Acta Materialia 52 (2004) 5157–5164

www.actamat-journals.com

Microstructure of APS alumina–titania coatings analysed using

artificial neural network
S. Guessasma *, C. Coddet
LERMPS-UTBM (Site de Sévenans), rue Leupe, 90 010 Belfort Cedex, France

Received 11 March 2004; received in revised form 29 June 2004; accepted 13 July 2004
Available online 24 August 2004

Abstract

An artificial neural network (ANN) methodology is developed to analyse and predict microstructure features of alumina–13 wt.%
titania coating. Implementation of ANN is detailed in the case of atmospheric plasma spray process parameters, which are related to
alumina, titania, porosity and unmolten particle contents.

2004 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.

Keywords: Artificial neural network; Coating; Plasma spraying; Ceramics; Microstructure

1. Introduction mechanical property stability and are directly related

to deposition efficiency. Porosity formed during the
Thermal spraying process is a technological process process is also an important microstructure feature that
for coating manufacturing implementing a wide variety must be controlled. This paper deals with microstructure
of materials and processes [1,2]. In this process, powder feature description of alumina–titania coating obtained
particles are injected in a plasma jet created by the pas- under various APS conditions. An ANN is chosen to
sage of a gas through an electric arc under high current quantify the effect of energetic process parameters on
conditions. Particles are heated and accelerated towards feature contents. This methodology proved to be an effi-
the workpiece where they flatten and solidify rapidly cient tool for experimental data analysis even when the
forming a stacking of lamellae with voids. The coating database size is small [10].
quality control of such technique generally considers
the monitoring of the molten feedstock particle charac-
teristics (i.e., velocity and temperature) before their 2. Experimental procedure
impingement onto the work piece to be covered [3].
These characteristics are intimately related to the parti- Atmospheric plasma spraying is carried out using a
cle semi-molten state and proved to be sensitive to proc- Sulzer Metco F4 gun operating at power levels up to
ess parameters [4–7]. These influence significantly the 50 kW. A gas mixture of hydrogen and argon is used
coating in-service properties [8] and microstructure fea- as a plasma gas. The argon gas is also considered as a
tures [9]. Among these features, phase contents and carrier gas for the feedstock material injection. Its flow
unmolten particle content are important indicators of rate is fixed to 3.2 SLPM.
Al2O3–(13 wt.%) TiO2 (Metco 130) powder is used
*
Corresponding author. Tel.: +33-3-84-58-32-06/31-29; fax: +33-3-
as a feedstock material exhibiting a size range of +15–
84-58-30-30/32-86. 53 lm. Powder injection is external to the torch and
E-mail address: sofi[email protected] (S. Guessasma). directed perpendicular to the plasma flow and parallel

1359-6454/$30.00
2004 Acta Materialia Inc. Published by Elsevier Ltd. All rights reserved.
doi:10.1016/j.actamat.2004.07.022
5158 S. Guessasma, C. Coddet / Acta Materialia 52 (2004) 5157–5164

Nomenclature

I arc current (A) N1 neuron number in the 1st hidden layer

H+A total plasma flow rate (SLPM) N2 neuron number in the 2nd hidden layer
H/A Hydrogen content (%) zi correct output corresponding to the ith
P porosity level (%) sample in the database
In unmolten particle content (%) O(Ii,W) predicted net output corresponding to the ith
A Alumina phase content (%) sample in the database as function of its input
T Titania phase content (%) Ii and the weight population W
ANN artificial neural network ETrn mean square training error
MLP multiLayer perceptrons ETst mean square test error
I number of network inputs NPtr size of the training database
O number of network outputs NPtst size of the test database

