Proceedings of ASME Turbo Expo 2013: Turbine Technical Conference and Exposition

GT2013
June 3-7, 2013, San Antonio, Texas, USA

GT2013-96008

THE APPLICATION OF THE NORTON-BAILEY LAW FOR CREEP PREDICTION

THROUGH POWER LAW REGRESSION

D. L. May, A. P. Gordon* D. S. Segletes

University of Central Florida Siemens Energy
4000 Central Florida Blvd, Orlando FL 32816 5101 Westinghouse Blvd, Charlotte NC 28273

ABSTRACT
Accurate determination of constitutive optimize these constants typically involve
modeling constants used in high value manual curve-fitting to creep data in order to
components, especially in electric power acquire best fits across several creep curves. If
generation equipment, is vital for related design the constants found were their true values, then
activities. Parts under creep are replaced after plotting the Norton-Bailey values versus time
extensive deformation is reached, so models, would result in a near-perfect match of the data.
such as the Norton-Bailey power law, support In some situations, the constant determination is
service life prediction and repair/replacement hampered by sparse data sets at intervals of
decisions. For high fidelity calculations, strain (e.g. 0.1%, 0.5%) or at constants times (1
experimentally acquired creep data must be hr, 10 hr).
accurately regressed over a variety of Research was conducted to develop a
temperature, stress, and time combinations. If formulation to identify power law creep
these constants are not precise, then engineers constants that would result in an optimal fit with
could be potentially replacing components with creep data across test variables of both stress
lives that have been fractionally exhausted, or and temperature. The purpose of this
conversely, allowing components to operate that investigation is to develop a reliable approach to
have already been exhausted. By manipulating regressing multivariate power law type data. A
the Norton-Bailey law and utilizing bivariate background look at creep deformation, other
power-law statistical regression, a novel method creep models, and general approaches to
is introduced to precisely calculate creep constant determination are discussed next.
constants over a variety of sets of data. The Following that, the methods being investigated
limits of the approach are explored numerically are derived and tested on both physical and
and analytically. simulated data and its limitations are discussed.

INTRODUCTION CREEP DEFORMATION

Material selection is a critical stage in Constitutive models have been developed to
mechanical design engineering of structural interpolate and predict the deformation behavior
components. Perhaps the most important of materials exhibiting time-dependent, inelastic
consideration for parts subjected to long term deformation. A model commonly applied for the
use are expected service life, acceptable primary and secondary creep regimes was
deformation rate, and the environment in which developed by Bailey and Norton [1], i.e.,
the material will be used. In order to accurately
determine creep rupture life, engineers use  cr  A nt m (1)
analytical approaches to simulate the primary
and secondary creep response. An example of where A, n, and m are temperature dependent
such a model is the Norton-Bailey model, which material constants that are generally
contains three temperature dependent independent of stress. While n and m are
regression constants. The methods used to unitless, the creep strain hardening coefficient,
A, has units that are consistent with those of
*corresponding author: [email protected]

1 Copyright © 2013 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 08/06/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 time, t, and stress, . The time-differentiated (14.5ksi) to 300MPa (43.5ksi). This model also
version of this expression is often referred to as represents a case in which constants are
the time-hardening formulation of power law determined for fixed levels of stress and
creep. Solving the equation above for t and temperature.
incorporating that into the time-differentiated
formulation yields the following: 0.18

0.16

 cr 
m1
 cr  mA 
1 n 0.14
m m m (2)
0.12

(%)
0.10

Strain,
This is called the strain hardening formulation of 0.08

power law creep. In practice, the time and strain 0.06

hardening formulations are used to predict the 0.04
creep strain histories at fixed stress and Sample Creep Data 1
0.02 Sample Creep Data 2
temperature levels. Experience indicates that Sample Creep Data 3

