Surface & Coatings Technology 197 (2005) 85 – 92

www.elsevier.com/locate/surfcoat

Neural computation applied to APS spray process: Porosity analysis

S. Guessasma*, C. Coddet
LERMPS-UTBM (Site de Sévenans) Rue Leupe, 90010 Belfort Cedex, France

Received 8 March 2004; accepted in revised form 14 January 2005

Available online 20 April 2005

Abstract

An artificial neural network (ANN) methodology was developed to analyse and predict porosity level in alumina – (13 wt.%) titania
coating. Background, mathematical formulation and implementation of ANN are detailed. Individual predicted effects of energetic and
injection parameters are presented. The use of ANN permitted to obtain the optimal parameter combinations for which porosity is lowered in
the considered parameter ranges.
D 2005 Elsevier B.V. All rights reserved.

Keywords: Artificial neural network; Thermal spray; Atmospheric plasma spray process; Porosity level

1. Introduction 2. Experimental procedure

Thermal spraying process is a technological process for Atmospheric plasma spraying was carried out using a
coating manufacturing implementing a wide variety of Sulzer Metco F4 gun operating at power levels up to 50 kW.
materials and processes [1,2]. The coating quality control A gas mixture of hydrogen and argon was used as a plasma
of such technique generally considers the monitoring of gas. The argon gas was also considered as a carrier gas for
the molten feedstock particle characteristics (i.e., velocity the feedstock material injection. Its flow rate was varied
and temperature) before their impingement onto the work from 2.2 to 4.2 SLPM.
piece to be covered [3]. These characteristics are intimately Al2O3 – (13 wt.%) TiO2 (Metco 130) powder was used
related to the particle semi-molten state and proved to be as a feedstock material exhibiting a size range of +15– 53
sensitive to processing parameters [4– 7]. These influence Am. Coatings produced with such feedstock material are
significantly the coating in-service properties [8] and related to industrial applications where good resistance to
microstructure features [9]. Among these features, porosity abrasive wear, sliding wear, friction and oxidation is
level is a key parameter describing the anisotropy of required. Powder injection was external to the torch and
sprayed coatings and controlling their properties [10]. In directed perpendicular to the plasma flow and parallel to
this study, the role of APS energetic and injection the torch trajectory. The powder feed rate was varied from
parameters on the evolution of porosity level of alu- 7 to 22 g min
1. The distance separating the injector tip
mina – titania coating is considered. An artificial neural from the geometric axis of the gun (i.e., the injection
network was developed to relate the process parameters to distance) was fixed at 6 mm. The injector diameter was
porosity level. 1.8 mm.
Experiments considered 5 parameters (Table 1): arc
current (I), argon primary plasma gas flow rate (A),
hydrogen secondary plasma gas flow rate (H), carrier gas
* Corresponding author. Tel.: +33 3 84 58 31 29; fax: +33 3 84 58 32 86. flow rate (CG) and powder feed rate (D m). The first three
E-mail address: [email protected] (S. Guessasma). parameters are known to significantly influence the plasma
0257-8972/$ - see front matter D 2005 Elsevier B.V. All rights reserved.
doi:10.1016/j.surfcoat.2005.01.124
86 S. Guessasma, C. Coddet / Surface & Coatings Technology 197 (2005) 85 – 92

Table 1 where a, b, c, d are parameters to be adjusted considering

APS processing parameters used in this study training procedure.
Feedstock powder Type size range Al2O3 – (13 wt.%)
TiO2 + 15 – 53 Am
For a DOE, the response is generally of the form:
Energy Nozzle type F4
Arc current A 350, 440,
530, 630, 750 y ¼ a1 x1 þ a2 x12 þ a3 x21 þ a4 x1 x22 þ a5 x1 x22 þ . . . ð2Þ
Total plasma SLPM 40, 54, 70
flow rate a 1, a 2, a 3, a 4, a 5 are parameters to be optimized.
Hydrogen ratio % 0, 13, 23, 33,
35, 43, 50
Injection Injector diameter mm 1.8 – There is no need to prior assumptions about the input/
Injection distance mm 6 output correlations, as these are obtained systematically
Carrier gas (Ar) SLPM 2.2, 2.6, 3.2, using experimental sets.
3.7, 4.2 – The number of experiments required for ANN
Powder feed rate g min
1 7, 17, 22
optimization is lower than what is required for
Spray configuration Spray distance mm 125
Spray angle Degrees 90 DOE analysis. Indeed for ANN, this number is a
polynomial function of process inputs an outputs
In bold character are shown parameter values corresponding to reference
condition. [17]:

