ARTICLE

Machine learning based modelling and optimization in hard turning

of AISI D6 steel with newly developed AlTiSiN coated carbide tool

Anshuman Dasa , Sudhansu Ranjan Dasb , Jyoti Prakash Pandaa , Abhijit Deyc ,
Kishor Kumar Gajranid , Nalin Somania , Nitin Kumar Guptaa
a
Department of Mechanical Engineering, DIT University, Dehradun, Uttarakhand, India;
arXiv:2202.00596v1 [[Link]] 30 Jan 2022

b
Department of Production Engineering, VSSUT, Burla, Odisha, India;
c
Department of Mechanical Engineering, NIT, Srinagar, Jammu and Kashmir, India;
d
Department of Mechanical Engineering, IIITDM, Kancheepuram, India

ARTICLE HISTORY
Compiled February 2, 2022

ABSTRACT
In recent times Mechanical and Production industries are facing increasing chal-
lenges related to the shift toward sustainable manufacturing. In this article machin-
ing was performed in dry cutting condition with a newly developed coated insert
called AlTiSiN coated carbides coated through scalable pulsed power plasma tech-
nique in dry cutting condition and a dataset was generated for different machining
parameters and output responses. The machining parameters are speed, feed, depth
of cut and the output responses are surface roughness, cutting force, crater wear
length, crater wear width and flank wear. The data collected from the machining
operation is used for the development machine learning (ML) based surrogate mod-
els to test, evaluate and optimize various input machining parameters. Different
ML approaches such as polynomial regression (PR), random forest (RF) regression,
gradient boosted (GB) trees and adaptive boosting (AB) based regression are used
to model different output responses in the hard machining of AISI D6 steel. The
surrogate models for different output responses are used to prepare a complex objec-
tive function for the germinal center algorithm based optimization of the machining
parameters of the hard turning operation.

KEYWORDS
Hard turning; AlTiSiN coating; Surface roughness; Flank wear; Cutting force;
Crater wear; Machine learning

1. Introduction

Numerous manufacturing industries worldwide are working on smart manufacturing

techniques for improving productivity. Various new techniques are being tested for en-
hancement of productivity according to the industrial revolution, Industry 4.0. Some of
these techniques are the Internet of Things (IoT), Machine Learning, Cyber-Physical
system, big data, machines with computer controls and sensors and additive manufac-
turing (Z. Zhang et al., 2020; Akhil et al., 2020; Mishra et al., 2020). By the applications
of the above mentioned techniques, the production level has shown increments. Among
all the mentioned techniques, IOT and big data are the most popular techniques Surler-
aux et al. (2020). The industries 4.0 are working on the information systems based on

Corresponding Author Email: jppanda@[Link]

