Study on Wear Analysis of Ni-20Al 2 O 3 HVOF Micron Layers Using Arti�cial Neural Network Technique

The paper presents the study of the erosion performance of stainless steel SS316L grade. The micron layers of Ni-20Al 2 O 3 powder were sprayed on the surface SS316L by utilizing a high-velocity oxygen–fuel (HVOF) thermal spraying process. Wear experiments were performed in a slurry pot tester (Ducom TR-41) at different speeds varying from 600 to 1500 rpm. During the experiments, the sand was used to erode the coated surface of SS316L. Multi-sized slurry having a solid concentration lies in the range of 30-60% (by weight) was prepared to perform the wear experiments. A neural network technique was implemented to predict the erosion data by training 75% of the total data. Results show that the microhardness of bare and Ni-20Al 2 O 3 HVOF coated SS316L was tested as 196±21 and 393.32±17 VHN respectively. The current ANN model yielded an R-value that was 0.98871 for training, 0.97209 for validation, and 0.96264 for testing accordingly. It was determined that the total R-value of the model was 0.98153. The ANN’s performance evaluated based on the mean squared error (MSE) was determined as 0.0176, 0.00265, and 0.00334 respectively for the training, validation, and testing.


Introduction
Nowadays, erosion caused by the ow of solid particles is a common issue in hydraulic equipment and machines, leading to signi cant maintenance and repair costs [1,2].In industries, minimizing the impact of erosion wear in hydraulic equipment requires a comprehensive approach.For example, using abrasionresistant materials, such as high-chromium alloys, can help to reduce the rate of wear.Additionally, the use of coatings can provide additional protection against abrasion and corrosion [3][4][5][6][7].For tribology applications, the Ni-based coatings are generally deposited utilizing a high-velocity oxy-fuel (HVOF) or Plasma spraying (PS).HVOF spraying has some speci c bene ts over PS coatings.The adherence strength of HVOF coatings is better than PS coatings.The porosity of PS coatings is about 10-20% whereas the HVOF coatings are less porous i.e. <1% porosity.Therefore, HVOF is much preferable for erosion applications.HVOF involves spraying a WC-based material in the form of powder onto the surface to be protected, resulting in a dense and uniform coating.Ni-based coatings are known for their high hardness, which makes them effective in resisting abrasive wear and erosion.
In this decade, machine learning (ML) has become an emerging technology in every engineering eld.ML can perform classi cation, curve tting, predictive, statistical, and image analyses.The ML-predicted results are effective in saving time as well as cost.In addition to statistical metrics, ML also provides the visualization of the predicted results alongside the actual data which was done using scatter plots.By comparing the predicted results with the actual data in a visual manner, it is often possible to identify patterns or trends that may not be apparent from the statistical metrics alone.Singh [8] performed an ANN-based study on erosion prediction of HVOF-deposited Ni-20Cr 2 O 3 , Ni-20Al 2 O 3 , and Al-20TiO 2 coatings on SS316L using a pot-type erosion tester.Input parameters used in this study were time duration, mass ux, and velocity.Results showed that the ANN model predicted erosion with R = 0.99525 and 0.99595 for the bottom and y ash respectively.Becker et al. [9] investigated the cavitation and slurry erosion resistance of Cr 3 C 23 7WC18M and WC20Cr 3 C 2 7Ni coatings sprayed by the HVOF method using the ANN technique.They employed input parameters namely coating material, combustible type, combustible ow, passes, standoff distance, stoichiometric and feed ratios, powder particle speed, and temperature.By using data from these parameters, they predicted outputs like microhardness, fracture toughness, coating thickness, porosity, cavitation rate, cavitation mass loss, slurry erosion rate, and erosive mass loss.ANN results showed relative errors from − 0.041% to + 0.14% of original values with R = 0.99.Szala et al. [10] developed a simpli ed ANN model that correlates atmospheric PS process parameters with the functional properties of Al 2 O 3 -13TiO 2 coatings.They predicted the hardness, porosity, and mean erosion depth with MSE of 98.4, 84.6, and 82.6.
Literature indicates that the high hardness of Ni-based HVOF coatings is also a key factor for the enhancement of erosion resistance.The high microhardness of the coating helps to resist erosion and prevents the wear of hydraulic components the underlying component.The literature gap suggests that there is a lack of neural computing-based studies on erosion.Neural computing can be effective in terms of saving erosion monitoring costs.Therefore, the present study has been performed to test the characteristics and erosion wear of Ni-20Al 2 O 3 coating sprayed by HVOF thermal spray.A novel arti cial neural network (ANN) technique was implemented to calculate the erosion of HVOF Ni-20Al 2 O 3 coating.

