Experimental Investigation and Machine Learning Modeling of Electrical Discharge Machining Characteristics of AZ31/B₄C/GNP_s Hybrid Composites

Abstract

Magnesium alloys, like AZ31, possess a desirable low weight and high specific strength, which make them favorable for aerospace and auto applications, yet their difficulty to machine limits their broader implementation for the industry. Electrical discharge machining (EDM) is an effective technology for machining difficult-to-machine materials, particularly when the materials are reinforced with ceramic and graphene-based fillers. This study examines the impact of reinforcement percentage (R) and different electrical discharge machining (EDM) parameters such as current (I), pulse on time (T_on) and pulse off time (T_off) on the material removal rate (MRR) and surface roughness (SR) of AZ31/B₄C/GNPs composites. The combined reinforcement range varies from 2 wt.% to 4 wt.%. The Taguchi design (L27) is utilized to conduct the experiments in this study. ANOVA of the experimental data indicated that current (I) significantly affects MRR and SR, exhibiting the greatest contribution of 44.93% and 51.39% on MRR and SR, respectively, among the variables analyzed. The surface integrity properties of EDMed surfaces are examined using SEM under both higher and lower material removal rate settings. Diverse machine learning techniques, including linear regression (LR), polynomial regression (PR), Random Forest (RF), and Gradient Boost Regression (GBR), are employed to construct an efficient predictive model for outcome estimation. The built models are trained and evaluated using 80% and 20% of the total data points, respectively. Statistical measures (MSE, RMSE, and R²) are utilized to evaluate the performance of the models. Among all the developed models, GBR exhibited superior performance in predicting MRR and SR, achieving high accuracy (exceeding 92%) and lower error rates compared to the other models evaluated in this work. This work demonstrated the synergy between techniques in optimizing EDM performance for hybrid composites using a statistical design and machine learning strategies that will facilitate greater use of hybrid composites in high-precision engineering applications and advanced manufacturing sectors.

1. Introduction

Metal Matrix Composites (MMCs) consist of a ductile metal matrix reinforced with hard, high-strength materials. MMCs have gained significant attention in many industries, such as automotive and aerospace, because of their many beneficial properties, such as high strength-to-weight ratio, improved wear resistance, and increased stiffness. Magnesium (Mg) has become one of the most popular materials as it is used in MMCs primarily because of its low density (1.738 g/cm³), which is extremely low compared to aluminum and titanium [1,2]. Magnesium’s low density and biocompatibility have led to the use of magnesium in biomedical applications, especially bioimplants [3]. However, magnesium does have considerable deficiencies that can limit its wider applications; these include a relatively low strength, relatively high susceptibility of magnesium to corrosion, and relatively low resistance to wear. To overcome these deficiencies and help improve the mechanical properties of magnesium, hard reinforcement materials are added into the matrix of magnesium, such as Al₂O₃, TiO₂, SiC, TiC, and graphene [4,5,6,7,8]. Adding hard reinforcement materials to magnesium MMCs will give improved strength and wear resistance, improvements to the overall potential of Mg to obtain even more advanced properties in extreme and demanding environments, including aerospace, automotive, and biomedical engineering.

1.1. Electrical Discharge Machining of Mg MMCs

Despite magnesium being a softer metal, the machining of Mg composites is challenging. Drilling micro-sized apertures, tool degradation, and fracture are prevalent issues linked to traditional machining methods when utilized on Mg composites. Consequently, there exists a substantial requirement for the development of economical and precise machining procedures for magnesium composite materials. A non-conventional machining process is essential to fulfill the aforementioned requirements, and among the available possibilities, Electric Discharge Machining (EDM) is an established method for machining Mg composite materials. It is a non-contact process, suitable for materials that cannot endure the stress of traditional machining. M. Somasundaram et al. [9] investigated the potential of EDM for AZ31 alloy, which is useful in medical, aerospace, and automotive applications due to its high strength-to-weight ratio. They focused on optimizing discharge current (I), pulse-on time (T_on), and pulse-off time (T_off) to maximize geometrical accuracy. Their work utilizes Taguchi’s L₁₆ orthogonal array for experimentation and observation and ultimately demonstrates that Ton and I affect geometrical features like overcut and circularity. Additionally, the regression models developed in their study provided effective predictive capacity for response parameters with a strong correlation with experimental results. Poovazhagan Lakshmanan et al. [10] examined the effect of three critical electrical parameters—I, T_on, and T_off—on the MRR during EDM of Mg-SiC-fly ash composites. The study observed that the I and T_on had a greater impact on MRR than did the T_off. The pulse current (I) not only has a direct impact on the total energy supplied each discharge event; when evaluating MRR, higher currents were seen to induce more material removal. Higher currents also caused more heat generation and surface damage. The longer T_on, the more time there is for material removal per cycle, and hence more MRR. Finally, the T_off is a major factor that influences cooling and possible electrode recharging merely has a minor role in MRR. The authors established that I and Ton both require optimization in order to secure a higher MRR outcome in EDM of these composites, while Toff merely controls the heat produced and surface quality. Alok Kumar et al. [11] employed powder-mixed EDM which is a modification of standard EDM in which nano-powders are suspended in the dielectric medium to improve the machining process by reducing surface defects such as cracks and craters. Here it was shown that PM-EDM treatment on Mg alloy AZ31B produced a surface with very few defects, a recast layer thickness of 2.4 to 2.7 μm, and surface roughness (SR) of Ra = 2.87 μm. The modified surface showed adequate corrosion resistance, hydrophobicity (CA 122°), and bioactivity after incubation in Ringer’s solution, indicating an increased degradation rate. From this study, it is observed that PM-EDM modifies the surface properties of magnesium alloy AZ31B, which may make the material suitable for use as a biodegradable implant material in adults and children.