the sensors attached with the machines (Bricher & Müller, 2020). The manufactur-
ing units are established and synchronized with various information systems through
IOT and utilized in different production management purposes (O’Donnell & Yoon,
2020). Cyber physical systems allow the data acquisition in the real world which sup-
ports smart manufacturing. Nowadays, the research in Machine Learning is flourishing
because of large amounts of data that have been accumulated by various industries
such as production, health, chemical, manufacturing and information technology (IT).
Machine learning plays a vital role for enhancement in demand for artificial intelli-
gence(AI) (Nagargoje et al., 2021). It is a sub branch of AI, which allows the equipment
to learn, improve and perform the specific task without disturbing the program. There
are five steps involved in machine learning to solve a problem: problem definition, data
collection, modelling, evaluation and result analysisJ. P. Panda & Warrior (2021).
There are three broad categories of machine learning algorithms: supervised learn-
ing, unsupervised learning, and reinforcement learning. In supervised learning, data
are trained for mapping between input and output. Knowing the ideal output for every
sample in the input data acts as a guide for the learning procedure, thus providing
supervision. In unsupervised learning, the ideal output is unknown and the model can
only learn patterns present in the input data. In reinforcement learning, a predefined
signal will be given, depending upon this, the machine will quantify the objectives.
These broad ML approaches can be used for various tasks, such as clustering, dimen-
sionality reduction, surrogate modeling for classification or regression, etc. ML based
data-driven surrogate models can act as Digital Twins (DT) for real-life systems and
these can be used for system exploration and optimization with a low computational
expense.
At present, industries are conducting smart machining rather than conventional ma-
chining. Because errors are more severe in conventional machining, production is not
effective and less efficient. In smart manufacturing, efficient production can be achieved
with a stipulated processing time without hampering the quality of the product. In
smart manufacturing, parameters can be optimized in real time through different sub-
systems and machines like controllers, sensors, systems, and machine tools. Numer-
ous researchers have applied data-driven analysis for machining and production-based
problems. The objective of these studies spans from using data to develop a deeper
understanding of the system, to developing ML based models to act as surrogates for
the system. Kothuru et al. (2018) determined the wear of the cutting tool and its
failure using sound signals with the application of Support vector machine (SVM).
The tool wear was predicted accurately through the ML algorithm for different ma-
chining conditions. Carrino et al. (2020) predicted the machining quality using ML
techniques. The convolutional neural network (CNN) was used to predict and monitor
the machined surface quality and the results showed that by combining both tech-
niques resulted in 94% accuracy of prediction. Madhusudana et al. (2016) studied the
fault diagnosis of face milling using the ML approach. For the identification of signifi-
cant parameters, a decision tree algorithm was used. Also, the Naive Bayes model was
used for the prediction of fault. It was observed that using the model, the results are
predicted with 96.9% accuracy. The ML models developed can be utilized for online
monitoring of tool conditions and diagnosis of the fault of the tool used in milling
operation. J. Wang et al. (2019) predicted the responses of wire electrical discharge
machining (WEDM) through voltage signal using AI technique. From the analysis, it
was found that the unsupervised AI techniques were predicted. Krishnakumar et al.
(2018) predicted the tool condition in a high-speed milling process using the ML tech-
nique. Four types of ML algorithms are employed, those are artificial neural network

2
(ANN), decision trees, naive Bayes, and SVM. The SVM was found to be the most
efficient method to predict the tool condition compared to other techniques. Q. Zhang
et al. (2020) predicted the output responses in laser machining using deep multi-task
learning. In this experiment, it was observed that Alex Net with multi-task learning
was found to be better than deeper models. J. Wang et al. (2020) detected the ge-
ometrical defects in WEDM using a physics-guided ANN model. The inconsistencies
in the model were eliminated using a physics guided loss function, from the analysis
it was observed that the predicted results matched well with the experimental results
with 80% accuracy. Cho et al. (2005) detected the tool breakage in a milling opera-
tion using a support vector-based machine learning approach. From the experimental
outcomes, it was confirmed that with the said technique, the machine downtime and
cost of production can be effectively reduced compared to other techniques. P. Wang
et al. (2019) predicted the tool condition through the data-driven hybrid ML ap-
proach during milling of H13 steel and Inconel 718. From the results, it was found
both the tool wear and surface roughness were predicted effectively with an accuracy
of 85% and 90%. Liu et al. (2020) predicted the specific cutting energy with a hybrid
approach of integrated ML technique. Both data-driven ML and process mechanics
have been hybridized to predict the response. The results are well predicted and the
authors claimed that the above-said model can be implemented in other cutting pro-
cesses. Chiu & Lee (2017) predicted the machining accuracy and surface quality using
a data-driven approach by (Adaptive Neuro-Fuzzy Interface System) ANFIS model.
From the simulation and experimental results, it was found that results can be effec-
tively predicted through the data-driven model for better quality and productivity.
McLeay et [Link] et al. (2020) detected the fault during the machining process
using an unsupervised ML approach. Through the PCA plot, the fault conditions are
observed. Parwal & Rout (2021) used the machine learning-based approach to predict
the tool wear during the machining operation. Various models are employed using
logistic regression. The model was found to be good and results are interpreted ef-
fectively. And the results are suitable for industrial application. Gouarir et al. (2018)
predicted the tool wear through the machine learning techniques. From the results,
it was confirmed that the tool wear was predicted effectively with machine learning
techniques and the accuracy was 90%. Wu et al. (2017) compared the tool wear predic-
tion results through different ML techniques. Three types of algorithms are used like
feed-forward back propagation(FFBP) based ANNs, random forest (RF) and support
vector regression (SVR). Better results are predicted using the RF algorithm. Fang et
al. (2016) used a new artificial intelligence approach for the prediction of machined
surface roughness in metal machining. ANN model was implemented. From the ANN
model, it was observed that the machined surface roughness can be effectively pre-
dicted through the ANN model. Ulas et al. (2020) predicted the surface roughness
of the machined surface of Aluminum alloy during WEDM with different machine
learning algorithms. Four types of machine learning algorithms are applied such as
ELM, W-ELM, SVR, and Q-SVR. Among four algorithms, weighted extreme learning
machines (WELM) performed better. Bustillo et al. (2021) predicted the flatness devi-
ation considering the wear of the face mill cutter teeth using a machine learning-based
algorithm. Four different types of machine learning techniques were proposed, out of
which, Random forest ensembles combined with the Synthetic Minority oversampling
technique performed better. Patange & Jegadeeshwaran (2021) predicted the health of
a vibration-based multi-point tool insert on a vertical machine center using a machine
learning approach. A tree-based algorithm was proposed by the authors. Various tree-
based algorithms are used, out of which, the J48 decision tree-based algorithm was