to the torch trajectory. The powder feed rate is fixed

to 22 g min
1. The distance separating the injector
tip from the geometric axis of the gun (i.e., the injec-
tion distance) is fixed at 6 mm. The injector diameter
is 1.8 mm.
Experiments considered three parameters (Table 1):
arc current (I), argon primary plasma gas flow rate (A)
and hydrogen secondary plasma gas flow rate (H). Each
of these parameters is varied individually keeping, at
each time, the other parameters to a reference condition.
Coatings are sprayed on button substrates (/ 25
mm · 10 mm). After metallographic preparation, coat-
ing characterization is performed on cross sections.
Four microstructure features are isolated (Fig. 1(a)).
These are porosity level, unmolten particles, alumina
and titania phases. Feature isolation was performed Fig. 1. A typical micrograph of alumina–13%wt titania coating
micrographs by applying a density slice operator. This microstructure.
can be expressed as follows:
where Cnew is the new grey level of a pixel from the im-
  age and Cold is its level before applying the operator.
 C old
C min 
C new ¼ 0 
if   6 1; ð1Þ Cmin and Cmax are the over which the pixel is attributed
C max
C min  a white level. The values of these limits depend on the

Table 1
APS processing parameters used in this study (Size range: +15–53 lm)
Feedstock powder Type Al2O3–(13 wt.%) TiO2
Energy Nozzle type F4
Arc current A 350, 530, 750
Ar + H2 flow rate Slpm 30, 50, 54
Hydrogen content % 0, 10, 35

Injection Injector diameter mm 1.8

Injection distance mm 6
Carrier gas (Ar) Slpm 3.2
Powder feed rate g min
1 22

Spray configuration Spray distance mm 125

In bold character are shown parameter values corresponding to reference condition.
 S. Guessasma, C. Coddet / Acta Materialia 52 (2004) 5157–5164 5159

selected feature. If the condition is fulfilled, the pixel is from layer k [11]. Okj is the output value of neuron j
attributed a black colour (grey level 0). of the k layer.
Eq. (1) was successfully applied for the isolation of In this calculation a two-hidden layer structure is
porosity, alumina and titania phases. Unmolten parti- considered. Thus, the maximum index layer is k = 4.
cles were identified using an ‘‘edge research’’ operator, Initially, weight parameters are not known and are
which permitted to isolate this feature by comparing tuned in order to decrease the difference between output
the produced contours to the original image. Feature layer values (yi) and real case values corresponding to a
ratios were measured on an effective area (area of submitted input case (Ii). Thus, calculation starts with
measurement) for which features touching any edge wijk ¼ 1; ð5aÞ
of the area were eliminated from the calculation,
assuming 1
af E ¼ ðri
O4i Þ2 ; i ¼ 1; . . . ; 4; ð5bÞ
F ð%Þ ¼ ; ð2Þ 2
A
ða
af Þ where E is the system energy which represents the quad-
where F is the ratio of the feature. A is the area of the ratic variation of the error between real output values
effective area. af is the area of the feature without select- (ri) and predicted value (O4i). The factor 1/2 is needed
ing parts touching the edges of the measurement area. a for derivation.
is the total area of the feature. Weight update is performed by back-propagating the
gradient of E from the output layer to the input layer,
following
3. Neural computation d4i ¼ O4i ð1
O4i Þðr
O4i Þ; k ¼ 4; i ¼ 1; . . . ; 4; ð6aÞ

