Materials Today: Proceedings xxx (xxxx) xxx

Contents lists available at ScienceDirect

Materials Today: Proceedings

journal homepage: www.elsevier.com/locate/matpr

ANN and RSM models approach for optimization of HVOF coating

R. Shankar, K.R. Balasubramanian ⇑, S.P. Sivapirakasam, K. Ravikumar
Department of Mechanical Engineering, National Institute of Technology, Tiruchirappalli 620 015, India

a r t i c l e i n f o a b s t r a c t

Article history: High-Velocity Oxygen Fuel (HVOF) coating process is employed in many industries, not only for extend-
Received 24 November 2019 ing the life of the material, but also to improve or restore the dimensions/surface properties of the com-
Received in revised form 4 January 2020 ponent, by spraying molten or semi-molten powder materials over the surface of the component.
Accepted 8 January 2020
Porosity and Hardness are the significant properties required to assess the quality of coatings. In this
Available online xxxx
research, response surface methodology (RSM) and artificial neural network (ANN) are used for optimiz-
ing the powder composition to obtain the desired response. Based on the mixture design, twenty-five
Keywords:
HVOF coatings were performed and the data were used for training and testing the ANN. The composition
HVOF coating
Porosity
of five powders, Chromium Carbide (Cr3C2), Nickel (Ni), Chromium (Cr), Boron (B) and Silicon (Si) were
Hardness varied and performed HVOF coating to obtain minimum porosity and maximum hardness to serve in high
RSM temperature oxidation environment. Optical Microscope, X-ray diffraction, Scanning electron micro-
ANN scope, and Vickers hardness tester were used to carry out the cross-section analysis on the coated sam-
ples. Optimized powder composition was identified to achieve a dense coating. Response value obtained
by RSM and ANN models indicate that the values obtained by ‘‘ANN Model” exhibit a better prediction
over ‘‘RSM Model”.
Ó 2020 Elsevier Ltd. All rights reserved.
Selection and Peer-review under responsibility of the scientific committee of the International Mechan-
ical Engineering Congress 2019: Materials Science.

1. Introduction are designed to provide required mechanical properties and with-

stand high temperature up to 800 °C. In light of the above, for this
Super-alloys/heat-resisting alloys/high-temperature alloys are research, Inconel 718 is selected as base material [5–8]. Nickel-
based on nickel, nickel-iron, cobalt – having a composite property based powder coatings were employed along with other alloying
of excellent strength and resistance to surface degradation. Super- elements, such as chromium, boron, and silicon. Chromium
alloys are widely used in space vehicles, heat exchangers, submari- increases resistance to corrosion and oxidation at high tempera-
nes and many other industries meant for nuclear, oil & gas and mil- tures and boron powder lowers the high-melting temperature.
itary sectors, where adverse operating conditions exist [1]. In a gas Coated substrate hardness can be increased by addition of boron
turbine, its blades are subjected to high thermal load, coupled with and chromium, which combination leads to the formation of hard
frequent oxidation and high-temperature corrosion attacks [2,3]. phases. Likewise, silicon is used to increase the self-fluxing proper-
During service, the materials undergo various types of time depen- ties. Also, Cr3C2 reduces the thermal stresses generated at a high
dent degradation due to long-time exposure to adverse operating temperature. Cr3C2, Ni, Cr, B, and Si powders are selected for ‘HVOF
environments, wherein it is advisable to resort to protective coat- coating’, in order to achieve protection against abrasion, wear,
ing on the surface of materials [4]. Furthermore, surface coatings, erosion, and corrosion [9–11].
including the thermal spray coatings (TSC) on components, with Design of Experiments (DOE) is a statistical method, used to
super-alloys, greatly improve surface-property, life and perfor- design experiments within the controlled intervals to attain the
mance of the material. Amidst various TSC methods, ‘HVOF coating desired response and help in investigating the complex relation-
technique’ offers a superior quality coating, coupled with lower ship among the sprayed coatings, using a mathematical model
porosity, higher hardness and bond strength. Super-718 alloys [12–14]. The primary advantage of employing DOE is to identify
the effect of input factors on response with less number of exper-
iments [15]. Therefore, in this research, ‘mixture design-based
⇑ Corresponding author. DOE’ is used to develop experimental-matrix, by varying the
E-mail address: [email protected] (K.R. Balasubramanian). composition of powders to conduct the experiments. Data

https://doi.org/10.1016/j.matpr.2020.01.211
2214-7853/Ó 2020 Elsevier Ltd. All rights reserved.
Selection and Peer-review under responsibility of the scientific committee of the International Mechanical Engineering Congress 2019: Materials Science.