3
found to be the effective one. Oleaga et al. (2018) predicted the chatter in heavy-duty
milling operation using machine learning techniques. Different machine learning mod-
els are used out of which, random forest performed better. And the authors concluded
that with the use of machine learning techniques, dynamic machining performance can
be enhanced in real working conditions. Peng et al. (2019) predicted the cutting forces
with consideration of the tool wear using machine learning methods. Two models are
used one is conventional linear regression and the other is a hybrid model with ma-
chine learning. From the results, it was observed that the hybrid model with machine
learning performed better. Mohanraj et al. (2021) developed a tool condition moni-
toring system at the end milling process using wavelet features and statistical features
based machine learning approach. Four different types of machine learning techniques
are used to predict the flank wear like SVM,, K-nearest neighbor (KNN), knowledge
based (KB), decision tree, and MLP, and among all, SVM and DT performed better.
Sanchez et al. (2018) predicted the thickness change of machined parts during WEDM
machining using machine learning techniques. Various models are used to predict the
results. Among various models, a first convolutional layer with two gated recurrent
units outperformed the other models. They also concluded that with the large data
set, better results can be predicted. Shen et al. (2020) predicted the tool wear size in
multi-cutting conditions using advanced machine learning techniques. Different models
are used to predict the tool wear and the results are compared with the experimental
results. It was observed that there was a good agreement between predicted and exper-
imental results. Moreover, the authors studied that machine learning techniques can
be used to predict the other machining responses. Shastri et al. (2021) optimized the
process parameters in the machining of Titanium alloy in an MQL environment using
an artificial intelligence-based algorithm. From the experimental results, it was ob-
served that better results in terms of cutting force, tool wear, tool-chip contact length,
and surface roughness were obtained using artificial intelligence-based optimization
compared to PSO and the experimental approach. Cheng et al. (2020) predicted the
tool wear based on machine learning during the turning of high-strength steel. Two
prediction models are used one is grid search algorithm-based support vector and
the other is genetic algorithm-based support vector regression. From the results, it
was found that with the tool wear, the stages of tool wear in a complete cycle can
be easily predicted through machine learning techniques. Moreover, they suggested
that through the machine learning method, tool wear monitoring can be made online.
Elsheikh et al. (2021) predicted the residual stress in turning of Inconel-718 using
fine tuned AI model using pigeon optimizer. Two models were incorporated like hy-
brid ML model and traditional ANN model. The traditional ANN was incorporated
with three optimization techniques, those are bio inspired optimization, pigeon and
Particle swarm optimization. The prediction accuracy of various models was exam-
ined through statistical measures. The traditional ANN was out performed by both
ANN-PSO and ANN-PAO. Jurkovic et al. (2018) compared different machine learning
methods for cutting parameter prediction in a high-speed turning process. Three types
of ML techniques are compared i.e. SVR, Polynomial regression, and ANN. Cutting
force, roughness, and machining time were the output responses. Better results in
terms of cutting force and roughness were achieved with polynomial regression. Better
machining time was predicted through ANN. The author also observed that for better
prediction of results, more data set are required. Cica et al. (2020) predicted vari-
ous machining responses under different cooling conditions in sustainable machining
of 1045 steel using machine learning techniques. From the results, they revealed that
an acceptable range of results can be predicted by using machine learning techniques