NEXP ¼ a* N W ¼ a* ðNI þ NOÞa ð3Þ

jet properties (enthalpy, temperature, velocity, etc.) and are
mostly related to in-flight particle characteristics. Each of where NEXP is the number of experimental sets
the considered parameters was varied individually keeping, required for ANN optimisation. NW is the number of
at each time, the other parameters to a reference condition. weight parameters. NI and NO are the number of
Coatings were sprayed on button substrates (/ 25  10 input and output units, respectively.
mm). After metallographic preparation, coating character-
ization was performed on cross sections. Porosity level For DOE, the required number is an exponential function
was determined by image analysis on 10 random micro- of the parameter levels:
structures for each condition.
NEXP ¼ ðNI ÞNL ð4Þ

3. Neural computation
where NL is the number of required levels for a given
3.1. Overview process parameter.
In addition, analysis may be conducted with a partial
The artificial neural network (ANN) methodology found database in the case of ANN methodology. This is not the
a wide implementation in several distinct fields, including case when a DOE is established. None of the experiments is
materials science [11], mathematics [12], image processing to be neglected.
[13], among others [14 – 16]. Its implementation is possible Some constrains are to be considered when using ANNs.
when a decision is needed or when an explicit mathematical These concern:
description of a problem become very complex, indeed im-
possible. The methodology is based on a learning procedure – The need to a database representing a good sampling of
with the aid of experimental sets. A successful optimization the correlations I / O;
process permits to predict input parameter correlations to the – The need to explain physical the phenomena represented
problem outputs. ANN offers the following advantages: by the predicted relationships.

– Discovering complex correlations with small structures. 3.2. Artificial neural network background
Indeed, non-linearity of a given problem is represented
correctly in the ANN formalism. When comparing the The artificial neuron is a mathematical concept [18],
response of DOE (Design of Experiments) analysis to that which can be viewed as a unit of decision receiving a flux of
of an ANN analysis, the last response is the most adequate numbers from neurons and emitting a number towards other
to represent complex relationships. For example, in the neurons (Fig. 1a). Neurons are connected together with
case of one output related to 3 inputs, the non-linear numbers called weights. These measure the strength of
response is obtained with few parameters when using an neuron connections. An artificial neural network is a
ANN [17]. Such response can be written as follows: structure of neurons permitting to relate input variables to
output responses (Fig. 1b). Input values are number fluxes
y ¼ d ð1 þ exp
ðax1 þ bx12 þ cx22 ÞÞ
1 ð1Þ feeding the network structure and permitting to obtain
 S. Guessasma, C. Coddet / Surface & Coatings Technology 197 (2005) 85 – 92 87

