The Performance Prediction of Electrical Discharge Machining of AISI D6 Tool Steel Using ANN and ANFIS Techniques: A Comparative Study

: AISI-D6 steel is widely used in the creation of dies and molds. In the present paper, ﬁrst the electrical discharge machining (EDM) of the aforementioned material is performed with a testing plan of 32 trials. Then, artiﬁcial neural networks (ANN) and adaptive neuro-fuzzy inference system (ANFIS) were applied to predict the outputs. The effects of some signiﬁcant operational parameters— speciﬁcally pulse on-time (Ton), pulse current (I), and voltage (V)—on the performance measures of EDM processes such as the material removal rate (MRR), tool wear ratio (TWR), and average surface roughness (Ra) are extracted. To lead the process operators, process plans (i.e., parameter–effect correlations) are created. The outcomes exposed the upper values of pulse on-time caused by higher amounts of MRR and Ra, and likewise lower volumes of TWR. Furthermore, growing the pulse current resulted in upper volumes of the material removal rate, tool wear ratio, and surface roughness. Besides, the higher input voltage resulted in a lower amount of MRR, TWR, and Ra. The estimation models developed by using experimental data recounting MRR, TWR, and Ra. The root means the square error was used to determine the error of training models. Furthermore, the estimated outcomes based on the models have been proven with an unseen validation set of experiments. They are found to be in decent agreement with the experimental issues. The investigation shows the powerful learning capability of an ANFIS model and its advantage in terms of modeling complex linear machining processes.


