Velocity)+(0.94402.*Temperature.*pH)+(357.1.*PCO2.*pH)-
(0.35984.*Velocity.*pH)+(0.0121.*Temperature.^2)-(188.67.*PCO2.^2)...
% -(0.02112.*Velocity.^2)+(0.59099.*Temperature.*PCO2.*Velocity)-
(20.312.*Temperature.*PCO2.*pH)-(0.037319.*Temperature.*Velocity.*pH)...
% +(9.3645.*PCO2.*Velocity.*pH)-(0.17275.*Temperature.*PCO2.*Velocity.*pH);
Temperature_01=test_set_01(:,1);
PCO2_01=test_set_01(:,2);
Velocity_01=test_set_01(:,3);
pH_01=test_set_01(:,4);
Corr_Rate_01=test_set_01(:,5);
%--Error Plots for Multiple Linear Regression -> Plots of Multiple Linear Corr Rate vs
Experimental Corr Rate--%
Corr_Rate_01_test=78.2+(0.9923.*Temperature_01)-
(2.0.*PCO2_01)+(0.9826.*Velocity_01)-(25.0.*pH_01)...
-(7.0e-3.*Temperature_01.^2)+(5.3011e-
3.*Temperature_01.*PCO2_01.*Velocity_01.*pH_01);
%display(Corr_Rate_01_test)
Error_test=Corr_Rate_01-Corr_Rate_01_test;
Error_test_omic_01=[-6.74 4.50 -1.81 -13.58 -4.72 -9.18 2.95 2.90 -4.67 -3.80 -6.58 -2.93
5.84 1.30 18.54];
%figure()
%bar(Error_test)
240
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
%xlabel('Test Data')
%ylabel('Error')
%figure()
X_001=linspace(0,60,15);
Y_001=X_001;
%plot(X_01,Y_01,'k-')
%hold on
%plot(Corr_Rate_01,Corr_Rate_01_test,'bx')
%xlabel('Experimental Corrosion Rate (mm/year)')
%ylabel('Multiple Linear Regression Model Corrosion Rate (mm/year)')
%legend('Blue Cross => Multiple Linear Regression')
hold on
plot(Corr_Rate_01,Corr_Rate_01_test,'bx','MarkerSize',7,'LineWidth',2.5)
Y_01_multi_uci=[8.27397568
11.27638841
14.32847528
17.47026027
20.78720942
24.45262515
28.69280475
33.53005299
38.74822417
44.15978721
49.66952557
241
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
55.23300039
60.82832323
66.44383875
72.07287628];
Y_01_multi_lci=[-8.27397568
-2.704988409
2.814324717
8.243939729
13.49839058
18.40437485
22.73559525
26.46974701
29.82297583
32.98281279
36.04447443
39.05239961
42.02847677
44.98436125
47.92672372];
plot(X_01_conf,Y_01_multi_uci,'k--','LineWidth',1.5)
plot(X_01_conf,Y_01_multi_lci,'k--','LineWidth',1.5)
xlim([0 60]);
ylim([0 60]);
xlabel('Experimental Corrosion Rate (mm/year)','FontWeight','Bold')
ylabel('Multiple Lin Regression Model Corr Rate (mm/year)','FontWeight','Bold')
legend(' Experiment=Model Line',' Datapoints',' 95% Confidence Interval',2)
text('Position',[33 17],'String','\it R^2 \rm\bf =
0.65','Background','w','EdgeColor','black','LineWidth',1.0,'FontName','Times New
Roman','FontWeight','Bold');
subplot(1,2,2)
bar(Error_test_omic_01)
set(gca,'FontName','Times New Roman')
xlabel('Test Data','FontWeight','Bold')
ylabel('Corrosion Rate Error (mm/year)','FontWeight','Bold')
set(gca,'XMinorTick','on','XTick',[0 5 10 15 20],'LineWidth',1)
%--Error Plots for Surface Model Plot -> Plots of Surface Model Corr Rate vs Experimental
Corr Rate--%
Sur_mod_01=[17.1679
7.8581
16.5356
5.2500
13.9275
7.2684
23.1771
242
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
15.4265
22.6577
15.2441
31.1528
23.1853
30.4165
25.0632
32.2945];
%figure()
Error_test_01=Corr_Rate_01-Sur_mod_01;
Error_test_omic_02=[-10.47 4.64 7.75 -10.06 2.07 -6.19 0.46 11.73 4.57 0.82 -4.42 4.34
14.76 4.71 28.85];
R_coef_01=corrcoef(Corr_Rate_01,Sur_mod_01);
R_squared_coef_01=(R_coef_01(1,2))^2;
display(R_squared_coef_01)
%figure()
%bar(Error_test_01)
%xlabel('Test Data')
%ylabel('Error')
hold on
plot(Corr_Rate_01,Sur_mod_01,'bx','MarkerSize',7,'LineWidth',2.5)
Y_01_sur_mod_uci=[8.942072077
11.58287086
14.34067466
17.33684729
20.83055263
25.15391021
30.28646651
35.88890152
41.71546102
47.65277063
53.6501071
59.68291495
65.7382078
71.80857065
243
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
77.88949395];
Y_01_sur_mod_lci=[-8.942072077
-3.011470859
2.802125339
8.377352713
13.45504737
17.70308979
21.14193349
24.11089848
26.85573898
29.48982937
32.0638929
34.60248505
37.1185922
39.61962935
42.11010605];
plot(X_01_conf,Y_01_sur_mod_uci,'k--','LineWidth',1.5)
plot(X_01_conf,Y_01_sur_mod_lci,'k--','LineWidth',1.5)
xlim([0 60]);
ylim([0 60]);
xlabel('Experimental Corrosion Rate (mm/year)','FontWeight','Bold')
ylabel('Response Surface Model Corrosion Rate (mm/year)','FontWeight','Bold')
legend(' Experiment=Model Line',' Datapoints',' 95% Confidence Interval',2)
text('Position',[33 17],'String','\it R^2 \rm\bf=
0.39','Background','w','EdgeColor','black','LineWidth',1.0,'FontName','Times New
Roman','FontWeight','Bold');
subplot(1,2,2)
bar(Error_test_omic_02)
set(gca,'XMinorTick','on','XTick',[0 5 10 15 20],'LineWidth',1,'FontName','Times New
Roman')
xlabel('Test Data','FontWeight','Bold')
ylabel('Corrosion Rate Error (mm/year)','FontWeight','Bold')
%--Error Plots for Mamdani FIS -> Plots of Mamdani FIS Corr Rate vs Experimental Corr
Rate--%
Mamdani_Corr_Rate_01=[9.9800
5.7500
5.8300
7.1200
14.1000
6.5400
20.8000
22.8000
23.5000
244
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
7.6600
28.9000
23.6000
23.8000
24.4000
24.7000];
Error_test_02=Corr_Rate_01-Mamdani_Corr_Rate_01;
Error_test_omic_03=[-3.28 6.75 5.88 -9.40 1.90 -6.60 11.17 12.46 -2.80 3.20 2.20 3.50 22.34
12.30 31.10];
hold on
plot(Corr_Rate_01,Mamdani_Corr_Rate_01,'bx','MarkerSize',7,'LineWidth',2.5)
Y_01_mamdani_uci=[8.454310412
11.05344341
13.83667082
17.00891198
20.93037816
25.79286637
31.31271579
37.15760942
43.15780204
49.23867426
55.36553781
61.52073541
67.69450053
73.88104499
80.07673936];
Y_01_mamdani_lci=[-8.454310412
-2.482043405
3.306129176
8.705288023
13.35522184
17.06413363
20.11568421
22.84219058
25.41339796
27.90392574
30.34846219
32.76466459
35.16229947
37.54715501
245
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
39.92286064];
plot(X_01_conf,Y_01_mamdani_uci,'k--','LineWidth',1.5)
plot(X_01_conf,Y_01_mamdani_lci,'k--','LineWidth',1.5)
xlim([0 60]);
ylim([0 60]);
xlabel('Experimental Corrosion Rate (mm/year)','FontWeight','Bold')
ylabel('Mamdani FIS Model Corrosion Rate (mm/year)','FontWeight','Bold')
legend(' Experiment=Model Line',' Datapoints',' 95% Confidence Interval',2)
text('Position',[33 17],'String','\it R^2 \rm\bf=
0.32','Background','w','EdgeColor','black','LineWidth',1.0,'FontName','Times New
Roman','FontWeight','Bold');
subplot(1,2,2)
bar(Error_test_omic_03)
set(gca,'XMinorTick','on','XTick',[0 5 10 15 20],'LineWidth',1,'FontName','Times New
Roman')
xlabel('Test Data','FontWeight','Bold')
ylabel('Corrosion Rate Error (mm/year)','FontWeight','Bold')
%--Error Plots for Sugeno FIS -> Plots of Sugeno FIS Corr Rate vs Experimental Corr Rate--
%
Sugeno_Corr_Rate_01=[5.3700
9.7200
26.5000
4.2300
6.0200
4.6800
8.0100
11.8000
12.0000
5.0200
29.4000
29.4000
30.0000
29.8000
30.0000];
Error_test_03=Corr_Rate_01-Sugeno_Corr_Rate_01;
figure()
subplot(1,2,1)
plot(X_001,Y_001,'k-')
hold on
plot(Corr_Rate_01,Sugeno_Corr_Rate_01,'bx')
xlim([0 60]);
246
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
ylim([0 60]);
xlabel('Experimental Corrosion Rate (mm/year)')
ylabel('Sugeno FIS Model Corrosion Rate (mm/year)')
%legend(' Experiment=Model Line',' Datapoints',' 95% Confidence Interval',2)
subplot(1,2,2)
bar(Error_test_03)
xlabel('Test Data')
ylabel('Corrosion Rate Error (mm/year)')
Error_test_Hougen=Corr_Rate_01-Corr_Rate_02_test;
Error_test_omic_04=[-6.41 4.18 9.38 0.20 9.76 3.66 5.53 12.81 12.98 0.26 -6.66 7.76 15.64 -
0.86 17.81];
subplot(1,2,1)
plot(X_001,Y_001,'k-','LineWidth',1.5)
set(gca,'FontName','Times New Roman')
hold on
plot(Corr_Rate_01,Corr_Rate_02_test,'bx','MarkerSize',7,'LineWidth',2.5)
Y_01_hougen_uci=[6.504021634
9.729126352
13.11129082
16.76113024
20.8202204
25.35997741
30.30121796
35.50227022
40.85542096
46.29788671
51.79505695
57.32742911
62.88351577
68.45622592
74.04098918];
Y_01_hougen_lci=[-6.504021634
-1.157726352
4.031509175
8.953069755
13.4653796
247
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
17.49702259
21.12718204
24.49752978
27.71577904
30.84471329
33.91894305
36.95797089
39.97328423
42.97197408
45.95861082];
plot(X_01_conf,Y_01_hougen_uci,'k--','LineWidth',1.5)
plot(X_01_conf,Y_01_hougen_lci,'k--','LineWidth',1.5)
xlim([0 60]);
ylim([0 60]);
xlabel('Experimental Corrosion Rate (mm/year)','FontWeight','Bold')
ylabel('Nonlinear Model Corrosion Rate (mm/year)','FontWeight','Bold')
legend(' Experiment=Model Line',' Datapoints',' 95% Confidence Interval',2)
text('Position',[33 17],'String','\it R^2 \rm\bf=
0.67','Background','w','EdgeColor','black','LineWidth',1.0,'FontName','Times New
Roman','FontWeight','Bold');
subplot(1,2,2)
bar(Error_test_omic_04)
set(gca,'XMinorTick','on','XTick',[0 5 10 15 20],'LineWidth',1,'FontName','Times New
Roman')
xlabel('Test Data','FontWeight','Bold')
ylabel('Corrosion Rate Error (mm/year)','FontWeight','Bold')
%--Error Plots for Neural Network Corrosion Model -> Plots of Neural Network Model Corr
Rate vs Experimental Corr Rate--%
Neural_Net_Corr_Rate_01=[13.4935
16.1991
13.8781
18.9542
14.7386
19.2873
25.0607
20.3173
30.9718
25.8677
36.0933
11.2166
17.6583
14.9526
19.9719];
248