1.2. Role of Machine Learning in EDM of MMCs

As a result of advances in computational techniques, several researchers have focused on creating effective prediction models to precisely predict the outcomes of machining. Fatih Aydin et al. [12] used various machine learning (ML) models, including Random Forests (RF), Support Vector Machines (SVM), and Artificial Neural Networks (ANN), to estimate the wear rate of AZ91. This study conducted a comparative comparison of the predicted accuracy of an ML model vs. experimental results. The study’s outcomes demonstrated that the ANN model had greater prediction accuracy (0.9845) relative to other models. Chander Prakash et al. [13] developed a novel method to produce hydroxyapatite (HA) coatings with a network of pores on biodegradable Mg-Zn-Mn alloys utilizing a hydroxyapatite powder mixed EDM (HAM-EDM) process. This research investigates the optimization of several process variables, including HA solution concentration, peak current, pulse-on time, and pulse-off time, with multi-objective particle swarm optimization (MO-PSO). The results showed the best conditions as follows: 5.28 g/L HA concentration, 3.48 A peak current, 40.33 μs pulse-on time, and 109.29 μs pulse-off time. Under these conditions, the SR of the HA coating was low (0.70 μm), recast layer thickness was low (11.85 μm), and the micro-hardness was high (246 HV). The coatings were characterized to confirm the calcium, phosphorus, and oxygen levels identified the HA coating and that the coatings also had some biocompatible phase present, which would result in improved mechanical properties, corrosion resistance, and potential to further improve osseointegration. These features of HA-coated Mg-Zn-Mn alloys make them promising candidates for biomedical applications. Dhanunjay et al. [14] focused on the role of EDM parameters on SR and MRR for Mg composites (AZ91 combined with graphene and aluminum oxide). In the manifestations it was noted that I had the greatest impact on MRR and SR. Voltage (V), flushing pressure (P), and T_ON were of next significance. MRR was maximized with the following parameters: material = AZ91 alloy, P = 0.75 kg/cm², V = 40 V, T_ON = 20 µs, and I = 30 A. The optimum SR was obtained with these parameters: material = AZ91 alloy, P = 0.50 kg/cm², V = 40 V, I = 10 A, and T_ON = 20 µs. The SEM images reveal the different surface characteristics at different MRRs. A hybrid ANN-TLBO model was developed to predict MRR and SR, and through confirmation tests it was determined that the model could predict the outcomes with a significant level of precision (±10% error for the majority of the data). The analysis of the hot deformation response of AZ31 alloy and the importance of temperature and strain rate to proper processing conditions to improve material performance was demonstrated by YU Jian-ming et al. [15]. In this work, the analysis of flow stress behavior was completed using the Gleeble-1500 thermal simulator with a developed prediction model leveraging BP neural networks to encapsulate the interrelationship between flow stresses, strain, strain rate, and temperature. The investigation revealed the presence of flow instability in two conditions: low temperature with a high rate of strain and moderate temperature with a low rate of strain. Additionally, it was found that an optimum processing zone existed in between 340 and 440 °C and a strain rate of 0.01–0.02 s⁻¹ due to optimal power dissipation. The prediction model produced flow stresses with good accuracy, with a maximal relative error of 6.67% concerning the flow stress under study, indicating the model’s validity. The investigation demonstrates the ability of BP neural networks to accurately predict flow stress behavior and may help develop more efficient processing methods for AZ31 magnesium alloy during hot working.

1.3. Research Gaps Identified

The literature clearly indicates that research on the EDM of AZ31 hybrid composites is significantly scarce in comparison to other composites. Additionally, enhancing the efficiency of machining variables has been the primary focus of many academic studies. Nevertheless, there has been a dearth of research on the evaluation of the effect of these various machining variables on the outcome and the building of an efficient predictive framework for EDM via ML methods. Additionally, it is apparent from the existing research that no other investigations have applied ML techniques to estimate the EDM features of AZ31/B₄C/GNPs composites under a variety of input parametric conditions, despite the fact that numerous studies have demonstrated that ML approaches have a superior prediction capability compared to conventional statistical methods. Consequently, it was determined to implement the same methodology in the present investigation. This research endeavors to examine the influence of a diverse array of EDM parameters on the material removal rate (MRR) and SR of AZ31/B₄C/GNPs composites. Additionally, it will utilize artificial intelligence (AI) techniques to establish the correlation between both the input and output parameters of EDM in hybrid composites.

1.4. Motivation of This Study

To resolve the gaps identified, this study is driven by the need for a thorough understanding of the EDM behaviors of AZ31/B₄C/GNPs hybrid composites. This study systematically examines the impact of varying reinforcement percentages (2–4 wt.%) and EDM parameters (current, pulse-on time, and pulse-off time) and their relation to important cutting outcomes such as MRR and SR. The work also incorporates a robust surface integrity assessment via SEM to reveal microstructural change under varying cutting conditions. Another motivation is to develop sophisticated predictive models beyond conventional statistical models that utilize complex machine learning processes, including linear regression (LR), polynomial regression (PR), Random Forest (RF), and Gradient Boosting Regression (GBR), in order to more accurately understand EDM performance for use in industry. The primary aims of this study are therefore to delineate scientific understanding as well as generate practical strategies for optimizing EDM processes of hybrid composites used within advanced manufacturing sectors such as aerospace and precision engineering.

2. Materials and Methods

2.1. Fabrication of Composites

Figure 1 depicts the methodology followed in this study. The present investigation utilizes the AZ31 alloy as a matrix to produce hybrid composites. Boron carbide (B₄C) and graphene nanoplatelets (GNPs) are employed to reinforce these composites.

Table 1. AZ31 alloy composition.

Element	Zn	Al	Si	Fe	Cu	Mn	Ni	Others	Mg
Content (%)	1.05	3.12	0.1	0.005	0.03	0.15	0.002	0.02	89.523

illustrates the chemical constituents of AZ31. Figure 2 illustrates the reinforcements using SEM pictures. Hybrid composites are made by employing the inert gas-assisted stir casting technique, which consists of reinforcements in variable proportions. A temperature of 750 °C is maintained in the furnace to dissolve AZ31 ingots. Upon the utilization of a stirrer to produce a rotating motion, a vortex is generated. Afterward, the molten material was integrated into the preheated nano powder once the temperature of the furnace approached the critical barrier of 750 °C using furnace (Swam Equip., Chennai, India). Stirring was performed for 15 min at a speed of 400 rpm to ensure that the elements that were reinforced were evenly assimilated. To prevent the oxidation of AZ31, a controlled gaseous condition was meticulously maintained throughout the experimental procedure, comprising a mixture of argon and SF6. The GNP wt.% in the hybrid composites was maintained at 1 wt.%, while the wt.% of the other component was varied between 1 wt.% and 3 wt.%. To lessen the detrimental effects of increased nano reinforcement percentages, including inadequate integration and congregating within the composite, the GNPs wt.% and the total wt.% were restricted to 1% and 4%, respectively. The details of the composites are summarized in

Table 2. Description of Hybrid Composite.