a inputs nonlinearly. In this study, a sigmoid function was

Flux of neuron Neuron Transfer Neuron
outputs input
used for each neuron. This proved to be an adequate
function output
transformation for problems related to function approxima-
tion [20].
The network structure can be defined by several hidden
layers. One hidden layer is generally not sufficient for non-
linear problems. Increasing significantly the number of
hidden layers is not interesting as the weight parameters
increase with the increase of this parameter and this requires
significant experimental sets to optimize the weight
population.
Initially, weight parameters are not known and are tuned
Neuron in order to decrease the difference between output layer
b values ( y i ) and real case values corresponding to a
Input pattern Hidden layers Output pattern submitted input case (I i ).
The calculation starts assuming a given weight popula-
neuron i tion and an error minimisation criterion:
wj1k-1
wijk wijk ¼ 1 ð7aÞ
wj2k-1
Oj 1
2
wjNk-1 E¼ rj
Okmaxj ð7bÞ
neuron j 2
where E is the system energy which represents the quadratic
variation of the error between real output values (r j ) and
predicted value (O kmaxj ). kmax is the index of the output
layer. j is the index of the neurons in the output layer. The
Fig. 1. (a) The artificial neuron, (b) The artificial neural network structure. factor 1/2 is needed for derivation.
Weight update is performed by back-propagating the
predicted output responses if weight parameters and neuron gradient of E from the output layer to the input layer,
number are optimized. following
Parameter values were introduced in the network
structure as formatted values according to: dkmaxi ¼ Okmaxi ð1
Okmaxi Þðri
Okmaxi Þ ð8aÞ
xi
ximin
Ii ¼ ð5Þ dki ¼ Oki ð1
Oki Þdkþ1j wkþ1ji ð8bÞ
ximax
ximin
t
1
where I i is the input formatted value of parameter i. x imax wtijk ¼ wt
1
ijk þ aIki dki ð8cÞ
and x imin are the maximum and minimum possible values
associated to parameter i. It is noticed also that Eq. (5) is where d kmaxi represents the error committed at the neuron
valid also for the output pattern. I from output layer kmax. d ki is the error at neuron i
Neuron inputs and outputs are related using the following from layer k. t is the cycle number. a is the training
equations: velocity.
It is noticed that Eq. (8a) is only valid for output layer
Iki ¼ wijk Oðk
1Þj þ wi0k ð6aÞ and Eq. (8b) is valid for the other layers. Eq. (8c) represents
the standard backpropagation algorithm [21], which was
1 considered only in the first cycle of weight update. For the
Oki ¼ ð6bÞ rest of cycles, weight update obeyed to the following
1 þ e
Iki
equation:
where k, i, j are subscripts designing the layer number,
forward and backward layer, respectively. I ki is the input dtki
Dwtki ¼  Dwt
1
ki ð9Þ
value of neuron i of the k layer. w ijk is the weight value dt
1
d t
ki ki
corresponding to the strength of the connection between
neuron j of the k
1 backward layer and neuron i of the k Eq. (9) is the basic equation of the quick propagation
forward layer. w i0k is the bias associated to neuron i from algorithm [22] which considers weight update as a function
layer k [19]. O kj is the output value of neuron j of the k of the previous and current cycles.
layer. The advantage of using Eq. (9) is related to the fact that
The function relating each neuron I / O in Eq. (6b) is training velocity is adapted in order to lower the error of
called a transfer function [19] and is required to transform weight optimization as shown in Fig. 2.
88 S. Guessasma, C. Coddet / Surface & Coatings Technology 197 (2005) 85 – 92

Standard Quickpropagation
backpropagation
– Test set maximal error (Tsmax);
Residual error  
Cycle (t) Cycle (t) 1
curve
Tsmax ¼ max ðri
OðIi ; W ÞÞ2 ð14Þ
Cycle (t+1) Cycle (t+1) NPtst
i ¼ 1; NPtst

Optimal residual Weight population

error 3.4. ANN implementation
Fig. 2. Quickpropagation principle.
The use of ANN in thermal spray process was illustrated
using porosity feature of alumina –titania coating. In the
3.3. Training and testing
case of this study, input parameters were: arc current (I),
total plasma gas flow rate (H + A), hydrogen ratio (H / A),
Weight update is performed during a stage called training
carrier gas (CG) and powder feed rate (D m). Each of these
procedure. Training is performed by submitting experimen-
parameters was labelled with one neuron (Fig. 3). The
tal sets, which are organized as formatted I / O samples.
unique output parameter was the porosity level. Input and
These include original sets representing a sampling of the
output parameters were formatted according to Eq. (5) using
parameter space. These sets are increased by considering
the data of Table 2.
standard deviation associated to each experimental con-
A network structure was built considering two hidden
dition, following:
layers. A previous calculation showed that the use of 2
new ¼ l þ GauðÞTr ð10Þ hidden layers permitted to obtain a correct optimization
[20].
where new is a new case. l and r are the mean and standard
In the case of the two hidden layer system, the k index of
deviation values associated to the original case. Gau() is a
Eqs. (6a) (6b) varies from 1– 4 referring to input, first and
random generator permitting to obtain a Gaussian distribu-
second hidden layer and output layer, respectively.
tion of numbers between
1 and + 1.
Input and output layers are characterized by the follow-
In order to validate ANN predictions, a test procedure
ing equations:
is run together with the training procedure, for which
neuron weights are not updated. In this procedure, weight O1i ¼ I1i i ¼ 1; . . . ; 5 ð15Þ
configuration is tested by experimental sets different from
those considered in the training procedure. Optimization P ¼ Pmin þ O41 ðPmax
Pmin Þ ð16Þ
of a network structure takes into account convergence
where Eq. (15) is valid for input layer and Eq. (16) for
criteria for both training and test procedure. These are
output layer. P is the porosity level and O 41 its related
[20]:
formatted value, P max and P min are the maximum and
minimum possible values for porosity value.
– Average training set error (Tr);
1
Tr ¼ ðri
OðIi ; W ÞÞ2 i ¼ 1; NPtr ð11Þ Inputs Hidden layers
NPtr
I
where N Ptr is the number of training samples, O is the
predicted response associated to a given weight popula-
tion W and input pattern I i corresponding to sample i. r i H/A
is the experimental response corresponding to sample i. Output
– Training set maximal error (Trmax);
H+A P
 