4
without conducting actual experiments. So both time and money can be saved by
using machine learning techniques. They also revealed that if a wide range of machin-
ing conditions will be adopted, better results can be predicted through this method.
Khoshaim et al. (2021) predicted both mechanical and micro-structural properties of
friction stir processed aluminium reinforced material using Grey Wolf optimizer. The
input parameters are rotational speed, linear processing speed and number of passes,
while the outputs were grain size, aspect ratio, micro-hardness and ultimate tensile
strength. The prediction accuracy of developed hybrid model was found to be more
accurate compared to standalone model.
Zhao et al. (2020) predicted the specific energy consumption of machine tool based
on tool wear in a dry milling operation. Three power characteristics are chosen like,
standby power, cutting material power and no load power. Three responses are chosen
spindle speed, MRR and tool wear. Both the process parameters and tool conditions
are optimized for the reduction in energy consumption of machine tool. Rašovi&39;c
(2021) recommended the layer thickness in a powder based additive manufacturing
using multi attribute decision support. Many attributes are responsible for the prod-
uct quality. But author has chosen two criteria one is strength and other is surface
roughness. The most suitable layer thickness was proposed by three techniques called
weighted features, DFAM knowledge and multiple attributes. Kahya et al. (2021) per-
formed the precision and energy efficient machining of Ti6Al4V alloy on a turn mill
machine tool. Three operations are conducted like face turning, rough flank milling
and finish milling. Three responses are analyzed known as specific cutting energy,
roughness and material removal rate. Different angle of inclinations (both lead and
tilt) was analyzed on the response. The optimization technique employed was Particle
swarm optimization. From the experimental outcomes, it was found that both lead
and tilt angle influenced surface roughness effectively. Guo et al. (2021) combined two
methods DPCA and IMODE to analyze the effect o9f input variables on the gear
quality in a hobbing process. The simulation of gear hobbing data was done and it
was compared with NSGA, NSGAII and MODE. From the outcomes it was found that
the proposed algorithm was very effective regarding diversity and optimization ability.
Ruiz et al. (2020) predicted the tensile strength of the steel rods by machine learning
algorithms those are manufactured in an electric arc furnace. Various ML algorithms
are proposed like multiple linear regression, K-Nearest Neighbors, Classification and
Regression Tree, Random forest, Adaboost gradient boost algorithms and ANN. Bet-
ter results are predicted through fine-tuned random forest. And chemical variables are
observed to be the most important variable affecting the material strength. Tian et al.
(2020) optimized the cutting parameters and processing sequence to minimize carbon
emissions and process time in a CNC machine using an integrated multi objective
optimization technique. The modelling of process parameters is done using NAGS-II
technique. From the results, it was found using integrated multi objective optimization
technique, both processing time and carbon emission can be minimized effectively.
Saidi et al. (2019) used response surface methodology for modeling the output ma-
chining responses and optimized the input parameters using a desirability function
[Link] et al. (2019) used a similar approach as of Saidi et al. (2019) for
modeling and optimization of the machining process parameters in ultra-high preci-
sion machining of optical lenses. Caggiano et [Link] et al. (2018) optimized the
tool life exploitation during CFRP composite drilling in aeronautical assembly using
ML techniques. From the experimental outcomes, it was found that both fractal and
statistical analysis can predict better results. Shukla & Priyadarshini (2019) applied
ML techniques for optimization in WEDM of Haste alloy C-276. From the experi-