3.1. Background dki ¼ Oki ð1
Oki Þdkþ1j wkþ1ji ; ð6bÞ

An artificial neural network is a structure of neurons dtki

Dwdtki ¼ Dwt
1
ki ; ð6cÞ
permitting to relate input variables to output responses dt
1
dtki
ki
[11]. Input values are number fluxes feeding the network
structure and permitting to obtain predicted output re- where Eq. (6a) is only valid for output layer and Eq. (6b)
sponses by sample learning. is valid for the other layers. t is the cycle number. a is the
In the case of this study, input parameters are: arc training velocity. Eq. (6c) is called the quick propaga-
current (I), total plasma gas flow rate (H + A) and tion [12] and considers error functions of previous and
hydrogen ratio (H/A). Each of these parameters is la- current cycles.
belled with one neuron. The output parameters are The advantage of using Eq. (6c) is related to the fact
porosity level (P), unmolten particle ratio (In), alumina that training velocity is adapted in order to lower the er-
(A) and titania (T) phase contents. ror of weight optimization.
Parameter values are introduced in the network struc-
ture as formatted values according to
3.2. Training and testing
xi
xi min
Ii ¼ ; i ¼ 1; 3; ð3Þ
xi max
xi min Weight update is performed during a stage called
where Ii is the input formatted value of parameter i. training procedure (Fig. 2(a)). Training is performed
xi max and xi min are the maximum and minimum possi- by submitting experimental sets, which are organized
ble values associated to parameter xi. These are summa- as formatted I/O samples. These included nine origi-
rized in Table 2. nal sets. These are increased by considering standard
Neuron inputs and outputs are related using the fol- deviation associated to each experimental condition,
lowing equations following
I ki ¼ wijk Oðk
1Þj þ wi0k ; ð4aÞ new ¼ l þ GauðÞ  r; ð7Þ
where new is a new case. l and r are the mean and
1
Oki ¼ ; ð4bÞ standard deviation values associated to the original case.
1 þ e
I ki Gau() is a random generator permitting to obtain a
where k, i, j are subscripts designing the layer number, Gaussian distribution of numbers between
1 and +1.
forward and backward layer, respectively. Iki is the input A case enlargement by a factor of 10 permitted to ob-
value of neuron i of the k layer. wijk is the weight value tain 90 new cases.
corresponding to the strength of the connection between In order to validate ANN predictions, a test proce-
neuron j of the k
1 backward layer and neuron i of the dure is run together with the training procedure, for
k forward layer. wi0k is the bias associated to neuron i which neuron weights are not updated (Fig. 2(a)). In this
5160 S. Guessasma, C. Coddet / Acta Materialia 52 (2004) 5157–5164

Table 2
Parameters of the ANN optimization process
Parameter Optimized values
Architecture Normal feed-forward MLP
Layer definition
Hidden layers 2
Learning rule Quick propagation
Number of neurons Inputs (I = 3) I (A) H/A (%) H + A (SLPM)
Min 300 0 20
Max 800 50 80
First layer (N1): varied
Second layer (N2): varied
Outputs (O = 4) A (%) T (%) I (%) P (%)
Min 0 0 0 0
Max 100 100 100 100
Weight population I*N1 + N1*N2 + N2*O
Input function Dot product between neuron outputs wijoj
Input preprocessing Variables between 0 and 1
x
xmin
xmax
xmin
Transfer function Sigmoid
1
1 þ expð
xÞ

Network error type (Mean square error) Training process

1
Etrn ¼ ðzi
OðI i ; W ÞÞ2 i ¼ 1; N Ptr
N Ptr
Test process
1
Etst ¼ ðzi
OðI i ; W ÞÞ2 i ¼ 1; N Ptst
N Ptst

Training and testing

Maximum iterations 1000
Tolerance 0.001 (not reached)
Sequencing training + test pass before weight update
Database size 45 training samples
45 test samples

procedure, weight configuration is tested by experimen- error is recorded. This permitted to obtain an optimal
tal sets different from those considered in the training configuration in which N1 = 10 and N2 = 5 neurons
procedure. Optimization of a network structure takes (Fig. 2(a)). For this structure, the residual average error
into account convergence criteria for both training and is 0.008 and 100% of the submitted cases are learnt cor-
test procedure as shown in Table 2 [13]. The first crite- rectly (output errors are less than 5%).
rion is related to the difference between the predicted Table 3 compares the predicted and experimental fea-
and the submitted response for both training and test ture ratios in the microstructure. It is remarked that pre-
sets. The second criterion is a fixed cycle number which dicted result approaches mean values when standard
avoid an over-training of the network structure deviation associated to the measurements is small. Aver-
(Fig. 2(a)). age relative scatter between predicted and experimental
results is less than 10%. It is the lowest in the case alu-
mina and titania contents whereas it is the highest in
4. Results and discussion the case of unmolten particle content.
The optimized ANN structure permitted to study
The ANN optimization process (training and testing) quantitatively the effect of each of the considered process
is conducted for 1000 cycles for which stabilization of parameters. Parameter values are varied at the network
the error is obtained. Neuron numbers in the first and input individually keeping the others at a reference value.
second layers are varied and the associated optimization For each case, parameter range is larger than experimen-
 S. Guessasma, C. Coddet / Acta Materialia 52 (2004) 5157–5164 5161