S. No	AZ31 (wt.%)	GNPs (wt.%)	B4C (wt.%)	Total (wt.%)
1	98.5	1	1	2
2	98	1	2	3
3	97	1	3	4

. Figure 3 depicts the SEM microstructures and XRD patterns the of hybrid composites using SEM (TESCAN, Visakhapatnam, India) and XRD (Bruker, Visakhapatnam, India). The SEM picture of the composite (a) is captured at low magnification, in contrast to other composites that are captured at high magnification. In addition to evaluating the compositional descriptions detailed in Table 2, this work examined the mechanical stability of the hybrid composites, focusing on the β-Mg₁₇Al₁₂ phase. The XRD pattern displayed in Figure 3 indicated that reflections of the β-Mg₁₇Al₁₂ were found at approximately 2θ ≈ 41.5° and 44.8° for both the base AZ31 alloy and the hybrid composites. From the Bayesian analysis of the peak shapes, it is evident that the β-Mg₁₇Al₁₂ phase was refined with reinforcement (B₄C and graphene), as well as other indications that suggested the distribution of Al atoms occurred while partially inhibiting the formation of the β-phase. The average sizes of the β-Mg₁₇Al₁₂ crystallites are about ~12–18 nm for all hybrid composites and approximately ~25 nm for the base AZ31 alloy (Scherrer equation). The Williamson–Hall analysis indicates a minor lattice strain (~0.15–0.20%), likely due to interactive effects of reinforcements and Mg matrix, while further refinement in the β-phase was evident with hybridization. The small crystallite size of the β-Mg₁₇Al₁₂ phase in conjunction with lattice strain contributed to microstructural stability by restricting dislocation slip, thereby improving strength and formability of hybridized composites. Figure 4 depicts the EDS mapping of the hybrid composite containing the highest amount of reinforcement (i.e., AZ31+1 wt.%GNPs+3 wt.%B₄C). The EDS mapping indicates a reasonably homogeneous distribution of elements Al, B, C, and Zn, including the reinforcements, with distinct representational clusters of the different elements shown in the mapping.

The distribution of grain sizes of the hybrid composites, presented in Figure 5 and measured by image analysis, shows a smooth decline in average grain size from Composite (a) (~14.7 µm) to Composite (c) (high magnification) (~7.9 µm) and a simultaneous decrease in standard deviation of the distribution. Composite (c) exhibited a relatively narrow distribution, which implies a homogeneous microstructure dominated by small, fine-grained microconstituents, potentially supporting strength through the Hall–Petch effect. By contrast, composite (a) exhibited a wide distribution, suggesting grain coarsening and inhomogeneity, most likely resulting from localized thermal gradients and selective particle agglomeration upon processing. Such grain size distribution differences have a direct impact on the mechanical soundness of the composites: smaller and more evenly distributed grain sizes engender enhanced capability for bearing loads and resistance against micro-crack growth, whereas larger irregularities lead to reduced ductility and increased micro-cracking susceptibility. Such findings correlate with the function of reinforcement components and the β-Mg₁₇Al₁₂ phase presented through XRD analysis, which collectively control the refinement and stability of the microstructure of the resulting composite materials.

2.2. Experimentation

The studies utilized a die-sinking EDM machine (Electronica, Nellore, India) as illustrated in Figure 6a,b. The present investigation considers current (I), pulse duration (Ton), and pulse off time (T_off) to be the EDM parameters, as per the literature [4,5,12].

Table 3. Factors and their levels (L).

S. No	Factor	Units	L1	L2	L3
1	Reinforcement percentage (R)	wt.%	2	3	4
2	Current (I)	A	3	6	9
3	Pulse on time (Ton)	µs	25	50	75
4	Pulse off time (Toff)	µs	5	10	15

delineates the parameters and their respective levels. An L27 orthogonal array (OA) was utilized to conduct experiments based on the specified parameters and their corresponding levels. The experimental setup employed a copper (Cu) rod with a 10 mm diameter as the electrode, and EDM oil was harnessed as the dielectric fluid. The experimental design was developed utilizing Minitab software (version 2021) with four factors at three levels each. In this non-traditional machining process, the tool and workpiece were submerged in a dielectric fluid, and controlled electrical discharges were applied across a small gap between them. These high-frequency sparks generated localized heat, melting and vaporizing the material from the workpiece surface in the form of microscopic craters, while the dielectric simultaneously flushed away debris to maintain the spark gap. In every trial, a uniform depth of incision was upheld at 1 mm. The material removal rate (MRR) was determined by measuring the weight loss of the specimen before and after machining.

The efficacy of the parameters was evaluated using measures such as SR and MRR. The MRR is calculated using Equation (1).

$$MRR={\displaystyle \frac{m_i-m_f}{t_m}}$$

(1)

where $m_i$—initial weight in grams, $m_f$—final weight in grams and $t_m$—machining time in minutes.

Measurements of the SR of the machined surfaces were conducted using the Talysurf surface analyzer (Mitutoyo, Vijayawada, India). Three measurements were employed to assess the SR, and the resulting average value was adopted as the output. The experimental results have been employed to develop a predictive model using machine learning techniques, as detailed in the next section.

3. Machine Learning (ML) Techniques

The principal attributes of EDM encompass productivity and surface integrity. Productivity is measured by the MRR, which indicates the speed at which material is extracted from the workpiece over a specified duration. Surface integrity is represented by SR. These two output characteristics are critical in every manufacturing process. This work aims to utilize a variety of ML methods to predict the MRR and SR of AZ31 Mg matrix composites when processing it. MRR and SR are dependent variables with four different key dependent parameters: reinforcement percentage, current, T_on, and T_off. These features were chosen from the existing literature [14,16] and have been documented to have a considerable effect on machining performance commonly. In order to model the output responses, the experimental data will be normalized using the min-max scaling method to bring all the features into the range [0–1] to ensure all features have equal contribution to whatever the models are learning and to better fit the models accurately in the existing training set. The proposed regression models include linear regression, polynomial regression, and tree-based methods, including decision trees and ensemble-based approaches. Based on the availability of various modules and libraries, the models will be developed in Python (version 3.12.6) using scikit-learn, which has a range of tools to provide bounded flexibility in building, training, and validating machine learning models. The predictive power of the models would be examined from the features dataset of respective input features and their output responses split for training and testing. To develop robust and generalizable models, the experimental data was separated into two sets that contained data for the training and testing set. The splitting of the data is based on standard practice in supervised machine learning. The training set should be used to allow the models to learn the underlying patterns and relationships between the input features (current, pulse on, pulse off, and percent reinforcement) and the output responses (MRR and SR). It is important that the testing data be separate from the training data in order to evaluate the predictive capabilities of the models on unseen data and ensure that the models will not overfit the training data and work in a practical application. Without the partitioning of the data, the performance metrics of the models could be inaccurate since the performance measurements would be misleading and give an overly optimistic estimate of accuracy when the same data is used to evaluate performance measurements.