5
 Element Fe C Si Mn P S Cr W

% Wt. 84.789 2.014 0.29 0.43 0.022 0.005 11.77 0.68

Table 1.: Chemical composition of AISI D6 alloy die steel

mental outcomes, it was observed that both surface roughness and kerf width were
influenced by pulse on time, pulse off time, and peak current. And the gradient descent
method was successfully implemented for the prediction of optimum results. Nain et
al. (2018) analyzed the cutting speed, wire wear ratio, and dimensional deviation of
WEDM machining of superalloy Udimet L605 using both support vector and grey
relational analysis. From the experimental results, it was observed that pulse on time
was the significant variable affecting the responses. The highest percentage of Copper
(8.66%) was obtained during machining at the highest cutting speed, whereas with
low cutting speed 7.05% of Copper was observed.
Although various works related to the application of RSM and ML techniques in
modeling and predictions of manufacturing processes are available in the literature,
there is very limited research published related to optimization of process parameters
of turning operation using ML-assisted meta-heuristic optimization algorithms. In this
article, the different output responses (such as surface roughness, cutting force, crater
wear length, crater wear width, and flank wear) during hard turning of AISI D6 steel
are modeled in terms of speed, feed, and depth of cut. We have used four different
ML algorithms for developing surrogate models to predict the output responses. The
four ML algorithms are polynomial regression (PR), Random Forests (RF), Adap-
tive Boosting (AB), and Gradient Boosting (GB). The predictive capability of the
ML models is assessed through a comparison of the different performance parameters
such as mean squared error (MSE), mean absolute error (MAE), and the calculated
value of coefficient of determination R2 . The surrogate model learned through PR
for five different output parameters is used to prepare a complex objective function
and its optimization for finding the optimal set of process parameters that minimizes
the outputs using system-based germinal center optimization [Link] detailed
framework used in study is shown in figure 1.
The rest of the article is arranged as follows: In section 2, the details of the exper-
imental procedure for the generation of data to be used in ML-based modeling are
discussed. In section 3, the coating characterisation of the cutting tool is discussed.
In section 4, 5, 6 and 7 the effect of input parameters on the output responses are
discussed. In section 8, the dataset used in the ML model development is discussed
in detail, including the features and levels used and the scaling procedure adopted.
The details of ML models are discussed in section 9. In section 10, the theory of the
germinal center algorithm is discussed. In section 11, the ML-based prediction of re-
sponses and multi-objective optimization are discussed and finally, in section 12, the
concluding remarks are presented.

6
Figure 1.: The detailed framework used in optimization of machining parameters.

7
2. Experimental setup of the turning operation and procedure of data
collection

The workpiece considered in this investigation corresponds to a shape of a cylindrical

bar having a specification of 350 mm length and 50 mm diameter. The designation of
the workpiece was AISI D6 steel. The particular grade of steel was selected because this
steel grade of steel has vast applications in particularly tool and die making industries.
And it has good dimensional stability with good wear and abrasion resistance. Table 1
shows the chemical composition of the material. As the machining is hard machining, to
retain the hardness, the workpiece was undergone heat treatment, called tempering and
quenching. Quenching was accomplished at 900◦ C and tempering was accomplished
at 420◦ C. Finally, 65HRC was achieved. Uncoated carbide inserts are procured from
Kennametal with a square shape having geometry SNMG 120408. Then the uncoated
inserts are coated using S3P technology called scalable pulsed power plasma. The thin
hard film coating which was deposited on the substrate was AlTiSiN coating. Before
coating, carbide inserts are properly cleaned and polished. Cleaning was done using
an ultrasonic bath and micro blasting. The scalable pulsed power plasma technique is
a combined methodology of magnetron sputtering and cathode arc evaporation. The
coating was done in Oreilikon Balzer, Pune, India. The Oerlikon Balzer make coating
machine having model INLENIA was employed for this scalable pulsed power plasma
coating. For coating, normally three operations are performed called, etching, heating
and coating. In the beginning stage of the coating, the temperature of the coating
chamber was maintained at 450◦ C and the vacuum pressure was maintained at 10−6
mill bar. The inserts are placed in the holder then it was rotated with a speed of 2
rpm speed. For the smooth deposition of coating on the substrate with nano layer
structure and uniform thickness, two fold motion was carried out. Two fold motion
was accomplished with motor. Both etching and thermal treatment on the substrate
are done. The etching was named as sputter etched. Sputter etching and thermal
treatment are done to remove the surface contaminants from the substrate. A voltage
of 450V and current of 20 A is normally maintained. The coating was accomplished
with the argon gas pressure of 350 MPa. Temperature was reduced gradually to 150℃
after completion of the coating. Then finally the inserts with coating was kept outside
with natural air cooling. The AlTiSiN coating deposition was completed using one
target each of AlTiN and TiSiN with 5 KW target power and 99.95% purity. The
thickness of AlTiN coating was 0.5± 0.1 micron and that of TiSiN was 0.2 ± 0.1 micron.
The final coating layer thickness was 1.3 ± 0.1 micron. Good adhesion strength was
observed between the coating material and substrate. Inserts (substrate) are placed in
a negatively biased coating chamber and from the positively charged source (AlTiSiN),
arc was produced. Then positively charged ions are attracted towards the substrate
and got deposited. After coating, the inserts are clamped on a tool holder called
PSBNR 2020K12 having an approach angle 75o and nose radius of 0.8 mm. The dry
hard turning on AISI D6 steel was performed on a high precision heavy duty lathe
(Model NH-26, make- HMT). The maximum spindle power was 11KW and the spindle
speed was in the range between 40-2040 rpm. For cutting force (Fz) measurement, 4-
component piezoelectric dynamometer was used, (Make- Kistler, Model-9272), The
Mitutoyo SJ-210 roughness tester was employed to measure the machined surface
roughness. Both optical and scanning electron microscope was employed to analyze
the machined surface morphology and tool wear (both at flank and rake face). In
the current experimental investigation, three input parameters are considered such as
speed, feed, and depth of cut. As the depth of cut is not an effective input variable