(a)
Start optimization

Define neuron number Network parameters

-Architechture
- Inputs –outputs parametrers
Save optimal structure - Input preporocessing
-Numberofhiddenlayers
Submit training samples - Learningrule
-Transferfunction
Update neuron number - Stoppingcriteria
1
1 Stop training
Database
-Training samples
To 0 0 -Test samples
Max. cycles

Correct weight values

1
0
Stop test Submit test samples

(b)
Weight

Neuron

H/A

In

H+A

Input layer
Output layer
nd
2 hidden layer

st
1 hidden layer

Fig. 2. (a) Flowchart of the optimization process. (b) An optimized neural network structure to predict microstructure feature relationships to APS
process parameters.

tal limits thus offering the possibility to use the generaliza- [14], momentum and thermal exchanges between powder
tion property of ANN [11] in a large parameter space. particles and plasma jet are improved [6,8,9,15,16]. The
Fig. 3 shows the predicted evolution of microstruc- consequent increase of in-flight particle velocity and
ture feature ratios with respect to arc current. temperature contribute to increase deposition efficiency
Alumina content is found to increase with the increase of alumina particles. This result is in good agreement
of arc current from 68% to 89%. As this parameter im- with that found by Ananthapadmanabhan et al. [17] in
proves plasma jet enthalpy, temperature and velocity the case of a microstructured powder and Gell et al.
5162 S. Guessasma, C. Coddet / Acta Materialia 52 (2004) 5157–5164

Table 3
Experimental and predicted microstructure feature contents
Process parameters Microstructure features (experimental predicted)
I (A) A + H (SLPM) H/A (%) A (%) T (%) In (%) P (%)
350 54 35 72 ± 7.2 10 ± 1.9 9 ± 0.7 9 ± 2.5
70 ± 1.6 11 ± 0.5 9 ± 0.0 10 ± 1.1
530 54 35 80 ± 5.3 7 ± 1.5 5 ± 0.5 8 ± 1.7
79 ± 1.3 8 ± 0.9 4 ± 0.4 9 ± 0.8
750 54 35 89 ± 3.6 5 ± 0.3 3 ± 0.8 3 ± 1.4
87 ± 1.4 6 ± 0.4 3 ± 0.3 4 ± 1.2
530 54 23 75 ± 4.1 12 ± 2.6 5 ± 0.9 8 ± 0.0
74 ± 0.4 13 ± 0.4 5 ± 0.1 8 ± 0.1
530 54 35 81 ± 5.8 7 ± 1.5 3 ± 0.3 8 ± 1.7
79 ± 2.5 8 ± 0.7 4 ± 1.1 9 ± 0.6
530 54 50 86 ± 5.5 7 ± 2.8 2 ± 1.0 6 ± 0.0
85 ± 1.5 7 ± 0.6 3 ± 0.9 6 ± 0.0
530 70 35 76 ± 3.9 13 ± 1.3 3 ± 0.3 8 ± 1.3
75 ± 1.1 12 ± 0.1 4 ± 1.0 8 ± 0.2
530 54 35 80 ± 5.3 7 ± 1.5 5 ± 0.5 8 ± 1.7
79 ± 1.3 8 ± 0.9 4 ± 0.4 9 ± 0.8
530 40 35 84 ± 1.8 6 ± 0.2 3 ± 1.0 7 ± 1.9
83 ± 1.2 6 ± 0.5 3 ± 0.9 8 ± 0.8

90
icant decrease of porosity level from 11% to 3%.
85
Depending titania powder content, different porosity
80 levels are reported, typically from 4% to 11% [21,22].
75 The decrease of porosity content is confirmed by several
70 studies on porosity relationships with plasma energy
Feature content (%)

