The International Journal of Advanced Manufacturing Technology (2025) 137:251–262

[Link]

ORIGINAL ARTICLE

Investigating the effect of process parameters on surface roughness

of AISI M2 steel in EDM using deep learning neural networks
Jabbar k. Abbas1 · Shukry H. Aghdeab2 · Amin Al‑Habaibeh3

Received: 11 June 2024 / Accepted: 4 February 2025 / Published online: 14 February 2025
© The Author(s) 2025

Abstract
This paper presents a unique and novel empirical study of electrical discharge machining (EDM) supported by artificial
intelligence and statistical analysis. Electrical discharge machining (EDM) is a general non-traditional machining process
for machining geometrically complicated parts or hard materials that are very difficult to machine by traditional machining
operations. EDM creates the material removal process by using electric spark erosion. This paper experimentally investi-
gates the process parameters of EDM on high-speed steel AISI M2 as a workpiece material with copper and brass as the
electrodes. The effect of various process parameters on machining performance is investigated in this study using AI and
statistical analysis where current, pulse on time and pulse off time are used for the experimental work and their effect on
surface roughness (Ra) are studied. The results of the present work show that the optimum Ra levels in copper and brass
electrodes are 2.16 µm and 3.43 µm, respectively, at current of 10 A, pulse on time of 100 µs, and pulse off time of 25 µs. The
high levels of Ra in copper and brass electrodes are found to be 6.37 µm and 7.93 µm, respectively, at current of 42 A, pulse
on time of 200 µs, and pulse off time of 4 µs. Deep learning neural networks and statistical analysis are used to evaluate the
results. It has been found that there is a significant correlation between the process current and average surface roughness,
and the pulsation time was not found significant. The use of deep learning neural networks has shown that AI could predict
the average Ra values with an average error of about 0.39% for copper and of 0.26% for brass indicating the benefits of using
AI in predicting the performance of manufacturing processes and the potential use of AI in future process modelling and
applications. The drive to increase productivity and enhance quality is attracting manufacturers into adopting Industry 4.0
and artificial intelligence in their facilities to increase flexibility, reduce waste, and enhance efficiency. EDM is considered
to be one of the most complex operations in manufacturing due to its high variability. Therefore, this paper suggests the use
of deep learning neural networks to model the process and to predict the surface roughness outcome with limited input data.
Statistical analysis was also used to test the statistical significance of each process parameter on the outcome.

Keywords AI · Electrical discharge machining · Surface roughness · Copper and brass electrodes · Neural Networks · Lean
Manufacturing

1 Introduction

Electrical discharge machining (EDM) is a non-conventional

material removal process which is widely used in industry
[1]. EDM sometimes named spark machining, spark erod-
* Amin Al‑Habaibeh ing, burning, die sinking, or wire erosion. It is a machining
[Link]-Habaibeh@[Link] process whereby a coveted shape is obtained using electri-
cal sparks. Material is taken away from the workpiece by a
1
Department of Scholarships and Cultural Relations, Ministry series of frequent periodic electric discharges between an
of Higher Education, Baghdad, Iraq
electrode and a workpiece that are separated by a dielec-
2
Nanotechnology and Advanced Materials Research Center tric solution and subjected to an electric voltage [2]. EDM
(NAMRC), University of Technology, Baghdad, Iraq
is basically a thermal process with a complicated metal
3
Product Innovation Centre, Nottingham Trent University, removal mechanism, which includes the formation of a
Nottingham NG1 4FQ, UK

Vol.:(0123456789)
252 The International Journal of Advanced Manufacturing Technology (2025) 137:251–262