Introduction
The difficulty in treating hard-to-cut materials impacted the start of numerous progressive machining approaches such as water jet machining, electric discharge machining, laser machining, and electrochemical machining. These processes, commonly termed nontraditional machining procedures, create energy to remove residual material from the stock to create the preferred portion. Of these procedures, electric discharge machining (EDM) has received ample attention from the nuclear, aerospace, and automobile subdivisions [1]. Regardless of the extraordinary price of the apparatus (e.g., electron beam processing, water jet, and laser) or the disadvantage of dangerous slope (e.g., electrochemical machining), it is still superior to other methods. The electrode modeling tool has a shape that is the opposite of the cut profile. The additional material is censored through a sequence of sparks among the tool and the workpiece in the form of a dielectric medium that works as an insulator to avoid the material declaration reserved externally to the device. The EDM procedure has found numerous uses such as fabrication of molds/dies and the cutting of holes in a variety of materials containing metals and composites [2] by vast distance to diameter ratio.
Various research has attempted to improve the quality attributes of the EDM process through a change in the process parameters. Boujelbene et al. [3] inspected the outcome of energy ignition on the tool wear ratio (TWR), material removal rate (MRR), and the depth of altered layer (i.e., the material re-solidified on the work-surface) in the machining of two steels: X200Cr15 and 50CrV5. They described the depth of the altered layer and the augmented MRR and the obstinately condensed TWR as the sparking energy was enlarged. Pradhan et al. [4] considered the performance of copper and aluminum electrodes on the EDM of EN-8 steel to display that the copper electrode executed better in terms of outward value. Besides, they found that the surface quality declined with a rise in the pulse-on-time (200 µs to 700 µs) over the reflected range of peak current (8 A to 24 A). Amorim et al. [5] examined the effect of polarized graphite and copper electrodes on the EDM performance of AISI P20 steel. As demonstrated by the outcomes, advanced MRR was achieved when the dispensation was prepared with the graphite electrode but good surface quality was appreciated when machining was approved with the copper electrode. Jefferson et al. [6] suggested that application of cryogenic usage on the electrode could improve the surface quality of the work piece.
Bhupinder Singh and J.P. Misra [7] used RSM and ANN modeling to study the apparent finish investigation of electric wire discharge machined samples through. This research determined the optimal morals of factors for the WEDM of nickel-based super alloys that will offer options to engineering production and to machine tool workers depending upon their occupation necessities. Vishnu P et al. [8] worked on the presentation forecast of electric discharge machining of inconel-718 by ANN. They used backpropagation algorithms to forecast performance characteristics, specifically MRR, SR, and TWR.
Tebni et al. [9] detected the influence of the difference in the pulse current, pulse on-time, and pulse off-time on the surface quality, MRR, and altered layer depth on two steels (50CrV4 and X200Cr15). They offered to work with a weak current pulse with short time if the surface quality was mainly neutral, and they consistently chose high sets of these factors if production speed was the main objective. Moreover, in order to reduce the depth of the replaced layer, low saving of input energy was proposed. The same results have been noted in the literature on 40CrMnNiMo864 steel [10]. Muthuramalingam and Mohan [11] observed the influence of a discharge current on machinability in EDM to show that the adapted ISO pulse achieved better surface quality than the prepared predictable pulse.
Chandramouli et al. [12] closely studied the outcome of MRR, TWR, and Ra by using input properties such as pulse-off time, current, and pulse-on time. Kibra et al. [13] described the outcome of diverse dielectrics on MRR, TWR, over-cut, and surface reliability during the micro-EDM of Ti-6Al-4V with a tungsten electrode. Retaining a copper electrode, Jeykrishnan et al. [14] operated on EN24 tool steel, optimizing procedural factors and expending the Taguchi technique. Yongfeng et al. [15] investigated an experimental study of EDM factors for zrb2-sic ceramics machining, and they discussed the effect of EDM factors on ZrB2-SiC ceramics EDM machining technique.
Employing a Cu-W electrode, Marafona [16] optimized the model through the design of an experimental approach. Guo and Tsai [17], using a genetic algorithm, developed an optimum model for processing BaTiO semi-conductive material. Shrivastava and Dubei [18] performed an intelligent modeling and multi-objective optimization of electric discharge diamond grinding. Baraskar et al. [19] applied a mixture of the genetic algo1ritm and the response surface methods to optimize the EDM of EN8 steel. Ramesh Raju et al. [20] worked on optimizing process parameters in the electrical discharge machining of haste alloy C276 using Taguchi's method. In this investigation, an attempt was made to determine the optimum process variables for obtaining better machining performance in terms of the material removal rate and the surface roughness. Basha et al. [21] worked on the experimental study of electrical discharge machining of Inconel X-750 using a tungstencopper electrode. The results showed that the MRR increased with an increase in the Crystals 2022, 12, 343 3 of 18 discharge current and pulse on-time, and the lower surface roughness was obtained at the initial conditions of the discharge current and pulse-on time.
Tsai and Wang [22] have proven that besides processing parameters, the material's physical properties such as heat conductivity, specific heat capacity, boiling point, melting point, and electrical conductivity play an essential role in determining the performance of the EDM process. From this research, it is understandable that EDM process performance is highly dependent on the material used. Thus, the setup parameters that are optimal for performance tend to vary between test materials. Therefore, a search should be made for each of the responsible parameters for each material.
Rajesh and Gagandeep [23] have investigated the effects of process parameters on the performance of electrical discharge machining of AISI M42 high-speed tool steel alloy. The objective of the study was to determine the effect of machining parameters on MRR such as the current, tool polarity, pulse on-time, and gap voltage for AISI M42 alloy. Singh et al. [24] instigated a mathematical model to predict MRR during gas-based EDM.
Khalid Al-Ghamdi and Osman Taylan [25] conducted a comparative study on modeling the material removal rate by ANFIS and on polynomial methods in the electrical discharge machining process. The results for this study showed that the ANFIS model with 21 rules was the best. Singh et al. [26] worked on predictive analysis of surface roughness in EDM using semi-empirical, ANN, and ANFIS techniques. In this research, a mathematical model was actuated to realize the SR by utilizing dimensional analysis hypothesis. Bobbili et al. [27] completed a comparative report on the wire electric discharge machining of materials used in defense for making arms with the dimensional method related to MRR and SR. In other investigations, Singh and Singh [28] put forward a mathematical model to estimate SR during gas-based EDM. Their findings revealed that the developed model predicted the responses with agreeable accuracy.
The AISI-D6 steel offers excellent corrosion resistance and high toughness at elevated temperatures. Thus, it is used extensively in the making of forming dies and injection molds. However, due to its high hardness, it is not an easy process to cut through typical mechanical machining. Therefore, different solutions with no mechanical machining means are to be used. In this study, the EDM process is used to cut AISI-D6 steel. To investigate the needed parameters offering the optimal performance, some essential input parameters-namely pulse on-time, pulse current, and voltage-are varied over tests. Their output effects such as MRR, TWR, and Ra are recorded. The drawn process maps provide guidelines for understanding the process users for quality EDM of AISI-D6 material. Moreover, adaptive neuro-fuzzy inference system (ANFIS) and artificial neural network (ANN) machine-learning approaches have been applied to the output estimation of the parameters. The control variables used for MRR, TWR, and Ra are pulse current, pulse on-time, and voltage.