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Error_test_neu_net_01=Corr_Rate_01-Neural_Net_Corr_Rate_01;
Error_test_omic_05=[-6.80 -3.70 -5.95 0.05 1.26 3.12 5.78 -0.29 -0.32 -1.06 8.34 -3.97 4.13
17.03 23.91];
R_coef_nn_01=corrcoef(Corr_Rate_01,Neural_Net_Corr_Rate_01);
R_sq_coef_nn_01=(R_coef_nn_01(1,2))^2;
display(R_sq_coef_nn_01)
%figure()
%bar(Error_test_01)
%xlabel('Test Data')
%ylabel('Error')
Y_01_neunet_uci=[6.895212776
9.865263005
12.89318747
16.04817034
19.51909385
23.68002714
28.62121496
33.97367112
39.49038024
45.07878625
50.70325877
56.34807536
62.00539645
67.67091947
73.34209999];
Y_01_neunet_lci=[-6.895212776
-1.293863005
4.249612535
9.666029655
14.76650615
19.17697286
22.80718504
26.02612888
29.08081976
32.06381375
35.01074123
249
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
37.93732464
40.85140355
43.75728053
46.65750001];
plot(X_01_conf,Y_01_neunet_uci,'k--','LineWidth',1.5)
plot(X_01_conf,Y_01_neunet_lci,'k--','LineWidth',1.5)
xlim([0 60]);
ylim([0 60]);
xlabel('Experimental Corrosion Rate (mm/year)','FontWeight','Bold')
ylabel('Neural Network Corrosion Rate (mm/year)','FontWeight','Bold')
legend(' Experiment=Model Line',' Datapoints',' 95% Confidence Interval',2)
text('Position',[33 17],'String','R^2 =
0.6444','Background','w','EdgeColor','black','LineWidth',1.0,'FontName','Times New
Roman','FontWeight','Bold');
subplot(1,2,2)
bar(Error_test_omic_05)
set(gca,'XMinorTick','on','XTick',[0 5 10 15 20],'LineWidth',1,'FontName','Times New
Roman')
xlabel('Test Data','FontWeight','Bold')
ylabel('Corrosion Rate Error (mm/year)','FontWeight','Bold')
%--Freecorp Model--%
Freecorp_Corr_Rate_01=[4.2
5.6
6.4
6.2
7.1
7.9
9.1
10.8
11.1
23.9
25.3
28.2
28.6
34.6
35];
Error_test_freecorp_01=Corr_Rate_01-Freecorp_Corr_Rate_01;
Error_test_omic_06=[2.50 6.90 6.80 -19.60 8.90 -11.20 10.60 11.10 9.20 14.90 -2.60 15.90
6.10 2.00 34.70];
250
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
R_coef_freecorp_01=corrcoef(Corr_Rate_01,Freecorp_Corr_Rate_01);
R_sq_coef_freecorp_01=(R_coef_freecorp_01(1,2))^2;
display(R_sq_coef_freecorp_01)
Y_01_freecorp_uci=[9.41690706
12.09367008
15.02585986
18.40373002
22.4606524
27.27686736
32.68102457
38.44260606
44.40643611
50.48737145
56.639728
62.83804391
69.06737523
75.31849927
81.58546403];
Y_01_freecorp_lci=[-9.41690706
-3.52227008
2.116940139
7.310469983
11.8249476
15.58013264
18.74737543
21.55719394
24.16476389
26.65522855
29.074272
31.44735609
33.78942477
36.10970073
38.41413597];
subplot(1,2,1)
plot(Corr_Rate_01,Freecorp_Corr_Rate_01,'bx','MarkerSize',7,'LineWidth',2.5)
set(gca,'FontName','Times New Roman')
hold on
plot(X_001,Y_001,'k-','LineWidth',1.5)
plot(X_01_conf,Y_01_freecorp_uci,'k--','LineWidth',1.5)
plot(X_01_conf,Y_01_freecorp_lci,'k--','LineWidth',1.5)
251
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
xlim([0 60]);
ylim([0 60]);
xlabel('Experimental Corrosion Rate (mm/year)','FontWeight','Bold')
ylabel('Freecorp Model Corrosion Rate (mm/year)','FontWeight','Bold')
legend(' Datapoints',' Experiment=Model Line',' 95% Confidence Interval',2)
text('Position',[33 17],'String','\it R^2 \rm\bf=
0.24','Background','w','EdgeColor','black','LineWidth',1.0,'FontName','Times New
Roman','FontWeight','Bold');
subplot(1,2,2)
bar(Error_test_omic_06)
set(gca,'XMinorTick','on','XTick',[0 5 10 15 20],'LineWidth',1,'FontName','Times New
Roman')
xlabel('Test Data','FontWeight','Bold')
ylabel('Corrosion Rate Error (mm/year)','FontWeight','Bold')
%--Norsok Model--%
Norsok_Corr_Rate_01=[11.921
11.865
16.8267
15.6008
21.7844
21.7068
35.7804
26.9424
31.1412
44.4108
76.0114
24.979
29.0564
30.7831
35.9498];
Error_test_norsok_01=Corr_Rate_01-Norsok_Corr_Rate_01;
Error_test_omic_07=[-5.22 0.64 -2.60 -15.78 -5.78 -7.98 0.17 -2.71 -6.94 -11.78 -3.06 -4.14 -
14.41 1.05 -16.01];
R_coef_norsok_01=corrcoef(Corr_Rate_01,Norsok_Corr_Rate_01);
R_sq_coef_norsok_01=(R_coef_norsok_01(1,2))^2;
display(R_sq_coef_norsok_01)
Y_01_norsok_uci=[ 6.059941559
10.86825435
15.74434531
252
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
20.73596666
25.92585717
31.42548738
37.30173923
43.50275517
49.91717948
56.45663953
63.0691332
69.72606932
76.41143338
83.11585138
89.83356939];
Y_01_norsok_lci=[-6.059941559
0.560345647
7.11285469
13.54983334
19.78854283
25.71751262
31.26986077
36.49744483
41.51162052
46.40076047
51.2168668
55.98853068
60.73176662
65.45594862
70.16683061];
X_01_norsok_conf=linspace(0,80,15);
X_002=linspace(0,80,15);
Y_002=X_002;
subplot(1,2,1)
plot(Corr_Rate_01,Norsok_Corr_Rate_01,'bx','MarkerSize',7,'LineWidth',2.5)
set(gca,'FontName','Times New Roman')
hold on
plot(X_002,Y_002,'k-','LineWidth',1.5)
plot(X_01_norsok_conf,Y_01_norsok_uci,'k--','LineWidth',1.5)
plot(X_01_norsok_conf,Y_01_norsok_lci,'k--','LineWidth',1.5)
xlim([0 80]);
ylim([0 80]);
xlabel('Experimental Corrosion Rate (mm/year)','FontWeight','Bold')
ylabel('Norsok Model Corrosion Rate (mm/year)','FontWeight','Bold')
253
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
legend(' Datapoints',' Experiment=Model Line',' 95% Confidence Interval',2)
text('Position',[33 17],'String','\it R^2 \rm\bf=
0.89','Background','w','EdgeColor','black','LineWidth',1.0,'FontName','Times New
Roman','FontWeight','Bold');
subplot(1,2,2)
bar(Error_test_omic_07)
set(gca,'XMinorTick','on','XTick',[0 5 10 15 20],'LineWidth',1,'FontName','Times New
Roman')
xlabel('Test Data','FontWeight','Bold')
ylabel('Corrosion Rate Error (mm/year)','FontWeight','Bold')
DeWaard_1991_Corr_Rate_01=[1.0969
2.1704
2.1704
2.7205
2.7205
4.012
4.012
6.0652
6.0652
16.2001
16.2001
10.6522
10.6522
10.8924
10.8924];
Error_test_dewaard_1991_01=Corr_Rate_01-DeWaard_1991_Corr_Rate_01;
Error_test_omic_08=[5.60 10.33 10.28 4.11 13.28 6.35 14.83 14.99 13.93 19.99 15.35 20.93
13.80 26.11 43.80];
R_coef_dewaard_1991_01=corrcoef(Corr_Rate_01,DeWaard_1991_Corr_Rate_01);
R_sq_coef_dewaard_1991_01=(R_coef_dewaard_1991_01(1,2))^2;
display(R_sq_coef_dewaard_1991_01)
Y_01_dewaard_1991_uci=[ 3.234752517
6.409148669
10.4951783
15.69246417
21.32898385
27.09907835
32.92067353
38.76674327
44.62621408
254
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
50.49378029
56.36659849
62.24301274
68.12199507
74.00287433
79.88519401];
Y_01_dewaard_1991_lci=[-3.234752517
2.162251331
6.647621697
10.02173583
12.95661615
15.75792165
18.50772647
21.23305673
23.94498592
26.64881971
29.34740151
32.04238726
34.73480493
37.42532567
40.11440599];
%X_01_norsok_conf=linspace(0,80,15);
%X_002=linspace(0,80,15);
%Y_002=X_002;
subplot(1,2,1)
plot(Corr_Rate_01,DeWaard_1991_Corr_Rate_01,'bx','MarkerSize',7,'LineWidth',2.5)
set(gca,'FontName','Times New Roman')
hold on
plot(X_001,Y_001,'k-','LineWidth',1.5)
plot(X_01_conf,Y_01_dewaard_1991_uci,'k--','LineWidth',1.5)
plot(X_01_conf,Y_01_dewaard_1991_lci,'k--','LineWidth',1.5)
xlim([0 60]);
ylim([0 60]);
xlabel('Experimental Corrosion Rate (mm/year)','FontWeight','Bold')
ylabel('The 1991 De Waard Model Corrosion Rate (mm/year)','FontWeight','Bold')
legend(' Datapoints',' Experiment=Model Line',' 95% Confidence Interval',2)
text('Position',[33 17],'String','\it R^2 \rm\bf=
0.54','Background','w','EdgeColor','black','LineWidth',1.0,'FontName','Times New
Roman','FontWeight','Bold');
subplot(1,2,2)
255
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
bar(Error_test_omic_08)
set(gca,'XMinorTick','on','XTick',[0 5 10 15 20],'LineWidth',1,'FontName','Times New
Roman')
xlabel('Test Data','FontWeight','Bold')
ylabel('Corrosion Rate Error (mm/year)','FontWeight','Bold')
DeWaard_1995_Corr_Rate_01=[10.4845
14.6379
19.1235
9.0386
14.4027
14.4644
24.5685
18.9571
22.8202
38.1299
65.4
20.0906
25.1775
22.9145
28.3876];
Error_test_dewaard_1995_01=Corr_Rate_01-DeWaard_1995_Corr_Rate_01;
Error_test_omic_09=[-3.78 -2.14 3.96 -7.91 1.60 -2.12 -3.09 4.54 1.04 -0.57 0.82 4.18 -8.13
8.61 -5.40];
R_coef_dewaard_1995_01=corrcoef(Corr_Rate_01,DeWaard_1995_Corr_Rate_01);
R_sq_coef_dewaard_1995_01=(R_coef_dewaard_1995_01(1,2))^2;
display(R_sq_coef_dewaard_1995_01)
Y_01_dewaard_1995_uci=[5.096154396
9.297678133
13.57457427
17.98218398
22.6110453
27.55929483
32.84441886
38.38226898
44.07394949
49.85414067
55.6865275
61.5512445
67.43702294
73.33715252
79.24744344];
256
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Y_01_dewaard_1995_lci=[-5.096154396
0.702321867
6.425425728
12.01781602
17.3889547
22.44070517
27.15558114
31.61773102
35.92605051
40.14585933
44.3134725
48.4487555
52.56297706
56.66284748
60.75255656];
X_01_dewaard_1995_conf=linspace(0,70,15);
X_003=linspace(0,70,15);
Y_003=X_003;
subplot(1,2,1)
plot(Corr_Rate_01,DeWaard_1995_Corr_Rate_01,'bx','MarkerSize',7,'LineWidth',2.5)
set(gca,'FontName','Times New Roman')
%set(gca,'YGrid','on')
hold on
plot(X_003,Y_003,'k-','LineWidth',1.5)
plot(X_01_dewaard_1995_conf,Y_01_dewaard_1995_uci,'k--','LineWidth',1.5)
plot(X_01_dewaard_1995_conf,Y_01_dewaard_1995_lci,'k--','LineWidth',1.5)
xlim([0 70]);
ylim([0 70]);
xlabel('Experimental Corrosion Rate (mm/year)','FontWeight','Bold')