1
Trmax ¼ max ðri
OðIi ; W ÞÞ2 ð12Þ
NPtr
CG
i ¼ 1; NPtr
Neuron
Dm
– Average test error (Ts);
1
Ts ¼ ðri
OðIi ; W ÞÞ2 i ¼ 1; NPtst ð13Þ weight
NPtst
Fig. 3. The optimized neural network structure to predict porosity level
where N Ptst is the number of test samples. related to APS process parameters.
 S. Guessasma, C. Coddet / Surface & Coatings Technology 197 (2005) 85 – 92 89

Table 2
Basic ANN parameters used in this study
Parameter Optimized values
Architecture Normal feed-forward MLP
Hidden layers 2
Learning rule Quick propagation
Number of neurons Inputs (I = 5) I (A) H + A (SLPM) H / A (%) CG (SLPM) D m (g min
1)
Min 300 20 0 1.5 5
Max 800 80 50 5 60
First layer (N 1): varied
Second layer (N 2): varied
Outputs (O = 1) Porosity level (%)
Min
0
Max
100
Weight population I * N1 + N1 * N2 + N2 * O

The error function defined in Eq. (7b), is rewritten – 22% for the testing of the generalization property of the
according to: optimized network (i.e., the ability to predict behaviours
without doing experiments). These concerned experi-
1
E¼ ðr
O41 Þ2 ð17Þ ments which were realised after optimisation process
2
(prediction stage).
The weight update at the output pattern Eq. (8a)
becomes: The first hidden layer of the optimized structure contains
d41 ¼ O41 ð1
O41 Þðr
O41 Þ; k¼4 ð18Þ 7 neurons and the second one contains 3 neurons. A total
cycle number of 2000 was required to obtain a stabilization
of the training and test errors. It was possible to predict
4. Results and discussion porosity level dependence on process parameters. In order to
compare predicted values to experimental ones, input
A database of 126 samples derived from 19 experimental parameters were varied individually keeping the other
sets permitted to obtain the optimized artificial neural parameters to reference value, as was the case of exper-
network structure shown in Fig. 3. The database was imental sets considered in this study. Figs. 4– 8 compare the
organized as follows: predicted and experimental porosity levels in the case of
individual process parameter increase. Experimental results
– 50% of the experimental sets were used to tune the used as training, test and prediction sets were indicated in
weight values (training stage); each case.
– 27% for the test of the network configuration (test Porosity level of deposits varied from 2% to 10%, depen-
stage); ding on spray conditions. Steeper et al. [23] suggested lower
values (4%). Chraska et al. [24] reported a porosity level
lower than 11% for a powder containing 4.6% of titania.

12 10 Test stage
Exp. 9
10 ANN
8
8 Training stage
7
P (%)

P (%)

6 6

5 Training stage Exp.

4 Test stage ANN
Prediction stage 4
2
3

0 2
300 400 500 600 700 800 20 30 40 50 60 70 80
I (A) H +A (SLPM)

Fig. 4. Porosity level related to arc current. Fig. 5. Porosity level related to total plasma gas flow rate.
90 S. Guessasma, C. Coddet / Surface & Coatings Technology 197 (2005) 85 – 92

10.5

Exp. 10.0
10 Exp.
Training stage ANN
9.5 ANN
9.0

8.5
8 Test stage

P (%)
8.0
P (%)

7.5 Test stage

7.0

6 6.5

6.0 Training stage

Prediction stage 5.5

0 10 20 30 40 50 0 10 20 30 40 50 60
-1
H/A (%) Dm (g.min )

Fig. 6. Porosity level related to hydrogen ratio. Fig. 8. Porosity level related to powder feed rate.