Please cite this article as: R. Shankar, K. R. Balasubramanian, S. P. Sivapirakasam et al., ANN and RSM models approach for optimization of HVOF coating,
Materials Today: Proceedings, https://doi.org/10.1016/j.matpr.2020.01.211
2 R. Shankar et al. / Materials Today: Proceedings xxx (xxxx) xxx

obtained, by performing experiments are analyzed using ‘‘Analysis (India), in the sheet form. Using a wire-cut EDM, specimens of size
of Variance” (ANOVA). With ANOVA in place, effect of input vari- 20
15
5 mm were prepared, to perform HVOF coating. In
ables on the response can be identified [16,17]. preparation for spraying by HVOF process, these specimens had
In recent years, researchers often resort to Artificial Neural Net- been polished with Silicon Carbide (SiC) papers (Grit 180) and then
work (ANN), in view of its capability to solve complex problems in grit-blasted with Alumina powders (Grit 45). Silicon (Si), Chro-
optimization. ANN is an efficient tool in predicting the process mium Carbide (Cr3C2), Boron (B), Nickel (Ni) and Chromium (Cr)
parameters and generation of outputs, for the set of demonstrative are the alloy powders selected to form the composite mixture for
inputs, which are considered during the training phase [18,19]. HVOF coating. Based on DOE, experimental design matrix is arrived
Even though the RSM is highly efficient; it also has some limita- which consists of twenty-five powder mixture composition (in
tions in the range of input parameters while obtaining the accu- percentage). The different powder mixture combinations were pre-
racy. These constraints are overcome in an ANN-based model, for pared by varying the quantity of the powders. Five powders Cr3C2
high range, large number of experiments, with a set of computa- (36–40%), Ni (31–35%), Cr (16–20%), B (2–4%) and Si (3–5%) termed
tional procedures. ANN performs like the human brain, as intu- as A, B, C, D and E respectively were varied to attain the desired
itions are made into decisions. The supremacy of ANN is the mixture. Cr3C2-Ni-Cr-B-Si powders were prepared for coating, by
ability in approximating any reasonable/irrelevant functions. This mechanical mixing for six hours.
capacity makes ANN a powerful modeling tool [20]. Since ANN HVOF Spray deposition of powders on the Inconel 718 substrate
has a wide range of usage in construction of algorithms and design had been carried out at M/s. Spray met Surface Technologies Pvt
of input-output relationships for complex and nonlinear systems, it Ltd., Bangalore, using DJH 2600 HVOF Gun with DJH 2603 nozzle.
is applied in many engineering disciplines. ANN approach com- The spray parameters employed were, Hydrogen pressure = 8 kgf/
prises input variables to predict output variables and initiated to cm2, Oxygen pressure = 12 kgf/cm2, Hydrogen flow rate = 55 l/
acquire the required outputs in designing the path and functions min, Oxygen flow rate = 30 l/min, Powder feed rate = 80 g/min,
of the complex nonlinear process by providing enough set of values Spray distance = 20 cm. During HVOF spray coating, all the coating
[21]. In ANN modeling, the set of input and output parameters are parameters were set to be fixed and samples were cooled using
trained using algorithms thus obtaining the optimum network. pressurized air jets. Coating thickness in the range of 150–
[22]. Hence, ANN tool is used for modeling the manufacturing pro- 200 lm had been obtained after performing HVOF coating. To
cesses with different input and output parameters [23]. Due to its study the cross-sectional details, each of the twenty-five coated
effectiveness in artificial intelligence modeling system, ANN specimens (20
15
5 mm) had been cut into two halves of size
Modeling is widely used in various fields of Mechanical 10
7.5
5 mm. Each cross sectioned specimen was mounted
Engineering. using a hot press. Then all specimens were mirror-polished and
Based on the literature review, it is observed that many subjected to microstructure examination. The morphology of the
researchers have done characterization of ‘‘HVOF sprayed coating” surface of the coatings was characterized by optical microscopy,
on Inconel 718 substrate by using a fixed composition of alloy XRD and SEM.
powders and not the combination of multiple powders. In this Porosity measurement on the test samples was done, using an
research, five powders are used for HVOF coating. RSM and ANN inverted optical microscope with image analysis software.
are used for optimizing the powder composition to obtain the Mounted samples were subjected to micro-hardness measure-
desired response for improving the longevity of super nickel mate- ment, using Vickers Indenter with Load  200 g, Dwell time:
rial at elevated temperatures. The main objective of this research 15 s, Angle between two faces: 136° with Indenter speed of
was to identify the best ANN model to predict the optimized pow- 60 lm/sec. Five readings were taken at different locations and
der composition to obtain dense coating with less porosity and the average value had been recorded. X-Ray Diffraction (XRD)
high hardness. Salient features of the ANN and RSM models are study was carried out, using the ‘‘Bruker AXS D 8 Diffracto-
compared and validated with the experimental results. meter” set up with CuKa radiation and Ni filter at the voltage of
35 KV and 20 mA. A scanning rate of 2°/min. over a 2h range from
2. Experimental details 10 to 70°. Fig. 1a shows the coated specimen area; thickness of the
coated area is found to be in the range of 150–200 lm. The figure
2.1. Material and method shows the complete melting of powders without any unmelted
particles. In some cases, unmelted particles with porous structure
Nickel based Inconel 718 substrate material selected for the was observed which leads to high porosity and low hardness.
study, had been obtained from M/s. Steel Mart India Ltd., Mumbai Fig. 1b details the plotting of XRD pattern for initial powder,