ylabel('The 1995 De Waard Model Corrosion Rate (mm/year)','FontWeight','Bold')
legend(' Datapoints',' Experiment=Model Line',' 95% Confidence Interval',2)
text('Position',[33 17],'String','\it R^2 \rm\bf=
0.88','Background','w','EdgeColor','black','LineWidth',1.0,'FontName','Times New
Roman','FontWeight','Bold');
subplot(1,2,2)
bar(Error_test_omic_09)
set(gca,'XMinorTick','on','XTick',[0 5 10 15 20],'LineWidth',1,'FontName','Times New
Roman')
xlabel('Test Data','FontWeight','Bold')
ylabel('Corrosion Rate Error (mm/year)','FontWeight','Bold')
257
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Neural Network Model Code
% Same neural network as the first except with a tansig transfer function in place
% of the logsig transfer function in the first hidden layer
%tic
%R_coef_nn_01=0;
%while R_coef_nn_01<0.85
%repeat
%[x_1,t_1]=size(Raw_01);
x_1=Raw_01;
CR_01_target=Raw_02;
%[trainInd,valInd,testInd]=dividerand(98,0.6,0.2,0.2);
net_01.trainParam.epochs=100;
%net_01=train(net_01,x_1,CR_01_target);
[net_01,tr]=train(net_01,x_1,CR_01_target);
net_01.layers{1}.transferFcn = 'tansig';%'tansig';'tansig';
net_01.layers{2}.transferFcn ='tansig';
net_01.layers{3}.transferFcn = 'tansig';
%view(net_01);
y_1=net_01(x_1);
%size(y_1)=196;
perf_01=perform(net_01,y_1,CR_01_target);
%x_2=[16 17 3 10 20 25 32 11 5 19];
y_2=sim(net_01,x_2); % Simulating the neural network, in terms of net_01 and x_1
%figure()
%plot(t_1,y_1,'k-x')
display (y_2)
X_01=linspace(0,60,15);
Y_01=X_01;
x_test_01=xlsread('G:\University Files and More\University of Newcastle Upon Tyne
Files\PhD\Dugstad_Data_From_Original_Journal_Article_01.xlsx','Test_Corr_Rates_01');
error_01=x_test_01-y_2;
E_01=sum(abs(error_01));
display(E_01)
R_coef_nn_01=corrcoef(x_test_01,y_2);
259
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
R_sq_coef_nn_01=(R_coef_nn_01(1,2))^2;
%else
%display(R_sq_coef_nn_01)
%end
%if R_sq_coef_nn_01<0.65
while R_sq_coef_nn_01<0.85
%[x_1,t_1]=size(Raw_01);
x_1=Raw_01;
CR_01_target=Raw_02;
%[trainInd,valInd,testInd]=dividerand(98,0.6,0.2,0.2);
net_01.trainParam.epochs=100;
%net_01=train(net_01,x_1,CR_01_target);
[net_01,tr]=train(net_01,x_1,CR_01_target);
net_01.layers{1}.transferFcn = 'tansig';%'tansig';'tansig';
net_01.layers{2}.transferFcn ='tansig';
net_01.layers{3}.transferFcn = 'tansig';
%view(net_01);
y_1=net_01(x_1);
%size(y_1)=196;
perf_01=perform(net_01,y_1,CR_01_target);
%x_2=[16 17 3 10 20 25 32 11 5 19];
y_2=sim(net_01,x_2); % Simulating the neural network, in terms of net_01 and x_1
%figure()
%plot(t_1,y_1,'k-x')
display (y_2)
X_01=linspace(0,60,15);
Y_01=X_01;
x_test_01=xlsread('G:\University Files and More\University of Newcastle Upon Tyne
Files\PhD\Dugstad_Data_From_Original_Journal_Article_01.xlsx','Test_Corr_Rates_01');
error_01=x_test_01-y_2;
E_01=sum(abs(error_01));
display(E_01)
R_coef_nn_01=corrcoef(x_test_01,y_2);
R_sq_coef_nn_01=(R_coef_nn_01(1,2))^2;
display(R_sq_coef_nn_01)
end
261
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
tic
figure()
subplot(1,2,1)
plot(X_01,Y_01,'k-')
hold on
plot(x_test_01,y_2,'bx')
xlabel('Experimental Corrosion Rate (mm/year)','FontName','Times New Roman')
ylabel('Neural Network Corrosion Rate (mm/year)','FontName','Times New Roman')
subplot(1,2,2)
bar(error_01)
xlabel('Test Data','FontName','Times New Roman')
ylabel('Error','FontName','Times New Roman')
figure()
plot(CR_01_target,y_1,'x')
%title('Neural Network Corrosion Rate Against Experimental Corrosion Rate for
Supercritical CO_2 (ScCO_2) Pressures')
xlabel('Experimental Corrosion Rate (mm/year)','FontName','Times New Roman')
ylabel('Neural Network Corrosion Rate (mm/year)','FontName','Times New Roman')
figure()
plotregression(CR_01_target,y_1)
%title('Neural Network Corrosion Rate Against Experimental Corrosion Rate for
Supercritical CO_2 (ScCO_2) Pressures')
xlabel('Experimental Corrosion Rate (mm/year)','FontName','Times New Roman')
ylabel('Neural Network Corrosion Rate (mm/year)','FontName','Times New Roman')
figure()
plotperform(tr)
X_01_conf=[0
4.2857
8.5714
12.8571
17.1428
21.4285
25.7142
29.9999
34.2856
38.5713
42.857
47.1427
262
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
51.4284
55.7141
59.9998];
Neural_Net_Corr_Rate_01=[12.3923
7.0497
11.1169
12.1380
15.7352
13.6138
17.2251
23.5501
37.0132
28.0422
44.6256
20.5530
27.3942
20.8959
29.3329];
Error_test_neu_net_01=x_test_01.'-Neural_Net_Corr_Rate_01;
R_coef_nn_01=corrcoef(x_test_01.',Neural_Net_Corr_Rate_01);
R_sq_coef_nn_01=(R_coef_nn_01(1,2))^2;
display(R_sq_coef_nn_01)
%figure()
%bar(Error_test_01)
%xlabel('Test Data')
%ylabel('Error')
figure()
subplot(1,2,1)
plot(X_01,Y_01,'k-','LineWidth',1.5)
hold on
plot(x_test_01.',Neural_Net_Corr_Rate_01,'bx','MarkerSize',7,'LineWidth',2.5)
Y_01_neunet_uci=[6.895212776
9.865263005
12.89318747
16.04817034
19.51909385
23.68002714
28.62121496
33.97367112
39.49038024
45.07878625
50.70325877
56.34807536
263
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
62.00539645
67.67091947
73.34209999];
Y_01_neunet_lci=[-6.895212776
-1.293863005
4.249612535
9.666029655
14.76650615
19.17697286
22.80718504
26.02612888
29.08081976
32.06381375
35.01074123
37.93732464
40.85140355
43.75728053
46.65750001];
plot(X_01_conf,Y_01_neunet_uci,'k--','LineWidth',1.5)
plot(X_01_conf,Y_01_neunet_lci,'k--','LineWidth',1.5)
xlim([0 60]);
ylim([0 60]);
xlabel('Experimental Corrosion Rate (mm/year)','FontWeight','bold','FontName','Times New
Roman')
ylabel('Neural Network Corrosion Rate (mm/year)','FontWeight','bold','FontName','Times
New Roman')
legend(' Experiment=Model Line',' Datapoints',' 95% Confidence Interval',2)
text('Position',[33 17],'String','R^2 =
0.6444','Background','w','EdgeColor','black','LineWidth',1.0,'FontName','Times New Roman');
subplot(1,2,2)
bar(Error_test_neu_net_01)
xlabel('Test Data','FontWeight','bold','FontName','Times New Roman')
ylabel('Corrosion Rate Error (mm/year)','FontWeight','bold','FontName','Times New Roman')
set(gca,'XMinorTick','on','XTick',[0 5 10 15 20],'LineWidth',1)
X_02_conf=[0
4.2857
8.5714
12.8571
17.1428
21.4285
264
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
25.7142
29.9999
34.2856
38.5713
42.857
47.1427
51.4284
55.7141
59.9998];
Neural_Net_Corr_Rate_02=[14.4904
20.3834
14.7004
16.2215
22.1838
20.5459
20.8682
18.7632
39.8937
29.00017
48.1738
17.4176
27.727
19.1816
34.316];
Error_test_neu_net_02=x_test_01.'-Neural_Net_Corr_Rate_02;
R_coef_nn_02=corrcoef(x_test_01.',Neural_Net_Corr_Rate_02);
R_sq_coef_nn_02=(R_coef_nn_02(1,2))^2;
%display(R_sq_coef_nn_02)
figure()
subplot(1,2,1)
plot(X_01,Y_01,'k-','LineWidth',1.5)
hold on
plot(x_test_01.',Neural_Net_Corr_Rate_02,'bx','MarkerSize',7,'LineWidth',2.5)
Y_02_neunet_uci=[6.270715157
9.608877368
12.98495857
16.42805428
19.99812187
23.81239276
28.02546799
32.67008875
37.60927109
42.70562029
265
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
47.88366413
53.10684567
58.35686651
63.62391643
68.90236576];
Y_02_neunet_lci=[-6.270715157
-1.037477368
4.15784143
9.28614572
14.28747813
19.04460724
23.40293201
27.32971125
30.96192891
34.43697971
37.83033587
41.17855433
44.49993349
47.80428357
51.09723424];
plot(X_02_conf,Y_02_neunet_uci,'k--','LineWidth',1.5)
plot(X_02_conf,Y_02_neunet_lci,'k--','LineWidth',1.5)
xlim([0 60]);
ylim([0 60]);
xlabel('Experimental Corrosion Rate (mm/year)','FontWeight','bold','FontName','Times New
Roman')
ylabel('Neural Network Corrosion Rate (mm/year)','FontWeight','bold','FontName','Times
New Roman')
legend(' Experiment=Model Line',' Datapoints',' 95% Confidence Interval',2)
text('Position',[33 17],'String','R^2 =
0.81','Background','w','EdgeColor','black','LineWidth',1.0);
subplot(1,2,2)
bar(Error_test_neu_net_02)
xlabel('Test Data','FontWeight','bold')
ylabel('Corrosion Rate Error (mm/year)','FontWeight','bold')
set(gca,'XMinorTick','on','XTick',[0 5 10 15 20],'LineWidth',1)
figure()
plot(X_01,Y_01,'k-','LineWidth',1.5)
hold on
plot(x_test_01.',Neural_Net_Corr_Rate_02,'bx','MarkerSize',7,'LineWidth',2.5)
plot(X_02_conf,Y_02_neunet_uci,'k--','LineWidth',1.5)
plot(X_02_conf,Y_02_neunet_lci,'k--','LineWidth',1.5)
xlim([0 60]);
266
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
ylim([0 60]);
xlabel('Experimental Corrosion Rate (mm/year)','FontWeight','bold','FontName','Times New
Roman')
ylabel('Neural Network Corrosion Rate (mm/year)','FontWeight','bold','FontName','Times
New Roman')
legend(' Experiment=Model Line',' Datapoints',' 95% Confidence Interval',2)
text('Position',[33 17],'String','R^2 =
0.81','Background','w','EdgeColor','black','LineWidth',1.0,'FontName','Times New Roman');
figure()
bar(Error_test_neu_net_02)
xlabel('Test Data','FontWeight','bold','FontName','Times New Roman')
ylabel('Corrosion Rate Error (mm/year)','FontWeight','bold','FontName','Times New Roman')
set(gca,'XMinorTick','on','XTick',[0 5 10 15 20],'LineWidth',1)
%t_1=linspace(20,90,15);
%figure()
%plot(t_1,y_1,'x')
%xlabel('Temperature (^oC)')
%ylabel('Corrosion Rate (mm/year)')