Work-Piece and Tool
The material for the electrode tool used for the EDM process needs to be electrically conductive. There is a wide range of materials that can be used to manufacture electrodes such as graphite, electrolytic copper, brass, tungsten carbides, silver-tungsten alloy, copper tungsten alloys, tellurium-copper alloys, and copper-graphite alloys. For the present study, an electrode made of electrolytic copper with a machining interface of 18 mm diameter and the AISI-D6 tool steel hardened to 60 HRC were, respectively, employed as the electrode and experimental materials. The test pieces and electrodes were precisely machined to the sizes given in Table 1 and Figure 1.   Table 2 shows the parametric conditions for conducting the experimental t controlling parameters pulse on-time (Ton) and pulse current (I) each was va four levels. The steel was usually cut in the range of 150 V to 250 V [29,30]; there specific parameter was varied over only two levels so that the number of tests minimized. The other parameters were set as: machining gap = 2 mm; duty cycle polarity = positive. Kerosene oil was used as the dielectric medium. A threefactorial experiment comprising 4 × 4 × 2 = 32 runs was conducted. In all, 32 exp as listed in Table 2, were performed on the Azarakhsh Ayzvpals CNC EDM syste in Figure 2.   Table 2 shows the parametric conditions for conducting the experimental trials. The controlling parameters pulse on-time (Ton) and pulse current (I) each was varied over four levels. The steel was usually cut in the range of 150 V to 250 V [29,30]; therefore, this specific parameter was varied over only two levels so that the number of tests could be minimized. The other parameters were set as: machining gap = 2 mm; duty cycle = 20; and polarity = positive. Kerosene oil was used as the dielectric medium. A three-level fullfactorial experiment comprising 4 × 4 × 2 = 32 runs was conducted. In all, 32 experiments, as listed in Table 2, were performed on the Azarakhsh Ayzvpals CNC EDM system shown in Figure 2.

Performance Measures in EDM
After the completion of every testing scenario, the relative output characteristics, such as MRR, TWR, and Ra were measured. The MRR and TWR weighing process was done by checking the difference in the masses of steel and copper tools before and after use in the machining process using a digital precision scale Mettler Toledo AB265 with an accuracy of ±0.00001 g. The MRR in mm 3 /min and TWR in g/g were calculated employing the following relationships:

10
(1) where MRR is the material removal rate; and TWR is the tool wear ratio. The weight of workpiece (in grams) before and after machining are are the weight of tool (in grams) before and after machining. The density of the workpiece in g/cm 3 is W ρ ; t is the machining time in minutes; and T ρ is the density of the Cu tool (i.e., 93.8 g/cm 3 ).

