International Journal on Interactive Design and Manufacturing (IJIDeM)

https://doi.org/10.1007/s12008-024-01923-x

ORIGINAL ARTICLE

Investigation and prediction of machining characteristics of aerospace

material through WEDM process using machine learning
Rupesh Chalisgaonkar1 · Sachin Sirohi1 · Jatinder Kumar2 · Sachin Rathore3

Received: 30 December 2023 / Accepted: 23 May 2024

© The Author(s), under exclusive licence to Springer-Verlag France SAS, part of Springer Nature 2024

Abstract
In this study, pure titanium is used as the workpiece material and a machine learning approach is used to forecast the material
removal rate (MRR) during wire electrical discharge machining (WEDM). The novelty of present research work was to
perform machining operation on pure titanium through the WEDM process as the pure titanium is majorly used in aviation
and aircraft industries. The machining industries could get help through this study to run the machining process based on
machine learning approach for increased productivity considering the scope of Industry 4.0. This study’s goal is to create a
precise prediction model that can anticipate MRR based on a variety of input factors, such as pulse-on time, pulse-off time,
wire feed rate, wire tension, servo voltage, and peak current. Experimental data was collected through a series of WEDM
experiments on pure titanium using an L-27 orthogonal array based on Taguchi’s design of experiments. MRR was selected
from the pre-processed data to train and evaluate the machine learning model. The prediction model was developed using
a variety of regression techniques, including Linear Regression using scikit-learn, support vector regression (SVR), random
forest, K-nearest neighbors regression using Python in Jupyter notebook. Coefficient of determination (R-squared) and root
mean squared error were used to assess the model’s performance. The results show that the linear regression using scikit-learn
and SVR algorithm performs better in terms of prediction accuracy than the other algorithms. The surface integrity analysis
was performed to determine the effects of process parameters on machined surface. The proposed study helps to increase
the efficacy and efficiency of WEDM operations by offering a trustworthy tool for MRR predictions. The proposed research
depicts a good agreement between experimental and predicted values.

Keywords Support vector regression · Pure titanium · Material removal rate · Wire electrical discharge machine · Machine
learning

1 Introduction Most of the engine parts of aircraft are made up of titanium

as it possesses strength at elevated temperatures. As the air-
Pure titanium is widely utilized in the aircraft industry due to craft is subjected to cyclic loading condition during landing
its exceptional strength-to-weight ratio which is very crucial and take-off, higher fatigue resistance of pure Use the "Insert
as per consideration of fuel efficiency and overall perfor- Citation" button to add citations to this document.
mance. The corrosion resistance of pure titanium helps to titanium helps to maintain its strength in such conditions.
protect the aircraft from extreme environmental conditions. Due to the above properties pure titanium is used in air-
frames, landing gear, and critical engine components such
as compressor blades, turbine blades, and engine casings.
B Sachin Sirohi
sachinsirohi2008@gmail.com Titanium is used in various fasteners and bolts throughout
the aircraft due to its lightweight yet robust nature. In this
1 Mechanical Engineering Department, SRM Institute of aspect, WEDM is the ideal option for pure titanium machin-
Science and Technology, DelhiNCR Campus, Ghaziabad,
ing and its application. Several factors contribute to pure
India
titanium’s poor machinability, including its strong chemical
2 Mechanical Engineering Department, National Institute of
reactivity, susceptibility to weld to the tool during machining,
Technology, Kurukshetra, Haryana, India
low Young’s modulus, which causes chatter and deflection
3 Sachin Rathore Mechanical Engineering Department, KIET
Group of Institutions, Delhi- NCR, Ghaziabad, India