plasma channel between the electrode and workpiece, result- the reference is surface roughness. MRR and TWR values
ing in the material removal, which also leads to variations in the experimental work are calculated using mathematical
in surface roughness, work hardness, hole magnification, methods and ANOVA analysis. The results of their study
white layer, heat effected layer, surface structural changes, have outlined the criticality of process parameters on the
metallurgical transformations, and the possibility of cracking product and process characteristics.
[3]. Hence, the process is extremely complicated to predict Titanium alloy machining was also investigated by Rah-
its expected outcome. These characteristics determine the man et al. [10]. They have studied an optimisation model
operational behaviour of the material and the process and to inspect the influence of peak current, pulse on time, and
govern surface integrity [3]. EDM process has an advantage pulse off time in EDM performance on material removal
over other material removal technologies due to its ability rate (MRR) of titanium alloys using copper tungsten with
to make complex parts with a higher accuracy especially positive polarity of the electrode. The findings indicate
for hard materials that are difficult to machine by other desirable MRR and economical industrial machining
means. Another advantage of EDM processes is their abil- results with optimized process parameters. Rajesh et al.
ity to machine parts on a very small scale. As a result, EDM [11] have investigated the discharge energy that is gener-
processes are normally used in the production of industrial ated during the EDM operation to characterize the pro-
tools, dies and moulds, and the machining of heat-treated ductivity of the process. The material removal process of
ceramic materials such as tungsten carbides [4]. EDM is modelled by obtaining the actual discharge energy
Significant research has been found to cover EDM charac- during the spark occurrence. This is achieved by analysing
terization. For example, Daneshmand et al. [5] have focused the real-time voltage and current waveforms acquired dur-
on the effects of EDM parameters such as the discharge cur- ing the EDM operation to determine the discharge power.
rent, voltage, pulse on time and pulse off time on the mate- The surface morphology of the workpiece and tool surface
rial removal rate, tool wear, and surface roughness of NiTi through SEM and EDS analysis shows the presence of dif-
alloy. The results of their research indicate that the current, ferent anomalies and confirms surface alloying. Electrode
voltage, and pulsation rate have a direct proportional effect wear rate (EWR) using different fluids was investigated by
on the material removal rate (MRR). The results also show Aiman et al. [12] where they have presented experimen-
that tool wear rate (TWR) decreases with the increase of tal investigations on the influence of electrical discharge
pulse off time. In terms of materials, Prajapati et. al. [6] have machining (EDM) input parameters on electrode wear rate
studied the performance of different electrode materials on (EWR) while using EDM oil and vegetable oil as dielectric
EN-9 workpiece with EDM process, where tools’ materials fluid. The feasibility of the EDM process is performed
are graphite, copper, and brass. The key process parameters using RBD palm oil. The results have shown that the EWR
are found to be peak current and the pulsation rates. The is comparable between the RBD palm oil and other stand-
results of their study have shown that graphite electrodes ard dielectric fluid.
give higher MRR when compared to other types; however, In relation to AI and statistical analysis, there are great
brass electrodes give a better surface finish in comparison needs for such advanced techniques as modern industrial
with the other material. Chikalthankar et al. [7] have investi- systems are becoming increasingly complex, optimized,
gated the effect of EDM machining parameters on WPS DIN and intelligent. Recent developments in artificial intelli-
1.2379/AISI D2 tool steel. Voltage, current, pulse on time, gence (AI) [13], especially machine learning (ML), have
and pulse off time are investigated to explore the effect on shown great potential to model, predict, and transform the
the material removal rate and surface roughness (Ra) when manufacturing domain through advanced analytics tools for
using copper electrode materials. The study has found that processing the vast amounts of generated manufacturing
the current is the most significant parameter affecting the data and process parameters. AI technology, therefore, has
material removal rate and surface roughness. For process been found to be used in EDM applications. For example,
optimisation, Chandramouli et al. [8] have discussed the Wang et al. [14] have used feedforward supervised neural
optimal process parameters of EDM on RENE80 nickel networks to extract relevant information from the voltage
super alloy material with aluminium as the electrode. The signal in the wire electrical discharge machining process. AI
input parameters for the experimental work are current and has also been used in other processes process monitoring;
pulsation time and their effect on MRR, TWR, and surface for example, Al-Azmi et al. [15] have used learning vector
roughness have been investigated. The results of the research quantisation neural network (LVQ) for the classification of
have shown that optimisation of input parameters play a crit- information in turning operations. It implements competitive
ical role in the EDM process optimisation. Tomadi et al. [9] neural network as an unsupervised neural network which
have also investigated the effect of the process parameters, uses associative learning rules which allow the network to
including peak current, power supply voltage, pulse-on time, learn the association between the inputs and the outputs in
and pulse off time. The surface quality that is considered in response to the data presented to them.
The International Journal of Advanced Manufacturing Technology (2025) 137:251–262 253