The performance evaluation would include the coefficient of determination (R²) and mean squared error (MSE) to determine the goodness of fit (R²) and the average of the squared differences between predicted and actual (MSE). The higher the R² value and lower the MSE values mean the model is more accurate and reliable. Overall, the approach of applying machine learning techniques to derive and predict the quantitative relationships between features and outputs provides a powerful way to model complex, and often nonlinear, relationships that exist in machining and provide accurate predictions of quality metrics relevant to optimizing composite manufacturing.

3.1. Linear Regression

Linear regression is an analytical technique that can be implemented to quantify the relationships between two variables (dependent and independent). In addition, linear regression incorporates a multitude of straightforward linear regressions. There are several linear regressions that are usually used to forecast outcomes. By employing the ordinary least squares (OLS) technique, the linear time series regression is modeled. The OLS approach was preferred, owing to its convenience, capacity for interpretation, and computational efficiency. OLS produces the unbiased and efficient estimators of the regression coefficients by minimizing the sum of squared residuals. Furthermore, OLS is useful for assessing if a linear approximation is sufficient to characterize the relationship between the machining parameters (current, pulse on time, pulse off time, reinforcement percentage) and performance variables (MRR and SR). OLS also acts as a benchmark model that can be compared to more complicated models like polynomial regression, Random Forest, and Gradient Boosting Regression. The framework for linear regression is illustrated in Equation (2).

$$y=c_o+\sum _{i=1}^n\left(c_i{x}_i\right)$$

(2)

where $c_o$—intercept, y—response variable, $c_i$—coefficient of feature $x_i$, $x_i$—independent variable or feature, n—number of independent variables or features and $i$ is a variable which represents the instant number of feature or independent variable.

3.2. Polynomial Regression

The polynomial regression technique is a method for evaluating regression that involves the introduction of more parameters and the elevation of each primary parameter to a power. A polynomial equation of degree n in x is used to carry this out in order to assess the impact of the prediction variable x on the outcome variable y. A statistical technique called PR can be used to evaluate how a predicting variable (x) affects an outcome variable (y). Cubic polynomial regression predictive variables are made up of x, x², and x³, whereas quadratic polynomial regression predictors are made up of x and x². Multiple LR is therefore considered to include polynomial regression as a subset. The link between independent and dependent variables is directly examined using this method. Equation (3) displays the model of polynomial regression.

$$Z=Y\lambda +\theta$$

(3)

where Z—output or predicted value, Y—polynomial term, lamda coefficient ($\lambda$) of Y, theta ($\theta$)—intercept.

3.3. Random Forest (RF) Regression

Random Forest (RF) regression is implemented with an ensemble learning approach and improves predictive performance by creating a large number of decision trees that are unpruned and built using multiple bootstrap samples of the original training data. Each tree in the RF is fitted on its own unique bootstrapped dataset (sampled with replacement from the original dataset), which adds a level of diversity in the ensemble. In addition to different bootstrapped samples, for every split in the tree, RF randomly selects a subset of features (of size K, where K < M, M is the total number of features in the data) to be considered, separating the trees from each other even further. This calculation of variance reduces correlation in the individual trees, improving robustness and decreasing the chance of the model overfitting. The trees are fitted with a growing parameter of ‘unpruned’, which means they are grown to full depth and complexity so they can learn the entire dataset accurately and learn any complex nonlinear relationships. The final prediction for a given input is obtained by taking the mean of all T predictions from the trees in the ensemble, which reduces the variance of the prediction while maintaining low bias. This averaging function is possible because of the law of large numbers theorem and provides a more stable and accurate estimator compared to single decision trees. The algorithm also provides a built-in method of model validation through the algorithm of out-of-bag (OOB) error estimates, where the unused samples within each bootstrap set are used for model performance evaluation. The RF framework inherently has a built-in process to assess feature importance and is highly resistant to noise with minor variations, making it a powerful and flexible regression model approach for a variety of data subjects. The formula for an RF model is illustrated in Equation (4).

$$g_{RF}^Z\left(y\right)={\displaystyle \frac{1}{Z}}{\sum }_{n=1}^Z{DT}_n\left(y\right)$$

(4)

where y is the input vector variable, Z indicates the total number of trees, and ${DT}_n\left(y\right)$ depicts an individual tree.

3.4. Gradient Boost Regression

Gradient boosting is a potent method for ensemble learning, or the combining of weak predictors (typically, shallow trees) into a strong prediction model in a sequential fashion using weak learners. The general principle of gradient boosting is to minimize a specified loss function iteratively by moving downhill (i.e., gradient descent) in function space. In effect, each new member is stacked onto all previous member(s) and corrects it by minimizing the residual errors by adding trees. The initial models start with a constant prediction (i.e., the mean of our target values for a regression). At each iterative step we train a new decision tree as a weak learner, which approximates the negative gradient (-g) of our loss function with respect to the predictions given by the current model. Essentially, this weak learner’s main role in the ensemble is to capture the errors (residuals) generated by the existing ensemble and suggest how to correct those errors before we add on the next member of the model. Each tree is added to the previous weak learners via a learning or shrinkage rate to control the amount of information contributed to the model by that weak learner, which also preserves some level of overfitting and allows for complex updates to the member ensemble building. The updated model at iteration m is formally expressed in Equation (5).

$$F_m\left(x\right)=F_{m-1}\left(x\right)+ʋ.h_m\left(x\right)$$

(5)

$F_{m-1}\left(x\right)$, $h_m\left(x\right)$ and $ʋ$ are old model, weak learner and learning rate, respectively.

This cycle is repeated until either a certain number of trees have been built or performance is constant. Gradient boosting supports differentiable loss functions (mean squared error for regression, log loss for classification) and is adaptive by nature through its iterative processes since it puts more focus on instances that are difficult to predict out of the gate. If ultimately constructed and tuned, gradient boosting has shown the ability to create highly predictive models with high generalization.