123
 International Journal on Interactive Design and Manufacturing (IJIDeM)

to the tool, poor heat conductivity, and narrow chip-tool con- as pulse on time (TON ), pulse off time (TOFF ), peak cur-
tact zones on the tool face [1, 2]. As titanium has a lower rent (IP), wire feed (WF), wire tension (WT), servo voltage
value of thermal conductivity (16.3 W/m K) versus medium (SV) were chosen as input variables during WEDM of pure
carbon steel (43 W/m K), WEDM process develops a signifi- titanium. Process parameters and their levels are shown in
cant amount of thermal energy in the machining zone causing Table 2. The Taguchi’s experimental design methodology
higher machining efficiency. has several benefits, including efficiency in experimentation,
WEDM process utilizes a thin metal wire as the electrode identification of significant variables, robustness to variabil-
in the manufacturing process to carve complicated shapes ity and systematic problem-solving method approach. Table
and curves in conductive materials. The wire is fed into the 3 shows the L-27 array design with mean values of MRR as
workpiece while being held under stress during the electri- process characteristics.
cal current passes through it. The material erodes and the Material Removal Rate (MRR) was calculated by follow-
required shape is cut when the wire comes near to the work- ing equation.
piece and sparks are produced between them. The spark is
carefully controlled to make the desired shape while the wire  
is continuously pushed into the machining zone. Most of the MRR mm3 /min  Cutting speed (mm/min)
wire materials used in WEDM process are copper, brass, × kerf (mm)
aluminium, molybdenum etc. having a diameter range of × thickness of plate (mm) (1)
0.05 to 0.3 mm as per requirement. The wire utilized in the
WEDM process must be straight, have a high melting point,
strong electrical conductivity, and good flush ability. Figure 1 Kerf width was calculated by subtracting the dimensions
exhibits the WEDM process. Literature review was summa- of punch from the cavity produced after machining.
rized in Table 1 which represents the various methodologies The flow chart shown in Fig. 2 depicts the details of the
used to predict the quality characteristics of WEDM process experimental and research methodology.
by different researchers.
It is revealed from literature that many authors have
applied the Artificial Neural Network (ANN) method to pre- 3 Implementation of supervised machine
dict the machining data. While machine learning methods learning (regression model)
are having more benefits in terms of interpretation, train-
ing speed, robustness, and control over feature engineering. The corelation behaviour is exhibited between dependent and
Machine learning is also comparatively better in terms of more then one independent variables by regression model.
computational efficiency and model transparency. The evaluation of degree of corelation is also shown by it.
The above review reveals that the various modelling Regression is found suitable to predict the machining char-
methods were used to predict the machining characteristics acteristics of manufacturing processes. Artificial intelligence
of WEDM process by various researchers. The novelty of and Machine Learning algorithms can do this task in a very
proposed research is focused on to predict the machining efficient manner.
outcome of aviation industry material (pure titanium). As There are two types of machine learning algorithms:
the modern manufacturing industries are leading toward the supervised learning and unsupervised learning. In supervised
era of Industry 4.0 which necessitates the need of integration learning, each data point is labelled with a known outcome,
of unconventional manufacturing with artificial intelligence and the algorithm is trained on this labelled data. The algo-
and machine learning. The proposed work exhibits the step rithm learns to anticipate the results for fresh unforeseen data
toward the inculcation of component of Industry 4.0 with based on the patterns discovered from the labelled data. In
wire cut electric discharge machining process. unsupervised learning, the algorithm learns to recognize pat-
terns and structure in the data on its own after being taught
on unlabelled data. Decision trees, support vector machines,
random forests, KNN classification, neural networks, and
2 Experimental procedure deep learning models are just a few examples of the many
kinds of machine learning algorithms. Regression techniques
In this study, pure titanium was machined using WEDM need a collection of input-output pairs, or training data, to be
process. The workpiece dimensions were 144.75 × 108.96 fed throughout the training process afterwards, the algorithm
× 24.25 mm3 . Electronica Machine Tool Limited, India’s makes use of this information to discover the fundamental
sprintcut (ELPULS-40A DLX) WEDM model was used for correlation between the input and output variables.
the study. In these experimental runs, 0.25 mm zinc-coated The following steps are involved to create machine learn-
brass wire was employed. Six machining parameters such ing regression models.

123
International Journal on Interactive Design and Manufacturing (IJIDeM)

Fig. 1 Schematic diagram of

WEDM

a. Importing the necessary packages 3.1 Linear regression using sckit learn in machine
b. Importing the dataset in Jupyter notebook learning
c. Creating a correlation heat map to visualize correlation
of input and output parameters of the dataset A linear regression model was built using sckit learn library.
d. Creating feature variables (Input and Output) The Python code is shown below.
e. Standardization of dataset
f. Splitting data into train and test sets
a. Importing the necessary packages
g. Create a regression model
h. Regression diagnostics