8
 Figure 2.: Figure SEM images showing (a) surface morphology and (b)
cross-sectional morphology of AlTiSiN coatings with EDS

compared to the speed and feed only four levels are considered for depth of cut,
whereas for both speed and feed seven levels were chosen. Speed (40, 50, 55, 60, 70,
80 and 90 m/min), feed (0.04, 0.06, 0.08, 0.1, 0.12, 0.14 and 0.16 mm/rev) and depth
of cut (0.2, 0.3, 0.4 and 0.5mm) to analyze different machining characteristics in dry
cutting condition. The above parameters and their levels are selected according to the
literature review and the tool manufacturer’s recommendation. The detail procedure
adopted in the present work with optimization is shown in Fig.1.

3. Coating characterization of cutting tool

If machining will be conducted with coated inserts, in dry cutting circumstances, coat-
ing should possess two important characteristics like anti- wear and anti-oxidant. The
surface morphology of AlTiSiN coated carbide insert is shown in Figure 2. The scan-
ning electron microscopic (SEM) image (figure 2) delineated that high dense structure
with droplet deposition was found. The cross section of AlTiSiN is illustrated in Fig.
columnar free structure was observed from the image. This was mainly due to the
existence of crystalline-amorphous phases of both AlTiN and TiSiN structures. Nano
crystalline TiN is dispersed to large amount of Si3N4 amorphous material in TiSiN
resulting a smoother surface with glassy appearance. TiSiN offers high oxidant resis-
tant because SiO2 forms an antioxidant layer resulting higher hardness. The hardness
mainly depends upon the following factors, a) robust covalent bond formation like
SiN compound, b) Strengthening the structure by incorporating various atoms and
c) Grain growth of nitrides due to strong amorphous surrounded SiN compounds in
grain refinement process.
Also higher hardness will be achieved with higher silicon content and lower crys-
talline size. Moreover, the higher volume of silicon content and smaller crystalline size
is responsible for the evanescement of columnar structure.

9
 Figure 3.: Optical micro-graphs illustrating the flank wear under different cutting
speeds at f = 0.16 mm/rev and d = 0.5 mm

4. Analysis of tool flank wear

As shown in figure 3, it was evident that flank wear is increasing with the cutting
speeds, four cutting speeds are chosen and feed and depth of cut were kept constant
as 0.16mm/rev and 0.5mm respectively. The increment trend of flank wear might be
due to the thermal softening near the cutting edge of the insert. Thermal softening
is mainly influenced by the temperature development or heat generation at tool-work
interface during machining. Moreover, heat is an energy normally flows from high
temperature region to low temperature region, but it needs some time. But when
machining will be accomplished at higher cutting speed, the time will be insufficient
for effective heat transfer. Thermal conductivity of the material is influenced by its
hardness. As the work piece was heat treated for the hardness enhancement, its thermal
conductivity might be hampered. That’s why chances for effective heat transfer to the
work piece will be less. As a result, inserts might be exposed to high temperature
resulting thermal softening near the cutting edge. For which more wear at the flank
surface of the insert was observed at higher speed values. As stated in the optical
microscopic images in Fig.3 Uniform wear land and abrasion marks are observed.
Smooth abrasion marks are observed when machining was accomplished at low and
medium speed shown in Fig.5. But when machining was done with high speed, thick
abrasion, chipping and plastic deformation observed which is shown in Fig.4. Adhesion
of chips and micro grooves are also observed which is shown in the SEM image (Fig.5).
The abrasion marks are observed due to abrasive nature of coarse grain structures and
different carbides precipitation, which might be developed due to heat treatment of the
test specimen for enhancement in hardness. Two main wear mechanisms are observed
adhesion and abrasion. Both the mechanisms are normally observed when continuous
machining is conducted for a long time. Because due to continuous machining, insert
is subjected to rapid cyclic loading.