Fig. 1. (a) HVOF spray coated sample, (b) Coated powder X-Ray Diffraction pattern.

Please cite this article as: R. Shankar, K. R. Balasubramanian, S. P. Sivapirakasam et al., ANN and RSM models approach for optimization of HVOF coating,
Materials Today: Proceedings, https://doi.org/10.1016/j.matpr.2020.01.211
 R. Shankar et al. / Materials Today: Proceedings xxx (xxxx) xxx 3

Fig 3. A typical multi-layered ANN structure.

Fig. 2. SEM analysis of HVOF coatings in as-deposited condition.

to as ‘synaptic weights’. Every signal is passing through the con-

showing the SiCrCr3C2Ni peaks. Fig. 2 depicts SEM analysis of HVOF
nections and is multiplied by the defined weights. ANN receives
coatings in ‘as-deposited condition’. Micro-structure revealed an
the input and generates the output. Fig. 3 shows the basic structure
unevenly shaped morphology and interconnected particles, which
of the neural network, which contains five input parameters, two
apparently leading to the conclusion that the coated specimen con-
output responses and hidden layer neurons, which are changing
tained porosity at black spots, inclusions and un-melted & spheri-
based on minimum RMSE.
cally melted particles.
In this research, five input factors (Powder components namely
Cr3C2 (A), Ni (B), Cr (C), B (D), Si (E) with various ranges) and output
2.2. Response Surface Methodology (RSM) factors (Porosity (R1) & Hardness (R2)) are used to train the net-
work. Training phase is commonly used to train the network - to
Response Surface Methodology is used to deduce the relation- minimize the errors in every iteration and re-arrange the network
ship between significant input factors and one or more measured weights. ‘‘Feed Forward-Back Propagation Neural Network” (FF-
responses. RSM is an optimization technique used to analyze the BPNN) is a suitable network model for common purpose networks,
performance of a process and to determine the optimal set of input which consists of input layers, hidden layers and output layers.
process parameters. It is a statistical method used to model and Node points presented in the hidden layer does not have any direct
predict the response influenced by input factors. Experimental connection with the outside nodal points. They exhibit the compu-
design and optimization are accomplished using the ‘‘Design tational calculations and transfer the information (input layer -
Expert 7” software - to predict the optimal powder composition hidden layers - output layers), from one node to another node.
in order to obtain low porosity and high hardness levels. Based Generally, FF-BPNN is a Bi-directional network where one is in
on the ANOVA analysis quadratic Eqs. (1) and (2) were derived Back-Propagation (BP) direction, and another one is in Feed For-
and furnished below. Accordingly, for various powder mixture ward (FF) direction. FF-BPNN is having an one input layer, one out-
compositions, the micro-hardness and porosity can be predicted. put layer and multiple hidden layers.