Mondal et al. [16] have conducted electrical discharge workpieces, particularly in relation to machining high-speed
machining (EDM) on AISI 304 stainless steel with artificial steel AISI M2 using both brass and copper tools as elec-
neural network (ANN) to predict surface roughness. It has trodes with variable process parameters and limited and
been found that ANN predicted surface roughness with some variable training data. The supervised deep learning neural
accuracy. The maximum percentage of error in the MRR networks with three inputs, as will be discussed in this paper,
prediction was 6.16% for the ANN model. This could be should allow learning from experience, and hence, this paper
argued to be a relatively high prediction error for the purpose investigates the use of deep learning to capture the experi-
of accurate modelling and forecasting. Other applications ence and enable the EDM modelling process for the selected
of AI include, for example, the use of AI in wire electri- material, process parameters, and electrodes.
cal discharge machining (WEDM) where aluminium-based As a result of the above, this work investigates the opti-
metal matrix using neural network and multiple regression misation of the EDM process when using high-speed steel
approaches are adopted for forecasting the aforementioned AISI M2 as a workpiece material with copper and brass
process variables [17]. Some work had attempted to com- as electrodes to define the effect of material type of elec-
pare between several neural network algorithms [18], and it trode on machining performance, hence, the smoothness
has been found that long short-term memory-based recur- of the machined surface (i.e. surface roughness of parts).
rent neural network provides more accurate prediction. The Furthermore, feed forward deep learning supervised neural
machining of Inconel 617 (IN617) and the prediction of networks (FNN) is utilized to model and predict the perfor-
MRR and surface roughness were investigated by [19]; ANN mance of the process with limited input data.
model is found to exhibit reasonable predictive capabilities.
However, one of the limitations of the work involves the data
hunger of the selected AI-based algorithm. ANNs is found 2 Methodology
to require large amounts of labelled data for training, which
can be a challenge when data is limited, time consuming, Figure 1 presents the suggested methodology of this paper.
or expensive to acquire. Another paper has focused also on As discussed above, the principle of EDM operation [9] is a
Inconel 718 [20] with artificial neural network (ANN), sup- non-conventional manufacturing process based on removing
port vector machine (SVM), and genetic algorithm (GA) material from the workpiece by means of a series of frequent
to model and optimize the surface roughness during wire electrical discharges (created by electric pulse generators at
electrical discharge machining (WEDM). The SVM predic- short intervals) between a tool, the electrode, and the work-
tions are found to be accurate among all the models studied, piece that is being machined in the presence of a dielectric
as determined from the R-value the least mean absolute per- fluid.
centage error (MAPE) of 0.0347%. The suggested methodology is to examine the effect of
In [21], researchers have used ANN to monitor EDM pro- EDM process parameters on the machined parts when using
cess of hot work steel DIN1.2344 using copper electrodes brass and copper tools. Statistical analysis is used to assess
with four inputs to the ANN, and the approach has been the significance of the parameters and then deep learning
found successful. However, there is still a knowledge gap in neural networks are used to learn from experience and to
the use of both copper and brass electrodes with high-speed model the EDM process using partial set of data to predict
steel AISI M2 workpiece material in EDM with AI model- the rest of the parameters. The EDM is a well-determined
ling. From the above, it can be argued that more work is machining choice for manufacturing geometrically of hard
still needed in understanding the EDM process and param- material parts that are highly difficult to machine by tra-
eters for different combination of materials for the tool and ditional machining processes [22]. Its unparalleled feature

Fig. 1  The suggested methodol-

ogy in this paper
254 The International Journal of Advanced Manufacturing Technology (2025) 137:251–262

of using thermal energy to machine electrically conductive same to that of the product desired with allowance for side
parts regardless of hardness has been its special advantage clearance and over cut. To maintain the gap between the tool
for manufacturing of mould, die, automotive, aerospace, and the workpiece, the distance is continuously monitored
and surgical components [23]. Appropriate selection and using a servo-controlled, as shown in Fig. 2.
optimisation of the machining parameters can result in a A stream of electrons during each electric discharge
higher material removal rate, better surface finish, and lower moves at a high velocity, and the electrons accelerate from
electrode wear ratio [24]. Many artificial intelligence tech- the cathode towards the anode creating compression shock
niques have been developed for manufacturing systems [15]; waves causing a rise in temperature. The forces of the elec-
a feed forward neural network is selected in this paper as one tric and magnetic fields produced by the spark create a ten-
of the simplest types of artificial neural networks devised sile force that tears off particles of metal from the focus
as it helps learning from experience as supervised neural spot on the machined surface. The experiments have used
networks [25–27]. The selected neural network has several kerosene as dielectric medium on EDM CM 323C (CNC
hidden layers which provide the structure for how the neural Machine), as shown in Fig. 2, with copper and brass elec-
network will interact with the data in a deep learning pro- trodes (4 mm diameter). The surface roughness of the work-
cess; this allows artificial intelligence system to draw con- pieces has been tested using Pocket surf PSI profile-metre
clusions independently from human intervention. The main as shown in Fig. 3.
reason the researchers use feedforward neural networks is The workpiece material used in this study is a high-
their ability to approximate functions, which involves mak- speed steel AISI M2 with nine specimen dimensions
ing predictions and supports process modelling. (60 × 45 × 3 mm). The chemical composition and mechanical
and physical properties of the workpiece material are given
in Tables 1 and 2, respectively.
3 Experimental work The materials of the electrode used in the experimental
study are copper and brass shaped as a rod with diameter
A schismatic of electro discharge machining (EDM) process (4 mm) and length (100 mm) as shown in Fig. 3. The pro-
is shown in Fig. 2. The tool and workpiece both need to cess parameters used in experimental work are shown in
be immersed in the dielectric fluid. The dielectric fluid is Table 3. In practice, the optimal or the expected values of
stored in a tank and circulated through a pump with the help EDM parameters are often a compromise between compet-
of a nozzle at the gap between the tool and workpiece. The ing factors, and surface roughness is the focus of this paper.
dielectric fluid is continuously flushed into the spark gap. The values of the parameters of the experiential work in
When the process starts, thousands of very short sparks are this paper are based on experience of using the material, the
produced per second via the gap between the tool and the electrode, and process parameters.
workpiece. When the spark comes in contact with the dielec- Table 4 presents the design of the experimental work and
tric fluid in the spark gap, the fluid gets ionized. It allows variables used and their levels, while Table 5 presents the
current to flow between the tools and workpiece. The tool average readings of the results. The implemented current
and workpiece are kept in a reservoir and are connected to levels were 10, 24, and 42 A. The Pulse on was selected
a DC power supply. Tool electrode has the shape basically to be 10, 150, and 200 µs, while pulse off is selected to be
4, 12, and 25 µs. This creates 27 permutations as shown in
Table 5. The experiment (run) was repeated three times to
characterize the process.