123
 Table 1 Summary of the literature review of AI/ML models applied on WEDM/EDM process

S. no. Author(s) Modelling method Quality characteristics Material Remarks

123
1 Narayanan and Vasudevan [3] Artificial Neural Network (ANN) Wear (W), Thickness (T) of job and Copper material Effectively used for predicting
time (Ti) copper material processing
characteristics
2 Devarasiddappa et al. [4] M-TLBO algorithm Material removal rate (MRR) Ti6Al4V alloy Several iterations of the M-TLBO
algorithm have shown accurate
and reliable performance. It also
required less computing time and
effort and converged more
quickly, often in fewer than 10
iterations for Ti6Al4V alloy
3 Thankachan et al. [5] Machine learning techniques and MRR and Surface Roughness Novel aluminum alloy and metal Multiple linear regression and
Neural Network models matrix composite ANN models was developed for
predicting MRR and Ra values
based on the significant input
parameters and the results
achieved from the models were
compared with experimental
values and was found efficient
4 Rees et al. [6] Inductive learning Surface Roughness 94% tungsten carbide and 6% Investigations were made into the
cobalt hybrid micro machining
technique known as WEDM for
conducting WEDG with a
submersible rotary head. A
prediction model for surface
finish using WEDG was created
by using inductive learning and
data collected through online
process monitoring
5 Shukla & Priyadarshini [7] Machine learning algorithm Surface roughness and kerf width HASTEALLOY C276 using The best values of response
namely, gradient descent method WEDM taken from the variables were effectively
as an optimization technique literature predicted using the gradient
descent approach
International Journal on Interactive Design and Manufacturing (IJIDeM)
 Table 1 (continued)

S. no. Author(s) Modelling method Quality characteristics Material Remarks

6 Nain et al. [8] Random forest, M5P tree MRR, SR, and DD Udimet-L605 In contrast to the M5P pruned tree
approaches model and the unpruned tree
model for MRR and SR of
WEDM, the RF model has
disclosed the necessary result.
For the dimensional deviation of
the WEDM of UdimetL605, the
M5P unpruned tree model
exhibits the significant findings
in contrast to the RF and M5P
unpruned tree model
7 Nain et al. [9] Support vector machine (SVM) Cutting speed, wire wear ratio Udimet-L605 When comparing the CS, WWR,
algorithm (WWR), and dimensional and DD of WEDM of
deviation (DD) Udimet-L605, Support Vector
Regression based on the RBF
model is better than the other two
models for predictions
8 Subrahmanyam and Sarcar Data mining approach MRR H13 steel By raising the amount of
International Journal on Interactive Design and Manufacturing (IJIDeM)

[10] automation and giving insight

into the process, this method can
be used for any manufacturing
process and may dramatically cut
cycle times
9 Pujara et al. [11] TLBO (Teaching–Learning Based MRR and surface roughness LM6 Aluminum MMC TLBO is a rapid method of
Optimization) technique optimization compared to other
time-consuming procedures that
are often used for the
optimization purpose since it
yields better results at a lesser
number of iterations

123
 Table 1 (continued)

S. no. Author(s) Modelling method Quality characteristics Material Remarks

123
10 Reddy et al. [12] TLBO (Teaching–Learning Based Energy consumption, kerf width, Al-Si metal matrix composite For the optimization of difficult
Optimization) technique metal removal rate and surface issues, TLBO is recommended.
roughness In the current work, it is intended
to maximize metal removal rate
during WEDM of Al-Si metal
matrix composite while setting
energy consumption, kerf width,
and surface quality at minimum
levels
11 Shadab et al. [13] Metaheuristic Techniques Material removal rate, cutting Al5083/7%B4C Composite RSM was utilized in the TLBO
speed, and surface roughness method that is to statistically
predict MRR, CS, and surface
roughness
12 Singh Nain et al. [14] SVM, GP and ANN methods Surface roughness Nimonic-90 The study’s best model, according
to findings, was the GP PUK
Kernel model
13 Paturi et al. [15] Artificial neural network (ANN), Surface roughness Inconel 718 When compared to a statistical
support vector machine (SVM), RSM model, machine learning
and genetic algorithm (GA) approaches (SVM and ANN)
were shown to have
exceptionally accurate prediction
capabilities
14 Ulas et al. [16] ELM, W-ELM, SVR and Q-SVR Surface roughness Al7075 The W-ELM model, with a value
of 0.9720 R2 , had the best
performance
15 Jatti et al. [17] Random Forest, Decision Tree, MRR Cryo-treated workpiece viz, The approach using gradient
Gradient Boosting and Artificial NickelTitanium (NiTi) alloys, boosting regression produced the
Neural Network developed by Nickel Copper (NiCu) alloys, best coefficient of determination
Python programming and Beryllium copper (BCu) value, whereas the technique
alloys using Random Forest
classification produced the
highest F1-Score
International Journal on Interactive Design and Manufacturing (IJIDeM)
 Table 1 (continued)