%[P_1,S_1]=polyfit(t_1,y_1,3);
%Y_1=polyval(P_1,t_1);
%hold on
%plot(t_1,Y_1,'k');
%y_2_pub_bar=mean(y_2_pub);
%y_2_pub_bar is the mean of y_2_pub
%y_2_pub_std=std(y_2_pub);
%y_2_pub_std is the standard deviation of y_2_pub
%n_pub=length(y_2_pub);
%n_pub is the number of NN predicted data-points in y_2_pub
%ci=0.95;
%ci is the confidence interval. In this case it's set at 95%
%alpha=1-ci;
%T_multiplier=tinv(1-alpha/2,n_pub-1);
267
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
%T_multiplier is part of the formulae for computing the confidence interval
%tinv is the matlab function which enables the computation of T_multiplier
%n_pub-1 is so because one degree of freedom is used in calculating the
%mean or standard deviation
%ci95=(T_multiplier.*y_2_pub_std)./sqrt(n_pub);
%conf_inf_95_01=[y_2_pub-ci95, y_2_pub+ci95];
%for n=1:length(y_2_pub)
%conf_int_95_01=y_2_pub-ci95;
%conf_int_95_02=y_2_pub+ci95;
%display(conf_int_95_01)
%display(conf_int_95_02)
%end
%display(conf_inf_95_01)
Y_02_net_pub_01=[4.347933016
7.977103099
11.6382261
15.35538299
19.17613313
23.18477432
27.47249051
32.03845381
36.79096383
41.64620233
46.5570312
51.49958737
56.46146411
61.43581576
66.41862842];
Y_02_net_pub_02=[-4.347933016
0.594296901
5.504573904
10.35881701
15.10946687
19.67222568
23.95590949
27.96134619
31.78023617
35.49639767
39.1569688
42.78581263
46.39533589
49.99238424
268
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
53.58097158];
figure()
plot(X_01,Y_01,'k-','LineWidth',1.5)
set(gca,'FontName','Times New Roman')
hold on
plot(x_test_01,y_2_pub,'bx','MarkerSize',7,'LineWidth',2.5)
plot(X_02_conf,Y_02_net_pub_01,'k--','LineWidth',1.5)
plot(X_02_conf,Y_02_net_pub_02,'k--','LineWidth',1.5)
%plot(conf_int_95_01,y_2_pub,'k--','LineWidth',1.5)
%plot(conf_int_95_02,y_2_pub,'k--','LineWidth',1.5)
xlim([0 60]);
ylim([0 60]);
xlabel('Experimental Corrosion Rate (mm/year)','FontWeight','bold','FontName','Times New
Roman')
ylabel('Neural Network Corrosion Rate (mm/year)','FontWeight','bold','FontName','Times
New Roman')
legend(' Experiment=Model Line',' Datapoints',' 95% Confidence Interval',2)
text('Position',[33 17],'String','\itR^2\rm\bf =
0.91','FontWeight','bold','Background','w','EdgeColor','black','LineWidth',1.0,'FontName','Tim
es New Roman');
%Error_pub_01=x_test_01-y_2_pub;
Error_pub_02=[-10.44 -3.42 -2.61 -6.14 -4.41 0.08 1.35 2.73 3.50 -4.52 -0.10 3.63 0.30 1.25
5.86];
figure()
bar(Error_pub_02)
xlabel('Test Data','FontWeight','bold','FontName','Times New Roman')
ylabel('Corrosion Rate Error (mm/year)','FontWeight','bold','FontName','Times New Roman')
set(gca,'XMinorTick','on','XTick',[0 5 10 15 20],'LineWidth',1)
figure()
subplot(1,2,1)
plot(X_01,Y_01,'k-','LineWidth',1.5)
set(gca,'FontName','Times New Roman')
hold on
plot(x_test_01,y_2_pub,'bx','MarkerSize',7,'LineWidth',2.5)
269
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
plot(X_02_conf,Y_02_net_pub_01,'k--','LineWidth',1.5)
plot(X_02_conf,Y_02_net_pub_02,'k--','LineWidth',1.5)
%plot(conf_int_95_01,y_2_pub,'k--','LineWidth',1.5)
%plot(conf_int_95_02,y_2_pub,'k--','LineWidth',1.5)
xlim([0 60]);
ylim([0 60]);
xlabel('Experimental Corrosion Rate (mm/year)','FontWeight','bold','FontName','Times New
Roman')
ylabel('Neural Network Corrosion Rate (mm/year)','FontWeight','bold','FontName','Times
New Roman')
legend(' Experiment=Model Line',' Datapoints',' 95% Confidence Interval',2)
text('Position',[5 37],'String','\itR^2\rm\bf =
0.91','FontSize',12,'FontWeight','bold','Background','w','EdgeColor','black','LineWidth',1.0,'Fo
ntName','Times New Roman');
subplot(1,2,2)
Error_pub_01=x_test_01-y_2_pub;
bar(Error_pub_02)
xlabel('Test Data','FontWeight','bold','FontName','Times New Roman')
ylabel('Corrosion Rate Error (mm/year)','FontWeight','bold','FontName','Times New Roman')
set(gca,'XMinorTick','on','XTick',[0 5 10 15 20],'LineWidth',1)
toc
270
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Monte Carlo Simulation Model Code
%The data in columns are low pressure CO_2 corrosion data from Dugstad 1994
%First column => Temperature (^oC)
%Second column => CO2 Pressure (MPa)
%Third column column => Velocity/flow (m/s)
%Fourth column => pH
%Fifth column => CO2 Corrosion rate (mm/year)
Dugstad_1994_data_01=...
[20.0000 0.0500 0.1000 4.1000 0.7000
20.0000 0.1300 0.1000 3.8000 0.9000
20.0000 0.1700 0.1000 3.7000 0.5000
20.0000 0.2800 0.1000 3.7000 1.9000
20.0000 0.3800 0.1000 3.7000 2.5000
20.0000 0.0500 13.0000 4.1000 3.5000
20.0000 0.0500 8.5000 4.1000 4.0000
20.0000 0.2800 13.0000 3.7000 13.0000
20.0000 0.2800 8.5000 3.7000 10.0000
20.0000 0.2800 3.1000 3.7000 9.0000
20.0000 0.1300 13.0000 3.8000 7.0000
20.0000 0.1300 8.5000 3.8000 8.0000
20.0000 0.1700 3.1000 3.7000 4.8000
20.0000 0.1700 8.5000 3.7000 5.4000
20.0000 0.1700 13.0000 3.7000 8.0000
40.0000 0.2000 3.1000 3.8000 1.0000
40.0000 0.2000 8.5000 3.8000 6.0000
40.0000 0.2000 13.0000 3.8000 18.0000
40.0000 0.0400 3.1000 4.1500 4.0000
40.0000 0.0400 8.5000 4.1500 5.0000
40.0000 0.0400 13.0000 4.1500 6.5000
40.0000 0.0530 3.1000 4.1000 4.5000
40.0000 0.0530 8.5000 4.1000 6.0000
40.0000 0.0530 13.0000 4.1000 8.5000
40.0000 0.0700 8.5000 4.0500 2.5000
40.0000 0.0700 13.0000 4.0500 9.0000
40.0000 0.7000 3.1000 3.6000 4.6000
40.0000 0.7000 8.5000 3.6000 17.0000
40.0000 0.7000 13.0000 3.6000 23.0000
40.0000 1.2000 8.5000 3.4000 15.0000
60.0000 0.2000 0.1000 3.8000 3.0000
60.0000 0.2100 0.1000 3.8000 5.0000
60.0000 0.2200 0.1000 3.8000 6.8000
60.0000 0.2200 3.1000 3.8000 16.0000
60.0000 0.2200 8.5000 3.8000 23.0000
60.0000 0.2200 13.0000 3.8000 32.0000
60.0000 0.2200 13.0000 3.8000 44.0000
60.0000 1.8000 3.1000 3.4000 33.0000
271
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
60.0000 1.8000 8.5000 3.4000 60.0000
60.0000 1.8000 13.0000 3.4000 70.0000
60.0000 0.0700 3.1000 4.1500 8.8000
60.0000 0.0700 13.0000 4.1500 20.0000
60.0000 0.4000 3.1000 3.7000 30.0000
60.0000 0.4000 8.5000 3.7000 34.0000
60.0000 0.4000 13.0000 3.7000 50.0000
90.0000 0.1600 3.1000 4.1000 1.9000
90.0000 0.1600 8.5000 4.1000 8.5000
90.0000 0.1600 13.0000 4.1000 19.0000
90.0000 0.3700 3.1000 3.8000 16.0000
90.0000 0.3700 13.0000 3.8000 32.0000
90.0000 0.3700 3.1000 3.8000 17.0000
90.0000 0.3700 8.5000 3.8000 21.0000
90.0000 1.4000 3.1000 3.6000 6.8000
90.0000 1.4000 3.1000 3.6000 15.0000
90.0000 1.4000 8.5000 3.6000 30.0000
90.0000 1.4000 13.0000 3.6000 41.0000
90.0000 1.4000 8.5000 3.6000 50.0000
90.0000 2.1000 3.1000 3.5000 15.0000
90.0000 2.1000 3.1000 3.5000 38.0000
90.0000 2.1000 13.0000 3.5000 65.0000];
%'Temperature ' 'CO_2 Partial Pressure (MPa)' 'Flowrate(m/s) ' 'pH
' 'Corrosion Rate (mm/year) '};
X_01=Dugstad_1994_data_01(:,1:3);
x_01=Dugstad_1994_data_01(:,1);
%x_01, x_02, x_03 and x_04 are column matrices comprising temp, pressure, velocity and
%pH values respectively
x_02=Dugstad_1994_data_01(:,2);
x_03=Dugstad_1994_data_01(:,3);
x_04=Dugstad_1994_data_01(:,4);
y_01=Dugstad_1994_data_01(:,5);
%y_01 represents a column matrix comprising the Corrosion rate values
%Temperature dataset%
%normal probability plot for temperature distribution
figure()
%comparing the fit of different probability distributions against the
%temperature dataset
subplot(3,2,1)
probplot('normal',x_01)
%hold on
%plot(conf_int_low_01,prob_frac_01,'k--')
%plot(conf_int_up_01,prob_frac_01,'k--')
272
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
xlabel('Temperature (^oC)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Probability','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
subplot(3,2,2)
histfit(x_01)
xlim([0 150])
xlabel('Temperature (^oC)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
%probplot('rayleigh',x_01)
subplot(3,2,3)
probplot('rayleigh',x_01)
%hold on
%plot(conf_int_low_01,prob_frac_01,'k--')
%plot(conf_int_up_01,prob_frac_01,'k--')
subplot(3,2,4)
histfit(x_01,4,'rayleigh')
xlim([0 150])
xlabel('Temperature (^oC)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
subplot(3,2,6)
histfit(x_01,4,'rayleigh')
xlim([0 150])
xlabel('Temperature (^oC)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
%PCO2 dataset%
%normal probability plot for PCO2 distribution
figure()
%comparing the fit of different probability distributions against the
%temperature dataset
subplot(3,2,1)
probplot('normal',x_02)
%hold on
%plot(conf_int_low_01,prob_frac_01,'k--')
%plot(conf_int_up_01,prob_frac_01,'k--')
subplot(3,2,2)
histfit(x_02)
xlim([0 3])
xlabel('CO_2 Partial Pressure (MPa)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
%probplot('rayleigh',x_01)
subplot(3,2,3)
probplot('rayleigh',x_02)
%hold on
%plot(conf_int_low_01,prob_frac_01,'k--')
%plot(conf_int_up_01,prob_frac_01,'k--')
274
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
subplot(3,2,4)
histfit(x_02,4,'rayleigh')
xlim([0 3])
xlabel('CO_2 Partial Pressure (MPa)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
subplot(3,2,6)
histfit(x_02,4,'rayleigh')
xlim([0 3])
xlabel('CO_2 Partial Pressure (MPa)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