4 Results and discussion

The experimental work included machining samples for

27 combination of parameters, with the experimental work
repeated three times, making the total conducted tests to be
162 tests between the two types of tools. This paper uses
AISI M2 steel as the workpiece material and two materials
for the electrodes, namely copper and brass electrodes.
Table 5 presents the experimental results of this experi-
mental work. Figures 3 and 4 present the graphical presen-
Fig. 2  A schismatic of electro discharge machining (EDM) process tation of the data and the average results for the copper and
used in this experimental work brass tools, respectively.
The International Journal of Advanced Manufacturing Technology (2025) 137:251–262 255

Fig. 3  The experimental work

components. EDM CM323C
CNC machine (a). Surface
roughness tester used in this
work (b). The copper and brass
electrodes used in the experi-
mental work with their associ-
ated workpieces (c, d)

Table 1  Chemical composition Material Weight (%) Table 3  Working factors used in the experimental work
of AISI M2 workpiece material
Working Factors Description
C 0.855
Si 0.305 Workpiece AISI M2 (60 × 45 × 3 mm)
Mn 0.28 Tool-electrode material Copper and brass (4 mm diameter)
P 0.001 Tool-electrode polarity Negative ( −)
S 0.001 Workpiece polarity Positive ( +)
Cr 4.71 Dielectric Kerosene
Mo 5.43 Dielectric temperature 40–80 °C
Ni 0.14 Input voltage 380 V (three phase) AC
Cu 0.175 Output voltage 140 V (two phase) DC
V 1.88 Current 10–42A
W 5.73 Pulse on time 100–200 µs
Fe Balance Pulse off time 4–25 µs

Table 2  Mechanical and physical properties of workpiece materials Table 4  Design of experiment (parameters and levels)
Properties AISI M2 Factors of consideration Symbols Levels

Bending strength (MPa) 4700 1 2 3

Modulus of elasticity (GPa) 207 Current (A) I 10 24 42
Specific heat capacity (J/g.°C) 17.2 × ­10−6 Pulse on time (µs) Pon 100 150 200
Thermal conductivity (W/m.K) 19 Pulse off time (µs) Poff 4 12 25
Compressive yield strength (MPa) 3250
Melting point (°C) 4680
Elastic modulus (GPa) 210 It can be noticed that the values of surface roughness of
Electrical resistivity (Ohm-cm) 54 × ­10−6 the machined part using the copper electrode have gradu-
Density (g/m3) 8.14 ally increased starting with Ra 2.41 µm and raised up to Ra
Hardness (HRB) 65 5.89 µm, as shown in Fig. 4. In contrast, the roughness of
Charpv impact (J) 36 the parts using the brass electrode is more than those when
Poisson’s ratio 0.30 using copper electrode with average values range from Ra
3.71 µm to Ra 7.75 µm, as shown in Fig. 5.
 256

Table 5  Experimental results of the machining parameters

No Current (A) Pulse on (µs) Pulse off (µs) Copper electrode Brass electrode Training/testing (Based
on the associate Training
Ra 1 (µm) Ra 2 (µm) Ra 3 (µm) Ra avg (µm) Ra 1 (µm) Ra 2 (µm) Ra 3 (µm) Ra avg (µm) Column)