Porosity level was found to decrease with the increase of few number of experimental points, a linear relationship
arc current (Fig. 4). Previous studies pointed out the should be used. The predicted result is explained by the fact
increase of coating density by increasing the electric energy that the network structure learns from the whole database
[8,9,25], especially those related to alumina – titania coatings and the behaviour suggested in Fig. 5 could not be
[16,26,27]. This is explained by the improvement of interpreted as a fitting result through 3 experimental points.
flattening process by increasing in-flight particle character- The increase of the porosity level can be explained by the
istics [8,9]. decrease of in-flight particle temperature [6,8,9]. Based on a
Experimental results were fitted by the polynomial design of experiments, Steeper et al. [27] reported such
relationship: increase for the same material.
A linear relationship is suggested to relate porosity level
Pð%Þ ¼ 6:26 þ 3:28T10
5 TI 2 ð AÞ
to total plasma gas flow rate:

5:14T10
8 TI 3 ð AÞ; R2
Pð%Þ ¼ 7:3 þ 0:01Tð H þ AÞðSLPMÞ; R2 ¼ 0:42 ð20Þ
¼ 0:83 ð19Þ
The low correlation factor is attributed to the scatter of
The predicted decrease of porosity level from 10% to 2% results around mean values.
corresponds to a factor of 68% between 350 and 750 A as ANN predicted a relative increase of less than 4% when
shown in Table 3. total plasma gas flow rate shifts by 11 from 55 SLPM
Fig. 5 shows the predicted porosity evolution as a (corresponding to the highest porosity level) towards lower
function of total plasma gas flow rate. It is remarked that the of higher flow rates.
effect of this parameter is minor compared to arc current. Fig. 6 shows a linear decrease of porosity level when
ANN suggests a parabolic relationship. However, due to the increasing hydrogen ratio. This result is in good agreement
with previous studies [9,27,28]. Despite of the fact that
14
ANN represented a nonlinear system, the predicted
13 response was linear. This can be viewed as an advantage
Exp. because a linear relationship was predicted using a non-
12
ANN linear system.
11 The decrease of porosity level was significant as in the
10
case of arc current intensity. This represented a factor of
P (%)

34% for an increase of hydrogen ratio from 13% to 50%

9 (Table 3). This decrease can be attributed to the same factors
8 as those related to arc current: improvement of in-flight
particle velocity and temperature [8,9].
7
Prediction stage A linear fitting is suggested to describe the effect of
6 Test stage hydrogen ratio:
Training stage
1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 5.5 Pð%Þ ¼ 10:13
0:09Tð H=AÞð%Þ; R2 ¼ 0:93 ð21Þ
VCG (SLPM)
The effect of carrier gas flow rate is predicted to be
Fig. 7. Porosity level related to carrier gas flow rate. significantly nonlinear (Fig. 7). Few studies reported carrier
 S. Guessasma, C. Coddet / Surface & Coatings Technology 197 (2005) 85 – 92 91

Table 3
Predicted and experimental effects of process parameters on porosity level
Parameter Arc current H +A H /A CG Dm
Range 350 – 750 A a
Opt. —T1.2*Opt. SLPM 13 – 50% a
Opt. —T1.2*Opt. SLPM 7 – 22 g min
1
Effect Experimental predicted Experimental predicted Experimental predicted Experimental predicted Experimental predicted
Porosity level (%) 57%„ 4%‘„ 40%„ 6%„‘ 33%‘
69%„ 4%‘„ 34%„ 9%„‘ 19%‘
a
Opt. is optimal parameter value corresponding to the lowest or highest porosity level.