Base Material and Coating
Thermal spraying Ni-20Al 2 O 3 powder was used to coat the base material, which was SS316L.It was possible to acquire granular forms of both the Ni (60 ± 15 µm) and the Al 2 O 3 (60 ± 15 µm). Figure 1 displays pictures of Ni and Al 2 O 3 powders taken with an SEM and an EDS.Blending was done to produce a wide variety of different covering particle compositions.As shown in Fig. 1, the appearance of the combined Ni-20Al 2 O 3 powder consisted of blocky-shaped particles.The high-velocity oxy-fuel (HVOF) application was carried out with the process speci cations derived from the previously published research [4].

Materials Characterization
Microstructure as well as phase characterization of erodent and coatings are very important parts of erosion wear assessment.The microstructure analysis provides the shape and morphological details which enables an analysis of the erodent and coating's physical characteristics.The microstructure and molecular composition of HVOF surfaces were analyzed with JEOL, JSM 6510LV apparatus (manufactured in the Netherlands) using SEM and EDS techniques, respectively.To get the sample ready for this kind of analysis, non-conducting granular samples like erodent need to be desiccated in an oven to get rid of any moisture, and then they need to be sputter-coated with gold for SEM-EDS analysis.The determination of Vickers microhardness number (VHN) was done on a digital microhardness tester, which utilized an indenter load of 1000 g for a 20-second dwell time.

Experimentation
Workpieces made of SS316L were subjected to erosion experiments using the measurements shown in Fig. 2.An apparatus known as an erosion pot tester was utilized to carry out the erosion tests.The various components of the pot tester are depicted in Fig. 2. The pot tester is made up of two pots: the inner pot is known as the slurry pot, and the outer pot is known as a water receptacle that serves the purpose of chilling the specimen.The experiments are carried out to determine the amount of mass loss in units of gm -2 min -1 that occurs despite differences in the characteristics of the sediment, the velocity of the particulates, and the contact angle.The sand was selected as the erodent for the experiments on erosion.

Arti cial Neural Network
NNTOOL, which is part of the MATLAB 2018a software program, was used to perform neural processing.To perform neural computing, the incoming data was run through demonstrations of neural networks on an Intel® machine that featured a core TM i5 2.67 GHz processor and 4 gigabytes of random access memory (RAM).To successfully implement an arti cial neural network (ANN) in MATLAB, one must have a sound comprehension of the problem at hand, a data collection that has been thoroughly prepared, and a meticulous selection of the architecture of the neural network as well as the training parameters.With these components in place, the MATLAB environment offers a robust set of tools for the construction and evaluation of neural networks.The data collection consisting of 52 values was utilized in this investigation.The data were structured in such a manner that 8 points would be used for the con rmation of the neural network (NN), and the NN itself would be evaluated on 8 points.As shown in Fig. 3, the design of a neural network consists of the number of levels, and neurons in each layer, and the sort of modi cation and learning function.The con guration of the network was created by taking into account several characteristics as well as the 12 neurons in the concealed layer that are shown in Fig. 4.
The data representing erosion rates are included in the result layer.The feed-forward transmission phase comes to a close when it reaches the output layer (Y), which is expressed as: The values f, Z q , and W iq are assigned to the activation function, the data collection, and the hiddenoutput layer weights, correspondingly, in Equation 1.To lower the error value, which can be represented as [11], a back-propagation learning method was implemented.
Term ΔW ij is the weight shift in the link relating the i th and j th neurons of two adjacent layers, where δ oi and δ hq stand for the native and fractional errors.W ij is denoted by the notation.The indicator for the learning rate is denoted by the letter , and the quantity of the supplied objective is denoted by the letter d.In this context, the term refers to the error gradient in relation to the weight W ij .
A total of 24 different input sources were used to train the network.The Levenberg-Marquardt [12] training procedure was utilized for this study's data collection.The NN was trained for a hundred epochs, but Fig. 5 demonstrates that its performance was at its peak after only six epochs.MATLAB's built-in algorithms, such as 'trainlm,' were utilized to carry out the training.The effectiveness of the neural network was assessed with the help of a calibration collection of data comprising 15% of the total.