the strain hardening formulation often produces 0.00

0 20 40 60 80 100
better agreement with the results of actual tests Time, t (hr)
under variable stress. Figure 1: Examples of primary and secondary dominant
In literature, the Norton-Bailey law has been creep.
expressed in another form [1]. The rate
formulation of the rule is given a modification of
Eq. (1), i.e. CONSTANT DETERMINATION
Creep constants embedded in the
 cr  A nt m (3)
frameworks of Eq. (1) or (3) are equivalently
determined by a number of approaches; creep
deformation history or creep deformation rate
where A′, n′, and m′ remain as temperature- data from experiments are needed for each of
dependent constants as in Eq. (1), but A′ has the respective history and rate forms. Constants
-n′ -m′-1 -n′ -m′-1
units of MPa hr % or MPa hr . The Eq. (3) are typically developed by regression curve
form of the Norton-Bailey law has been used fitting against time first or with respect to stress.
with m′ equal to zero [2]. This form also has the There are no techniques that have been
restrictions that A’ must be greater than zero. established that help to analytically identify
These models are suitable when primary and creep constants across various stress levels
secondary creep dominate the history, as seen when data is sparse.
in Fig. 1. For the history formulation of Norton-Bailey
Another notable constitutive model that has Eq. (1), ideally the constants of the formation
been applied for the prediction of creep is the can be determined with a minimum of three
theta-project approach [3]. In this approach, the creep points (termed 1, 2, and 3). The points,
complete creep curve is simulated by which take the form of ordered triples [i.e., (time,
stress, strain)] could be derived from one curve
 cr  1 1  e  t    3 e t   1
2 4
(4)
at a given creep stress or points could be taken
from separate experiments, such as the time to
a given level of strain for creep curves at
Here the  terms are regression constants that
different stress level. The latter collection is
preferred. Identical temperature is assumed. For
allow the formulation to interpolate the primary
either case, Eq. (1) is manipulated to develop a
and tertiary regimes of creep. This is a plausible
series of linear equations, i.e.,
model for primary dominant creep if  3 is set to
zero. Secondary creep is not explicitly
accounted for, but is generally predicted well if
the initial and final responses are closely curve
fit. In separate studies, Parker (1985) and
Ghosh and Chaudhuri (1994) developed
modeling constants for 2.25Cr-1Mo, a low alloy
steel, at 538C at stresses ranging from 100MPa

2 Copyright © 2013 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 08/06/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

  tr ,2      cr ,2
 log  m   log 2  n  log
3.5

 tr ,1   1   cr ,1 3.0

(5)
 tr ,3   3   cr ,3 2.5

 log  m   log  n  log

 2   cr ,2

(%)
 tr ,2  2.0

Strain,
1.5
Without loss of generality, point pairs 1-2 and 2-
3 are used; however, other combinations can be 1.0 Strain-based

used equivalently. One restriction placed on Eq.

(5) is that at least one strain level in the 0.5
Time-based
collection of points must be unique. Similarly, at Sample Copper Data
0.0
least one stress level and one time coordinate in 0 50 100 150 200 250 300
the collection must be unique. Otherwise, the
Time, t (hr)
coefficient matrix of the system becomes Figure 2: Creep data sampling methods
singular or the lines become non-intersecting.
Constants m and n are derived from this linear Since the Norton-Bailey is a power law, the
system and used to develop an approximation of equation for general power law regression fitting
the creep strain coefficient, e.g. is used:

 cr ,i y  Bx c (7)
A
 i   ti 
n m (6)

where B and c are found through the regression

equations:
where i corresponds to either point 1, 2, or 3.
This approach is repeated for each temperature
level for which data exists. A tacit assumption for k k k
this two-step approach to lead to valid material k  (ln xi ln yi )   (ln xi ) (ln yi )
properties is that candidate points must be
derived from the primary or secondary regime of c i 1
k
i 1
k
i 1
(8)
the creep curves.
k  (ln xi )  ( ln xi )
2 2

i 1 i 1

REGRESSION ANALYSIS k k

Experiments generally require measuring a  (ln yi )  c  (ln xi )

i 1 i 1
dependent and independent variable. To
simulate similar data, the relationship between Be k
(9)
the two variables needs to be precisely
calculated. Regression analysis provides a Where k is the number of points used. Clearly, y
method for which constants can be calculated can take the form of strain and x can stand for
that allow a given function to best-fit the data [4]. time or stress; however, this is only applicable
Experimental creep deformation and data for time-based data. Strain-based data could
extracted from standard creep experiments can require x as strain and y as time or stress. The
be reduced in two ways. The first technique limitation of these equations when applying them
involves keeping the time increments constant to the Norton-Bailey power law is that it only
and measuring the strain at each point across models primary and secondary creep since
multiple stresses. The second technique tertiary creep is of an exponential nature. Also,
involves measuring the time it takes to reach set xi’s cannot be identical, as the denominator in
increments of strain. Both of these techniques Eq. (8) would become zero. Regression has also
can be better understood in Fig. 2. been applied to other curve-fitting techniques
such as the Coffin-Manson equation [5].