%Velocity dataset%
%normal probability plot for Velocity distribution
figure()
%comparing the fit of different probability distributions against the
%temperature dataset
subplot(3,2,1)
probplot('normal',x_03)
%hold on
%plot(conf_int_low_01,prob_frac_01,'k--')
%plot(conf_int_up_01,prob_frac_01,'k--')
subplot(3,2,2)
histfit(x_03)
xlim([0 15])
275
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
xlabel('Velocity (m/s)','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
%probplot('rayleigh',x_01)
subplot(3,2,3)
probplot('exponential',x_03)
%hold on
%plot(conf_int_low_01,prob_frac_01,'k--')
%plot(conf_int_up_01,prob_frac_01,'k--')
subplot(3,2,4)
histfit(x_03,4,'exponential')
xlim([0 15])
xlabel('Velocity (m/s)','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
subplot(3,2,6)
histfit(x_03,4,'extreme value')
xlim([0 15])
xlabel('Velocity (m/s)','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
%pH dataset%
%normal probability plot for pH distribution
figure()
%comparing the fit of different probability distributions against the
276
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
%temperature dataset
subplot(3,2,1)
probplot('normal',x_04)
%hold on
%plot(conf_int_low_01,prob_frac_01,'k--')
%plot(conf_int_up_01,prob_frac_01,'k--')
subplot(3,2,2)
histfit(x_04)
xlim([3 5])
xlabel('pH','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
%probplot('rayleigh',x_01)
subplot(3,2,3)
probplot('rayleigh',x_04)
%hold on
%plot(conf_int_low_01,prob_frac_01,'k--')
%plot(conf_int_up_01,prob_frac_01,'k--')
subplot(3,2,4)
histfit(x_04,4,'rayleigh')
xlim([3 5])
xlabel('pH','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
277
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
set(gca,'FontName','Times New Roman','LineWidth',1)
xlabel('pH','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
ylabel('Probability','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
subplot(3,2,6)
histfit(x_04,4,'rayleigh')
xlim([3 5])
xlabel('pH','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
subplot(3,2,2)
histfit(y_01,10,'normal')
xlim([0 80])
xlabel('Corrosion Rate (mm/year)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
%probplot('rayleigh',x_01)
subplot(3,2,3)
probplot('rayleigh',y_01)
%hold on
%plot(conf_int_low_01,prob_frac_01,'k--')
%plot(conf_int_up_01,prob_frac_01,'k--')
subplot(3,2,4)
histfit(y_01,10,'rayleigh')
xlim([0 80])
xlabel('Corrosion Rate (mm/year)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
subplot(3,2,6)
histfit(y_01,10,'weibull')
xlim([0 80])
xlabel('Corrosion Rate (mm/year)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
tic
toc
%Multiple linear regression equation for CO2 Corrosion rate with T, PCO2,
%Vel and pH as the independent variables
Corr_Rate_01_test=(78.2+(0.9923.*x_01)-(2.0.*x_02)+(0.9826.*x_03)-(25.0.*x_04)...
279
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
-(7.0e-3.*x_01.^2)+(5.3011e-3.*x_01.*x_02.*x_03.*x_04));
%Corr_rate_01_test is the original multiple linear equation derived from
%regression analysis
Corr_Rate_01_monte=abs(78.2+(0.9923.*temp_01)-(2.0.*pCO2_01)+(0.9826.*vel_01)-
(25.0.*pH_02)...
-(7.0e-3.*temp_01.^2)+(5.3011e-3.*temp_01.*pCO2_01.*vel_01.*pH_02));
figure()
subplot(1,2,1)
histfit(Corr_Rate_01_monte,100,'normal')
set(gca,'FontName','Times New Roman','LineWidth',1)
xlabel('Corrosion Rate (mm/year)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
xlim([0 80])
sample_01=wblrnd(15.9056,0.8561,1e5,1);
subplot(1,2,2)
histfit(Corr_Rate_01_monte,100,'weibull')
set(gca,'FontName','Times New Roman','LineWidth',1)
xlabel('Corrosion Rate (mm/year)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
xlim([0 80])
%sample_01=wblrnd(12.34,1.56,1e5,1);
figure()
%comparing the fit of different probability distributions against the
%Corrosion Rate dataset
%subplot(3,2,1)
subplot(1,2,1)
pp_normal=probplot('normal',y_01);
%probplot('normal',y_01)
%hold on
%plot(conf_int_low_01,prob_frac_01,'k--')
%plot(conf_int_up_01,prob_frac_01,'k--')
title('')
%axes;
set(gca,'FontName','Times New Roman','LineWidth',1)
%line;
%set(gca,'Color','b--','LineWidth',2)
subplot(1,2,2)
pp_weibull=probplot('weibull',y_01);
%probplot('weibull',y_01)
%hold on
%plot(conf_int_low_01,prob_frac_01,'k--')
%plot(conf_int_up_01,prob_frac_01,'k--')
%axes;
set(gca,'FontName','Times New Roman','LineWidth',1)%,'XTick',0:20:80)
%xlim([0 80])
%line;
%set(gca,'Color','b--','LineWidth',2)
%figure()
%subplot(121)
%sample = wblrnd(15.9056,0.8561,100,1);
%histfit(sample,100,'wbl')
%title('100 draws')
%subplot(122)
%sample = wblrnd(15.9056,0.8561,1e5,1);
%histfit(sample,100,'wbl')
%title('100,000 draws')
title('')
xlabel('Corrosion Rate (mm/year)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Probability','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
set(pp_weibull(1),'Color','b') % line
set(pp_weibull(2),'Color','k') % markers
%display(mean_03)
%display(std_dev_03)
281
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
fprintf('The mean and standard deviation of the Experimental CO2 Corr rate data are %.2f
mm/year and %.2f mm/year respectively \n',mean_exp_corr,std_exp_corr)
fprintf('The mean and standard deviation of the Monte Carlo CO2 Corr model are: %.2f
mm/year and %.2f mm/year respectively \n',mean_03,std_dev_03)
fprintf('The percentage errors in the means and standard deviations of experimental data and
the Monte Carlo model are: %.2f and %.2f respectively
\n',mean_per_error_01,std_per_error_01)
%--Sensitivity Analysis--%
temp_ave=(min(x_01)+max(x_01))./2;
pCO2_ave=(min(x_02)+max(x_02))./2;
vel_ave=(min(x_03)+max(x_03))./2;
pH_ave=(min(x_04)+max(x_04))./2;
for i_1=temp_sens_01
corr_rate_01_temp_change=(78.2+(0.9923.*temp_sens_01)-
(2.0.*pCO2_ave)+(0.9826.*vel_ave)-(25.0.*pH_ave)...
-(7.0e-3.*temp_sens_01.^2)+(5.3011e-
3.*temp_sens_01.*pCO2_ave.*vel_ave.*pH_ave));
end
for i_2=pCO2_sens_01
corr_rate_01_pCO2_change=(78.2+(0.9923.*temp_ave)-
(2.0.*pCO2_sens_01)+(0.9826.*vel_ave)-(25.0.*pH_ave)...
-(7.0e-3.*temp_ave.^2)+(5.3011e-3.*temp_ave.*pCO2_sens_01.*vel_ave.*pH_ave));
end
282
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
for i_3=vel_sens_01
corr_rate_01_vel_change=(78.2+(0.9923.*temp_ave)-
(2.0.*pCO2_ave)+(0.9826.*vel_sens_01)-(25.0.*pH_ave)...
-(7.0e-3.*temp_ave.^2)+(5.3011e-3.*temp_ave.*pCO2_ave.*vel_sens_01.*pH_ave));
end
for i_4=pH_sens_02
corr_rate_01_pH_change=(78.2+(0.9923.*temp_ave)-(2.0.*pCO2_ave)+(0.9826.*vel_ave)-
(25.0.*pH_sens_02)...
-(7.0e-3.*temp_ave.^2)+(5.3011e-3.*temp_ave.*pCO2_ave.*vel_ave.*pH_sens_02));
end
display(corr_rate_01_temp_change)
figure()
plot(percent_change_01,corr_rate_01_temp_change,'k-p','LineWidth',2.0)
hold on
plot(percent_change_01,corr_rate_01_pCO2_change,'g-o','LineWidth',2.0)
plot(percent_change_01,corr_rate_01_vel_change,'b-*','LineWidth',2.0)
plot(percent_change_01,corr_rate_01_pH_change,'r-x','LineWidth',2.0,'MarkerSize',10.0)
xlim([-50 50])
xlabel('Percentage Change in Input Magnitude (%)','FontName','Times New
Roman','FontWeight','Bold')
ylabel('Corrosion Rate (mm/year)','FontName','Times New Roman','FontWeight','Bold')
legend('Temperature','CO_2 Partial Pressure','Velocity','pH')
grid on
ylim([0 80])
figure()
subplot(1,2,1)
histfit(y_01,10,'normal')
set(gca,'FontName','Times New Roman','LineWidth',1)
xlim([0 80])
xlabel('Corrosion Rate (mm/year)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
subplot(1,2,2)
283
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
%figure()
histfit(y_01,10,'weibull')
xlim([0 80])
xlabel('Corrosion Rate (mm/year)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Frequency','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
figure()
%comparing the fit of different probability distributions against the
%Corrosion Rate dataset
%subplot(3,2,1)
subplot(1,2,1)
pp1=probplot('normal',y_01);
%probplot('normal',y_01)
%hold on
%plot(conf_int_low_01,prob_frac_01,'k--')
%plot(conf_int_up_01,prob_frac_01,'k--')
title('')
%axes;
set(gca,'FontName','Times New Roman','LineWidth',1)
%line;
%set(gca,'Color','b--','LineWidth',2)
subplot(1,2,2)
pp1=probplot('weibull',y_01);
%probplot('normal',y_01)
%hold on
%plot(conf_int_low_01,prob_frac_01,'k--')
%plot(conf_int_up_01,prob_frac_01,'k--')
%axes;
set(gca,'FontName','Times New Roman','LineWidth',1)
%line;
%set(gca,'Color','b--','LineWidth',2)
title('')
xlabel('Corrosion Rate (mm/year)','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1)
ylabel('Probability','FontWeight','Bold','FontName','Times New Roman','LineWidth',1)
set(pp1(1),'Color','k') % line
284
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
set(pp1(2),'MarkerSize',4,'Color','b') % markers
%figure()
%[f,x_values,flo,fup] = ecdf(y_01);
%F=plot(x_values,f);
%set(F,'LineWidth',2,'k');
%hold on
figure()
%[H_01,STAT_01]=cdfplot(y_01);
%G=plot(x_values,normcdf(x_values,0,1),'b-');
%set(G,'LineWidth',2);
%legend([F G],...
% 'Empirical CDF','Standard Normal CDF',...