1 10 100 4 2.40 2.43 2.42 2.41 3.73 3.76 3.64 3.71 Training
2 10 100 12 2.39 2.34 2.38 2.37 3.64 3.45 3.59 3.56 Testing
3 10 100 25 2.15 2.18 2.15 2.16 3.41 3.49 3.39 3.43 Training
4 10 150 4 2.65 2.57 2.67 2.63 3.82 3.72 3.80 3.78 Testing
5 10 150 12 2.57 2.61 2.53 2.57 3.69 3.56 3.81 3.62 Training
6 10 150 25 2.37 2.51 2.50 2.46 3.47 3.55 3.45 3.49 Testing
7 10 200 4 3.20 2.84 2.93 2.99 4.11 3.84 3.90 3.95 Training
8 10 200 12 2.79 2.75 2.95 2.83 3.84 3.81 4.02 3.89 Testing
9 10 200 25 2.62 2.60 2.70 2.64 3.78 3.84 3.87 3.83 Training
10 24 100 4 4.66 4.62 4.58 4.62 5.85 5.90 5.80 5.85 Testing
11 24 100 12 4.56 4.54 4.49 4.53 5.69 5.62 6.03 5.78 Training
12 24 100 25 4.37 4.35 4.39 4.37 5.35 5.51 5.40 5.42 Testing
13 24 150 4 4.73 4.65 4.66 4.68 5.96 5.87 5.90 5.91 Training
14 24 150 12 4.71 4.65 4.50 4.62 5.71 5.77 5.68 5.72 Testing
15 24 150 25 4.49 4.65 4.60 4.58 5.49 5.56 5.69 5.58 Training
16 24 200 4 4.79 4.89 4.78 4.82 6.35 6.43 6.30 6.36 Testing
17 24 200 12 4.72 4.84 4.75 4.77 6.23 6.12 6.16 6.17 Training
18 24 200 25 4.60 4.68 4.67 4.65 5.96 6.30 5.98 6.08 Testing
19 42 100 4 5.88 5.97 5.94 5.93 7.45 7.40 7.56 7.47 Training
20 42 100 12 5.73 5.74 5.66 5.71 7.33 7.45 7.30 7.36 Testing
21 42 100 25 5.11 5.10 5.15 5.12 7.33 7.31 7.23 7.29 Training
22 42 150 4 5.80 5.89 5.77 5.82 7.61 7.71 7.57 7.63 Testing
23 42 150 12 5.73 5.75 5.80 5.76 7.58 7.64 7.52 7.58 Training
24 42 150 25 5.30 5.28 5.38 5.32 7.40 7.42 7.35 7.39 Testing
25 42 200 4 6.39 6.33 6.39 6.37 7.97 7.90 7.92 7.93 Training
26 42 200 12 6.20 6.06 6.22 6.16 7.79 7.88 7.85 7.84 Testing
27 42 200 25 5.99 5.86 5.82 5.89 7.69 7.79 7.77 7.75 Training
The International Journal of Advanced Manufacturing Technology (2025) 137:251–262
The International Journal of Advanced Manufacturing Technology (2025) 137:251–262 257

Fig. 4  Surface roughness of the machined parts by EDM using copper electrode (a) and brass electrode (b)

4.1 Correlation coefficient

Correlation is a measure of the association between vari-

ables. In correlated data, the change in the magnitude of
one variable is associated with a change in the magnitude of
another variable, either in the same (positive correlation) or
in the opposite direction (negative correlation). Most often,
the term correlation is used in the context of a linear rela-
tionship between two continuous variables and expressed
as Pearson product-moment correlation. In this paper, the
correlation coefficients between the input parameters and Ra
output are calculated, as shown in Table 5. The correlation
coefficient between of the process current (A) and surface
roughness (Ra) is found to be 0.955 and 0.98 for the copper
and brass electrodes, respectively. Other parameters are not
highly correlated, as shown in Table 6 and Fig. 6.
Moreover, Fig. 7 presents the relationship between the
Fig. 5  Average surface roughness of the machined part using copper current and surface roughness for both electrodes, where the
and brass electrodes
surface roughness depends on current value and the material
of the electrode. This leads to the conclusion that that the
As shown in Fig. 5, the electrode material has signifi- surface roughness increases with the increase in the process
cant effect on the surface roughness. Also, other param- current.
eters play an important role in the values of surface rough- It can be observed that the surface roughness of work-
ness. Consequently, it is important to select optimum piece using the copper electrode produces much better
conditions to achieve the required product’s surface finish surface roughness in comparison to the brass electrode.
with maximum productivity (machining time). Figure 6 This could be explained by the characteristics of cop-
presents a summary of all the variables and the results in per when compared to brass such as better thermal con-
graphical format. You can see from Fig. 6 that the cur- ductivity, and the homogeneity of copper as pure metal
rent and material of the electrode have the most effect on creates a better spark stability and lower wear rate [28].
surface roughness. In statistical hypothesis testing, p-value can be defined
It can again be observed from Fig. 6 that the current has the statistical probability of getting the experimentally
the most significant effect on the surface roughness. observed results, when assuming that the null hypothesis
258 The International Journal of Advanced Manufacturing Technology (2025) 137:251–262

Fig. 6  Data normalization of the

machining parameters and the
effect of machining parameters
on the performance of machin-
ing (surface roughness)

Table 6  The correlation coefficient between input parameters and Ra

output
Copper electrode (Ra) Brass electrode (Ra)
Correlation coefficient Correlation coefficient