4. Results and Discussion

Electrical Discharge Machining (EDM) is a non-conventional machining method used to manufacture parts or components from hard and complex materials. EDM removes material from a rotating workpiece by way of controlled electrical discharges. The MRR is one of the primary indicators of electronic discharge machining productivity as it calculates the volume of material removed from the workpiece in a time frame. Higher MRR values indicate that the EDM process can reduce production times and enhance overall productivity, and efficiency. In addition, the MRR metrics can be influenced by numerous parameters, which include discharge energy, electrode configuration, and dielectric fluids (to flush and cool the workpiece and electrode). Consequently, as an EDM process developer, it will be important to calibrate these parameters to develop a safe process that is effective at removing material. Surface integrity also plays an important role in EDM development. Surface integrity can often be described by way of SR—which describes the texture of the surface finish and quality of the part produced. Lower roughness values indicate a smoother and higher quality surface finish. In EDM, surface roughness is significantly affected by the interaction of the process parameters: pulse on time (Ton), pulse off time (Toff), and discharge current (I). This is because they affect the energy delivered per discharge and thus affect the thermal mechanism at the surface of the bulk material. As either the pulse on time or current increases, discharge energy increases, resulting in deeper and wider craters on the machined surface because more material is melted and ejected when the discharge occurs. This leads to larger surface roughness values. On the other hand, by lowering either the pulse on duration or the current (or both), the energy per pulse is decreased, producing shallower craters, resulting in improved surface finish. Additionally, the incorporation of reinforcing particles like B₄C and GNPs is also going to affect the heat dissipation and electrical features of the composite, thereby changing the energy dissipation and melting characteristics in the adjacent area. The particles may sometimes act as thermal barriers or sinks, change the plasma channel characteristics and rate of material removal, and modify the surface roughness. The combination of these various effects influences the ultimate roughness of the surface once the process has been completed. The primary objective of this investigation is to evaluate the impact of EDM process variables on the operational properties, MRR, and SR of composite materials. The experimental outcomes are summarized in

Table 4. Experimental results at different input parameters (L27 OA).

S. No	Reinforcement	Toff	Ton	I	MRR	SR
	wt.%	µs	µs	A	g/min	µm
1	2	5	25	3	0.08250 ± 0.002	2.247 ± 0.170
2	2	5	50	6	0.21560 ± 0.008	3.433 ± 0.179
3	2	5	75	9	0.36900 ± 0.018	4.987 ± 0.179
4	3	10	25	3	0.09724 ± 0.029	2.600 ± 0.401
5	3	10	50	6	0.12903 ± 0.005	4.751 ± 0.152
6	3	10	75	9	0.36666 ± 0.007	5.149 ± 0.253
7	4	15	25	3	0.08343 ± 0.019	4.007 ± 0.502
8	4	15	50	6	0.14810 ± 0.007	4.722 ± 0.394
9	4	15	75	9	0.17640 ± 0.012	6.720 ± 0.427
10	3	15	25	6	0.26920 ± 0.015	4.473 ± 0.438
11	3	15	50	9	0.47360 ± 0.014	5.127 ± 0.245
12	3	15	75	3	0.12100 ± 0.046	2.967 ± 0.432
13	4	5	25	6	0.11240 ± 0.010	3.487 ± 0.213
14	4	5	50	9	0.20400 ± 0.012	4.973 ± 0.278
15	4	5	75	3	0.17890 ± 0.011	3.160 ± 0.236
16	2	10	25	6	0.16580 ± 0.009	3.032 ± 0.156
17	2	10	50	9	0.38400 ± 0.025	4.533 ± 0.432
18	2	10	75	3	0.22850 ± 0.017	1.872 ± 0.117
19	4	10	25	9	0.29660 ± 0.021	4.493 ± 0.413
20	4	10	50	3	0.08910 ± 0.006	3.357 ± 0.220
21	4	10	75	6	0.23100 ± 0.018	4.383 ± 0.264
22	2	15	25	9	0.45380 ± 0.007	3.633 ± 0.211
23	2	15	50	3	0.13040 ± 0.043	2.727 ± 0.260
24	2	15	75	6	0.66660 ± 0.007	4.400 ± 0.296
25	3	5	25	9	0.10869 ± 0.008	3.007 ± 0.301
26	3	5	50	3	0.11570 ± 0.024	3.393 ± 0.395
27	3	5	75	6	0.27340 ± 0.026	4.080 ± 0.344

. The data was then statistically analyzed with Minitab software to determine the significance and correlation of the input variables with the outcomes (MRR and SR). The levels of parameters are taken based on the literature [4,5,12].

4.1. Analysis on the Impact of EDM Parameters on MRR

The Taguchi approach aids in determining the best factors and their levels while reducing variation and increasing robustness. Taguchi’s methodology improves quality through the reduction in the influences of factors out of control in order to increase reliability and consistency of the process of manufacturing. The response table as depicted in

Table 5. Response table for SN ratios of MRR.

Level	R	Toff	Ton	I
1	−11.84	−15.42	−16.04	−18.36
2	−14.69	−13.66	−14.33	−14.00
3	−16.15	−13.60	−12.32	−10.33
Delta	4.31	1.82	3.73	8.03
Rank	2	4	3	1
Larger is better

shows that current (I) has the rank 1 and therefore is the most significant factors. The ANOVA table as depicted in

Table 6. Analysis of variance for SN ratio of MRR.

Source	DF	Seq SS	Adj SS	Adj MS	F	p	Contribution (%)
R	2	86.41	86.41	43.206	18.60	0.003	13.34
Toff	2	19.15	19.15	9.575	4.12	0.075	2.96
Ton	2	62.59	62.59	31.295	13.47	0.006	9.66
I	2	291.04	291.04	145.520	62.65	0.000	44.93
Toff*Ton	4	52.98	52.98	13.245	5.70	0.030	8.18
Toff*I	4	64.82	64.82	16.206	6.98	0.019	10.01
Ton*I	4	56.85	56.85	14.212	6.12	0.026	8.78
Residual Error	6	13.94	13.94	2.323			2.14
Total	26	647.78
Model Summary
S		R-Sq		R-Sq(adj)
1.5240		97.85%		90.68%

shows that current (I) and reinforcement percentage (R), factors are significant factors with current (I) being the most significant, followed by reinforcement percentage (R) with both having very low p-values. T_on and T_off are also important factors; however, T_off has a smaller effect on the S/N ratio. The significant interactions between the factors, including the interactions T_off × T_on, T_off × I, and T_on × I, all provide significant contributions to the final S/N ratio outcome and show that the response of these systems is complex, so it is important to consider the combined factor effects. The data show a great fit, with 97.85% of the variability in the model explained. The significant interactions further emphasize that improving the signal-to-noise ratio can take either or multi-factored approaches, with current (I) and reinforcement percentage (R) being the focus of any adjustment; however, it is also important to consider factor interactions.