3 Copyright © 2013 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 08/06/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 k k k
BIVARIATE REGRESSION: TIME-BASED
The Norton-Bailey equation is distinct from
k  (ln  i ln  i )   (ln  i ) (ln  i )
m
Eq. (7) because of the t term. The regression n i 1 i 1 i 1
2
(18)
k
  k
k  (ln  i ) 2    ln  i 
approach of Eqs. (8) and (9) is plausible for this
creep model since it is separable, i.e.
i 1  i 1 
  f (T ) g ( )h(t ) (10)
Determination of the creep-strain coefficient, A,
For adequate power law regression, terms in the requires rearranging the Norton-Bailey equation
Norton-Bailey equation had to be factored into the form:
together. For constant temperature, i.e. f(T) = 1,
by grouping the left most terms in the first   A( nt m ) (19)
regression equation, the following results,
where A equals B and

  ( A n )t m (11)
( nt m )1  xc (20)
Comparing this equation to the general power
law of Eq. (9) it is observed that
Inspecting Eq. (11) in this way requires using the
A n  B (12) previously found constants of n and m and then
finding the coefficient, A, by substituting the
corresponding values into Eq. (9).
t m  xc (13) k k
 (ln  i ) (1)  ln( nt m )i
Thus, substituting the corresponding values: t for i 1 i 1

x and  for y, into Eq. (8), the regression Ae k (21)

equation to determine m becomes:
Thus, the three coefficients in Eq. 1 can be
calculated when time is the dependent variable.
k k k
k  (ln ti ln  i )   (ln ti ) (ln  i )
BIVARIATE REGRESSION: STRAIN-BASED
m i 1 i 1 i 1
2 (14) This method involves solving the Norton-
k
 k 
k  (ln ti ) 2    ln ti  Bailey equation for time and reevaluating it to
match the regression equation. The
i 1  i 1  transformation of the Norton-Bailey equation
takes on the form:
To find n, the Norton-Bailey power law was 1

  
rearranged into the form: m

  ( At ) m n t n  (22)
 ( A ) 
(15)

where At m  B (16) Solving for m first requires the transformation of

Eq. (22) into the form:

and  n  xc (17)
 
 1  m1
And after substituting σ for x and ε for y, into Eq. t ( ) (23)
(8), the regression equation to determine n  1 

 ( A ) 
n m
becomes:

1

where ( A ) n m
B (24)

4 Copyright © 2013 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 08/06/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 1 1
n

and  x m c
(25) and (  m m 1
)  xc (33)

Thus, after equating the reciprocal of m with c After relating Eq. (32) to Eq. (9) and substituting
and substituting in ε for x and t for y, the in the corresponding values, A is found to be:
equation becomes: 1
k k n

k k
 (ln ti )  (1)  ln( m  m )
i
i 1 i 1
k  (ln  i )  ( ln  i )
2 2 m
m i 1 i 1 Ae k
(34)
k k k (26)
k  (ln  i ln ti )   (ln  i ) (ln ti ) Thus, the three coefficients in Eq. 1 can be
i 1 i 1 i 1
calculated when strain is the dependent
variable.
To find n, Eq. (22) must be rearranged to:

1 SIMULATED RESPONSE
n In order to validate the time-based
  m
t    ( m approach, simulated data was constructed. This
) (27) was done by assigning values to A, n, and m,
 A and then choosing predetermined time and
1 stress intervals as seen in Fig. 2. Then, strain
values were calculated using those times and
  m stresses with the constants by using the Norton-
 A   B
where (28)
Bailey power law. In essence, this was an ideal
data set for which the value of the constants can
be reversed out using the regression method. To
n accomplish this, every factor in Eq. (14), Eq.
and  m
 xc (29)
(18), and Eq. (21), i.e. the natural logarithms of
the times, stresses, and strains and the
combinations between them, was calculated. In
total there were 24 points, 12 for 150 MPa and
Equating Eq. (29) to Eq. (8), n is found to be:
12 for 200 MPa, so the summation range was
from 0 to 24. Using those equations and the
 k k k

 k  (ln  i ln ti )   (ln  i ) (ln ti ) 
calculated values, the constants were calculated
and tabulated in Table 1. Substituting these
n  m  i 1 k i 1
k
i 1
 values into the Norton-Bailey power law and
 k  (ln  i )  ( ln  i )
2 2  (30) graphing the resulting values netted creep
 i 1 i 1
 curves that matched the constructed data
perfectly, that is, with a correlation coefficient of
1.0, as expected since the data we regressed
Determination of the creep-strain coefficient, A, from was crafted from the Norton-Bailey
with this method, involves manipulating the equation, hence, Method 1 is proven to work.
equation into the form: The results are shown in Fig. 3.