S. no. Author(s) Modelling method Quality characteristics Material Remarks

16 D Srinivasan et al. [18] Regression models based on Material Removal Rate (MRR) and SS304 The multilayer perception model
machine learning surface roughness was shown to have the greatest
correlation coefficient (0.999) for
MRR and surface roughness,
respectively
17 Chou and Hwang [19] ANN Wire rupture Data from past research For WEDM, wire rupture may be
predicted using ML algorithms.
Although the basic artificial
neural network, which may not
be the finest one, was utilized in
this study, it nevertheless
performed well enough to
highlight the application’s
potential for machine learning
18 Saha et al. [20] Linear regression, regression trees, Data driven analysis Data driven analysis The WEDM process’s performance
support vector machines, and metrics are mainly sensitive to
Gaussian process regression) the parameters pulse on time
(Ton) and peak current (IP)
International Journal on Interactive Design and Manufacturing (IJIDeM)

19 Jiang and Yen [21] MTF-CLSTM Surface roughness SKD61 Steel In terms of predicting mean
absolute percentage error,
MTF-CLSTM performs
noticeably better than the
currently used approach
20 Walia et al. [22] Decision tree, random forest, Out-of-roundness of the tool tip EN31 tool steel The random forest methodology
generalized linear model, and emerged as the most successful
neural network in predicting the outcome among
the machine learning methods
used

123
 International Journal on Interactive Design and Manufacturing (IJIDeM)

Table 2 Process parameters with

their levels Process parameters (unit) Parameter designation Level 1 Level 2 Level 3

TON (µs) A 0.5 0.7 0.9

TOFF (µs) B 7 9.5 14
IP (Amp.) C 80 140 200
WF (m/min) D 6 8 10
WT (gm) E 850 1200 1600
SV (Volts) F 30 50 70

b. Importing the dataset in Jupyter: The data set was It is shown from Fig. 3 that TON and SV are more influen-
imported using pandas. tial process parameters for MRR rather than IP and TOFF . WF
and WT are the least influential process parameter depicted
from the Fig. 3.

c. Creating a correlation heat map to visualize correlation

of input and output parameters of the dataset
A scatter plot matrix is used to visualize the relationships
between many variables in a dataset. It shows the compre-
Heatmap matrix is a visual presentation exhibiting the hensive visualization of the relationship between different
values in different colours in a matrix in machine learning. parameters in a matric format. Scatterplot is used for pattern
The level of correlation between variable could be identi- identification, corelation assessment and data exploration in
fied by the heatmap matrix. In this research work analysis machine learning. From scatterplot (Fig. 4), it is evident
and visualization for the datasets was done by correlation that MRR increases with TON and IP while decreases with
heatmap matrix to find out the co-relation between the MRR increase value of TOFF and SV. The frequency distribution
and process parameters of WEDM (Fig. 3). From the co- of MRR values is also shown in scatter pair plot diagram.
relation matrix of MRR, the input parameters such as pulse
on time (Ton), Peak current (IP) and wire feed rate (WF) d. Creating feature variables (Input and Output)
are positively co-related to MRR while pulse off time (Toff),
wire tension (WT) and servo voltage (SV) are negatively
co-related to MRR. It interprets that if Ton, IP and WF is
increased then MRR will also be in increasing manner while
increasing values of Toff, WT and SV will lead to decrease
in MRR.

123
International Journal on Interactive Design and Manufacturing (IJIDeM)

e. Standardization of dataset 3.2 Support vector regression

In this method after creating X and Y features in pre-
ceding method, standard scaling method was used to Regression analysis is often performed using the well-liked
standardize both features. machine learning technique i.e., Support Vector Regression
(SVR). The Support Vector Machines (SVM) method serves

f. Splitting data into train and test sets as the foundation for this supervised learning approach. In
The X and Y datasets were split into training and test- SVR, hyperplane is created allowing a specific degree of
ing sets consisting of 33% data in to testing and Linear error or margin which best matches the training data in the
Regression method was used from sklearn.linear_model. best way. In SVR, the hyperplane is defined as the collec-
Training data was fittted in to regression model. tion of points where constant value is equal to dot product
of a weight vector and an input vector plus a bias term

g. Create a linear regression model (Fig. 5). There are two hyperparameters named C which con-
trols the error term and e controlling the width of margin

The following equation was formed using the above tech- (Fig. 5). SVR has the benefit of utilizing a kernel function to
nique. address non-linear interactions between the input and output
variables. The hyperplane may be linearly separable in the
MRR  −0.333 + 0.887 × TON − 0.535 × TOFF + 0.372 × IP higher-dimensional feature space where the kernel function
− 0.0188 × WF − 0.066 × WT − 0.736 × V (2) translates the input space from. The kernels which are used
generally are Linear, Non-Linear, Polynomial, Radial Basis
Function (RBF) and Sigmoid.
SVR Regression analysis was done using python. The
above steps of python code defined in Sect. 3.1 (a) to (f) were
implemented and Kernel linear model was used for training
the data.