Material Removal Rate (MRR) is influenced by a number of input parameters, including discharge current (I), pulse off-time (T_off), pulse on-time (T_on), and reinforcement content (R). Each has an independent impact on MRR, as illustrated in Figure 7. There is a direct correlation between the discharge current and MRR in EDM, which is observed in numerous studies [16,17]. The discharge current (I) is clearly the most potent control variable to increase material removal rate (MRR), supported by the experimental results and the physical mechanisms of the EDM process. Increasing the discharge current will increase the energy per pulse since energy is a function of current multiplied by pulse duration (E ∝ I × Ton) and also the intense electrical discharge made possible by a higher current; therefore, generating more intense plasma channels with a higher localized temperature at the electrode-workpiece interface that melts and vaporizes the workpiece material more effectively results in more material being removed per discharge. The ANOVA results of this study quantitatively validated this concept, as the current was responsible for the largest percentage (44.93%) of MRR response compared with other input factors, including pulse on time, pulse off time, and reinforcement percentage. An increase in T_on results in a proportionate rise in the energy supplied, which subsequently leads to greater energy discharge. An increase in discharge energy results in a greater volume of metal removal from the workpiece, hence enhancing the MRR [18]. The MRR of the composite was strongly influenced by the reinforcing content. An increase in B₄C content from 1 to 3 (wt.%) led to a significant decrease in MRR. The noted decrease in MRR of the composites can be ascribed to an enhancement in the strength and hardness of the composite material resulting from the integration of B₄C alongside a constant proportion of GNPs. According to Figure 7a, the ideal parameters for increased Material Removal Rate (MRR) are identified as follows: R = 2 wt.%, T_off = 15 µs, T_on = 75 µs, and I = 9 A. The interaction plot as shown in Figure 7b showcases the effects of the parameters Toff (off-time), Ton (on-time), and the current (I) on the Signal-to-Noise (S/N) ratios. As T_off is increased, the S/N ratio improves over all levels of T_on and I, indicating that as the off-time increases the performance improves [19]. Additionally, as T_on increases, the S/N ratio is reduced, indicating that less on-time improves performance. Current (I) has a clear positive influence on the S/N ratio. The greater the current, the more improved the performance, especially associated with longer T_off and T_on. It is critical to optimize the parameters of interest to achieve the maximum performance effect. Figure 7c illustrates the residual plots of MRR, which are employed to evaluate the data’s normality and randomness. In Figure 7c, the normal probability plot illustrates that the data is in proximity to the center line, suggesting high normality and minimal error. It is evident that the data points fail to demonstrate an identifiable trend in the context of the versus-fits and versus-order plots, which implies that the variations are uniform and time-independent.

4.2. Analysis on the Effect of EDM Parameters on SR

Surface roughness (SR) is an additional substantial output response in each machining process. The objective for SR data in this endeavor is to minimize SR.

Table 7. Response table for SN ratios of SR.

Level	R	Toff	Ton	I
1	−10.300	−10.985	−10.522	−9.124
2	−11.679	−11.187	−12.093	−12.127
3	−12.585	−12.392	−11.949	−13.313
Delta	2.285	1.407	1.570	4.189
Rank	2	4	3	1
Smaller is better

depicts the response table of S/N at an aggregate of three levels and for four variables, namely R, T_off, T_on, and I, where lower values are better. Current (I) has the greatest impact on SR (rank 1), followed by others. The ANOVA analysis depicted in

Table 8. Analysis of variance for SN ratios of SR.

Source	DF	Seq SS	Adj SS	Adj MS	F	p	Contribution (%)
R	2	23.836	23.836	11.9178	42.13	0.000	14.59
Toff	2	10.417	10.417	5.2085	18.41	0.003	6.38
Ton	2	13.567	13.567	6.7833	23.98	0.001	8.31
I	2	83.922	83.922	41.9610	148.32	0.000	51.39
Toff*Ton	4	10.137	10.137	2.5344	8.96	0.011	6.20
Toff*I	4	4.280	4.280	1.0699	3.78	0.072	2.62
Ton*I	4	15.445	15.445	3.8613	13.65	0.004	9.46
Residual Error	6	1.697	1.697	0.2829			1.05
Total	26	163.301
Model Summary
S		R-Sq		R-Sq(adj)
0.5319		98.96%		95.50%

showed that the model was a good fit for the variance in the dependent variable, with an R² of 98.96%, indicating that nearly 99% of the variance in the dependent variable could be attributed to the factors and their interactions. The important factors R, T_off, T_on, and the interactions all have p-values below 0.05, indicating they affect SR significantly. In particular, R, T_off, and T_on are all the most influential contributions, based on their high F-values and low p-values; the F statistically signals that the factors have importance in the model. The interaction, T_off * I, has a p-value of 0.072 which is above 0.05. This indicates it does not affect the SR significantly. Overall, the fit of the model was strong, as reflected in an adjusted R² = 95.50%, which compensates for the number of predictors used in its model structure; this supports that the model generalizes well.

Figure 8 also illustrates the impact of numerous factors on SR. The mean effects diagram of SR indicates that the SR of the machined surface increases as I and T_on increase. The higher I, the more spark energy is generated, which results in deeper cavities on the workpiece surface and a deeper penetration of the material [5,6]. The increase in surface roughness (SR) due to longer pulse off time (T_off) can be rationalized through an understanding of the discharge stability of the EDM process and the thermal dynamics involved. Toff is the time between discharges; during this time the dielectric surrounding the spark gap can recover, and debris from the previous discharge can be eliminated from the tool-workpiece gap. Increasing the Toff decreases the discharge frequency over the given time, which may initially suggest it results in a more stable discharge. However, if the Toff is too long, it will diminish the average energy input over time, which will significantly diminish material removal per unit time and consistent sparking conditions. This discontinuity can cause more inconsistent melting and ejection of the melted material and could create unstable plasma channels for subsequent discharges, therefore contributing to inconsistent surface topographies and subsequently higher surface roughness. Still, the effect of the T_off on SR is less significant than the effect of current (I), because current increases the corresponding energy delivered per pulse and has a greater immediate change in input energy or intensity of the event. In current terms, if the current is higher, the plasma channel produced is wider, and the energy transported to the workpiece from the discharge will have a higher energy transfer, with the overall effect changing both the MRR and the SR of the machined surface. The ANOVA analysis of the SR data substantiates that current showed the largest contribution to the variation in SR (51.39%). In comparison to the other two compositions, the composite with a high R demonstrates a higher level of surface irregularity. An increase in reinforcement (wt.%) leads to a corresponding increase in SR. The primary reason for this is the enhancement in the composite’s hardness as a result of the reinforcement volume. The high discharge energy of the EDM technique caused the AZ31 alloy to melt, which was subsequently removed using the dielectric fluid [20]. Due to the high melting temperature of the reinforcements (B₄C and GNPs), their disintegration and disposal during machining while working with the AZ31 are not feasible. Similar reinforcement behavior was observed in the previous investigations [21]. From Figure 8a, it is observed that the ideal parameters for minimizing the SR are R = 2 wt.%, T_off = 5 µs, T_on = 25 µs, and I = 3 A. The interaction effects illustrated in Figure 8b indicate that I and T_on have the strongest combined effects on SR, with moderate increases in both factors resulting in significantly worse surface quality (more roughness) as SNR decreases (becomes more negative). The interaction effect from T_off with I is also weak and negative, resulting in rougher surfaces when the factor levels are increased as well. The interaction between T_off and T_on also contributes to increases in roughness but the effect is the most uniform and linear as the levels of both factors are increased together. From Figure 8c, it is observed that the regression model was properly specified and that the regression assumptions of linearity, homoscedasticity, and independence were not violated. Figure 8d shows the microgeometry profiles of the composites at the surface level. From the figure it is noticed that Composite 1 shows the smoothest profile, which includes relatively low variation in peaks and valleys, meaning finer roughness and more consistent machining. Composite 2 shows moderate variation in peaks and valleys with slightly larger surface irregularities, leading to a balance of material detached and surface finish being achieved. Conversely, Composite 3 shows the highest variation in peaks and valleys, which corresponds to a coarser surface and more microstructure damage at the machining level. These differences indicate the role of reinforcement content and electrical discharge machining (EDM) parameter levels on the finish of the surface; in particular, higher reinforcement levels of Composite 3 tend to raise localized levels of spark release, leading to a deeper crater formation and poor surface finish relative to Composites 1 and 2.