Result and Discussion
In experimental applications, the effect of the input parameters on the output performance of electronic discharge machining differs from theory. Thus, it is recommended to set the I and the Ton to high values and the voltage to low values if productivity is the prime objective during the electronic discharge machining of AISI-D6 steel. However, contrarily, the Ton and the I need to be set to high values and the voltage must be kept low when the tool wear ratio is the major objective. To obtain less surface roughness, low values for the pulse on-time and the pulse current and large values for the voltage are recommended.

Machine Learning Algorithms
Machine learning is a branch of artificial intelligence. In this research, artificial intelligence teaches computers to do what a human operator may do, i.e., regressive learning. As the learning samples increase, the algorithm's performance improves adaptively [31]. Deep learning began from artificial neural networks which is a subcategory of machine

Performance Measures in EDM
After the completion of every testing scenario, the relative output characteristics, such as MRR, TWR, and Ra were measured. The MRR and TWR weighing process was done by checking the difference in the masses of steel and copper tools before and after use in the machining process using a digital precision scale Mettler Toledo AB265 with an accuracy of ±0.00001 g. The MRR in mm 3 /min and TWR in g/g were calculated employing the following relationships: where MRR is the material removal rate; and TWR is the tool wear ratio. The weight of workpiece (in grams) before and after machining are M W1 , M W2 , and M T1 , M T2 are the weight of tool (in grams) before and after machining. The density of the workpiece in g/cm 3 is ρ W ; t is the machining time in minutes; and ρ T is the density of the Cu tool (i.e., 93.8 g/cm 3 ).

Result and Discussion
In experimental applications, the effect of the input parameters on the output performance of electronic discharge machining differs from theory. Thus, it is recommended to set the I and the T on to high values and the voltage to low values if productivity is the prime objective during the electronic discharge machining of AISI-D6 steel. However, contrarily, the T on and the I need to be set to high values and the voltage must be kept low when the tool wear ratio is the major objective. To obtain less surface roughness, low values for the pulse on-time and the pulse current and large values for the voltage are recommended.

Machine Learning Algorithms
Machine learning is a branch of artificial intelligence. In this research, artificial intelligence teaches computers to do what a human operator may do, i.e., regressive learning. As the learning samples increase, the algorithm's performance improves adaptively [31]. Deep learning began from artificial neural networks which is a subcategory of machine learning. Machine learning usually implements neural network-based operations such as deep learning. The application of deep learning is available in all industries from automated driving to medical devices [32]. Wuest and unsupervised machine learning algorithm. Supervised machine learning was suitable for most manufacturing applications mainly because manufacturing applications provided labeled data [33].
Machine learning (ML) applications are used in all areas of industry. Machine learning approaches are implemented in procedural compliance, documentation of process and orientation, and risk and quality frameworks in the manufacturing industry. Machine learning is utilized in cloud computing, data science, and IoT. The ability of machine learning approaches to anticipate failure in advance of its occurrence is a cost and risk minimizing approach that is being implemented by most industries [34].