Porosity 1:91641  Cr3 C2 þ 2:86248  Ni þ 11:37239  Cr  25:45713  B þ 30:90109

Si  0:087101  Cr3 C2  Ni  0:12361  Cr3 C2  Cr þ 0:19801  Cr3 C2  B
ð1Þ
0:33557  Cr3 C2  Si  0:22727  Ni  Cr þ 0:24328  Ni  B  0:094669  Ni
Si þ 0:26272  Cr  B  0:95375  Cr  Si  0:084494  B  Si

Hardness  180:30998  Cr3 C2  219:47947  Ni þ 60:24500  Cr þ 33:24079

B  1552:45522  Si þ 9:87881  Cr3 C2  Ni  2:14703  Cr3 C2  Cr þ 4:15877
ð2Þ
Cr3 C2  B þ 16:10689  Cr3 C2  Si þ 1:68944  Ni  Cr  0:51627  Ni  Bþ
5:70301  Ni  Si  1:10038  Cr  B þ 45:36820  Cr  Si þ 15:03134  B  Si

PN 2
2.3. Artificial Neural Network (ANN) ðActual  PredictedÞ
R2 ¼ 1  P i¼1 ð3Þ
i¼1 ðActual  Av gofactualÞ
N 2

An ANN is the best method suitable for modelling of the modern

industrial process, where no analytical approach is available. ANN 1 XN
2
models consist of a broad number of simple processing entities MSE ¼ ðActual  PredictedÞ ð4Þ
N n¼1
called ‘neurons’ that are structured in layers and are further linked
to each other, to form a network of huge connections. These con- In the BP direction, Root Mean Square Error (RMSE) is calculated
nections again consist of weights combined with them, referred by the outputs and target values of the network. BP algorithm

Please cite this article as: R. Shankar, K. R. Balasubramanian, S. P. Sivapirakasam et al., ANN and RSM models approach for optimization of HVOF coating,
Materials Today: Proceedings, https://doi.org/10.1016/j.matpr.2020.01.211
4 R. Shankar et al. / Materials Today: Proceedings xxx (xxxx) xxx

