Exploring Artificial Neural Network Techniques for Modeling Surface Roughness in Wire Electrical Discharge Machining of Aluminum/Silicon Carbide Composites

Abstract

Understanding wire-cut electrical discharge machining (WEDM) parameters’ impact on surface roughness (Ra) is crucial for optimizing processes. This study uses artificial neural network (ANN) techniques to estimate the surface roughness of Al/SiC composites during WEDM, examining how process parameters affect the roughness. The experiment used a stir casting aluminum alloy with a 7.5% silicon carbide metal matrix composite (MMC), adjusting parameters like the wire tension (WT), servo voltage (SV), peak current (IP), pulse on time (TON), and pulse off time (TOFF). An ANN model was created to forecast the surface roughness. The study developed an ANN model to forecast surface roughness in Al/SiC composites during WEDM, demonstrating its accuracy in identifying the link between surface finish and input parameters, thereby improving the surface quality. The ANN model accurately predicted the surface roughness based on WEDM parameters, with strong correlations between predictions and actual data, demonstrating its ability to estimate surface quality accurately.

1. Introduction

The growing need for high-performance materials across different engineering fields has prompted the investigation of advanced composite materials, including aluminum/silicon carbide (Al/SiC) composites. These materials are recognized due to their exceptional mechanical properties, such as increased strength and lighter weight, making them particularly suitable for industries like aerospace and automotive. However, machining these composites presents significant challenges due to their unique material characteristics, especially when employing techniques like wire-cut electrical discharge machining (Wire EDM).

Surface roughness is a critical parameter in the machining process, influencing not only the esthetic quality of the finished product but also its functional performance. Traditional methods of predicting surface roughness often rely on empirical models, which may not fully capture the complex interactions between machining parameters and material properties. In this context, artificial neural networks (ANNs) have emerged as a powerful tool for modeling and predicting surface roughness in various machining processes. By leveraging the capabilities of ANNs to learn from data, researchers can develop more accurate predictive models that account for the inherent nonlinear relationships in Wire EDM of Al/SiC composites.

Mangesh R. Phate et al. [1,2] conducted WEDM on Al/SiCp composites with SiC contents ranging from 0% to 20%, utilizing advanced modeling methods such as dimensional analysis (DA), principal component analysis (PCA), and artificial neural networks (ANNs) to forecast and enhance machining outcomes. Process variables including the pulse on/off time (TON/TOFF), current (IP), wire tension (WT), and SiC percentage were found to significantly impact the surface roughness, material removal rate, and kerf width. Among these methods, ANNs proved to be the most accurate and reliable, achieving a high prediction performance (R² up to 99.99%) and demonstrating effectiveness in WEDM process optimization. V. Kavimani et al. [3] used the Taguchi L27 design to evaluate the influence of WEDM process parameters on the surface roughness, kerf width, and material removal rate for Mg–Li–Sr alloys, produced using inert gas-assisted stir casting. A hybrid multi-objective optimization method combining CRITIC–WASPAS with artificial neural network (ANN) modeling was used to improve the prediction accuracy. The ANN model achieved a high R² value of 99.9% and showed lower prediction errors compared to conventional regression models, demonstrating its effectiveness in optimizing the WEDM process for lightweight alloys. Yusliza Yusoff et al. [4] utilized the ortho ANN approach to enhance the WEDM method while processing Inconel 718. By employing a Taguchi orthogonal array for systematic parameter selection, the authors minimized trial-and-error experimentation and demonstrated that the cascade-forward back-propagation neural network (CFNN) significantly enhances the prediction accuracy and efficiency, contributing to cost reductions and improved operational effectiveness in machining processes.

Rahim Jafari et al. [5] utilized micro-WEDM to make oxygen-free copper microchannel heat sinks with controlled surface textures, which play a key role in enhancing heat dissipation in compact electronic systems. Taguchi’s experimental design was used along with ANOVA to examine and optimize the surface roughness outcomes. An artificial neural network (ANN) model was developed to predict surface characteristics, achieving a high prediction accuracy with an R² of 99.5%, confirming its effectiveness in determining suitable µ-WEDM parameters for improved thermal performance. Pragya et al. [6] used RSM and ANN modeling methodologies to determine the cutting speed (average) in the WEDM of Al 6061/Si metal matrix composites. They found that artificial neural network models produced more accurate predictions than response surface methodology models. This highlights the significance of the servo voltage as a key parameter affecting the cutting speed.