% 'Location','SE');
%display(STAT_01)
figure()
%[f_monte,x_values_monte,flo_monte,fup_monte] = ecdf(Corr_Rate_01_monte);
%plot(x_values_monte,f_monte)
%hold on
%[H_03,STAT_03]=cdfplot(Corr_Rate_01_monte);
%display(STAT_03)
%test_cdf= [y_01,cdf('normal',17.03,17.02,1)];
%[H,P,KSSTAT,CV] = kstest(y_01,'cdf',test_cdf,0.05,'unequal');
285
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
[H,P,KSSTAT,CV] = kstest(y_01); %Carrying out a Kolmogorov-Smirnov (kstest) normality
test on the experimental CO2 corrosion rate data(y_01)
display(H) %H is the hypothesis, when H=0, accept the nul hypothesis; when H=1, reject the
null hypothesis (hence data does not follow Normal distribution)
display(P) %P is the p-value of the test (asymptotic p-value)
display(KSSTAT) %KSSTAT is the magnitude of the kstest statistic
display(CV) %CV is the critical Value
display(H_02) %H is the hypothesis, when H=0, accept the nul hypothesis; when H=1, reject
the null hypothesis (hence data does not follow Normal distribution)
display(P_02) %P is the p-value of the test (asymptotic p-value)
display(KSSTAT_02) %KSSTAT is the magnitude of the kstest statistic
display(CV_02) %CV is the critical Value
%R_01=wblrnd(31.30,2.236,[10000,1]);
%[H_03,P_03,KSSTAT_03] = kstest2(Corr_Rate_01_monte,R_01,0.05,'unequal');
%display(H_03) %H is the hypothesis, when H=0, accept the nul hypothesis; when H=1,
reject the null hypothesis (hence data does not follow Normal distribution)
%display(P_03) %P is the p-value of the test (asymptotic p-value)
%display(KSSTAT_03) %KSSTAT is the magnitude of the kstest statistic
%display(CV_03) %CV is the critical Value
figure()
% Output fitted probablility distributions: PD1,PD2,PD3
% Prepare figure
clf;
hold on;
LegHandles = []; LegText = {};
xlim([0 80])
% --- Create fit "fit 2"
287
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
% Adjust figure
box on;
hold off;
figure()
% Output fitted probablility distributions: PD1,PD2,PD3
% Prepare figure
%clf;
hold on;
LegHandles_01 = []; LegText_01 = {};
% Adjust figure
box on;
hold off;
%Corr_rate_01_monte_weib=wblrnd(17.136,1.0149,[sample_num_01 1]);
%figure()
%histfit(Corr_rate_01_monte_weib,100,'weibull')
%display(mean(Corr_rate_01_monte_weib))
%display(std(Corr_rate_01_monte_weib))
Sim_run_01=[1e7 1e6 1e5 1e4 1e3]; %Simulation Run to determine best number for Monte
Carlo simulations
figure()
semilogx(Sim_run_01,Time_01,'b-*','LineWidth',2)
set(gca,'FontName','Times New Roman','LineWidth',1)
xlabel('Number of Monte Carlo Simulations','FontWeight','Bold','FontName','Times New
Roman','LineWidth',1');
ylabel('Time Elapsed (s)','FontWeight','Bold','FontName','Times New Roman','LineWidth',1')
hold on
text(0.3e4,0.3,'Chosen Number of Simulation Runs','FontName','Times New
Roman','FontWeight','Bold')
annotation('arrow',[0.4289 0.5125],...
[0.3143 0.1310]);
figure()
% Output fitted probablility distributions: PD1,PD2
% Prepare figure
clf;
hold on;
LegHandles_02 = []; LegText_02 = {};
% Adjust figure
box on;
hold off;
figure()
% Output fitted probablility distributions: PD1,PD2
% Prepare figure
clf;
291
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
hold on;
LegHandles_03 = []; LegText_03 = {};
% Adjust figure
box on;
hold off;
292
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
%Optimising the Mean and Standard Deviation of the Monte Carlo dataset
%Soln_01 is the difference between the mean of the Monte Carlo dataset and
%that of the Experimental corrosion rate dataset
a_01=mean(y_01);
a_02=mean(Corr_Rate_01_monte);
b_01=std(y_01);
b_02=std(Corr_Rate_01_monte);
%Soln_01=fzero(@mean_optim_01,5,optimset('TolFun',1e-12));
%Soln_02=fzero(@stand_dev_optim_01,5,optimset('TolFun',1e-12));
%figure()
%histfit(Corr_Rate_01_monte,100,'Weibull')
%[x_pH,fval,exitflag,output] = Auto_gen_01(x0)
%% This is an auto generated MATLAB file from Optimization Tool.
display(x_vel)
display(fval_vel)
display(exitflag_vel)
display(output_vel)
293
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
display(x_vel_01)
display(fval_vel_01)
display(exitflag_vel_01)
display(output_vel_01)
294
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
References
Ali, S., Smith, K. A. (2006). On Learning Algorithm Selection for Classification, Applied
Soft Computing, Vol. 6, pp. 119-138.
Alsmadi, M. K. S., Omar, K. B., Noah, S. A. (2009). Back Propagation Algorithm: The Best
Algorithm among the Multi-layer Perceptron Algorithm, International Journal of Computer
Science and Network Security, Vol. 9(4), 2009, pp. 378 – 383.
Al-Fattah, S.M. and Startzman, R.A. (May 2003). Neural Network Approach Predicts U.S.
Natural Gas Production, SPE production and Facilities, Vol. 18, No. 2, pp. 84-91.
American Petroleum Institute. (1981). Recommended Practice for Design and Installation of
offshore Production Platform Piping Systems, American Petroleum Institute, Recommended
Practice 14E (API RP 14E), Washington DC, USA.
Baboian, R. (eds.), Dean Jr., S. W. D. (eds.), Hack, H. P. (eds.), Hibner, E. L. (eds.), Scully, J.
R. (eds.) (2005). Corrosion Tests and Standards: Application and Interpretation, 2nd Edition,
ASTM International, 100 Barr Harbor Drive, PO Box C700, West Conshohocken, PA 19428-
2959, Chapter 5: Statistical Treatment of Data, Data Interpretation and Reliability, Pages 84,
85.
295
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Beale, M. H., Hagan, M. T., Demuth H. B. (2014). Neural Network Toolbox User’s Guide
R2014b. The MathWorks, Inc., 3 Apple Hill Drive Natick, MA 01760-2098. 2014, Pages 2-4,
2-5, 2-17, 2-18, 2-19.
Belohlavek, R. (ed.) and Klir, G. (ed.) (2011). Concepts and Fuzzy Logic, Massachusetts
Institute of Technology, USA.
Benning, L. G., Wilkin, R. T., Barnes, H. L. (2000). Chemical Geology, 167, 25‐51.
Bondos, J. C., Reddy, R. V., Pugh, D. V. and Norman, D. A., Pacheco, J. L. and Nelson, J. L.
(2007). Accurate Corrosion Prediction through an Integrated Approach, SPE Production &
Operations: 2007, Vol. 22, No. 2, pp. 161-167.
Campbell, B. [Future Refining Capacity Changes, Chapter 3] [in Annual Outlook for Oil and
Gas 1991], Energy Information Administration (EIA), Office of Oil and Gas, U.S.
Department of Energy (DOE), Washington DC 20585, USA. Page 42.
Carberry, J., J. (2001). Chemical and Catalytic Reaction Engineering, Dover Edition, General
Publishing Company, Ltd., 30 Lesmill Road, Don Mills, Toronto, Ontario, Canada, p. 383.
Chatterjee, S., Hadi, A., S. (1986). Influential Observations, High Leverage Points, and
Outliers in Linear Regression, Statistical Science, Vol. 1, No. 3, pp. 379-416.
Choi, Y-S., Nesic, S. (2009). Corrosion Behavior of Carbon Steel in Supercritical CO2-Water
Environments, NACE International, Corrosion 2009, Paper 09256, Atlanta, Georgia, USA.
296
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Choi, Y-S, Nešić, S. (2011). Determining the Corrosive Potential of CO2 Transport Pipeline
in High pCO2-Water Environments, International Journal of Greenhouse Gas Control,
Volume 5, Issue 4, pp. 788-797.
Choi, Y-S., Nesic, S. and Young, D. (2010). Effect of Impurities on the Corrosion Behavior of
CO2 Transmission Pipeline Steel in Supercritical CO2-Water Environments, Environmental
Science and Technology, Vol. 44, Issue 23, pp. 9233-9238.
Clarke, L., Edmonds, J., Jacoby, H., Pitcher, H., Reilly, J., Richels, R. (2007). Scenarios of
Greenhouse Gas Emissions and Atmospheric Concentrations. Sub-report 2.1A of Synthesis
and Assessment Product 2.1 by the U.S. Climate Change Science Program and the
Subcommittee on Global Change Research. Department of Energy, Office of Biological &
Environmental Research, Washington, 7 DC., USA, 154 pp. 59-86.
Criaud, A., Fouillac, C. and Marty, B. (1989). Low Enthalpy Geothermal fluids from the Paris
basin. 2—Oxidation‐reduction state and consequences for the prediction of corrosion and
sulfide scaling, Geothermics, Vol. 18, Issues 5‐6, 711‐727.
Cottis, R. A., Qing, L., Owen, G., Gartland , S. J., Helliwell, I. A., Turega, M. (1999).
Neural Network methods for Corrosion Data Reduction, Materials and Design: 1999, Vol. 20,
No. 4, pp. 169-178.
297
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
CO2 pipeline Infrastructure: An analysis of global challenges and opportunities report,
International Energy Agency Greenhouse Gas (IEA GHG), Conference /2010, (Cambridge,
UK, 2010).
Crolet, J. L., Bonis, M. R. (1991). Prediction of the Risks of CO2 Corrosion in Oil and Gas
Wells, SPE Production Engineering, Society of Petroleum Engineers (SPE), Volume 6,
Number 4, pp. 449-453.
Crolet, J. L. (1994). “Which CO2 Corrosion, Hence Which Prediction?,”in Predicting CO2
Corrosion in the Oil and Gas Industry, European Federation of Corrosion Publication no. 13
(London, U.K.: Institute of Materials, 1994), p 1.
Crolet, J-L., Thevenot, N., Dugstad, A. (1999). Role of Free Acetic Acid on the CO2
Corrosion of Steels, CORROSION 99, Paper 99024, NACE International, San Antonio,
Texas, USA.
Crolet, J. L. (2002). Corrosion in Oil and Gas Production, in Corrosion and Anticorrosion,
eds. G. Beranger, H. Mazille (Paris, France: Hermes Science, 2002).
Crolet, J. L., Thevenot, N., Nesic, S. (1996). Role of Conductive Corrosion Products on the
Protectiveness of Corrosion Layers, Corrosion 96, paper no. 4 (Houston, TX: NACE, 1996).
Cui, Z. D., Wu, S. L., Zhu, S. L., Yang, X. J. (2006), Study on Corrosion Properties of
Pipelines in Simulated Produced Water with Supercritical CO2, Journal of Applied Surface
Science, Volume 252, Issue 6, pp.2368-2374.
Demuth, H., Beale, M., Hagan, M. (2009). Neural Network Toolbox 6 User’s Guide, The
MathWorks, Inc., 3 Apple Hill Drive Natick, MA 01760-2098, Pages 2-4, 2-5, 2-17, 2-18, 2-
19.
298
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
De Waard, C., Lotz, U. and Dugstad, A. (1995). Influence of Liquid Flow Velocity on CO2
Corrosion: A Semi-Empirical Model, CORROSION/95, Paper 128, NACE International,
Houston, Texas, USA.
De Waard, C., Lotz, U. and Milliams, D. E. (1991). Predictive model for CO2 Corrosion
Engineering in Wet Natural Gas Pipelines, Corrosion: 1991, Vol. 47, No. 12, pp. 976-985.