Current (A) 0.955 0.98

Pulse on (µs) 0.1302 0.111
Pulse on (µs) − 0.105 − 0.06

is true. p-value shows how likely it is that your set of

observations could have occurred under the null hypoth-
esis. Hence, p-values are used in statistical hypothesis
testing to determine whether to reject the null hypothesis
(the null hypothesis indicates no relationship between Fig. 7  The relationship between the current and surface roughness for
variables). The smaller the p-value, the stronger the likeli- both electrodes, where x is the value of the current
hood that you should reject the null hypothesis (i.e. there
is a relationship). In this research, when testing the null 4.2 Artificial intelligence for modelling the EDM
hypothesis which assumes that there is no relationship processes
between the observed phenomena (Ra) and the process
variables, p-value of less than 0.05 is considered statisti- To model the process using deep learning supervised neural
cally significant, as shown in Table 7. From Table 7, it networks, the suggested neural network in Fig. 8 includes
is observed that only the current is considered statisti- three inputs (current, pulse on, and pulse off). The data is
cally significant to reject the null hypothesis and assume fed into 30 neurons with sigmoid functions in the first hidden
a relationship between the current and surface roughness.
The International Journal of Advanced Manufacturing Technology (2025) 137:251–262 259

Table 7  The observed phenomena (Ra) and (null hypothesis)

p-value of cop- p-value of brass
per electrode electrode (Ra)
(Ra)

Current (A) 0 0 Reject null

hypothesis
Pulse on (µs) 0.517 0.581 Accept null
hypothesis
Pulse on (µs) 0.602 0.74 Accept null
hypothesis

layer and a second hidden layer with 10 neurons with also

sigmoid hidden functions. The output layer includes 1 output
Fig. 9  The results of the suggested deep learning neural network (for
(estimated Ra) with linear function for data estimation. The
the copper electrode)
rows that are used for training or testing (prediction) of the
neural networks are indicated in Table 5.
Figure 9 presents the output for copper electrode when experimental work. The data presents the experimental three
50% of the average Ra data is used for training the neural runs, the experimental average, and the estimated Ra by the
network in comparison to the experimental work. The data neural network. Figure 12 presents, for the brass electrode,
presents the experimental three runs, the experimental aver- the average error of the estimates Ra when using 50% of the
age, and the estimated Ra by the neural network. Figure 10 data from run 1, run 2, run 3, and the average for Fig. 12a–d,
presents, for the Copper electrode, the average error of the respectively. We notice that the minimum error occurs in the
estimates Ra when using 50% of the data from run 1, run training run or the average run as shown in Fig. 12.
2, run 3, and the average for Fig. 10a–d, respectively. We From Figs. 10 and 12, it can be concluded that the worst
notice that the minimum error occurs in the training run or average error levels in µm using the ANN are estimated to be
the average run as shown in Fig. 10. 0.078 µm for cooper and 0.09 µm for brass. The use of deep
The ANN predicted error (Ei ) in μm is described in learning neural network has shown that AI could predict
Eq. (1), where Ti is the expected target (actual value), Pi is the average Ra values with average absolute error of about
the predicted value by the neural network, and i is the test 0.39% for copper and of 0.26% for brass indicating the ben-
number. The average error in µm for any training and testing efits of using AI in predicting the performance of manufac-
process is described in Eq. (2), where n is the total number turing processes and the potential use of AI in future process
of samples. modelling and applications. The results show that a minimal
or limited training should provide reasonable ANN results to
(1)
( ) ( )
The predicated error Ei = ||Ti − Pi || 𝜇m model and predict the EDM process using a fraction of the
experimental work. The paper also confirms and provides
further evidence of the benefits of using AI to model EDM
∑n
Ei
The average error = i=1
𝜇m (2) processes for different workpiece materials and electrodes.
n
The results of this paper show novelty in terms of the
Similarly, for the brass electrode, Fig. 11 presents the selected deep learning ANN structure with three inputs,
output for brass electrode when 50% of the average Ra data using both brass and copper materials. This paper also uses
is used for training the neural network in comparison to the AISI M2 steel, which is a material that has not been well

Fig. 8  The implemented deep

learning supervised neural
network
260 The International Journal of Advanced Manufacturing Technology (2025) 137:251–262

Fig. 10  The prediction error of the deep learning neural networks of EDM surface roughness (copper)

discussed in literature. Moreover, the paper utilizes statisti-

cal hypothesis testing and analysis to find the relationship
between the variables that was confirmed independently by
the ANN.
Overall, the results show a reasonable prediction of the
ANN when compared with the actual values.