5. Analysis of cutting force

Out of different machining characteristics, cutting force is one of the most important
characteristics. Power consumption, dimensional deviation and tool life were signif-
icantly affected by the cutting force. Cutting force was measured with the help of
dynamometer. Experiments are carried out in seven different speeds (40, 50, 55, 60,
70, 80 and 90m/min). From the graphical representation shown in Fig. 6, it was ob-
served that at four different levels of speed such as 40, 50, 55 and 60m/min, the

10
Figure 4.: Microscopic view of the flank surface of AlTiSiN coated insert

11
Figure 5.: Flank wear evaluation of SPPP-AlTiSiN coated carbide tool under dry
cutting conditions at v = 70 m/min, d = 0.4 mm, f = 0.04 mm/rev

12
 Figure 6.: Cutting force vs. cutting speed (low and medium range)

decrements in the cutting force was observed. This might be due to the thermal soft-
ening of the work piece. Because at high speed machining, time will be limited for
the effective heat transfer to the surroundings from the machining zones. So there
is a chance of more heat transfer to the work piece. Moreover, at high speed, more
temperature is generated due to tool-work and tool-chip friction. Due to which shear
strength of the specimen diminished and grain boundary disruption occurred resulting
depletion of cutting force. As discussed earlier (in the section Analysis of tool flank
wear), tool wear was highly influenced by the cutting speed. Particularly for three
ranges of cutting speed i.e., 70, 80 and 90m/min, rapid wear on the flank surface was
observed. That’s why there was a rapid increment in the cutting force as shown in
Fig.7. Cutting force increment was also observed with the depth of cut shown in Fig.8
this might be due to the increment of shear plane area, chip thickness and material
removal rate. Moreover, similar trend was observed for cutting force with axial feed
rates as shown in Fig.10. With lower feed values, less force was observed and with
higher feed values, more force was observed. This might be due to the decrements of
contact length between tool tip and test specimen, for which sharpness of the tool
retained resulting less force. But as feed increased, contact length increased between
tool tip and work piece for which sharpness of the cutting edge demolished resulting
rough machining and generation of more cutting force.

6. Analysis of surface roughness

From the figure 10 it was found that with the increment in the depth of cut, surface
roughness increased by keeping speed and feed constant. This might be due to the
increment of the contact area between the tool and work piece at the machining zone
resulting extreme ploughing force. That leads to the side flow of material which is
responsible for the machined surface degradation. Another reason might be chatter
or self-induced vibration due to tool wear. Moreover, more elastic deformation and
squeezing of materials was observed at tool-work contact zone in lower value of depth of
cut compared to higher value of depth of cut. That’s why more plasticized material flow
might be occurred through insert’s side way resulting machined surface deterioration.

13
Figure 7.: Cutting force vs. cutting speed (higher range of speed)

Figure 8.: Cutting force vs. depth of cut

Figure 9.: Cutting force vs. feed rate

14
 Figure 10.: Effect of flank wear on machined surface morphology under varying
depth of cut at cutting speed = 40 m/min, feed = 0.16 mm/rev

Self-induced chatter or vibration is more susceptible in machining of hard materials.

In current experimental investigation, material having hardness 65 HRC was used.
So chatter was noticed. Moreover, with the increment in depth of cut, contact region
between tool flank face and work piece surface increased. For which both temperature
and pressure might be increased at the contact zone resulting severe abrasion, burnt
marks, plastic deformation at the cutting edge was observed. That’s why deeper and
wider feed marks on the machined surface were observed as noticed in the figure. The
roughness profiles generated by surface roughness tester with varying depth of cut
conditions are shown in figure 11 below where speed and feed were kept constant as
40 m/min and 0.16 mm/rev respectively.