Yunn-Shiuan Liaouses et al. [7] studied integrated specific discharge energy (SDE) as a material property to streamline WEDM process planning across different work piece types. By combining neural networks and genetic algorithms, the authors developed a predictive model that achieved a less than 7% error rate in estimating machining characteristics and identified optimal parameter settings. This approach reduced the need for extensive experimentation, making it a practical tool for efficient process planning. Titus Thankachan et al. [8] developed aluminum alloys with varying tin contents (5, 10, 15, and 20 wt%) and the addition of 5 wt% SiC particles composite materials. Microscopic analysis showed changes in grain structure and a decrease in hardness with increased tin contents. The WEDM process parameters were examined using Gray Relational Analysis and Taguchi’s method. To predict the surface roughness (Ra) and MRR, an artificial neural network (ANN) model was created, and its effectiveness was confirmed through experimental validation. Harmesh Kumar et al. [9] investigated the process parameters of the wire-cut electrical discharge machine while producing silicon carbide-reinforced aluminum metal matrix composites (Al/SiCp-MMC). They investigated how various input parameters affect the surface roughness, cutting speed, and spark gap by building quadratic regression models using a Box–Behnken Design of Experiments (DOE) and response surface technique. The authors identified the optimal conditions for effective machining and validated these through confirmation experiments, demonstrating the reliability of the developed models.

Gurupavan H R et al. [10] assessed the surface roughness, dimensional accuracy, electrode wear, and volumetric material removal rate during Wire EDM of aluminum/silicon nitride composites by varying key parameters such as the pulse on time (TON), pulse off time (TOFF), current, and bed speed. The surface characteristics were measured using Surfcom Flex and a micrometer, while artificial neural network (ANN) techniques were employed for advanced analysis across different machining conditions. Reza Kashiry Fard et al. [11] investigated WEDM of Al–SiC composites under dry conditions, using gas instead of liquid dielectric. Using Taguchi’s method and ANOVA, critical factors influencing the cutting velocity (CV) and surface roughness (SR) were identified, indicating that the best results are obtained when brass wire and oxygen gas are utilized. The combination of ANFIS and an artificial bee colony algorithm effectively optimized the process parameters, which was validated by confirmatory experiments.

Amresh Kumar et al. [12] investigated the machinability of aluminum (Al-6061)-grade composites that were reinforced with a combination of graphite, silicon carbide, and iron oxide using WEDM. Through a controlled experimental design and statistical analyses, it was found that optimized input parameters improved the spark gap (SG) width and MRR by 27.28% and 33.72%, respectively. Mangesh R. Phate [13] showed that aluminum/silicon carbide metal matrix composites (AlSiC MMCs) offer excellent mechanical and thermal properties but pose machining challenges. The authors used a Taguchi-based hybrid gray-fuzzy grade (GFG) approach to identify the optimal wire electric discharge machining (WEDM) parameters for AlSiC (20%) composites. Using Taguchi’s L9 array, they analyzed factors such as the pulse on/off time (TON/TOFF), wire feed rate (WF), and peak current (IP), finding that the best settings are a pulse on time (TON) of 108 µs, pulse off time (TOFF) of 56 µs, wire feed rate of 4 m/min, and peak current of 11 amps. An ANOVA revealed that the pulse on time (TON) significantly impacts performance, contributing 52.61% to the gray-fuzzy reasoning grade (GFRG).Mohammad Azad Alam et al. [14] synthesized pure aluminum reinforced with different percentages of silicon carbide micro particles (5%, 7.5%, and 10%) using powder metallurgy. The artificial neural networks (ANNs) and microhardness were assessed using a response surface methodology (RSM), and microstructural analyses were carried out using SEM, EDS, and XRD. The findings indicated that the composite with 7.5% SiC reached the highest sintered density and Vickers microhardness, attributed to its uniform distribution of fillers. Additionally, the ANN model proved to be more accurate in predicting the hardness than the RSM method. Vijayabhaskar et al. [15] explored the optimization of wire-cut electrical discharge machining (WEDM) parameters for magnesium metal matrix nanocomposites (MMNCs) that were reinforced with silicon carbide nanoparticles. Employing a D-optimal design using four factors and a response surface methodology, they assessed the impact of parameters like the servo voltage (SV) and pulse on /off times (TON/TOFF) on the surface roughness (R_a) and material removal rate (MRR). Second-order quadratic models were created and optimized using desirability analysis, showing good alignment between predicted and experimental values, with the surface morphology being analyzed via SEM. Naik et al. [16] studied electrical discharge machining of Al-SiC metal matrix composites with an emphasis on the material removal rate, surface finish, and dimensional accuracy. Using a Box–Behnken experimental setup with 46 runs, the process was optimized through a response surface methodology and evolutionary algorithms. The discharge current emerged as a key factor affecting all major outputs. Additionally, a sustainable machining strategy was explored by using a biodegradable vegetable oil-based dielectric fluid, highlighting its environmental benefits without compromising the machining performance. Wuyi Ming et al. [17] utilized Radial Basis Function Neural Networks (RBFNNs), Back-Propagation Neural Networks (BPNNs), and multiple regression analyses to explore the relationship between machining parameters and EDM performance in Al/SiC composites. The BPNN models achieved average prediction errors of 10.61% for surface roughness and 12.77% for material removal rate, outperforming the regression model, which showed a 13.93% error rate. Additionally, a graphical user interface was developed to assist users in selecting the optimal machining settings based on the desired surface quality. Additionally, the system incorporates genetic algorithms and a desirability function to support multi-objective optimization. Hariharan Sree Ram et al. [18] used analysis of variance (ANOVA) to generate predictive equations based on the material hardness and discovered patterns for MRR and surface roughness (Ra) when machining Al6061-based composites. The results demonstrated that the current coefficients varied significantly, which had an impact on both output parameters. The derived equations closely matched the experimental results, with average discrepancies of 9.3% for MRR and 7.2% for Ra, indicating their effectiveness for validation.