5 Conclusion

This paper has presented a unique and novel empirical

study of EDM machining supported by artificial intelli-
gence and statistical analysis. The paper has included an
experimental work for HSS AISI M2 material using EDM.
Brass and copper electrodes have been tested. It can be
concluded that the surface roughness of the workpiece
using the copper electrode produces much better surface
roughness in comparison to the brass electrode. However,
Fig. 11  The results of the suggested deep learning neural network the highest results of Ra in copper and brass electrodes are
(for the brass electrode) found to be 6.37 µm and 7.93 µm, respectively, at current
The International Journal of Advanced Manufacturing Technology (2025) 137:251–262 261

Fig. 12  The prediction error of the deep learning neural network of EDM surface roughness (brass)

42 A, pulse on time of 200 µs, and pulse off time of 4 µs. processes with predicting process outcomes for selected
This result will help in the prediction and optimisation of workpiece and electrode materials using a suitable ANN
the EDM process for machining HSS AISI M2 by esti- structure and functions with different training regimes.
mating process time, energy use, and quality of the sur- The results show that limited training samples are able to
face roughness of the workpiece. Deep learning neural provide reasonable ANN results to the selected model and
networks and statistical analysis are used to evaluate the predict the EDM process using a fraction of the experi-
results. The data from the experimental work has indicated mental work. This paper also provides further confirma-
that there is a significant correlation between the process tion of the benefits of using AI to model EDM processes
current and average surface roughness; and the pulsation for different workpiece materials and electrodes.
time was not found significant. The use of deep learning
neural networks has shown that the worst average error
Author contribution Jabbar K. Abbas: conceptualization, methodol-
levels in all training and testing in µm using the ANN ogy, validation, formal analysis, investigation, resources, data cura-
are estimated to be 0.078 µm for copper and 0.09 µm for tion, writing—original draft, writing—review and editing, supervi-
brass. However, in percentage overall average error, ANN sion, project administration; Shukry H. Aghdeab: conceptualization,
is found to be able to predict the average Ra values with methodology, formal analysis, investigation, resources, data curation,
writing—original draft, supervision, project administration; Amin
an average absolute error of about 0.39% for copper and Al-Habaibeh: conceptualization, methodology, software, validation,
of 0.26% for brass indicating the benefits of using AI in formal analysis, writing—original draft, writing—review and editing,
predicting the performance of manufacturing processes visualization, funding acquisition.
with limited training data. The paper proves the benefits
Funding The authors would like to thank the British Council, Ishtar
of using AI and statistical modelling of manufacturing Light Project (Ref: RE2023-114), for partially funding this research
262 The International Journal of Advanced Manufacturing Technology (2025) 137:251–262