Dlugokencky, E. J., Hall, B. D., Montzka, S. A., Dutton, G., Muhle, J., and Elkins, J. W.
(2014): [Global climate] Long-lived greenhouse gases [in State of the Climate in 2013].
Bulletin of the American Meteorological Society, Volume 95, No.7, S33–S34.
Downie, M., Race, J., Seevam, P. (2007). Transport of CO2 for Carbon Capture and Storage
in the UK, Offshore Europe 2007, Society of Petroleum Engineers (SPE), Aberdeen,
Scotland, UK.
Dugstad, A., Lunde, L. and Nesic, S. (1994a). Control of Internal Corrosion in Multi-phase
Oil and Gas Pipelines, Prevention of Pipeline Corrosion Conference, Pipe Line Industry and
Pipes & Pipelines International, Houston, Texas, USA.
Dugstad, A., Lunde, L. and Videm, K. (1994b). Parametric Study of CO2 Corrosion of
Carbon Steel, Corrosion/94, Paper no. 14, (Houston, TX: NACE International, 1994).
299
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Eriksrud, E., Sφntvedt, T. (1983). Effect of Flow on CO2 Corrosion Rates in Real and
Synthetic Formation Waters, Paper 44, NACE International.
Fujino, J., Nair, R., Kainuma, M., Masui, T. and Matsuoka, Y. (2006). Multi-gas Mitigation
Analysis on Stabilization Scenarios Using Aim Global Model, The Energy Journal, Special
Issue: Multi-Greenhouse Gas Mitigation and Climate Policy, Volume 27, pp. 344-354.
Fuzzy Logic Toolbox User’s Guide Matlab R2015b, (2015). The MathWorks, Inc., 3 Apple
Hill Drive Natick, MA 01760-2098.
Giourntas, L., Hodgkiess, T., Galloway, A. M. (2015). Enhanced Approach of Assessing the
Corrosive Wear of Engineering Materials under Impingement, Wear, Volumes 338-339, pp.
155-163.
Grau, R., M. (2000). Mechanisms for Improved Corrosion Resistance of Steels in Carbonic
Acid Environments, PhD Thesis, University of Leeds.
Graves, A., Liwicki, M., Fernandez, S., Bertolami, R., Bunke, H., Schmidhuber, J. (2009). A
Novel Connectionist System for Improved Unconstrained Handwriting Recognition, IEEE
Transactions on Pattern Analysis and Machine Intelligence, Volume 31, No. 5, pp. 855-868.
300
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Gronhölm, T., Annila, A. (2007). Natural Distribution, Mathematical Biosciences, Volume
210, Issue 2, pp. 659-667.
Guo, X. P., Chen, Z. Y. and Liu, D. (2005). The Effect of Acetic Acid and Acetate on CO2
Corrosion of Carbon Steel, CORROSION 2005, Paper 05306, NACE International, Houston,
Texas, USA.
https://co2corrosionchem409.wikispaces.com/Background+of+CO2+Corrosion
Hajizadeh, Y. (2006). Fuzzy Corrosion Modeling, New Horizons, New Opportunities, SPE
Eastern Regional Meeting, Paper 104298, Society of Petroleum Engineers (SPE), Canton,
Ohio, USA.
Hajizadeh, Y. (2007). Neural Networks Can Enhance Fuzzy Corrosion Modeling, Society of
Petroleum Engineers, SPE Annual Technical Conference, Society of Petroleum Engineers
(SPE), Anaheim, California, USA.
Halvorsen, A. M. K., Sφntvedt, T. (1999). CO2 Corrosion Model for Carbon Steel Including a
Wall Shear Stress Model for Multiphase Flow and Limits for Production Rate to Avoid Mesa
Attack, Corrosion 1999, Paper 99042, NACE International, San Antonio, Texas, USA.
301
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Hays, G. F. (2011). Now is the Time, World Corrosion Organization (WCO), pp.1-2.
Helliwell, I., Turega, M., Cottis, R. A. (1996). Neural networks for corrosion data reduction,
Corrosion/96, paper 379, NACE, 1996.
Hernandez, S., Nesic, S., Weckman, G. and Ghai, V. (2006). Use of Artificial Neural
Networks for Predicting Crude Oil Effect on Carbon Dioxide Corrosion of carbon Steels,
Corrosion: 2006, Vol. 62, No. 06, pp. 467-482.
Hesjevik, S. M., Olsen, S., Seiersten, M. (2003). Corrosion at High CO2 Pressure, StatOil
and Institute for Energy Technology in Norway, NACE International, Corrosion 2003, Paper
03345, San Diego, California, USA.
Hodgkiess, T., Neville, A., Shrestha, S. (1999). Electrochemical and Mechanical Interactions
during Erosion-Corrosion of a High-Velocity Oxy-fuel Coating and a Stainless Steel, Wear,
Volumes 233-235, pp. 623-634.
Hu, X., Barker, R. J., Neville, A., Gnanavelu, A. (2011). Case Study on Erosion-Corrosion
Degradation of Pipework Located on an Offshore Oil and Gas Facility, Wear, Volume 271,
Issues 9-10, pp. 1295-1301.
International Energy Agency (IEA), 2010: World Energy Outlook 2010, Chapter 14: The
energy transformation by sector, pp. 418.
International Energy Agency (IEA), 2012: World Energy Outlook 2012, Chapter 2: Energy
Trends to 2035, pp. 49-80.
302
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Ikeda, A., Mukai, S., Ueda, M. (1984). Prevention of CO2 Corrosion of Line Pipe and Oil
Country Tubular Goods, Corrosion 84, paper no. 289 (Houston, TX: NACE, 1984).
IPCC, 2014: Climate Change 2014: Synthesis Report. Contribution of Working Groups I, II
and III to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change
[Core Writing Team, R.K. Pachauri and L.A. Meyer (eds.)]. IPCC, Geneva, Switzerland, 151,
pp. 46, 47.
IPCC, 2007: Climate Change 2007: Synthesis Report. Contribution of Working Groups I, II
and III to the Fourth Assessment Report of the Intergovernmental Panel on Climate Change
[Core Writing Team, Pachauri, R.K and Reisinger, A. (eds.)]. IPCC, Geneva, Switzerland,
104, pp. 17, 60.
Jackson, J. E., A User's Guide to Principal Components, John Wiley and Sons, 1991, p. 592.
James, H. L., Fleischer, M. (ed.). (1966). [Chemistry of the Iron-rich Sedimentary Rocks,
Chapter W] [in Data of Geochemistry], Sixth Edition, U.S. Geological Survey Professional
Paper 440-W, United States Government Printing Office, Washington DC 20402, USA. Pages
W9, W10 and W11.
Johnson, M. L., & Tomson, M. B. (1991). Ferrous Carbonate Precipitation Kinetics And its
Impact on CO2 Corrosion, NACE International, Corrosion 1991, Paper No. 268, Houston,
Texas, USA.
303
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Jolliffe, I. T., Principal Component Analysis, 2nd edition, Springer, 2002.
Kermani, M. B. and Morshed, A. (2003). Carbon Dioxide Corrosion in Oil and Gas
Production – A Compendium, Corrosion: August 2003, Vol.59, No.8, pp. 659-683.
Kermani, M. B., Gonzales, J. C., Linne, C., Dougan, M., Cochrane, R. (2001). Development
of Low-Carbon Cr-Mo Steels with Exceptional Corrosion Resistance for Oilfield
Applications, Corrosion /2001, paper no. 01065 (Houston, TX: NACE, 2001).
Kermani, M. B., Gonzales, J. C., Turconi, G. L., Edmonds, D., Dicken, G., Scoppio, L.
(2003). Development of Superior Corrosion Resistance 3%Cr Steels for Downhole
Applications,” Corrosion/ 2003, Paper no. 03116 (Houston, TX: NACE, 2003).
Kermani, M. B., Smith, L. M. (1994). Predicting CO2 Corrosion in the Oil and Gas Industry,
European Federation of Corrosion Publication no. 13 (London, U.K.: Institute of Materials,
1994).
Koch, G. H., Brongers, M. P. H., Thompson, N. G., Virmani, Y. P., Payer, J. H. (2002).
Corrosion Costs and Preventive Strategies in the United States, U.S Department of
Transportation: Federal Highway Administration (FHWA), CC Technologies, NACE
International. Publication No. FHWA-RD-01-156.
Krogh, A. (2008). What are artificial neural networks?, Nature Biotechnology, Nature
Publishing Group, Vol. 26, No. 2, pp. 195-197.
304
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Lee, E. T., Wang, J. W. (2003). Statistical Methods for Survival Data Analysis, Third Edition,
Wiley-Interscience, A John Wiley & Sons, Inc., Publication, Hoboken, New Jersey, USA,
Chapters 6, 9, Pages 138, 143, 144, 221.
Lyle, F., Schutt, H. U. (1998). CO2/H2S Corrosion under Wet Gas Pipeline Conditions in the
Presence of Bicarbonate, Chloride and Oxygen, CORROSION 98, Paper 98011, NACE
International, San Diego, California, USA.
Mamdani, E., H. and Assilian, S. (1975). An experiment in linguistic synthesis with a fuzzy
logic controller, International Journal of Man-Machine Studies, Vol. 7, No. 1, pp. 1-13.
Man, H. C., and Gabe, D. R. (1981). The Determination of Pitting Potentials, Corrosion
Science, Volume 21, Issue 4, Pages 323-326.
Meinshausen, M., Smith, S. J., Calvin, K., Daniel, J. S., Kainuma, M. L. T., Lamarque, J-F.,
Matsumoto, K., Montzka, S. A., Raper, S. C. B., Riahi, K., Thomson, A., Velders, G. J.M.,
van Vuuren, D. P. P. (2011). The RCP Greenhouse Gas Concentrations and their Extensions
from 1765 to 2300, Climatic Change, Volume 109, Issue 1, pp. 213-241.
Mohamed, M. F., Nor, A. M., Suhor, M. F., Singer, M., Choi, Y. S., Nesic, S. (2011). Water
Chemistry For Corrosion Prediction in High Pressure CO2 Environments, Corrosion 2011,
Paper 11375, NACE International, Houston, Texas, USA.
Naiman, B. (1948). The Brönsted Concept of Acids and Bases in Quantitative Analysis,
Journal of Chemical Education, Volume 25, Issue 8, pp. 454-456.
305
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Nesic, S. and Vrhovac, M. (1999). A Neural Network Model for CO2 Corrosion of Carbon
Steel, The Journal of Corrosion Science and Engineering (JCSE), Vol. 1, No. 6, ISSN 1466-
8858.
Nesic, S., Cai, J., Lee, K. L. J. (2005). A Multiphase Flow and Internal Corrosion Prediction
Model for Mild Steel Pipelines, Corrosion 2005, Paper no. 05556, NACE International,
Houston, Texas, USA.
Nesic, S., Li, H., Sormaz, D. and Huang, J. (2009). An Open Source Mechanistic Model for
Prediction of Mild Steel Corrosion, CORROSION 2009, Paper 09572, NACE International,
Atlanta, Georgia, USA.
Nesic, S., Postlethwaite, J. and Olsen, S. (1996). An Electrochemical Model for Prediction of
Corrosion of Mild Steel in Aqueous Carbon Dioxide Solutions, Corrosion: April 1996,
Volume 52, No. 04, pp. 280-294.
Nesic, S., Wang, S., Fang, H., Sun, W. and Lee, J. K-L. (2008). A New Updated Model of
CO2/H2S Corrosion in Multiphase flow, CORROSION 2008, Paper 08535, NACE
International, New Orleans, Louisiana, USA.
Nordsveen, M., Nesic, S., Nyborg, R. and Stangeland, A. (2003). A Mechanistic Model for
Carbon Dioxide Corrosion of Mild Steel in the Presence of Protective Iron Carbonate Films-
Part 1: Theory and Verification, CORROSION, NACE International, Volume 59, No. 5, pp.