123
 International Journal on Interactive Design and Manufacturing (IJIDeM)

Table 3 L-27 array design with

experimental values of MRR S. no. TON TOFF IP WF WT SV MRR (mm3 /min)

1 0.5 7 80 6 850 30 15.923

2 0.5 7 140 8 1250 50 15.265
3 0.5 7 200 10 1600 70 9.723
4 0.5 9.5 80 8 1250 70 5.278
5 0.5 9.5 140 10 1600 30 14.935
6 0.5 9.5 200 6 850 50 11.311
7 0.5 14 80 10 1600 50 5.573
8 0.5 14 140 6 850 70 4.549
9 0.5 14 200 8 1250 30 11.647
10 0.7 7 80 8 1600 50 18.326
11 0.7 7 140 10 850 70 14.284
12 0.7 7 200 6 1250 30 19.285
13 0.7 9.5 80 10 850 30 18.62
14 0.7 9.5 140 6 1250 50 17.889
15 0.7 9.5 200 8 1600 70 13.347
16 0.7 14 80 6 1250 70 6.682
17 0.7 14 140 8 1600 30 20.14
18 0.7 14 200 10 850 50 16.254
19 0.9 7 80 10 1250 70 16.789
20 0.9 7 140 6 1600 30 23.789
21 0.9 7 200 8 850 50 34.033
22 0.9 9.5 80 6 1600 50 19.617
23 0.9 9.5 140 8 850 70 17.221
24 0.9 9.5 200 10 1250 30 27.191
25 0.9 14 80 8 850 30 20.763
26 0.9 14 140 10 1250 50 20.384
27 0.9 14 200 6 1600 70 13.815

123
International Journal on Interactive Design and Manufacturing (IJIDeM)

Fig. 2 Flow chart of the experimental and research methodology

3.3 Random forest regression In a random forest regression model, several decision trees
are trained on different subsets of the training data (Fig. 6).
Machine learning methods for regression challenges include For building each decision tree, a random subset of the char-
random forest regression. It is a member of the ensemble acteristics that are accessible are selected, and the data is
technique family, which combines many models to increase then split based on the best feasible split for the selected cri-
forecast stability and accuracy. terion (e.g., minimising mean square error). In the prediction

123
 International Journal on Interactive Design and Manufacturing (IJIDeM)

Fig. 3 Correlation heatmap matrix of WEDM process parameters and MRR

phase, the model generates a final forecast by averaging the 3.4 K-nearest neighbors (KNN regression)
predictions from each decision tree.
One advantage of random forest regression is that it can K-Nearest Neighbours (KNN) machine learning approach
handle missing data and both continuous and categorical uses the K closest data points in the feature space to predict
variables. It is also less prone to overfitting than individual a continuous target variable in regression applications. The
decision trees since the combination of several trees lessens value of K (i.e., the number of nearest neighbours to consider)
the impact of noise and outliers in the data. Moreover, fea- is specified. K closest data points in the feature space are
ture significance scores that may be used to identify the key determined for each new data point based on their Euclidean
components of the model may be produced using random distance from the new point. The forecast value for the new
forests. data point is the average of the target values of the K closest
The Random Forest Regressor function of the sklearn data points (Fig. 7).
package was used for training the random forest regression
model.

123
International Journal on Interactive Design and Manufacturing (IJIDeM)

Fig. 4 Scatter pairplot matrix for WEDM process parameters and MRR

123
 International Journal on Interactive Design and Manufacturing (IJIDeM)

Fig. 5 Generalized view of support vector regression

4 Result and discussion

The correlation between experimental value of MRR and

machine learning model is shown as follows. Fig. 7 Generalized view of K-nearest neighbours (KNN) regression

4.1 Regression diagnostic for linear regression for training and testing data was found 0.23 and 0.57 respec-
model tively as shown in Fig. 8a, b.