1 1
n
t  ( A) (  m m m
) (31)

1

where A B m
(32)

5 Copyright © 2013 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 08/06/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 Table 2: Creep constants for second set of formulated data
0.10
Material: Simulated Fe-Alloy using strain-based method
Data Dispersion: Evenly Spaced -6 -n′ -m′-1
0.08
Data Type: Time-based A 1.85(10) MPa hr %
n 2.5
(cm/cm)

0.06 150 MPa m 0.3

200 MPa
150 MPa Model
Strain,

200 MPa Model

0.04
Substituting those constants along with the data
into the Norton-Bailey power law resulted in
0.02
curves that perfectly matched the formulated
data, again proving the validity of using
0.00 regression with the rearranged Norton-Bailey
0 100 200 300 400 500 600
equation, Eq. (22).
Time, t (hr)
Figure 3: Comparison of simulated data and regression
model for time-based conditions.
ACTUAL RESPONSE
Table 1: Creep constants for first set of formulated data The model is applied to experimental data to
found using time-based method see how strongly the models correlate to the
-10 -n′ -m′-1 actual data. The first set of data that is to be
A 1.05(10) MPa hr %
looked at is experimental creep data from
n 3.5
copper [6]. This data is included in Fig. 5 along
m 0.3 with the calculated constants used. Modeling
with these constants, the correlation coefficient
To prove a strain-based method, the was 0.996 when modeling the 40 MPa data and
constants were assigned different values, and 0.987 when modeling the 50 MPa data.
intervals of stress and strain were chosen. The The next set of data was creep data from
time values were calculated using that data and SUS316 stainless steel [7]. This data is included
Eq. (23). This data can be found in Fig. 4, below. in Fig. 6 along with the model comparison and
1.2 the constants found. When used in modeling,
Material: Simulated Fe - Alloy
Data: Strain-based
these constants resulted in a correlation
1.0 coefficient of 0.998 when modeling 245 MPa
and 0.988 when modeling 265 MPa.
0.8 Another material analyzed with the
(%)

regression model was arc-cast tungsten [8]. The

0.6
experimental data and the predicted model are
Strain,

150 MPa
160 MPa shown in Fig. 7 with the constants found using
0.4
170 MPa
180 MPa
the time-based regression method. The model
0.2
150 MPa Model
160 MPa Model
had an R-squared value of 0.9983 when
170 MPa Model
180 MPa Model
correlated against the 460 MPa data and 0.9956
0.0 when correlated against the 560 MPa data.
0 2.0x105 4.0x105 6.0x105 8.0x105 106 1.2x106

Time, t (hr) Fi
gure 4: Comparison of simulated data and regression model
for strain-based conditions.

All factors in Eq. (26), Eq. (30), and Eq. (34),

are then calculated. For this data set, there were
a total of 16 data points, meaning the
summation range was from 0 to 16. Using those
equations and the calculated values, the
constants were calculated and tabulated in
Table 2.

6 Copyright © 2013 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 08/06/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 3.0
as few points as possible. The ideal data set has
Material: Tough Pitch Copper Tube near continuously recorded data which would
Data Type: Time-based
2.5 o confer the most accurate constants. Either of
Temperature: 250 C
these approaches can be applied with much less
2.0 A: 5.26E-05 MPa-nhr-m-1% data that has some restrictions. Theoretically,
n: 2.0416
constants can be determined with these
(%)

m: 0.5188
1.5 methods with as little as 4 points. This is
Strain,

because the denominator of the equation for n

1.0 and m becomes zero when only one x-value, i.e.
time and stress, are used in method one. For
40 MPa [6]
0.5 50 MPa [6] method two, n and m become zero when one
40 MPa Model strain or stress is used. This means that, not
50 MPa Model
0.0 only must there must be at least two different
0 50 100 150 200 250
times and stresses or strains, but they must be
Time, t (hr) the same two times and stresses or strains.
Figure 5: Comparison of experimental data and modeled
creep data for tough pitch copper at 250 oC
Applying this to physical data resulted in the
following observations. Clearly, more data points
3.0
Material: Stainless Steel SUS316
result in a better fit. Determining the number of
Data Type: Time-based points to use depends on how accurate of a fit is
2.5 Temperature: 550°C needed and how much physical time is available
to collect data e.g. multiple experiments over the
2.0 course of months, or one experiment done in a
(%)