4.3. SEM Analysis of EDMed AZ31 MMCs

Based on the results of the experiments, the greater MRR was obtained at R = 2 wt.%, T_off = 10 µs, T_on = 75µs, and I = 6 A, and while the lower MRR was obtained at R = 2 wt.%, T_off = 5 µs, Ton = 25 µs, and I = 3 A. In the SEM images as shown in Figure 9a,b, both representing surfaces generated with high MRR of EDM, significant surface damage including craters, melted areas, and vaporized remnant material is observed. These parameters indicate significant thermal damage due to high energy discharge processes from the EDM cutting process. The mechanisms present include melt erosion and vaporization [14]. The localized temperatures in the discharge can increase rapidly to a very high value and melt the material in contact or vaporize it. The discharge ejected the melted material, thus being responsible for the craters and the edge of the craters where the molten, now solidified, droplets exist along the edge of the crater. The thermal energy of the energized material is very high, and once the material cools, it solidifies rapidly, thus causing micro-cracking and poor surface finish from the stress of the release energy as it was melting and solidifying. In addition, the electrolyte chemical effects likely also have an effect on surface finish; the high current density in the EDM process could ionize the workpiece surface to remove material electrochemically as it was electrically discharged at high energy densities. All of these processes and thermal energies lead to a rougher and less smooth surface finish from the thermal and mechanical stresses. For SEM images as depicted in Figure 9c,d—which both depict low MRR surfaces—a smoother surface finish with smaller craters and very little melting or signs of melting material is observed. The surface appears less damaged by the extremes of thermal cycles with still high energy densities. Therefore, it is reasonable to conjecture that the thermal energy per discharge must have been less in the low MRR than in the high MRR conditions; hence, less melting and vaporization occurred in the material of the workpiece. In low MRR conditions characterized by low discharge current and short pulse on-time, the energy delivered per discharge is not sufficient to compensate for the thermal conductivity and heat dissipation properties of the AZ31/B₄C/GNPs composite. The addition of B₄C particles and graphene nanoplatelets (GNPs) increases the thermal conductivity of the composite compared to the base AZ31 matrix. The better thermal conductivity allows heat to dissipate away from the discharge site over a shorter period of time and therefore increases the localized temperature and decreases the extent of melting and vaporization of the material. Also, the GNPs were shown to affect the stability of the plasma channel due to their electrical conductivity. In the low-energy zone, the plasma channel does not fully stabilize during/after the discharge pulse, resulting in less effective material removal per discharge and a lower MRR. In the high MRR zone, with higher current and longer pulse time, it generates stronger plasma channels and larger heat-affected zones, which overcomes the heat dissipation effect and leads to more effective erosion of the material. Therefore, the differences in thermal and electrical properties of the composite and process parameters applied explain the differences in MRR and surface morphology seen with the different experimental conditions. The main mechanisms in the low MRR conditions were localized heating and micro-ablation. The total energy from the discharge was necessarily a lower order of magnitude than high MRR—this means that only a small amount of material would be melted and removed from the initial surface with every discharge, forming smaller craters. The overall smoother surface path was due to the temperature-affected areas being smaller and the volume of molten material being less. Even in the low MRR discharge conditions, the amount of cracking was less, and thermal stresses that contribute to surface roughness (on the original surface) would be reduced. The reduced turbulence would lead to a finer surface finish. The performance evaluation of the diverse ML models is examined in the subsequent section.

4.4. Performance Analysis of ML Models

The evaluation of the precision and variability of the ML models that are presented in the previous sections to estimate the EDM characteristics of AZ31 hybrid composites was conducted utilizing statistical metrics. This study employs 21 data points (80% of the dataset) for training and the other six points (20% of the dataset) for testing to evaluate the effectiveness of ML approaches. In ML, regression modeling approaches can be applied to predict a continuous numerical dependent variable from one or more independent variables or features. In the regression modeling context, performance is systematically assessed using a range of performance metrics such as MSE, RMSE, and R², because they provide complementary insights into model performance. Additional performance evaluation methods for regression models include methods like cross-validation or residual analysis. Cross-validation assesses whether or not a model may overfit or underfit the data and then generalizes to a level of performance beyond the data it was developed on, providing a broader context to its performance level. Residual analysis provides insight into the extent to which the model estimations depart from the actual values and helps to determine the best subsequent steps for refining the model. In the present work, the ML models are evaluated using the following measures [22].

$$MSE=\sqrt{{\displaystyle \frac{1}{k}}\sum _{i=1}^k{\left(z_p-z_a\right)}^2}$$

(6)

where ${\mathit{z}}_{\mathit{p}}$—predicted outcome, ${\mathit{z}}_{\mathit{a}}$—actual outcome, k—total number of data points.

$$R^2=1-{\displaystyle \frac{\sum _{i=1}^k{\left(z_p-z_a\right)}^2}{\sum _{i=1}^k{\left(z_p-{\overline{z}}_a\right)}^2}}$$

(7)

where ${\overline{\mathit{z}}}_{\mathit{a}}$—average of actual outcome values,

$$MSE={\left(RMSE\right)}^2.$$

(8)

The present investigation employs various algorithms to construct predictive models. Linear and polynomial are normal regression models. Further, other models, such as RF and GBR, utilize the combination of multiple decision tree models where allowed. To efficiently establish a relationship between input and output data, these algorithms employ a combination of decision trees. Using random samples from the dataset, the bagging approach is employed by the RF algorithm to generate each tree independently. After the tree formation process is complete, the average of the unique outputs produced by each tree is calculated to determine the outcome. In contrast, GBR implements a boosting methodology that involves the iterative enhancement of feeble decision trees to become robust learners by learning from the errors of their predecessors.