65
A [8,9,21,23–25].
T
60 IN The decrease of unmolten particle ratio when increas-
P ing arc current can be related to the above considera-
10
tions concerning phase content variations. Jordan
et al. [26] used a parameter exhibiting the ratio of arc
power to the argon flow rate and related it to unmolten
5
particle content. They found a decrease of this feature
with the increase of this parameter. However, the use
0
300 400 500 600 700 800 of heat plasma energy instead of arc power would be a
I (A) better criterion as this represents truly the available en-
ergy for particle heating.
Fig. 3. Effect of arc current on feature percentages of alumina–titania
Fig. 4 shows the predicted ANN responses when var-
coating. A + H = 54 SLPM, H/A = 35%.
ying hydrogen content in the plasma jet. Hydrogen is
[18] for a nanostructured powder. The decrease of known to improve enthalpy of the plasma jet because
titania phase in the microstrucrure from 12% to 6% is of its relative high ionization and molecular dissociation
explained by the high evaporation of titania particles rates compared to argon gas [14,23]. This parameter im-
[17]. Despite of the fact that titania liquid state is stable proves particle temperature and velocity due to
over a larger temperature range than that of alumina, improvement of transfer coefficient [8,9]. Therefore,
titania relative low melting temperature, low latent heat the increase of alumina and decrease of titania, unmol-
of fusion and low specific heat molten state drop its ten particle and porosity ratios can be attributed to
powder particle temperature beyond evaporation the same phenomena related to arc current effect.
temperature [17]. Fig. 4 shows the quantified effect of total plasma gas
Improvement of coating density (decrease of porosity flow rate on microstructure feature contents. This proc-
content) is related to improvement of flattening process ess parameter increases plasma jet velocity and decreases
as particle surface tension decreases and particle weta- the arc diameter [14,27]. Consequently, the core region,
bility increases for a high particle temperature. The where plasma temperature is the highest, becomes smal-
reduction of particle dimensions during flattening is also ler. Particles are not well heated during their flight as
increased by the increase of particle velocity [19,20]. For they have a lower probability to reach the core region.
these considerations, the predicted result shows a signif- In addition, high particle velocity decreases their resi-
 S. Guessasma, C. Coddet / Acta Materialia 52 (2004) 5157–5164 5163

90
[9,23,25,30]. Unmolten particle content exhibited a
slight increase from 5% to 7%, which is within the exper-
80 imental standard deviation.
The discussion of energetic parameter effects on
70
microstructure attributes is regarded as pure effect of
A
Feature content (%)

60 T process parameters. This description can be completed

IN by considering the role of injection parameters. These
P
can correct significantly the predicted correlations. For
30 example, a large powder feed rate was found to de-
crease the spray efficiency [9,14,31] due principally to
20
the decrease of in-flight particle temperature [9,32,33]
10 and velocity. This has the consequence to increase
the porosity level [4,9,23]. Carrier gas flow rate is an-
0
0 10 20 30 40 50
other important factor influencing the injection veloc-
H/A (%)
ity and thus particle trajectories. Injection velocity has
to be large enough to overcome plasma jet viscosity
Fig. 4. Effect of hydrogen content on feature percentages in alumina– and not significant to permit to the particles to cross
titania coating. I = 530 A, A + H = 54 SLPM.
the plasma jet. Parabolic relationships were reported
to describe the effect of this parameter on the spray
dence time in the plasma jet and thus their temperature efficiency [31]. In addition this parameter is to be cor-
[6,8,9,16,28]. rected when varying the energetic parameters because
These arguments explain the decrease of alumina of the variation of core region and length of the
content and the increase of titania in the microstructure. plasma jet.
Jordan and coworkers [26,29] found a stability of
unmolten particle and alumina phase contents in the
case of Metco130 powder and found an increase of alu- 5. Conclusions
mina content and a decrease of unmolten particle con-
tent in the case of a nanostructured powder. Their The use of artificial neural network permitted to
result concerning Metco130 powder can be attributed quantify the relationships between APS process param-
to the fact that they used an argon plasma gas where eters and microstructure feature contents of alumina–
the available net plasma energy is lower than that when titania coating.
using argon–hydrogen plasma gas (Fig. 5). Porosity le- Alumina and titania phase contents are found to de-
vel is found to be less sensitive to plasma gas flow rate pend significantly on process parameters, especially tita-
and exhibited a slight increase at low levels. Previous nia content. Porosity and unmolten particle contents are
studies showed a more significant increase of porosity le- less sensitive to process parameters and this can be
vel with the increase of this process parameter attributed to the fact that the available plasma energy
is sufficiently high by using argon–hydrogen plasma
gas mixture.
95
Arc current and hydrogen content are the control fac-
tors of alumina–titania coating microstructure. Alumina
90
A content is found to increase with the increase of these
T
85
IN
parameters whereas the other microstructure features
80
P exhibited an opposite variation with the same parame-
ters. The role of total plasma gas flow rate opposed that
Feature content (%)