443-456.
NORSOK STANDARD M-506 Rev. 2:CO2 Corrosion Rate Calculation Model, Standards
Norway, Norway (June 2005).
Nyborg, R. (2002). Overview of CO2 Corrosion Models for Wells and Pipelines, Corrosion
2002, Paper no. 02233, NACE International, Houston, Texas, USA.
306
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Nyborg, R. (2005). Controlling Internal Corrosion in Oil and Gas Pipelines, Business
Briefing: Exploration and Production, Oil and Gas Review, Issue 2, pp. 70-74.
Nyborg, R. (2010). CO2 Corrosion Models for Oil and Gas Production Systems, Corrosion
2010, Paper no. 10371, NACE International, San Antonio, Texas, USA.
Ohio University (2008), Department of Chemical Engineering, Institute for Corrosion and
Multiphase Technology, Ohio University’s Research and Enterprise Park, 342 West State
Street, Athens, Ohio 45701, USA. Accessed from:
http://www.corrosioncenter.ohiou.edu/software/freecorp/
Ohio University (2015), Department of Chemical Engineering, Institute for Corrosion and
Multiphase Technology, Ohio University’s Research and Enterprise Park, 342 West State
Street, Athens, Ohio 45701, USA. Accessed from:
https://www.ohio.edu/mechanical/thermo/Intro/Chapt.1_6/Chapter2a.html
Olsen, S, Halvorsen, A. M., Lunde, P. G., Nyborg, R. (2005). CO2 Corrosion Prediction
Model – Basic Principles, CORROSION 2005, Paper 05551, NACE International, Houston,
Texas, USA.
Perez, T. E. (2013). Corrosion in the Oil and Gas Industry: An Increasing Challenge for
Materials, The Journal of Minerals, Metals and Materials Society (TMS), Vol. 65, No. 8, pp.
1033-1041.
Pots, B. F. M., John, R. C., Rippon, I. J., Thomas, M. J. J. S., Kapusta, S. D., Girgis, M. M.,
Whitham, T. (2002). Improvements on De Waard-Milliams Correlation Prediction and
Applications to Corrosion Management, Corrosion 2002, Paper no. 02235, NACE
International, Houston, Texas, USA.
307
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Rao, A., Rubin, E. (2002). A Technical, Economic and Environmental Assessment of Amine-
based CO2 Capture Technology for Power Plant Greenhouse Gas Control, Environmental
Science & Technology, Volume 36, pp. 4467-4475.
Riahi, K., Grubler, A., Nakicenovic, N. (2007). Scenarios of Long-term Socio-economic and
Environmental Development under Climate Stabilization, Technological Forecasting and
Social Change, Volume 74, Issue 7, pp. 887-935.
Rice, J. A. (2007). Mathematical Statistics and Data Analysis, Third Edition, Thomson
Brooks/Cole Publishing, Thomson Higher Education, 10 Davis Drive, Belmont, California
94002-3098, USA, Chapter 5, Pages 181-183.
Sak, H., Senior, A. W., Beaufays, F. (2014). Long Short-Term Memory Recurrent Neural
Network Architectures for Large Scale Acoustic Modeling, INTERSPEECH-2014, Singapore,
pp. 338-342.
Satter, A., Iqbal, G. M., Buchwalter, J. L. (2007). Practical Enhanced Reservoir Engineering,
PennWell Corporation, 1421 South Sheridan Road, Tulsa, Oklahoma 74112-6600, USA,
Chapter 17, Page 556.
Schmitt, G., Feinen, S. (1983). Effect of Anions and Cations on the Pit Initiation in CO2
Corrosion of Iron and Steel, Corrosion 83, NACE, Paper no. 43, Houston, Texas, USA.
Schmitt, G., Engels, D. (1988). SEM/EDX Analysis of Corrosion Products for Investigations
on Metallurgy and Solution Effects in CO2 Corrosion, Corrosion 98, NACE, Paper no. 149,
Houston, Texas, USA.
Schmitt, G., Feinen, S. (2000). Effect of Anions and Cations on the Pit Initiation in CO2
Corrosion of Iron and Steel, Corrosion 2000, Paper no. 1 (Houston, TX: NACE, 2000).
308
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Schmitt, G., Schütze, M., Hays, G. F., Burns, W., Han, E-H, Poubaix, A. and Jacobson, G.
(2009). Global Needs for Knowledge Dissemination, Research and Development in Materials
Deterioration and Corrosion Control, The World Corrosion Organization (WCO).
Smith, S. N., Joosten, M. W. (2006). Corrosion of Carbon Steel by H2S in CO2 Containing
Oilfield Environments, Corrosion 2006, Paper No. 06115, NACE International, San Diego,
California, USA.
Sobol, I. M. (1994). A Primer for the Monte Carlo Method, CRC Press Inc., 2000 Corporate
Boulevard, N. W., Boca Raton, Florida 33431, USA, Pages 1-16.
Song, F. M., Kirk, D. W., Cormack, D. E. (2005). A Comprehensive Model for predicting
CO2 Corrosion in Oil and Gas Systems, Corrosion 2005, Paper No. 05180, NACE
International, Houston, Texas, USA.
Spycher, N., Pruess, K., Ennis-King, J. (2003). CO2-H2O mixtures in the geological
sequestration of CO2. I. Assessment and calculation of mutual solubilities from 12 to 100°C
and up to 600 bar, Geochimica et Cosmochimica Acta 2003, Vol. 67, No. 16, pp. 3015-3031.
Statistics and Machine Learning Toolbox User’s Guide R2015b, (2015). The MathWorks,
Inc., 3 Apple Hill Drive Natick, MA 01760-209, Pages 4-10, 4-11, 4-15, 4-16, 5-3, 5-8, 5-9,
5-10, 5-13 and 5-17.
309
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Stewart, W. J. (2009). Probability, Markov Chains, Queues, and Simulation: The
Mathematical Basis of Performance Modeling, Princeton University Press, 6 Oxford Street,
Woodstock, Oxfordshire OX20 1TW, Chapter 5, Page 105.
Sun, W. and Nesic, S. (2007). A Mechanistic Model of H2S Corrosion of Mild Steel,
Corrosion 2007, Paper No. 07655, NACE International, Nashville, Tennessee, USA.
Sun, W., Nesic, S. and Papavinasam, S. (2006). Kinetics of iron sulfide and mixed iron
sulfide/carbonate scale precipitation in CO2/H2S corrosion, Corrosion 2006, Paper No.
06644, NACE International, Houston, Texas, USA.
Sun, Y. and Nesic, S. (2004). A Parametric Study And Modeling on Localized CO2 Corrosion
in Horizontal Wet Gas Flow, Corrosion 2004, Paper 04380, NACE International, New
Orleans, Louisiana, USA.
Suryanarayana, T. M. V., Mistry, P.,` B. (2016). Principal Component Regression for Crop
Yield Estimation, 1st Edition, SpringerBriefs in Applied Sciences and Technology, Springer
Nature Publishing, Springer Science+Business Media Singapore Plc Ltd., Springer, Chapter
2, pages 17-25.
Tan, J., Chan, K. S. (2011). Understanding Advanced Physical Inorganic Chemistry: The
Learner’s Approach, World Scientific Publishing Co. Ltd., Chapters 5 and 6, Pages 159-160,
201, 211, 221-22.
Tanupabrungsun, T., Brown, B., Nešić, S. (2013). Effect of CO2 Corrosion of Mild Steel at
Elevated Temperatures, Corrosion 2013, Paper No. 2348, NACE International, Orlando,
Florida, USA.
310
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Thomson, W. (1848). On an Absolute Thermometric Scale Founded on Carnot’s Theory of
the Motive Power of Heat, and Calculated from Regnault’s Observations, Proceedings of the
Cambridge Philosophical Society, Volume 1, pp. 66-71.
Ukil, A. (2010). Intelligent Systems and Signal Processing in Power Engineering, 2nd Edition,
Springer-Verlag Berlin Heidelberg, Chapters 2 and 3, Pages 5, 6, 13, 14, 93 and 95.
van Vuuren, D. P., den Elzen, M. G. J., Lucas, P. L., Eickhout, B., Strengers, B. J., van
Rujiven, B., Wonink, S., van Houdt, R. (2007). Stabilizing Greenhouse Gas Concentrations
at Low Levels: An Assessment of Reduction Strategies and Costs, Climatic Change, Volume
81, Issue 2, pp. 119-159.
Videm, K., Dugstad, A. (1989). Materials Performance, Volume 28, No. 3, Page 63.
Vogl, T. P., J.K. Mangis, A. K. Rigler, W. T. Zink, and D. L. Alkon. (1988). Accelerating the
Convergence of the Backpropagation Method, Biological Cybernetics, Volume 59, pp. 257-
263.
Wise, M., Calvin, K., Thomson, A., Clarke, L., Bond-Lamberty, B., Sands, R., Smith, S. J.,
Janetos, A., Edmonds, J. (2009). Implications of Limiting CO2 Concentrations for Land Use
and Energy, SCIENCE, Volume 324, pp. 1183-1186.
311
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Woollam, R. C., and Hernandez, S.E. (2006). Assessment and Comparison of CO2 Corrosion
Prediction Models, SPE International Oilfield Corrosion Symposium, May 2006, Society of
Petroleum Engineers, Aberdeen, UK.
World Energy Council, WEC (2007). Carbon Capture and Storage: A WEC “Interim Balance”,
World Energy Council, pp.17.
Yap, K.-M., Srinivasian, S. Honeywell International Inc. (2010). Key Factors In Development
of a CO2/H2S Corrosion Prediction Model for Multiphase Oil/Gas Production Systems,
Corrosion 2010, NACE International, San Antonio, Texas, USA.
Yin, Z. F., Feng, Y. R., Zhao, W. Z., Bai, Z. Q., and Lin, G. F. (2009). Effect of Temperature
on CO2 Corrosion of Carbon Steel, Journal of Surface and Interface Analysis, Volume 41,
Issue 6, pp. 517-523.
Yoon-Seok, C., Nesic, S., (2009). Corrosion Behaviour of Carbon Steel in Supercritical CO2-
Water Environments, Corrosion 2009, NACE International, Atlanta, Georgia, USA.
Zadeh L.A. (1965). Fuzzy sets, Information and Control, Vol. 8, Issue 3, pp. 338–353.
Zadeh, L., A. (1973). Outline of a new approach to the analysis of complex systems and
decision processes, IEEE Transactions on Systems, Man, and Cybernetics, Vol. 3, No. 1, pp.
28-44.
Zadeh, L. A. (1975). The Concept of a Linguistic Variable and its Application to Approximate
Reasoning – I, Information Sciences, Volume 8, Issue 3, pp. 199-249.
312
Newcastle University Muhammad Hashim Abbas
School of Marine Science and Technology (MAST) 15 December 2016
__________________________________________________________________________________
Zevenhoven, R. and Kilpinen, P. (2004). Control of Pollutants in Flue Gases and Fuel Gases,
3rd Edition, Sponsors: The Nordic Energy Research Programme, Solid Fuel Committee, Ås,
Norway and Helsinki University, Espoo, Finland. Chapter 2: Flue gases and fuel gases, Pages:
2-2, 2-3.
Zhang, Y., Gao, K., Schmitt, G. and Hausler, R. H. (2012a). Modeling Steel Corrosion under
Supercritical CO2 Conditions, Journal of Materials and Corrosion, Volume 64, Issue 6,
pp.478-485. DOI: 10.1002/maco.201106382
Zhang, Y., Pang, X., Qu, S., Li, X., Gao, K. (2012b). Discussion of the CO2 corrosion
mechanism between low partial pressure and supercritical condition, Corrosion Science:
2012, Vol. 59, pp. 186-197.
313