Table 1 Based on the different neural network combinations analysed, it

ANN parameters for the prediction of porosity and hardness of the HVOF coated is observed that RMSE is minimum for the network with one hid-
specimen.
den layer and 11 neurons in it. In this network (5-11-2), the ‘log-
No. of neurons in the input layer 5 sigmoid’ transfer function is used in the hidden layer, and ‘Purelin’
No. of hidden layers 1 transfer function is used in the output layer, to predict the Porosity
No. of neurons in the hidden layers 11
Number of neurons in the output layer 2
and Hardness of the coated specimen. The transfer function (log-
Total data samples 25 sigmoid) used in the hidden layer is given in Eq. (5).
Training data samples 75%
Testing data samples 25% 1
Hidden & Output layer activation functions Logsig & purelin FðxÞ ¼ ð5Þ
Performance goal fixed 109 1þex
Number of Epochs 1000
where x is the weighted sum of the input. General form of a transfer
function used in the output layer (Purelin), is given in Eq. (6)
modifies the weights on its own to attain the minimum RMSE. Lin-
ear function is used in the output layer (for responses). Neural Net-
FðxÞ ¼ x ð6Þ
work in the Feed Forward direction is trained using the input data In ANN models, optimum neurons for the hidden layer are cal-
and output data. The ANN model performance is evaluated based culated based on the Eq. (7)
on RMSE and Co-efficient of Determination (R2). Coefficient of
determination and RMSE can be calculated by given Eqs. (3) and I þ O pffiffiffi
NN ¼ þ T ð7Þ
(4). FF-BPNN is the most suitable method to predict the output 2
responses because of it is simple in nature and widely applicable where, NN represents the number of neurons in the hidden layer, I
to predict all process parameters. Multi-layer FF-BPNN demon- represents number of input parameters, O represents number of
strates the network architecture and learning algorithm. Experi- output parameters, T takes the value of a number of training data
mental data is separated into two - one set for training the used in the ANN model. In order to identify the best network, fol-
network and other one for testing the neural network. lowing two parameters are required the number of neurons in the
Training data samples play a significant role to estimate the hidden layer and RMSE performance indicator. Optimum neurons
performance of the neural network. In Mat lab, various combina- to be used in the hidden layer can be predicted based on the Eq.
tions of networks had been analysed based on the parameters (7). To identify the best network, neurons in the hidden layer may
listed in Table 1 and by altering the number of hidden layers and be varied by ±5 from the optimum neurons in the hidden layer.
the number of neurons in it. Finally, a network (5-11-2) as shown Figs. 4 and 5 represent the plots between percentage of error for
in Fig. 3, is selected which consist input layer with 5 neurons, hid- training and testing data for porosity and hardness. Relative error
den layer with 11 neurons and output layer with 2 neurons. of the ANN model for both training and testing shows a very low

Fig. 4. Error percentage of trained data (a) porosity; (b) Hardness.

Fig. 5. Error percentage of tested data (a) porosity; (b) Hardness.

Please cite this article as: R. Shankar, K. R. Balasubramanian, S. P. Sivapirakasam et al., ANN and RSM models approach for optimization of HVOF coating,
Materials Today: Proceedings, https://doi.org/10.1016/j.matpr.2020.01.211
 R. Shankar et al. / Materials Today: Proceedings xxx (xxxx) xxx 5

value. It is seen that the error obtained from the ANN model and it is According to RMSE, R2 performance indications, ANN models for
not crossing 10%. It indicates that the current neural network model porosity and Hardness values were found to perform with low
is the best to suit for predicting the porosity and hardness of the error and high accuracy. With these models, Porosity and Hardness
HVOF coated sample. Output of each neuron is a result of a can be predicted with high accuracy without any additional exper-
weighted set of inputs. Sum of weighted inputs formed by neurons imental work for the selected powder. ‘‘Tested Predicted Values”
is given as: and ‘‘Actual Measured Values” are compared and shown in Figs. 6
and 7 for ANN model and RSM model respectively, Co-efficient of
X
N Determination (R2) obtained during testing the network (5-11-2)
X¼ ðWij ai þ bjÞ ð8Þ for Prediction of Porosity  0.992167 and Prediction of Hardness
n¼1
 0.962556 the following figures are found to be satisfactory.

Coating performance is calculated using the equations obtained

from the NN. In the equation, Xi and Fi are the weighted sums of 3. Results and discussions
input and log-sigmoid activation function, respectively for every
neuron of hidden layer and these values are obtained from the Effect of coated power composition with various propositions
low RMSE trained the neural network. Trained network is evalu- had been experimentally investigated, and optimized powder com-
ated with a set of training data (eighteen input data) and found position was obtained, with the help of RSM model. Moreover,
that the network gives the best result. From the network, weights same input parameters were used to predict the output parame-
and bias are obtained, which are used to make the mathematical ters, like porosity and hardness. XRD and SEM analysis had been
formulations. As eleven neurons are used in the hidden layer, ele- performed to identify which composition of coating powder and
ven equations are formulated for the hidden layer. Mathematical micro-porous value of HVOF coating. Fig. 1a shows the image of
Eqs. (9) and (10) derived from the output layer for porosity and coated Inconel 718 substrate and measured thickness of coating,
hardness is furnished below. which is in the range of 150–200 lm. XRD results had been anal-
ysed to observe the current powder components. This analysis
was carried out with X’Pert High Score Plus software. From
Porosity 3:3064 þ 0:050012F1  0:44665F2  1:8488F3  0:73087F4 Fig. 1b, it is confirmed that Si, Cr, Cr3C2, Ni are present in the cur-
0:01933F5 þ 1:035F6  1:2324F7  0:81074F8  1:2183F9 rent coated HVOF sample. SEM analysis of coated specimens as
0:72396F10  0:51942F11 shown in Fig. 2. It shows that micro-porous surface structures
ð9Þ are present in the HVOF coated sample.
The porosity and hardness obtained from 25 HVOF coating trails
Hardness 1:6819F1  0:77661F2 þ 2:9523F3  1:2471F4 þ 0:52672F5 were used to model the RSM. Initially RSM was developed and
0:93162F6 þ 0:90204F7 þ 0:62809F8 þ 0:73071F9  1:1751F10 regression equations were generated for deriving the porosity
0:87595F11  1:0991 and hardness values of the coated sample. In the ANN modelling,
18 trial data were used for training the network and 7 data were
ð10Þ
used for testing the neural network. Percentage error of ANN