75
of arc current intensity and hydrogen content, especially
in the case of alumina and titania contents.
20

15

10 References
5
[1] Pfender E. Fundamental studies associated with plasma spray
0 process. In: Houck DL, editor. Thermal spray advances in
20 30 40 50 60 70 80 coatings technology. Materials Park (OH): ASM International;
H+A (SLPM) 1987. p. 14.
[2] Fauchais P, Vardelle M. Pure Appl Chem 1994;66:1247.
Fig. 5. Effect of total plasma gas flow rate on feature percentages in [3] Moreau C, Gougeon P, Lamontagne M, Lacasse V, Vaudreuil G,
alumina–titania coating. I = 530 A, H/A = 35%. Cielo P. On-line control of the plasma spraying process by
5164 S. Guessasma, C. Coddet / Acta Materialia 52 (2004) 5157–5164

monitoring the temperature, velocity and trajectory of in-flight [17] Ananthapadmanabhan PV, Thiyagarajan TK, Sreekumar KP,
particles. In: Berndt CC, Sampath S, editors. Thermal spray Satpute RU, Venkatramani N, Ramachandran K. Surface Coat-
industrial applications. Materials Park (OH): ASM Interna- ings Technol 2003;168:231.
tional; 1994. p. 431. [18] Gell M, Jordan EH, Sohn YH, Goberman D, Shaw L, Xiao TD.
[4] Refke A, Barbezat G, Loch M. The benefit of an on-line Surface Coatings Technol 2001;146–147:48.
diagnostic system for the optimization of plasma spray [19] Madjeski J. Int J Heat Mass Transfer 1976;19:1009.
devices and parameters. In: Berndt CC, Khor KA, Lugsh- [20] Trapaga G, Szekely J. Metall Trans B 1991;222:901.
eider EF, editors. Thermal spray 2001: new surfaces for a [21] Steeper TJ, Rotolico AJ, Nerz JE, Riggs WL, Varacalle DJ,
new millennium. Materials Park (OH): ASM International; Wilson GC. A Tagushi experimental design study of plasma
2001. p. 765. sprayed alumina–titania coatings. In: Bernecki TF, editor. Pro-
[5] Vaidya A, Bancke G, Sampath S, Herman H. Influence of process ceedings of the Thermal Spray Coatings: Properties, Processes
variables on the plasma sprayed coatings: an integrated study. In: and Applications. Materials Park (OH): ASM International;
Berndt CC, Khor KA, Lugsheider EF, editors. Thermal spray 1991. p. 13.
2001: new surfaces for a new millennium. Materials Park [22] Chraska P, Brozek V, Kolman BJ, Ilavsky J, Neufuss K, Dubsky
(OH): ASM International; 2001. p. 1345. J, et al. Porosity control of thermally sprayed ceramic deposits.
[6] Döring JE, Vassen R, Stöver D. The influence of spray In: Coddet C, editor. Proceedings of the Thermal Spray: Meeting
parameters on particle properties. In: Lugscheider E, Kammer the Challenges of the 21st Century. Materials Park (OH): ASM
PA, editors. Proceedings of International Thermal Spray International; 1998. p. 1299.
Conference and Exposition. Düsseldorf: DVS-Verlag GmbH; [23] Chen SL, Sitonen P, Kettunen P. Experimental design and
2002. p. 440. parameter optimization for plasma spraying of alumina coatings.
[7] Lugscheider E, Papenfuß-Janzen N. Simulation of the influence of In: Berndt CC, editor. Proceedings of the Thermal Spray:
spray parameters on particle properties in APS. In: Lugscheider Advances in Coatings Technology. Materials Park (OH): ASM
E, Kammer PA, editors. Proceedings of International Thermal International; 1992. p. 51.