gas effect on porosity level. Most of these related this fit non-linear relationships. When choosing high polynomial
process parameter to in-flight particle characteristics [6,29]. degrees, these fit better the experimental results. However,
Depending on the range of carrier gas flow rate, in-flight they do not represent a meaningful functions and requires
particle velocity and temperature increase or decrease. An generally more that three experimental points for general-
optimal flow rate can be defined for which improvement of ization of the correlations. The use of ANN is interesting in
in-flight particle characteristics would produce a low the sense that the functions representing the relationships
porosity coating. This is suggested by the predicted curve between process parameters and porosity level are encoded
of Fig. 7. The effect of carrier gas flow rate is minor in its structure. Thus, there is no need to perform regression
compared to arc current and hydrogen ratio (Table 3). analysis for each of the correlations. In addition, it is
A nonlinear fitting is suggested to represent the effect of possible to vary all the process parameters through their
carrier gas flow rate: range to obtain the predicted optimal porosity level.
Assuming a given sampling step (d), the porosity level at
Pð%Þ ¼ 16:58
6:34TCGðSLPMÞ þ 1:10TCG2 ðSLPMÞ;
each point of the process variable space can be calculated
R2 ¼ 0:38 assuming:
Fig. 8 shows the increase of porosity level with respect to maxi
mini
powder feed rate. This result is in good agreement with di ¼ ð24Þ
Ni
previous studies [4,9,28]. The increase of porosity level with
the increase of powder feed rate can be attributed to charge where d i is the step size (resolution) of parameter i. maxi
effect which can lower particle velocity and temperature [4]. and mini are relative to maximum and minimum values of
The effect of powder feed rate is larger than carrier gas parameter i according to Table 2. N i is the number of
flow rate and the total plasma gas flow rate. This result is subdivisions. This was assumed to be the same for all the
also in good agreement with that of Steeper et al. [27] for considered parameters and was equal to 5. Thus, the
feed rates increasing from 22 to 38 g min
1. optimization of porosity level was undertaken considering
A linear relationship is suggested to relate this parameter 3125 parameter combinations. These combinations pro-
to porosity level: duced the predicted distribution shown in Fig. 9. Only 4%

of the parameter combinations produced less than 4% of
Pð%Þ ¼ 5:45 þ 0; 1TDm g min
1 ; R2 ¼ 0:93 ð23Þ porosity level. Thus, this states a reduced parameter range
When comparing ANN result to fitting procedure result, for which porosity level can be optimized.
it is noticed that when performing regression analysis using In order to determine the optimal conditions, minimum
simple relationships (small number of function parameters), and maximum values of each process parameter are
correlation factors are small which translate the difficulty to calculated. These concern the combinations for which
porosity level is less than 2%. It seems that the optimal
conditions are:
Number of parameter combinations (-)

40 0

35 0 – Arc current larger than 425 A

30 0
– Hydrogen fraction larger than 37.5%

25 0
It is noticed also that the frequency of parameter
20 0 combinations is larger for an increased arc current or
15 0 hydrogen fraction and remains stable with the increase of
total plasma gas flow rate.
10 0

50

0 5. Summary
0 5 10 15 20 25
Porosity level (%)
An artificial neural network methodology was applied in
Fig. 9. Number of parameter combinations vs. porosity level. thermal spray process as a tool of prediction. APS process
92 S. Guessasma, C. Coddet / Surface & Coatings Technology 197 (2005) 85 – 92