Fig. 6. Correlation coefficient values with ANN Model (a) porosity; (b) Hardness.

Fig. 7. Correlation coefficient values with RSM Model (a) porosity; (b) Hardness.

Please cite this article as: R. Shankar, K. R. Balasubramanian, S. P. Sivapirakasam et al., ANN and RSM models approach for optimization of HVOF coating,
Materials Today: Proceedings, https://doi.org/10.1016/j.matpr.2020.01.211
6 R. Shankar et al. / Materials Today: Proceedings xxx (xxxx) xxx

trained data and tested data is comparatively low (<10%), as shown CRediT authorship contribution statement
in Figs. 4 and 5. Also, in the ANN predicted samples, minimum
error deviation had been noted. Network with minimum RMSE R. Shankar: Methodology, Software, Investigation, Writing -
(5-11-2 network) was identified as the best network for predicting review & editing. K.R. Balasubramanian: Conceptualization, Data
the porosity and hardness. curation, Supervision, Writing - original draft. S.P. Sivapirakasam:
From the RMSE graph it is observed that ANN model is best sui- Supervision. K. Ravikumar: Software, Validation, Writing - review
ted to predict the porosity and hardness of the current coating & editing, Visualization.
method. ANN is more precise to predict the coating powder com-
position than the RSM method. From Fig. 6a and b, it is observed Declaration of Competing Interest
that prediction of porosity and hardness by the ANN model closely
agrees with the experimental values. The authors declare that they have no known competing finan-
It is also found that the R2 value of porosity and hardness for cial interests or personal relationships that could have appeared
testing data of ANN model as 0.9921 and 0.9625 respectively to influence the work reported in this paper.
whereas the correlation coefficient R2 value of porosity and hard-
ness (Fig. 7a and b) for the RSM model is 0.9746 and 0.9577 respec- References
tively, Thus, the ANN model is quite effective in predicting the
porosity and hardness when compared to the RSM model. From [1] Enes Akca, Ali Gursel, Periodicals Eng. Natural Sci. 3 (1) (2015).
[2] T.S. Sidhu, S. Prakash, R.D. Agrawal, Surf. Coat. Technol. 201 (2006) 273–328.
the ANN modeling, it is observed that 5-11-2 network is having
[3] N.K. Mishra, S.B. Mishra, R. Kumar, Surf. Coat. Technol. 260 (2014) 23–27.
the minimum RMSE value. This network is found to possess [4] H. Singh, S. Prakash, D. Puri, Mater. Sci. Eng., A 444 (2007) 242–250.
weights and bios from the input layer to the hidden layer and also [5] Subhash Kamal, R. Jayaganthan, S. Prakash, Sanjay Kumar, J. Alloys Compounds
from hidden layer to output layer. Using this data, porosity and 463 (2008) 358–372.
[6] Harpreet Singh, D. Puri, S. Prakash, Surf. Coat. Technol. 192 (2005) 27–38.
hardness equations could be developed to predict further [7] Atikur Rahman, R. Vipin Chawla, Ramesh Chandra Jayaganthan, R. Ambardar,
experimentation. Mater. Chem. Phys. 138 (2013) 49–62.
[8] K. Premkumar, K.R. Balasubramanian, Eng. Failure Anal. 97 (2019) (2019) 635–
644.