Spray Conference and Exposition. Düsseldorf: DVS-Verlag [24] Varacalle DJ, Herman H, Bancke GA, Riggs WL. Surface
GmbH; 2002. p. 42. Coatings Technol 1992;54/55:1992.
[8] Prystay M, Gougeon P, Moreau C. J Thermal Spray Technol [25] Steeper TJ, Varacalle DJ, Wilson GC, Riggs WL, Rotolico AJ,
2001;10:67. Nerz JE. A design of experimental study of plasma sprayed
[9] Friis M, Persson C, Wigren J. Surface Coatings Technol alumina–titania coatings. In: Berndt CC, editor. Proceedings of
2001;141:115. the Thermal Spray: Advances in Coatings Technology. Materials
[10] Guessasma S, Montavon G, Coddet C. Neural Networks, Park (OH): ASM International; 1992. p. 415.
design of experiments and other optimizations methodologies to [26] Jordan EH, Gell M, Sohn YH, Goberman D, Shaw L, Jiang S,
quantify parameter dependence of atmospheric plasma spray- et al. Mater Sci Eng A 2001;301:80.
ing. In: Marple R, Moreau C, editors. Proceedings of the [27] Fauchais P, Coudert JF, Vardelle A, Vardelle M, Grimaud A,
Thermal Spray 2003: Advancing the Science and Applying the Roumilhac P. State of the art for the understanding of the physical
Technology. Materials Park (OH): ASM International; 2003. p. phenomena involved in plasma spraying at atmospheric pressure.
939. In: Houck DL, editor. Proceedings of the First National Thermal
[11] Nelson MM, Illingworth WT. A practical guide to neural nets. Spray Conference, Thermal spray: Advances in Coatings Tech-
3rd ed.. New York, NY: Addison-Wesley Publications; 1991. nology. Materials Park (OH): ASM International; 1987. p. 11.
[12] Patterson D. Artificial neural networks. Singapore: Prentice- [28] Lynn M, Lynn S, Varacalle DJ, Riggs WL. An experimental/
Hall; 1996. analytical investigation of plasma – sprayed alumina coatings. In:
[13] Guessasma S, Montavon G, Gougeon P, Coddet C. On the Berndt CC, editor. Proceedings of the Thermal Spray: Research
neural network concept to describe the thermal spray deposi- Design and Application. Materials Park (OH): ASM Interna-
tion process: correlation between in-flight particles characteris- tional; 1993. p. 315.
tics and processing parameters. In: Lugscheider E, Kammer [29] Goberman D, Sohn YH, Shaw L, Jordan E, Gel M. Acta Mater
PA, editors. Proceedings of International Thermal Spray 2002;50:1141–52.
Conference and Exposition. Düsseldorf: DVS-Verlag GmbH; [30] Steeper TJ, Riggs WL, Rotolico AJ, Nerz JE, Varacalle DJ,
2002. p. 453. Wilson GC, et al. Optimizing plasma sprayed alumina–titania
[14] Fisher IA. Int Metall Rev 1972;17:117. coatings using statistical methods. In: Berndt CC, editor.
[15] Vardelle M, Fauchais P. Pure Appl Chem 1999;71:1909. Proceedings of the Thermal Spray: Research Design and Appli-
[16] Nylén P, Wigren J, Idetjärn J, Pejryd L. On-line microstruc- cation. Materials Park (OH): ASM International; 1993. p. 37.
ture and property control of thermal sprayed abrasive coat- [31] Marsh DR, Weare NE, Walker DL. J Metals 1961;2:473.
ings. In: Berndt CC, Khor KA, Lugsheider EF, editors. [32] Fauchais P, Vardelle P, Vardelle M, Vardelle A, Coudert JF. Met
Thermal spray 2001: new surfaces for a new millen- Trans B 1989;20:263.
nium. Materials Park (OH): ASM International; 2001. p. [33] Proulx P, Mostaghimi J, Boulos MI. Int J Heat Mass Transfer
1213. 1985;28:1327.