Table 9. Performance evaluation of ML models for test data.

Data	ML Model	MSE	RMSE	R2
MRR	LR	0.02583	0.1607	74.58
	PR	0.01639	0.1280	82.36
	RF	0.00582	0.07629	89.24
	GBR	0.00264	0.05138	94.68
SR	LR	0.2473	0.4973	72.83
	PR	0.1568	0.3959	78.92
	RF	0.0837	0.2893	85.53
	GBR	0.0574	0.2396	92.68

illustrates the statistical metrics of the ML models generated for the output responses of the test data (i.e., MRR and SR). The models exhibit predictive performance in estimating MRR and SR, ranging from 74.58% to 94.68% and 72.83% to 92.68%, respectively. Moreover, the values of MSE and RMSE are significantly low. The comparison of performance results of the developed predictive models shows that there are evident benefits of advanced machine learning techniques relative to simple regression methods. While the simplest methods (LR and PR) allow for direct acceptance and interpretation of results, LR assumes a linear relationship, and PR can approximate some nonlinearity through polynomial predictor variables. Simple linear and polynomial methods fall short of displaying the complex, nonlinear nature of interactions that exist between EDM parameters and outcomes (MRR and SR). Random Forest (RF) Regression is advantageous in that it provides better predictive capabilities than single-source regression techniques because it constructs an ensemble of decision trees that combine predictions. It retains the ability to represent complex interactions while modeling nonlinear effects without a required functional equation. In this way it is more robust; by using the mean of all predictions from the different trees, the errors from the model trees are canceled out. Deterring the overfitting associated with a single-source tree and generalization to new datasets during classification. In this work, Gradient Boosting Regression (GBR) performed best due to the predictive capabilities of additional trees. GBF builds trees sequentially, meaning the tree model captures observations that reduce the residual error missing from the previous trees. Thus, the sequential model identifies subtle and complex interactions present in the total datasets. As illustrated in the finding, GBR outperformed all other models, achieving accuracies above 92% and the lowest errors predicting MRR and SR.

Figure 10a,b illustrate the errors encountered in the prediction of test data across different models. According to the statistical indices, the Gradient Boosting Regressor (GBR) exhibited superior predictive accuracy for the MRR and SR, exceeding 92%, as seen in Figure 10c. The improved performance of RF and GBR over simpler regression techniques comes from their use of ensemble learning, which reduces variance and thus improves the accuracy of predictions. RF and GBR both comprise a number of decision trees (DT) built and trained using the provided data. Each of these methods builds trees in a different way. In RF, if a tree is built and provided with training data randomly selected from the original data, the overall final prediction is determined by the mean of all the predictions, which drastically reduces the influence of any one tree that has overfitted some noise or outliers in training data. This averaging effect reduces the prediction variance of a single decision tree model. GBR works slightly differently, as it builds a tree that minimizes the residual error of the previous tree; thus, each tree is built to build and improve upon the previous tree. By doing so, the resultant model does a good job modeling complex nonlinear relationships between parameters and target responses, particularly where those responses can be impacted by material-specific issues, as with EDM processes. As a result, RF’s ensemble strategy and GBR’s boosting capability will improve model robustness, decrease prediction variance, and have overall lower error and higher R² statistics than a simpler model (LR or PR). These advanced methodologies are particularly advantageous in predicting complex manufacturing processes, offering significant improvements in predictive reliability and practical applicability.

5. Conclusions

The main objective of this study is to examine the impact of various EDM parameters, including R, T_off, T_on, and I, the SR and MRR of AZ31/B₄C/GNPs composites, combining experimental analysis with advanced machine learning modeling. The Taguchi method was applied to attain an optimized experimental design. Numerous ML techniques (LR, PR, RF, and GBR) have been employed to develop an effective predictive model that integrates the optimal parameter values. The principal conclusions obtained from this investigation are detailed below.

• According to the study’s findings, the current (I) is the most significant factor in relation to the material removal rate (MRR) with 44.93% of contribution, with reinforcement percentage (R), pulse on time, and pulse off time following in that order.
• The main effect S/N graphs of MRR demonstrate that the optimal parameters for attaining a greater MRR are R = 2 wt.%, T_off = 15 µs, T_on = 75 µs, and I = 9 A.
• The study’s findings demonstrate that the current (I) possesses the greatest significance concerning SR, with 51.39% of contribution, succeeded by reinforcement percentage (R), T_on, and T_off. The ideal parameters for minimizing the SR are R = 2 wt.%, T_off = 5 µs, T_on = 25 µs, and I = 3 A.
• Prediction models were developed using LR, PR, RF, and GBR methods to predict MRR and SR of AZ31 composites. The GBR model provided the best predictive performance, outperforming the other methods with an accuracy of 94.68% and 92.68% for MRR and SR, respectively.

In summary, this study provides two contributions: the overall understanding of how EDM parameters impact hybrid composite machining characteristics and a robust prediction tool via GBR modeling, allowing the EDM process to be quickly optimized for industry application. The overall benefit is the ability to accurately model process outcomes with GBR without the need for trial-and-error experimentation, which saves time, cost, and material when producing advanced composites.

6. Outlook

The results of this study provide valuable insights into how EDM process parameters and composite reinforcement affect the machining performance of AZ31/B₄C/GNPs hybrid composites; however, there are still many paths that can be developed further. First, while the current study only focused on MRR and SR as the output responses, future research should include tool wear, dimensional accuracy, microhardness, and residual stress, etc., in order to holistically evaluate the EDM performance. These factors play important roles in industrial applications where, while productivity matters, surface integrity and tool life also matter greatly. Second, the predictive models created in this research were trained on a limited dataset, which was produced using a Taguchi L27 design of experiments. A greater sample size with a wider variety of input parameters and material compositions could enable broader generalizability and greater robustness and reliability of the machine learning model predictions, particularly for industrial applications. In addition, more advanced techniques such as deep learning or other hybrid modeling techniques that combine physics-based simulations with data-driven models could potentially lead to improved accuracy in predictions and better insight into the physical phenomena being modeled. Future work will also focus on developing mathematical models that link EDM responses to structural parameters, enabling predictive optimization for industrial applications.