day. The caveat of this approach is that the

1.5 A: 2.559E-06 MPa-nhr-m-1% closer the time intervals used between each
Strain,

n: 1.6
m: 0.494 stress, the better the fit. In most cases, using
1.0 vastly different times between the stresses
265 MPa [7]
resulted in negative constants and therefore no
0.5 245 MPa [7] model could be constructed. Even in simulated
265 MPa Model
245 MPa Model data where the constants were simulated and
0.0 the strains were formulated with the Norton-
0 2000 4000 6000 8000 10000 12000 14000 16000 18000 20000
Bailey power law, using different times across
Time, t (hr)
Figure 6: Comparison of experimental data and modeled
each stress, even by very small amounts,
creep data for SUS316 at 550 oC resulted in less than ideal constants. Using this
method on data extracted from graphs will be
0.14
less accurate than analyzing the exact data
Material: Arc-Cast Tungsten itself.
Data Type: Time-based
0.12
Temperature: 2400 oC

CONCLUSION
0.10
(cm/cm)

0.08 A: 4.46E-14 MPa-nhr-m-1% Limited experimental data on materials used

n: 4.123
m: 0.786
in pressure turbines makes accurate creep
Strain,

0.06 prediction difficult. The focus of this investigation

is to determine a method for determining
0.04
temperature-dependent creep constants for
460 MPa [8]
0.02 560 MPa [8] modeling creep fatigue when there is limited
460 MPa Model
560 MPa Model data. Two methods were developed and
0.00 validated: one for use when the data assumes
0 5 10 15 20 25 30
that strain is the independent variable and the
Time, t (hr)
other when the data assumes that time is the
Figure 7: Comparison of experimental data and modeled
creep data for arc cast tungsten at 2400 oC independent variable. The limitations of these
techniques were also analyzed, showing that
while this method could be used on as little as
four points that if accuracy was desired that as
BIVARIATE ANALYSIS many points as possible should be used. Use of
A notable characteristic of regression these methods on available creep data showed
analysis is that it finds the best possible fit with

7 Copyright © 2013 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 08/06/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use

 accurate prediction and will be used on future
experimental data.

ACKNOWLEDGMENT
The authors are thankful for the support of
Siemens Energy in conducting this research.
David May and Ali Gordon acknowledge the
support of a CAE Fellowship in preparing this
publication.

REFERENCES
[1] Betten, J., 2008. Creep Mechanics,
Springer, Berlin, pp. 52.
[2] Brathe, L., and Josefson, L., 1979.
“Estimation of Norton-Bailey parameters
from creep rupture data,” Materials Science,
13(12): 660-664.
[3] Evans, R. W., and Wilshire, B., 1985. Creep
of metals and alloys, London: Institute of
Metals.
[4] Dally, James W. "Statistical Analysis of
Experimental Data." Springer Handbook of
Experimental Solid Mechanics. Berlin:
Springer, 2008. 259-80. Print.
[5] Daubenspeck, Brian R. "Extrapolation
Techniques for Very Low Cycle Fatigue
Behavior of a Ni-base Superalloy." Journal
of Engineering Materials and
Technology 133, 2011: n. pag. Print.
[6] Murakami, S., Sanomura, Y., and Saitoh K.,
1986. “Formulation of cross-hardening
in creep and its effect on the creep damage
process of copper,” J. Eng. Mat. Tech. 108:
167-173.
[7] Takahashi, Y.,1998. “Evaluation of Creep-
Fatigue Life Prediction Methods for Low-
Carbon/Nitrogen-Added SUS316,” J. Soc.
Mat. Sci. 47: 527-533
[8] Conway, J.B. and Flagella, P.N., 1971.
Creep-Rupture Data for the Refractory
Metals to High Temperatures, Gordon and
Breach, New York, NY pp. 34.

8 Copyright © 2013 by ASME

Downloaded From: http://proceedings.asmedigitalcollection.asme.org/ on 08/06/2015 Terms of Use: http://www.asme.org/about-asme/terms-of-use