work for the UK author, Professor Amin Al-Habaibeh, to support 12. Supawi A, Ahmad S, Talib N, WoonKiow L (2024) Electrode wear
improving the research environment in Iraq’s Higher Education sector. rate of RBD palm oil as dielectric fluids on electrical discharge
machining (EDM) at different peak current and pulse duration of
Declarations titanium alloys. J Adv Res Fluid Mech Thermal Sci 112(2):54–63
13 Arinez JF, Chang Q, Gao RX, Xu C, Zhang J (2020) Artificial
Conflict of interest The authors declare no competing interests. intelligence in advanced manufacturing: current status and future
outlook. J Manuf Sci Eng 142(11):110804
Open Access This article is licensed under a Creative Commons Attri- 14. Wang J, Sánchez JA, Iturrioz JA, Ayesta I (2019) Artificial intel-
bution 4.0 International License, which permits use, sharing, adapta- ligence for advanced non-conventional machining processes. Proc
tion, distribution and reproduction in any medium or format, as long Manuf 41:453–459
as you give appropriate credit to the original author(s) and the source, 15. Al-Azmi A, Al-Habaibeh A, Abbas J (2023) Sensor fusion and the
provide a link to the Creative Commons licence, and indicate if changes application of artificial intelligence to identify tool wear in turning
were made. The images or other third party material in this article are operations. Int J Adv Manuf Technol 126:429–442
included in the article’s Creative Commons licence, unless indicated 16. Mondal N, Nishant, Ghosh S, Mandal MC, Pati S, Banik S (2023)
otherwise in a credit line to the material. If material is not included in ANN and RSM based predictive model development and EDM
the article’s Creative Commons licence and your intended use is not process parameters optimisation on AISI 304 stainless steel. Mate-
permitted by statutory regulation or exceeds the permitted use, you will rials Today: Proceedings 2023. [Link]
need to obtain permission directly from the copyright holder. To view a 2023.​01.​322
copy of this licence, visit [Link] 17. Surendranatha GM, Naidu BVV, Gangaraju M, Sarapure S,
Hemanth KR (2022) Development of predictive models for wire
electrical discharge machining of aluminium metal matrix com-
posites by using regression analysis and neural network. Mater
References Today: Proc 68 (part 5):1581–1587
18. Dey K, Kalita K, Chakraborty S (2022) Prediction performance
1. Kumar S, Gupta T (2023) A review of electrical discharge machining analysis of neural network models for an electrical discharge turn-
(EDM) and its optimization techniques, Materials Today: Proceed- ing process. Int J Interact Des Manuf 17:827–845
ings, ISSN 2214-7853. [Link] 19. Sana M, Farooq MU, Anwar S, Haber R (2023) Predictive model-
2. Chaitanya S, Kumar S, Sharma A, Kumar S, Goyal R (2023) ling framework on the basis of artificial neural network: a case
Experimentation on HCHCr steel using vegetable oil as dielectric of nano-powder mixed electric discharge machining. Heliyon
in EDM, Materials Today: Proceedings. [Link] 9(12):e22508. [Link]
matpr.​2023.​07.​375 20 Paturi UMR, Cheruku S, Uma Pasunuri VP, Salike S, Reddy NS,
3. Paswan K, Sharma S, Prakash Dwivedi S, Abdulameer MK, Li C, Cheruku S (2021) Machine learning and statistical approach in
Yasin Y, Abbas M, Tag-Eldin EM (2023) An analysis of micro- modeling and optimisation of surface roughness in wire electrical
structural morphology, surface topography, surface integrity, discharge machining. Mach Learn Appl 6(15):100099
recast layer, and machining performance of graphene nanosheets 21. Amini S, Atefi R, Solhjoei N (2011) The influence of EDM
on Inconel 718 superalloy: investigating the impact on EDM char- parameters in finishing stage on surface quality of hot work steel
acteristics, surface characterizations, and optimization. J Mater using artificial neural network. Am Inst Phys. [Link]
Res Technol 27:7138–7158 1063/1.​35523​50
4. Wong YS, Lim LC, Lee LC (1995) Effect of flushing on electro 22. Ho KH, Newman ST (2003) State of the art electrical discharge
discharge machined surface. J Mater Process Technol 48:299–305 machining (EDM). Int J Machine Tools Manufac 43:1287–1300
5. Daneshmand S, Kahrizi EF, Abedi E, Abdolhosseini MM (2013) 23. Ponappa K, Aravindan S, Rao PV, Ramkumar J, Gupta M (2010)
Influence of machining parameters on electro discharge machining The effect of process parameters on machining of magnesium
of NiTi shape memory alloys. Int J Electrochem 8:3095–3104 nano alumina composites through EDM. Int J Adv Manufac Tech-
6. Prajapati HB, Patel VA, Prajapati HR (2012) Experimental inves- nol 46:1035–1042
tigation of performance of different electrode materials in electro 24. Rajyalakshmi G, Venkata Ramaiah P (2013) Multiple process
discharge machining for material removal rate and surface rough- parameter optimisation of wire electrical discharge machin-
ness. International Journal of Engineering Research and Applica- ing on Inconel 825 using Taguchi grey relational analysis. Int
tions (IJERA) 2(3):3072–3076. www.​ijera.​com J Adv Manuf Technol 69:1249–1262. [Link]
7. Chikalthankar SB, Nandedkar VM, Borde SV (2013) Experi- s00170-​013-​5081-z
mental investigations of EDM parameters. Int J Eng Res Dev 25. Al-Habaibeh A, Gindy N (2000) A new approach for systematic
7(5):31–34 design of condition monitoring systems for milling processes. J
8. Chandramouli S, Shrinivas UB, Eswaraiah K (2014) Optimisa- Mater Process Technol 107: 243–251. ISSN 0924–0136
tion of electrical discharge machining process parameters using 26. Al-Habaibeh A, Gindy N (2001) Self-learning algorithm for auto-
Taguchi method. Int J Adv Mech Eng 4:425–434 mated design of condition monitoring systems for milling opera-
9. Tomadi SH, Hassan MA, Hamedon Z (2009) Analysis of the influ- tions. Int J Adv Manuf Technol 18(6): 448–459. ISSN 0268–3768
ence of EDM parameters on surface quality, material removal rate 27. Elgargni M, Al-habaibeh A, Lotfi A (2015) Cutting tool tracking
and electrode wear of tungsten carbide. In: Proceedings International and recognition based on infrared and visual imaging systems
Multi Conference of Engineers and Computer Scientists, vol. 2 using principal component analysis (PCA) and discrete wave-
10. Rahman MM, Khan MAR, Noor KK, Bakar RA (2010) World let transform (DWT) combined with neural networks. Int J Adv
Academy of Science, Engineering and Technology (47):567–581. Manuf Technol 77(9):1965–1978. ISSN 0268–3768
[Link] 28. Jameson EC (2001) Electrical discharge machining. Society of
doi=​275e7​a169b​c9abb​04fcf​26c8b​aa185​f1eb3​c589d Manufacturing Engineers
11. Sahoo R, Singh NK, Bajpai V (2023) A novel approach for mod-
eling MRR in EDM process using utilized discharge energy. Mech Publisher's Note Springer Nature remains neutral with regard to
Syst Signal Process 185(15):109811 jurisdictional claims in published maps and institutional affiliations.