Artificial Neural Network (ANN)
A subcategory of statistical machine learning, neural network (NN), is often used in various kinds of prediction tasks. Artificial neural networks is the most commonly used branch of neural networks; they work very similarly to brain neurons. Due to their accuracy in predicting output over other methods of prediction of non-linear input variables, ANNs have recently been emerging as a forecasting solution.
In this study, a MATLAB computer program was employed to plan the best ANN structure. The information layer was identified with pulse current, pulse on-time, and voltage. The yield layer was compared to the MRR, TWR and R a . In the proposed model, the data layer was identified with a hidden layer neuron and the concealed layer was related to yield layers. After expansive fundamentals and based on investigation of the network, the ANN model for the MRR, TWR, and R a was created.
The ANN architecture is defined by the way in which the neurons are interconnected. The network is fed with a set of input-output pairs, and it is trained to reproduce the output. The structure of each ANN is represented as (i, j, k), where i expresses the number of nodes in the input layer, j the nodes in the hidden layer, and k the nodes in the output layer. In Figure 3, the typical structure of a multi-layer ANN model is presented. In this example a model of an ANN (3-10-3) structure is presented with three variables (T on , I, and V) in the input layer, 10 nodes in the hidden layer, and 3 nodes (MRR, TWR, and R a ) in the output layer [35]. learning. Machine learning usually implements neural network-based operations such as deep learning. The application of deep learning is available in all industries from automated driving to medical devices [32]. Wuest, Weimer, Irgens & Thoben (2016) distinguished the supervised and unsupervised machine learning algorithm. Supervised machine learning was suitable for most manufacturing applications mainly because manufacturing applications provided labeled data [33]. Machine learning (ML) applications are used in all areas of industry. Machine learning approaches are implemented in procedural compliance, documentation of process and orientation, and risk and quality frameworks in the manufacturing industry. Machine learning is utilized in cloud computing, data science, and IoT. The ability of machine learning approaches to anticipate failure in advance of its occurrence is a cost and risk minimizing approach that is being implemented by most industries [34].

Artificial Neural Network (ANN)
A subcategory of statistical machine learning, neural network (NN), is often used in various kinds of prediction tasks. Artificial neural networks is the most commonly used branch of neural networks; they work very similarly to brain neurons. Due to their accuracy in predicting output over other methods of prediction of non-linear input variables, ANNs have recently been emerging as a forecasting solution.
In this study, a MATLAB computer program was employed to plan the best ANN structure. The information layer was identified with pulse current, pulse on-time, and voltage. The yield layer was compared to the MRR, TWR and Ra. In the proposed model, the data layer was identified with a hidden layer neuron and the concealed layer was related to yield layers. After expansive fundamentals and based on investigation of the network, the ANN model for the MRR, TWR, and Ra was created.
The ANN architecture is defined by the way in which the neurons are interconnected. The network is fed with a set of input-output pairs, and it is trained to reproduce the output. The structure of each ANN is represented as (i, j, k), where i expresses the number of nodes in the input layer, j the nodes in the hidden layer, and k the nodes in the output layer. In Figure 3, the typical structure of a multi-layer ANN model is presented. In this example a model of an ANN (3-10-3) structure is presented with three variables (Ton, I, and V) in the input layer, 10 nodes in the hidden layer, and 3 nodes (MRR, TWR, and Ra) in the output layer [35].  In this investigation, the ANN prediction model is trained for each component using the Levenberg-Marquardt algorithm which shows a stable and a fast convergence. Figure 4 reveals the design of this ANN: 3 layers with full connection and 3 input nodes are logged into the input layer to describe 3 outputs. The input nodes include pulse-on time, pulse current, and voltage. The output of this design is MRR, TWR, and R a . Additionally, a total of seventy percent (70%) of the experimental data was used for training, with 15% used for validation and testing, respectively. The training algorithm used was the Levenberg-Marquardt algorithm [36]. In this investigation, the ANN prediction model is trained for each component using the Levenberg-Marquardt algorithm which shows a stable and a fast convergence. Figure 4 reveals the design of this ANN: 3 layers with full connection and 3 input nodes are logged into the input layer to describe 3 outputs. The input nodes include pulse-on time, pulse current, and voltage. The output of this design is MRR, TWR, and Ra. Additionally, a total of seventy percent (70%) of the experimental data was used for training, with 15% used for validation and testing, respectively. The training algorithm used was the Levenberg-Marquardt algorithm [36].