Saravanan et al. [19] employed stir casting to improve the properties of Al7075 nanocomposites that were reinforced with nanosilicon carbide and nanographite. WEDM process parameters were adjusted to reduce the kerf width. The investigation, carried out using a response surface methodology (RSM), demonstrated that the best performance is attained when using brass wire with a zinc coating, resulting in the smallest kerf width at the following optimized settings: pulse on time (TON): 117 µs; input current (IP): 160 A; pulseoff time (TOFF): 60 µs; and gap voltage (V): 10. V. D. Devarasiddappa et al. [20] formulated an ANN-based prediction model to estimate surface roughness during WEDM of an Inconel 825 alloy. They examined the impact of the pulse on time (TON), pulse off time (TOFF), peak current (IP), and servo voltage using a Box–Behnken design. The results showed that a lower pulse on time (TON) and servo voltage (SV) led to an improved surface finish. The optimized ANN architecture (4-16-1) demonstrated reliable prediction performance, with a 93.62% accuracy and a 6.38% average error. The statistical analysis indicated that the pulse on time (TON) had the highest influence on the surface roughness, followed by the servo voltage (SV) and pulse off time (TOFF). Sable Y.S. et al. [21] used a back-propagation algorithm of an ANN-based predictive model for the material removal rate (MRR) during WEDM of WC-Co composites. The model was trained and tested using data from experiments designed with five control factors—peak current (IP), wire tension (WT), pulse on time (TON), pulse off time (TOFF), and servo voltage (SV). The resulting ANN model demonstrated high accuracy, with an error rate below 0.05 and R² values of 0.9587 (testing) and 0.9968 (validation). Sener Karabulut et al. [22] investigated the milling performance of Al7075 and its open-cell SiC foam composite using an uncoated carbide tool. Machining was optimized for surface roughness using a Taguchi L27 design, ANOVA, and ANN-based modeling. The feed rate was identified as the most influential factor. The prediction models developed through regression and ANN showed high accuracy, with mean squared errors of 1.6% for Al7075 and 0.24% for the composite, demonstrating the high reliability of the ANN model. The open-pore SiC foam structure improved the surface quality by restricting matrix deformation during milling.

This paper aims to utilize the ANN method for surface roughness modeling in the Wire EDM process. By examining different ANN architectures and training methodologies, we seek to provide insights into their effectiveness and applicability in this niche area. Ultimately, the findings will contribute to optimizing machining parameters and enhancing the overall quality of finished components, paving the way for broader applications of Al/SiC composites in advanced manufacturing.