The R2 score for training and testing data was found 0.917
and 0.775 respectively, while root mean square error (RMSE)

Fig. 6 Generalized view of

random forest regression

123
International Journal on Interactive Design and Manufacturing (IJIDeM)

Fig. 8 a Correlation between experimental results and linear regression using sckit learn model predictions for training data set. b Correlation
between experimental results and linear regression using sckit learn model predictions for testing dataset

4.2 Regression diagnostic for support vector 4.3 Regression diagnostic for support Random
regression model Forest Regression model

The R2 score training and testing data was found 0.951 and The R2 score training and testing data was found 0.931 and
0.753 respectively, while root mean square error for training 0.296 respectively, while root mean square error for training
and testing data was found to be 0.21 and 0.84 respectively and testing data was found to be 0.22 and 0.99 respectively
as shown in Fig. 9a, b. (Fig. 10a, b).

123
 International Journal on Interactive Design and Manufacturing (IJIDeM)

Fig. 9 a Correlation between experimental results and support vector regression model prediction for training data set. b Correlation between
experimental results and support vector regression model predictions for testing dataset

4.4 Regression diagnostic for support KNN 4.5 Comparative analysis of ML models
Regression model
The comparative analysis of correlation between experimen-
The R2 score training and testing data was found 0.78 and tal value of MRR and machine learning model adopted in the
0.12 respectively, while root mean square error for training study are shown in bar chart (Figs. 12 and 13) in terms of R2
and testing data was found to be 2.07 and 7.95 respectively score and RMSE values. It can be inferred from comparison
Fig. 11a, b. that support vector regression and linear regression model

123
International Journal on Interactive Design and Manufacturing (IJIDeM)

Fig. 10 a Correlation between experimental results and random forest regression model prediction for training data set. b Correlation between
experimental results and random forest regression model predictions for testing dataset

must be preferred as compared to other models such as Ran- made it possible to understand how the WEDM procedures
dom Forest and K-Nearest Neighbor algorithm models due affected structural modifications. Figure 14 displays the sam-
to higher accuracy found in current research study (Figs. 12 ple’s SEM picture. A thicker coating of debris, deep craters,
and 13) larger-sized spherical droplets (SPD), and a substantial num-
ber of fissures are all shown in Fig. 14’s surface texture.
Cracks are sometimes referred to as microcracks because
5 Microstructure analysis they are the result of internal stresses that occurred because
of a rapid local heating cycle. However, the length and den-
Machined surface after WEDM process was examined to sity of fractures are influenced by the discharge energy as
investigate the surface integrity with the help of scanning well as the material’s thermal characteristics. Peak current,
electron microscope (Make Zeiss EV040). This analysis pulse on and off times, and spark gap voltage all affect the

123
 International Journal on Interactive Design and Manufacturing (IJIDeM)

Fig. 11 a Correlation between experimental results and K-nearest neighbors (KNN) regression model prediction for training data set. b Correlation
between experimental results and random K-nearest neighbors (KNN) Regression model predictions for testing dataset

discharge energy. The development of high thermal stresses formation of spherical droplets on the machined surface may
that are greater than the material’s ability to withstand frac- be the result of the molten work material solidifying during
ture in combination with plastic deformation is indicated by the machining process. The reduction of surface energy dur-
the presence of microcracks [23]. Higher thermal stresses ing solidification is related to the spherical form of droplets
may have evolved at the machined surface because of the [24].
strong heat discharge (longer pulse length and peak current)
followed by quick cooling (due to short pulse off time). The

123
International Journal on Interactive Design and Manufacturing (IJIDeM)

1. The prediction capability of machine learning techniques

(Linear Regression using scikit-learn and Support Vector
Regression (SVR)) was found to be very accurate when
compared to other models such as Random Forest (RF),
K-Nearest Neighbors Regression (KNN).
2. The high R2 score for training and testing data between
model predictions and experimental data was found 0.917
and 0.775 respectively by applying linear Regression
using scikit-learn while it was found 0.951 and 0.753
for training and testing respectively using support vec-
tor regression (SVR). The above model shows a good
agreement between experimental and modeling results.
A low RMSE value (0.23 and 0.21 for training data set
and 0.57 and 0.75 for testing data set) for above model
also validate it.
3. The analysis and visualization for the datasets was done
Fig. 12 Comparative bar chart (R2 score) of ML models
by correlation heatmap matrix and scatter pairplot to find
out the co-relation between the MRR and process param-
eters of WEDM which reveals that MRR increases as
pulse on time (Ton ), Peak current (IP) and wire feed rate
(WF) are increases due to positive correlation while MRR
decreases as pulse off time (T0ff ), wire tension (WT) and
servo voltage (SV) are increased due to negative cor-
relation. Pulse on time (TON ) and servo voltage (SV)
are more influential process parameters for MRR rather
than Peak current (IP) and Pulse off time (TOFF) . SEM
analysis explains the formation of debris, deep craters,
larger-sized spherical droplets (SPD), and a substantial
number of crack formation due to influencing effect of
process parameter of WEDM during machining of pure
titanium.
4. The above study has utilized the four machine learn-
ing models out of which two models (Linear Regression
using scikit-learn and Support Vector Regression (SVR))
could be used for MRR data prediction efficiently. As far
Fig. 13 Comparative bar chart (RMSE score) of ML models
as future scope is concerned more modification and iter-
ations will be required in Random Forest and K-Nearest
Neighbors (KNN Regression) models. Apart from it
6 Conclusion deep learning models could also refine the application of
machine learning in the above said manufacturing pro-
In this research work, machine learning regression mod- cess.
els were developed using python in Jupyter notebook after
machining of pure titanium using WEDM process. MRR
was predicted using scikit-learn, Support Vector Regression
(SVR), Random Forest (RF), K-Nearest Neighbors Regres-
sion (KNN). The important conclusions are summarized
below:

123
 International Journal on Interactive Design and Manufacturing (IJIDeM)

Fig. 14 Microstructure of the surface of pure Titanium machined after WEDM process (CRT crater, CR crack, DB debris, SPD spherical deposit)
(TON  0.9 µs, TOFF  9.5 µs, IP  200A, SV  50 V, WT  1200 g, WF  10 m/min) a ×3000, b ×5000

Data availability The data that support the findings of this study are Forum, 969. MSF, (2019). https://doi.org/10.4028/www.scientific.
available from the corresponding author upon reasonable request. net/MSF.969.800
8. Nain, S.S., Garg, D., Kumar, S.: Performance evaluation of the
WEDM process of aeronautics super alloy. Mater. Manuf. Process.
(2018). https://doi.org/10.1080/10426914.2018.1476761
9. Nain, S.S., Garg, D., Kumar, S.: Evaluation and analysis of cutting
References speed, wire wear ratio, and dimensional deviation of wire electric
discharge machining of super alloy Udimet-L605 using support
1. Ezugwu, E.O., Wang, Z.M.: Titanium alloys and their machinabil- vector machine and grey relational analysis. Adv. Manuf. (2018).
ity—a review. J. Mater. Process. Technol. 68(3), 262–274 (1997) https://doi.org/10.1007/s40436-017-0192-7
2. Ezugwu, E.O., Booney, J., Yamane, Y.: An overview of machin- 10. Subrahmanyam, S.V., Sarcar, M.M.M.: Application of machine
ability of aeroengine alloys. J. Mater. Process. Technol. 134(2), learning techniques in the process modeling of WEDM. Int. J. Sci.
233–253 (2003) Res. Publ. 3(12), 1–7 (2013)
3. Narayanan, G.S., Vasudevan, D.: A mathematical model for wire 11. Pujara, J.M., Kothari, K.D., Gohil, A.V.: Process parameter opti-
cut electrical discharge machine parameters using artificial neural mization for MRR and surface roughness during machining LM6
network. Adv. Mater. Res. 984–985, 9–14 (2014). https://doi.org/ Aluminum MMC on WEDM. Adv. Eng. Forum (2017). https://doi.
10.4028/www.scientific.net/AMR.984-985.9 org/10.4028/www.scientific.net/aef.20.42
4. Devarasiddappa, D., Chandrasekaran, M., Ravikumar, M., Thirug- 12. Reddy, M.C., Venkata Rao, K., Suresh, G.: An experimental investi-
nanasambandam, M.: Modified teaching learning based optimiza- gation and optimization of energy consumption and surface defects
tion for maximization of MRR in wire-cut EDM of Ti6Al4V alloy in wire cut electric discharge machining. J. Alloys Compd. (2021).
for sustainable production. In: AIP Conference Proceedings, vol. https://doi.org/10.1016/j.jallcom.2020.158582
2128 (2019). https://doi.org/10.1063/1.5117969 13. Shadab, M., Singh, R., Rai, R.N.: Multi-objective optimization
5. Thankachan, T., Soorya Prakash, K., Malini, R., Ramu, S., Sun- of wire electrical discharge machining process parameters for
dararaj, P., Rajandran, S., Rammasamy, D., Jothi, S.: Prediction Al5083/7% B 4C composite using metaheuristic techniques. Arab.
of surface roughness and material removal rate in wire electri- J. Sci. Eng. (2019). https://doi.org/10.1007/s13369-018-3491-9
cal discharge machining on aluminum based alloys/composites 14. Singh Nain, S., Sai, R., Sihag, P., Vambol, S., Vambol, V.: Use of
using Taguchi coupled Grey Relational Analysis and Artificial machine learning algorithm for the better prediction of SR pecu-
Neural Networks. Appl. Surf. Sci. (2019). https://doi.org/10.1016/ liarities of WEDM of Nimonic-90 superalloy. Arch. Mater. Sci.
j.apsusc.2018.06.117 Eng. (2019). https://doi.org/10.5604/01.3001.0013.1422
6. Rees, A., Brousseau, E., Dimov, S.S., Bigot, S., Griffiths, C.A.: 15. Paturi, U.M.R., Cheruku, S., Pasunuri, V.P.K., Salike, S., Reddy,
Development of surface roughness optimization and prediction for N.S., Cheruku, S.: Machine learning and statistical approach in
the process of wire electro-discharge grinding. Int. J. Adv. Manuf. modeling and optimization of surface roughness in wire electrical
Technol. 64(9–12), 1395–1410 (2013). https://doi.org/10.1007/s0 discharge machining. Mach. Learn. Appl. 6, 100099 (2021). https://
0170-012-4110-7 doi.org/10.1016/j.mlwa.2021.100099
7. Shukla, S.K., Priyadarshini, A.: Application of machine learning 16. Ulas, M., Aydur, O., Gurgenc, T., Ozel, C.: Surface roughness pre-
techniques for multi objective optimization of response variables in diction of machined aluminum alloy with wire electrical discharge
wire cut electro discharge machining operation. Materials Science machining by different machine learning algorithms. J. Mater. Res.