Adaptive Neural Fuzzy Inference System (ANFIS)
The word ANFIS means an adaptive neuro-fuzzy inference system that uses a specified input/output data collection; the ANFIS toolbox creates a fuzzy inference system (FIS), and its subscription includes calibrated (tuned) variables) either using a backpropagation method on its own or in combination with a least squares type approach. This modification allows the fuzzy structures to learn from the data they are processing. The neuro-adaptative learning approach works in a similar way to that of neural networks. Neuro-adaptative learning methods provide a mechanism for a fuzzy modeling system to learn details about data collection. The Mamdani fuzzy inference system's fundamental structure is a model that maps input features to input membership functions, input membership functions to principles, rules to a set of output features, output features to output membership functions, and output membership functions to a single-valued output or output-related decision. Such a system uses fixed membership functions that are chosen arbitrarily and a rule structure that is essentially predetermined by the user's interpretation of the characteristics of the variables in the model. The fuzzy inference style utilized in this paper contains three inputs, three MFs for every input, and two rules. The Takagi-Sugeno fuzzy design aimed to be consistent with the two IF rules constructed as follows [37]. The ANFIS utilized in this investigation was settled with MATLAB. Figure 12 demonstrates the view of the developed ANFIS model. Figure 11. The plot of all data regression for the data set of R a .

Adaptive Neural Fuzzy Inference System (ANFIS)
The word ANFIS means an adaptive neuro-fuzzy inference system that uses a specified input/output data collection; the ANFIS toolbox creates a fuzzy inference system (FIS), and its subscription includes calibrated (tuned) variables) either using a backpropagation method on its own or in combination with a least squares type approach. This modification allows the fuzzy structures to learn from the data they are processing. The neuro-adaptative learning approach works in a similar way to that of neural networks. Neuro-adaptative learning methods provide a mechanism for a fuzzy modeling system to learn details about data collection. The Mamdani fuzzy inference system's fundamental structure is a model that maps input features to input membership functions, input membership functions to principles, rules to a set of output features, output features to output membership functions, and output membership functions to a single-valued output or output-related decision. Such a system uses fixed membership functions that are chosen arbitrarily and a rule structure that is essentially predetermined by the user's interpretation of the characteristics of the variables in the model. The fuzzy inference style utilized in this paper contains three inputs, three MFs for every input, and two rules. The Takagi-Sugeno fuzzy design aimed to be consistent with the two IF rules constructed as follows [37]. The ANFIS utilized in this investigation was settled with MATLAB. Figure 12 demonstrates the view of the developed ANFIS model.
The planned ANFIS designs for the output parameters is shown in Figure 13. It includes 3 nodes in the input layer, 100 nodes in the hidden layer, and 3 nodes (MRR, TWR, and R a ) in the output layer. Figures 14-16 exhibit the contour and the 3D graph of MRR, TWR, and Ra values with different input parameters. The results of the graphs show that MRR increases with the increase in the pulse current and the pulse on-time. On the other hand, as the pulse current decreases, the MRR decreases. Figure 17 shows the graph of the estimated values versus the actual values for MRR, TWR, and R a . The results prove that the estimated values are in good agreement with the actual responses. The root mean square error (RMSE) of the ANFIS training model for MRR, TWR, and R a are 0.81, 0.28, and 0.17, respectively. The precision of the ANFIS model relies on a couple of essential variables, which are listed in Table 3. The planned ANFIS designs for the output parameters is shown in Figure 13. It includes 3 nodes in the input layer, 100 nodes in the hidden layer, and 3 nodes (MRR, TWR, and Ra) in the output layer. Figures 14-16 exhibit the contour and the 3D graph of MRR, TWR, and Ra values with different input parameters. The results of the graphs show that MRR increases with the increase in the pulse current and the pulse on-time. On the other hand, as the pulse current decreases, the MRR decreases. Figure 17 shows the graph of the estimated values versus the actual values for MRR, TWR, and Ra. The results prove that the estimated values are in good agreement with the actual responses. The root mean square error (RMSE) of the ANFIS training model for MRR, TWR, and Ra are 0.81, 0.28, and 0.17, respectively. The precision of the ANFIS model relies on a couple of essential variables, which are listed in Table 3.
.   The planned ANFIS designs for the output parameters is shown in Figure 13. It includes 3 nodes in the input layer, 100 nodes in the hidden layer, and 3 nodes (MRR, TWR, and Ra) in the output layer. Figures 14-16 exhibit the contour and the 3D graph of MRR, TWR, and Ra values with different input parameters. The results of the graphs show that MRR increases with the increase in the pulse current and the pulse on-time. On the other hand, as the pulse current decreases, the MRR decreases. Figure 17 shows the graph of the estimated values versus the actual values for MRR, TWR, and Ra. The results prove that the estimated values are in good agreement with the actual responses. The root mean square error (RMSE) of the ANFIS training model for MRR, TWR, and Ra are 0.81, 0.28, and 0.17, respectively. The precision of the ANFIS model relies on a couple of essential variables, which are listed in Table 3.
. Figure 13. The design of the ANFIS model for output parameters (MRR, TWR, and Ra). Figure 13. The design of the ANFIS model for output parameters (MRR, TWR, and R a ).