Assessment of coating properties
Microhardness is an important parameter that contributes majorly to erosion wear.The microhardness of the base material (SS316L) was tested as 196 ± 21 VHN whereas the microhardness of Ni-20Al 2 O 3 HVOF coating was found as 393.32 ± 17 VHN.Using the water immersion method [13], the porosity of Ni-20Al 2 O 3 HVOF coating was found as 1.13%.The XRD patterns, which can be seen in Fig. 6, provide evidence of the crystalline nature of the HVOF surfaces after they have been sprayed.The XRD structure of Ni-20Al 2 O 3 coating reveals the presence of four different phases: α-Al 2 O 3 , γ-Al 2 O 3 , Ni, and NiO.In the earlier investigation [14], the same structures were observed to be found in Ni-30Al 2 O 3 plasma spray coating.

Variation of rotational speed
The velocity was replaced by the different rotational speeds (N).At a solid concentration (C) of 30% (by weight), impact angle (α) of 0, and a time (T) of 180 minutes, erosion tests were carried out to see how the rotating speed would affect the results.As illustrated in Fig. 7, the erosion rate of the sample increases at a rate that is not linear when the tester's shaft spins faster.At a rotating speed of 1500 rpm, the greatest amount of material deterioration was observed resulting in a maximum erosion rate.This occurs because the target material exerts high impact energy by striking sand particles moving at higher kinetic energy.Higher impact energy was converted into the material deformation or cutting rate.

Variation of concentration
The rate of erosion that Ni-20Al 2 O 3 HVOF coating experiences as a function of the concentration of solids present are seen in Fig. 8.In Fig. 8, we have a visual representation of the association between the amount of solids and the proportional amount of erosion.It was revealed that the rate of erosion increases was not linear.The erosion testing was carried out at an impact angle (α) of 0, time (T) of 180 min, and a rotational speed (N) of 1500 rpm as shown in Fig. 8.The general inference that can be derived from the graphical representation of the data is that the rate of erosion increase was larger at the start of the testing.It was revealed that the rate of erosion slows down as the time of the process increases.This happened as a result of the erosion process changing the properties of the particle [15][16][17][18][19][20][21][22][23].Additionally, a change in the outward appearance of the particles is a contributor to the slower rate of erosion that occurs over time [24].According to Fig. 9, the greatest erosion rate of SS316L occurred at an impact angle of α = 30, whereas for Ni-20Al 2 O 3 HVOF coating, this angle ranged from 30̊ to 45 .Hence, the erosion behavior of SS316L

Variation of impact angle
re ects the ductile nature, and Ni-20Al 2 O 3 HVOF coating brings the semi-ductile orientation to the erosion behavior.

Variation of particle diameter
The studies on erosion wear were carried out for several particle size distributions (75 m, 75-106 m, 106-150 m, and > 150 m) to determine the in uence of the mean diameter of particles.The mean diameter of sand particles was evaluated as 45.6, 93.4,121.7, and 257.8 µm respectively for size distributions of < 75 µm, 75-106 µm, 106-150 µm and > 150 µm.Slurry erosion experiments were performed at C = 60% (by weight), T = 180 min, α = 0° and N = 1500 rpm.According to Fig. 10, when the weighted mean diameter increases, the amount of material that is eroded by the slurry also increases.The erosion of Ni-20Al 2 O 3 was risen by 2.52 times with an increase in the mean diameter of sand particles from 45.6 to 257.8 µm (as observed in Fig. 10).