2. Materials and Methods

A wire-cut electrical discharge machine with a 5-axis Numerically Controlled Computer (Model: SPRINT CUT, Manufacturer: Electronica Machine Tools Ltd., Kolkata, India) was employed for the experiments (Figure 1). The dielectric fluid was pure water, and the tool electrode was a 0.25 mm diameter brass wire electrode. The base material of the composite was the aluminum alloy LM 25, and the matrix material was 7% SiC.

Table 1. Details of material composition.

Component	Percentage
Cu	0.20
Mg	0.20–0.60
Si	6.5–7.5
Fe	0.5
Mn	0.3
Ni	0.1
Zn	0.1
Pb	0.1
Sn	0.01
Ti	0.2
SiC	7.5
Al	Remainder

shows the list of aluminum alloy specs that were used in the experiments.

A Mitutoyo Surface Roughness SJ-410 tester (Mitutoyo Corporation, Kawasaki-shi, Kanagawa, Japan) was used to detect the surface roughness in µm. The five WEDM process parameters—servo voltage (SV), peak current (IP), wire tension (WT), pulse on time (TON), and pulse off time (TOFF)—were chosen based on the outcomes of pilot tests. Each of these parameters had five levels, as indicated in

Table 2. Process parameters and levels.

Factor	Level
	−2	−1	0	1	2
Pulse on Time (TON)	106	112	118	124	130
Pulse off Time (TOFF)	52	54	56	58	60
Peak Current (IP)	190	200	210	220	230
Servo Voltage (SV)	15	20	25	30	35
Wire Tension (WT)	4	6	8	10	12

.

3. Data Processing

A central composite design (CCD) of response surface approaches provides the smallest number of machining trials required to evaluate the impact of individual parameters and their interactions in experiments comprising five parameters at five different levels. All 32 experiments were conducted, and the combinations of experimental runs and the associated surface roughness values for each trial are shown in

Table 3. Experimental DOE with Ra values.

Sr. No.	TON	TOFF	IP	SV	WT	Ra
1	112	54	200	20	10	1.118
2	124	54	200	20	6	1.365
3	112	58	200	20	6	1.521
4	124	58	200	20	10	1.116
5	112	54	220	20	6	1.109
6	124	54	220	20	10	2.17
7	112	58	220	20	10	1.094
8	124	58	220	20	6	2.051
9	112	54	200	30	6	1.355
10	124	54	200	30	10	1.393
11	112	58	200	30	10	0.961
12	124	58	200	30	6	1.414
13	112	54	220	30	10	1.114
14	124	54	220	30	6	1.39
15	112	58	220	30	6	1.252
16	124	58	220	30	10	1.602
17	106	56	210	25	8	1.309
18	130	56	210	25	8	1.379
19	118	52	210	25	8	1.5
20	118	60	210	25	8	1.725
21	118	56	190	25	8	1.452
22	118	56	230	25	8	1.391
23	118	56	210	15	8	1.318
24	118	56	210	35	8	1.743
25	118	56	210	25	4	1.599
26	118	56	210	25	12	1.903
27	118	56	210	25	8	1.212
28	118	56	210	25	8	1.687
29	118	56	210	25	8	1.017
30	118	56	210	25	8	1.063
31	118	56	210	25	8	1.46
32	118	56	210	25	8	1.573

.

Studying the structure and function of the human brain resulted in the creation of computational models known as artificial neural networks. They consist of multiple layers of interconnected nodes, or neurons, such as an input layer, one or more hidden layers, and an output layer. ANNs can learn from input data using a technique known as training, which involves modifying the weights of connections based on the input–output relationships. This allows them to identify trends, make forecasts, and solve challenging issues in a variety of fields, including financial forecasting, picture identification, and natural language processing. They are effective tools in artificial intelligence and machine learning because of their capacity to generalize from instances. Figure 2 depicts the five parameters that served as input layer neurons in this study: wire tension (WT), servo voltage (SV), peak current (IP), pulse on time (TON), and pulse off time (TOFF). The surface roughness functioned as an output layer neuron.

4. Results and Discussion

The artificial neural network was trained using 70% of the complete dataset and the K-fold cross-validation method in order to increase performance. After then, 30% of the data were used for testing and validation. Training was carried out using the Levenberg–Marquardt back-propagation method (trainlm), and calculations were carried out using the tansig transfer function. The linear fit model for the surface roughness model training, validation, testing, and overall datasets is shown in Figure 3.