4. Conclusion [9] T.S. Sidhu, R.D. Agrawal, S. Prakash, Surf. Coat. Technol. 198 (2005) 441–446.
[10] J.M. Miguel, J.M. Guilemany, S. Vizcaino, Tribol. Int. 36 (2003) 181–187.
[11] Sukhpal Singh Chatha, Hazoor S. Sidhu, Buta S. Sidhu, Surf. Coat. Technol. 206
The present study focused on the optimization of powder mix-
(2012) 3839–3850.
ture for HVOF coating on Inconel 718 material using design of [12] P. Dinesh Babu, G. Buvanashekaran, K.R. Balasubramanian, Int J Adv Manuf
experiments. RSM and ANN are used to determine the optimal Technol 67 (2013) 1883–1897.
powder mixture composition. This research work leads to the fol- [13] Nanda Naik Korra, M. Vasudevan, K.R. Balasubramanian, Proc. Inst. Mech. Eng.
L: J. Mater.: Des. Appl. 230 (2016) 837–846.
lowing major conclusions. [14] Tanya J. Levingstone, Malika Ardhaoui, Khaled Benyounis, Lisa Looney, Joseph
T. Stokes, Surface & Coatings Technology, doi: 10.1016/j.surfcoat.2015.10.044.
 Input and output parameters are analyzed using neural net- [15] Seyed Hadi Seyedin, Mehdi Ardjmand, Ali Akbar Safekordi, Shahram Raygan,
Ehsan Zhalehrajabi, Nejat Rahmanian, Surf. Coat. Technol. 334 (2018) 43–49.
works model (NN) with different network configurations. ANN [16] P. Kardar, M. Ebrahimi, S. Bastani, M. Jalili, Prog. Org. Coat. 64 (2009) 74–80.
modelling is done to identify the best model for the prediction [17] R. Shankar, K.R. Balasubramanian, S.P. Sivapirakasam, Materials Science
of response. Effects of input parameters on coating hardness Forum, Vol. 969, pp 48-54.
[18] Pragya Shandilya, P.K. Jain, N.K. Jain, Proc. Eng. 64 (2013) 767–774.
and porosity are investigated. A suitable neural network model [19] M. Muthu Krishnan, J. Maniraj, R. Deepak, K. Anganan, Mater. Today:. Proc. 5
(5-11-2) with best R2 value is arrived at. Expressions for poros- (2018) 716–723.
ity and hardness of coated samples are extracted from the neu- [20] Kiran A Patel, Pragnesh K Brahmbhatt, Proc. Technol. 23 (2016) 391–397.
[21] G.R. Kannan, K.R. Balasubramanian, R. Anand, Int. J. Automot. Technol. 14 (4)
ral network model. (2013) 507–519.
 It is observed that the porosity is in the range of 1.07–4.2% and [22] K.R. Balasubramanian, G. Buvanashekaran, K. Sankaranarayanasamy, CIRP J.
micro-hardness of the coatings in the range of 510–671 HV. Manuf. Sci. Technol. 3 (2010) 80–84.
[23] K.R. Balasubramanian, K. Sankaranarayanasamy, G.N. Buvanashekaran,
 It was also observed that Chromium, Nickel, and Chromium
Chapter 6, in: Laser in Manufacturing, ISTE-Wiley, U.K., 2012, pp. 203–246.
Carbide are the major factors playing a crucial role to determine
the micro-hardness and porosity followed by Boron and Silicon.
 Co-efficient of determination values (R2) of ANN models are
found to be greater than RSM and justifies the acceptability of
its prediction capability.

Please cite this article as: R. Shankar, K. R. Balasubramanian, S. P. Sivapirakasam et al., ANN and RSM models approach for optimization of HVOF coating,
Materials Today: Proceedings, https://doi.org/10.1016/j.matpr.2020.01.211