Prediction Error
The precision of the prediction model was evaluated by using the root mean square error (RMSE) [38]. The accompanying condition can be utilized to get the RMSE.
where N is the complete training data, p i is the estimation of the deliberate information, and q i is the worth, anticipated by the ANFIS model. The adequacy of the created model was checked by the mean square error (MSE), root mean square error (RMSE), and standard deviations [39], and it is depicted in Tables 4-6. From these evaluations, it very well may be surmised that the prompted model has the dynamically accurate expectation.

Prediction of Output Responses
With the help of an artificial neural network (ANN), an adaptive neural fuzzy inference system (ANFIS), and the prophetic value of MRR, TWR and R a were detected for observation in training data sets as depicted in Tables 4-6. Two parameters such as MSE and RMSE were used to examine the outcome of models for the better judgment of the MRR, TWR, and R a value obtained through the ANN and ANFIS methodologies. The values of the MSE and RMSE for all the models were specified in Table 7. The outcomes show that the actual responses are in near agreement with the predicted responses. The root mean square error (RMSE) of the ANN and the ANFIS models are 1.03 and 0.81 for MRR; 1.17 and 0.28 for TWR; and 0.33 and 0.17 for R a , which proves that ANFIS models are relatively superior to other ML techniques.

Conclusions
In the current investigation, two models-artificial neural network (ANN) and adaptive neuro-fuzzy inference system (ANFIS)-have been utilized to estimate the MRR, TWR, and R a, in the EDM of AISI-D6 tool steel. The important conclusions of the study are summarized as follows: (1) As the pulse current and pulse on-time increase, the material removal rate increases; and, conversely, it decreases as the voltage increases, especially when the pulse current is above 30 A. (2) The tool wear ratio increases as the pulse current increases, it decreases when the voltage increases; and as the impact on-time increases, it decreases in contrast to the increase in the metal removal rate. (3) In terms of roughness, surface quality improves with decreasing pulse current and pulse on-time, and it improves as the voltage is increased. (4) Both the accuracy of prediction and the suitability for use of these models are considered to support the forecast. The results indicate that the ANFIS approach is relatively superior to other ML techniques, providing more reliable and accurate results in terms of lower RMSE (0.81, 0.28, and 0.17) for output parameter requirements in electric discharge machining. (5) The trends presented in this study are expected to act as guidelines for users to set the parameters in order to achieve their desired objective. Furthermore, from the above findings, it follows that the effect of input parameters on various performance measures of the process are opposing in nature. Therefore, to acquire a trade-off among all of the considered measures, the parameters should be set to intermediate values of the settings employed in this study. (6) In the future, hybrid models can be applied to further increase the accuracy of forecasts.
Hybrid models combine machine learning techniques with optimization algorithms. They are more powerful than single models as they commonly incorporate the advantages and they compensate for the weaknesses of the individual techniques involved, improving forecasting accuracy. Hybrid models can be created with one or more phases, corresponding to different problem-solving goals.