The equation for the Levenberg–Marquardt back-propagation algorithm is as follows

$$Levenberg Marquardt=\left(J^{T}J+\mathsf{\lambda }\mathrm{I}\right)\mathsf{\Delta }\mathsf{\theta }=-J^{T}e$$

(1)

where

$J^T$ = Jacobian matrix (derivatives of the errors with respect to the parameters).

λ = damping factor (controls the balance between Gauss–Newton and gradient descent).

e = error vector (difference between predicted and actual values).

Δθ = update to the parameters (weights and biases).

I = identity matrix.

The following transfer function was used:

$$\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{s}\mathrm{i}\mathrm{g}\left(\mathrm{N}\right)=2/(1+\mathrm{e}\mathrm{x}\mathrm{p}(-2\mathrm{*}\mathrm{N}\left)\right)-1$$

where

N = the net input to a neuron OR input matrix.

The values of the experimental results and the response variable predictions made by the artificial neural network model are contrasted in Figure 4. As shown in

Table 4. Statistical evaluation of the ANN model.

Performance Parameter	Data	MSE	R2
Surface Roughness	Testing	0.016298	0.94542
	Validation	0.012141	0.85402

, statistical analysis was used to assess the ANN model’s fit quality.

The normalization of input parameters was performed using the min-max normalization approach prior to training the ANN model to ensure that all features contributed equally to the learning process and to avoid dominance of features with larger numerical ranges. The obtained correlation coefficients, R²,for the surface roughness ANN models were as follows: training = 0.94542; testing = 0.87502; and validation = 0.85402. Meanwhile, the mean squared error (MSE) for surface roughness was 0.012141 for validation, calculated from the validation results using validated data, and 0.016298 for training, as shown in Figure 5. The statistical analysis, along with Figure 4 and

Table 5. Experimental values and predicted values from surface Roughness ANN model.

Experimental	ANN Model	Error
1.118	1.119688	−0.00169
1.365	1.365095	−0.00009
1.521	1.521738	−0.00074
1.116	1.11662	−0.00062
1.109	1.107762	0.001238
2.17	2.169901	0.00009
1.094	1.097702	−0.0037
2.051	2.05061	0.00039
1.355	1.355345	−0.00035
1.393	1.392857	0.000143
0.961	0.962165	−0.00116
1.414	1.412847	0.001153
1.114	1.114992	−0.00099
1.39	1.382968	0.007032
1.252	1.250735	0.001265
1.602	1.600847	0.001153
1.309	1.309103	−0.0001
1.379	1.379478	−0.00048
1.5	1.498642	0.001358
1.725	1.726094	−0.00109
1.452	1.451054	0.000946
1.391	1.391073	0.000073
1.318	1.319575	−0.00158
1.743	1.742181	0.000819
1.599	1.598635	0.000365
1.903	1.903048	−0.000048
1.212	1.316917	−0.10492
1.687	1.316917	0.370083
1.017	1.316917	−0.29992
1.063	1.316917	−0.25392
1.46	1.316917	0.143083
1.573	1.316917	0.256083

, compared the predicted response variable values from the artificial neural network model with the experimental values, showing that they are closely aligned. This suggests that the experimental and predicted values in the ANN model have a strong correlation. The validation performance curves for the surface roughness (Ra) response variable are shown in Figure 5.

5. Conclusions

The surface roughness of aluminum/silicon carbide (Al/SiC) composites during the WEDM process may be accurately predicted using artificial neural networks (ANNs), as this study shows. The ANN effectively links key input elements such as the wire tension (WT), servo voltage (SV), peak current (IP), pulse on time (TON), and pulse off time (TOFF) to the resulting surface roughness with a high degree of accuracy in predictions that closely match the experimental data. The ANN model can greatly improve the accuracy of machining processes, as evidenced by the low mean squared error of 0.016298 and the strong correlation coefficient of 0.94542 by normalization. Better surface finishes in Al/SiC composites can be achieved by optimizing WEDM parameters using an ANN method, as demonstrated in this study. Additionally, the findings show how different machining parameters affect the surface quality, which promotes advancements in manufacturing techniques.