123
International Journal on Interactive Design and Manufacturing (IJIDeM)

Technol. 9(6), 12512–12524 (2020). https://doi.org/10.1016/j.jmrt. 23. Lee, S.H., Li, X.: Study of the surface integrity of the machined
2020.08.098 workpiece in the EDM of tungsten carbide. J. Mater. Process. Tech-
17. Jatti, V.S., Dhabale, R.B., Mishra, A., Khedkar, N.K., Jatti, V.S., nol. 139, 315–321 (2003)
Jatti, A.V.: Machine learning based predictive modeling of elec- 24. Huang, C.A., Hsu, F.Y., Yao, S.J.: Microstructure analysis of the
trical discharge machining of cryo-treated NiTi, NiCu and BeCu martensitic stainless steel surface fine cut by the wire electrical
alloys. Appl. Syst. Innov. (2022). https://doi.org/10.3390/asi506 discharge machining (WEDM). Mater. Sci. Eng. A 371, 119–126
0107 (2004)
18. Srinivasan, D., Ganesh, N., Ramakrishnan, H., Balasundaram, R.,
Sanjeevi, R., Chandran, M.: Surface topography: metrology and
properties investigation of surface roughness and material removal
Publisher’s Note Springer Nature remains neutral with regard to juris-
rate of WEDM of SS304 using ANOVA and regression models.
dictional claims in published maps and institutional affiliations.
Surf. Topogr. Metrol. Prop. 10, 025014 (2022). https://doi.org/10.
1088/2051-672X/ac6c9e
Springer Nature or its licensor (e.g. a society or other partner) holds
19. Chou, P.H., Hwang, Y.-R.: The study of machine learning for wire
exclusive rights to this article under a publishing agreement with the
rupture prediction in WEDM. Int. J. Adv. Manuf. Technol. (2021).
author(s) or other rightsholder(s); author self-archiving of the accepted
https://doi.org/10.21203/rs.3.rs-755393/v1
manuscript version of this article is solely governed by the terms of such
20. Saha, S., Gupta, K.M., Dey, S.: Data-driven probabilistic perfor-
publishing agreement and applicable law.
mance of Wire EDM: a machine learning based approach. Proc.
Inst. Mech. Eng. Part B J. Eng. Manuf. (2021). https://doi.org/10.
1177/09544054211056417
21. Jiang, J.R., Yen, C.T.: Product quality prediction for wire electrical
discharge machining with Markov transition fields and convolu-
tional long short-term memory neural networks. Appl. Sci. (2021).
https://doi.org/10.3390/app11135922
22. Walia, A.S., Srivastava, V., Rana, P.S., Somani, N., Gupta, N.K.,
Singh, G., Pimenov, D.Y., Mikolajczyk, T., Khanna, N.: Predic-
tion of tool shape in electrical discharge machining of EN31 steel
using machine learning techniques. Metals (2021). https://doi.org/
10.3390/met11111668

123