Neural computing results
In the current work, an ANN was used to make predictions about the erosion rate of a pump and coating material under varying conditions of process parameters.An ANN was trained and validated on the provided data set and was applied to the task of determining how accurate its predictions were.The most prevalent methods that were used in this research were statistical measures like mean squared error (MSE) and correlation coe cient (R) to assess the gap that existed between the values that were predicted and those that were observed.The current ANN model yielded an R-value that was 0.98871 for training, 0.97209 for validation, and 0.96264 for testing accordingly.It was determined that the total Rvalue of the model was 0.98153.
Figure 11 displays the R-value charts for the currently implemented NN model.Figure 11 is a comparison of the actual erosion rate to the one that was anticipated.Each point in the diagram represents a unique event.The diagonal line illustrates the best-case scenario, in which the rate of erosion that was predicted was identical to the rate of erosion that occurred.The effectiveness of the ANN is said to improve in direct proportion to the points' proximity to the diagonal line.The performance was evaluated based on the mean squared error (MSE), which was determined to be 0.0176, 0.00265, and 0.00334 for the training, validation, and testing phases, respectively.This is the sum of the squared errors of all of the forecasts taken together.A histogram, which can be seen in Fig. 12 illustrates the inaccuracy that the ANN model generates.

Erosion mechanisms
The microstructure of erosion mechanisms occurred on the surface of Ni-20Al 2 O 3 coating by sand slurry at α = 0, N = 1500 rpm, and C = 60% (by weight) is shown in Fig. 13.Ploughing, intact patches, craters, and cracks were all seen in the degraded Ni-20Al 2 O 3 coating.It was reported that craters could be seen on the Ni-20Al 2 O 3 coating at different locations.Intact patches on the coating's surface developed simultaneously as a result of the fracturing of brittle regions.In addition to this, fractures were seen on the surface, which is an indication of the brittle behavior of the reinforcement i.e.Al 2 O 3 .Since nickel is the ductile phase in the coating, the deformation is the most evident fundamental mechanism that occurred [25].Research has shown that the high toughness and microhardness of alumina help to reduce the amount of erosion and wear that occurs [10].

Conclusions
The coating characterization and erosion results signify the following conclusions: • VHN respectively.
• The porosity of Ni-20Al 2 O 3 HVOF coating was found as 1.13% using the water immersion method.
• Experimental outcomes indicate that as the mean diameter increases the erosion rate also increases.
• The maximum erosion rate of SS316L occurred at an impact angle of α = 30, whereas for Ni-20Al 2 O 3 HVOF coating, this angle ranged from 30̊ to 45 .This indicates the ductile nature of the erosion of SS316L and Ni-20Al 2 O 3 HVOF coating brings the semi-ductile orientation to the erosion behavior.
• The current ANN model yielded an R-value that was 0.98871 for training, 0.97209 for validation, and 0.96264 for testing accordingly.
• The total R-value of the model was determined as 0.98153.
• The ANN's performance evaluated based on the mean squared error (MSE) was determined to be 0.0176, 0.00265, and 0.00334 for the training, validation, and testing phases, respectively.
• Morphologically, the ploughing, intact patches, craters, and cracks were all seen in the SEM images of eroded Ni-20Al 2 O 3 coating.This con rms the semi-ductile orientation of its erosion behavior.Error histograms of the ANN model

Figure 9
Figure 9 illustrates the results of an erosion rate test conducted on a Ni-20Al 2 O 3 HVOF coating under the following conditions: C = 30% (by weight), T = 180 minutes, and N = 1500 revolutions per minute.According to Fig.9, the greatest erosion rate of SS316L occurred at an impact angle of α = 30, whereas for Ni-20Al 2 O 3 HVOF coating, this angle ranged from 30̊ to 45 .Hence, the erosion behavior of SS316L The XRD structure of Ni-20Al 2 O 3 coating reveals the presence of four different phases: α-Al 2 O 3 , γ-Al 2 O 3 , Ni, and NiO.• Microhardness of bare and Ni-20Al 2 O 3 HVOF coated SS316L was tested as 196 ± 21 and 393.32 ± 17

Figure 3 Basic
Figure 3

Figure 4 Detailed
Figure 4

Figure 5 Performance
Figure 5