parameters were related to porosity level. An optimized [10] C.-J. Li, A. Ohmori, J. Therm. Spray Technol. 11 (2002) 365.
structure permitted to obtain an average scatter of less than [11] S. Guessasma, G. Montavon, C. Coddet, Proc. of MRS Fall Meeting,
vol. 2001, MRS, Boston, MA, 2001, (S8.2).
15% from experimental points. This scatter is mainly [12] J. Meade, A.A. Fernadez, Math. Comput. Model. 9 (1994) 19.
attributed to experimental standard deviations. [13] M. Isard, A. Blake, Int. J. Comput. Vis. 29 (1998) 5.
Porosity level was found to decrease with the increase of [14] J.B. Tenenbaum, W.T. Freeman, M.C. Mozer, M.I. Jordan, T. Petsche
arc current, hydrogen fraction and with the decrease of (Eds.), Proc. of Advances in Neural Processing Information Systems,
vol. 9, The MIT Press, Cambridge, MA, USA, 1996, p. 662.
powder feed rate. Parabolic relationships were predicted in
[15] H.H. Thodberg, IEEE Trans. Neural Netw. 7 (1996) 56.
the case of total plasma gas flow rate and carrier gas flow [16] J.D. Hirst, R.D. King, M.J.E. Sternberg, J. Comput.-Aided Mol. Des.
rate. The predicted control factors were arc current, hydro- 8 (1994) 405.
gen fraction and powder feed rate when varying individually [17] S. Guessasma, PhD thesis, UTBM, n-006, 2003.
process parameters. [18] W.C. Mc Culloch, W.H. Pitts, Bull. Math. Biophys. 5 (1943) 115.
When considering the possible combinations of process [19] M.M. Nelson, W.T. Illingworth, A Practical Guide to Neural Nets, 3rd
edition, Addison-Wesley Pub., New York, NY, USA, 1991.
parameters, the control factors were the arc current and [20] S. Guessasma, G. Montavon, P. Gougeon, C. Coddet, in: E.
hydrogen fraction. Lugscheider, P.A. Kammer (Eds.), Proc. 2002 International Thermal
Spray Conference and Exposition, DVS-Verlag GmbH, Düsseldorf,
Germany, 2002, p. 453.
References [21] D.E. Rumelhart, G.E. Hinton, R.J. Williams, Nature 323 (1986) 533.
[22] D. Patterson, Artificial Neural Networks, Prentice Hall, Singapore,
1996.
[1] E. Pfender, in: D.L. Houck (Ed.), Thermal Spray: Advances in [23] T.J. Steeper, A.J. Rotolico, J.E. Nerz, W.L. Riggs, D.J. Varacalle Jr.,
Coatings Technology, vol. 14, ASM International, Materials Park, OH, G.C. Wilson, in: T.F. Bernecki (Ed.), Proc. Thermal Spray Coatings:
1987.
Properties, Processes and Applications, ASM International, Materials
[2] P. Fauchais, M. Vardelle, Pure Appl. Chem. 66 (1994) 1247.
Park, OH, USA, 1991, p. 13.
[3] C. Moreau, P. Gougeon, M. Lamontagne, V. Lacasse, G. Vaudreuil, P. [24] P. Chraska, V. Brozek, B.J. Kolman, J. Ilavsky, K. Neufuss, J. Dubsky,
Cielo, in: C.C. Berndt, et al., (Eds.), Thermal Spray Industrial K. Volenik, in: C. Coddet (Ed.), Proc. Thermal Spray: Meeting the
Applications, ASM International, Materials Park, OH, 1994, p. 431.
Challenges of the 21st Century, ASM International, Materials Park,
[4] A. Refke, G. Barbezat, M. Loch, in: C.C. Berndt, et al., (Eds.), OH, USA, 1998, pp. 1299.
Thermal Spray 2001: New Surfaces for a New Millennium, ASM [25] S. Growri, G.U. Shankar, K. Narayanassamy, R. Krishnamurthy, J.
International, Materials Park, Singapore, 2001, p. 765. Mater. Process. Technol. 63 (1997) 724.
[5] A. Vaidya, G. Bancke, S. Sampath, H. Herman, in: C.C. Berndt, et al.,
[26] D.J. Varacalle, H. Herman, G.A. Bancke, W.L. Riggs, Surf. Coat.
(Eds.), Thermal Spray 2001: New Surfaces for a New Millennium, Technol. 54/55 (1992) 19.
ASM International, Materials Park, Singapore, 2001, pp. 1345. [27] T.J. Steeper, D.J. Varacalle Jr., G.C. Wilson, W.L. Riggs, A.J.
[6] J.E. Döring, R. Vassen, D. Stöver, in: E. Lugscheider, et al., (Eds.), Rotolico, J.E. Nerz, in: C.C. Berndt (Ed.), Thermal Spray: Advances
Proc. of International Thermal Spray Conference and Exposition,
in Coatings Technology, ASM International, Materials Park, OH,
DVS-Verlag GmbH, Düsseldorf, 2002, p. 440. USA, 1992, p. 415.
[7] E. Lugscheider, N. Papenfuh-Janzen, in: E. Lugscheider, et al., (Eds.), [28] S.L. Chen, P. Sitonen, P. Kettunen, in: C.C. Berndt (Ed.), Thermal
Proc. of International Thermal Spray Conference and Exposition,
Spray: Advances in Coatings Technology, ASM International,
DVS-Verlag GmbH, Düsseldorf, 2002, p. 42. Materials Park, OH, USA, 1992, p. 51.
[8] M. Prystay, P. Gougeon, C. Moreau, J. Therm. Spray Technol. 10 [29] T. Zhang, D.T. Gawne, B. Liu, Surf. Coat. Technol. 132 (2000) 233.
(2001) 67.
[9] M. Friis, C. Persson, J. Wigren, Surf. Coat. Technol. 141 (2001) 115.