7. Effect of chip-tool contact length on crater wear

When machining was conducted at higher feed values continuous spiral chips are pro-
duced. But one interesting phenomena was observed that chips were up curl in nature.
There was a contact between the chips and the work piece at their end and chips
rolling were observed for more than one revolution. Straightened chips with increased
curvature radius are observed. This might be due to the end contact of the chips with
the specimen. The longitudinal grooves are observed due to the chip sliding at the
tool-chip interface. In the contact region, the grooves are observed but with the chip
flow direction. Due to the continuous chip sliding on the rake surface an irregular side
flow of chips is observed resulting a long tool-chip contact length. Moreover, more
wear on the rake surface was observed. From the Fig.12, it was clear that crater wear
increased with the feed rate because of increment in the tool-chip contact length. Low-
est tool wear was observed with low speed and feed. The microscopic view of the rake
surface of the AlTiSiN coated insert was shown in Fig.13. Various characteristics are
observed like abrasion marks, chipping, multi hot spots and burn spots. With high
speed and feed, these characteristics are observed. Both chip sticking and chip sliding
were the influencing factors for the above said characteristics. Abrasion marks were
mainly observed due to the chip sliding. But hot spots, burn spots and chipping might
be influenced through chip sticking. Due to machining a hard material, hot and hard
chips are produced. As discussed in the above section, more heat might be transferred
to tool material compared to work piece material. So heat will be retained on the tool
surface. So if in this situation, hot and hard chips will pass, due to high temperature,

15
 (a)

(b)

(c)

(d)

Figure 11.: Roughness profile with different depth of cut a) 0.2mm b) 0.3mm c)
0.4mm d) 0.5mm

16
 Figure 12.: Crater wear due to tool-chip sliding contact length with varying feed at
v= 90 m/min; d= 0.2 mm

chip welding or sticking will be occurred resulting hot spots or burn spots. And chip-
ping was observed on the rake surface due to collision of hot and hard chips with the
rake face of the insert.

8. Description of the dataset used in ML model development

The dataset used in this article has eight columns as shown in the table. 2. In this
tabular representation, each column represents a characteristic attribute and each row
represents a specific sample. The first three columns, s, f, d are used as input to the
ML models. These represent features that the algorithm uses to model the target
output. The next five columns are the outputs of the ML models. The total number
of samples is 49. The input features were normalized to the range -1 to 1 to ensure
efficient gradient-based learning. This min-max scaling method for normalization can
be defined as follows:
φ − φmin
φs = (1)
φmax − φmin

where, φs is the new scaled data, φ is the original cell data, φmin and φmax are the
minimum and maximum value in corresponding column. The importance of scaling of
input data is to make the learning process numerically stable. The inputs/ features
of the ML models are s, f and d and the outputs/ levels are Ra, F , CWL , CWW and
F W . We have prepared two train-test datasets (D1 and D2 ) for the development of
regression models. The training dataset of D1 consist of all the data as presented in
table 2, except sl. no 8-14, those correspond to 50 m/min speed. The data correspond
to 50 m/min was used for testing of the developed ML regression models. The train-
test dataset D2 consist of arbitrary testing data from various speed from table 2. Those
are (40, 0.04, 0.2), (50, 0.06, 0.2), (55, 0.16, 0.4), (60, 0.14, 0.5), (70, 0.12, 0.2), (80,
0.1, 0.5) and (90, 0.06, 0.4) in the order (s, f, d). The dataset except the following
combinations were using in model training.

9. Machine learning models for predictive modeling of machining

parameters

Machine Learning is a sub-field of AI that has its focus on enabling computer-based sys-
tems to develop predictive models purely from data. The learning aspect in the name
represents the ability of such algorithms to improve their performance at a task by

17
 Figure 13.: Microscopic view of the rake surface of AlTiSiN coated insert

processing data, without needing any updated human or domain expert instructions.
The data used in ML occurs in pairs of features and targets in supervised learning. By
use of machine learning, a function relating the inputs features to the target outputs
can be developed by using suitable datasets. There are different ML algorithms avail-
able through which a surrogate model can be developed for a set of input and output
data, those are linear regression (LR), polynomial regression (PR), support vector
regression (SVR), and different tree-based algorithms such as Adaptive Boosting (Ad-
aBoost), Gradient-Boost (GB) and Random Forests (RF). The linear regression (LR)
is suitable for simple type linear relationships and is a subset of polynomial regression.
In this investigation, polynomial regression, Random Forests, Gradient Boosting, and
Adaptive Boosting algorithms for predictive modeling of the machining parameters is
considered. Fig. 14 shows the machine learning frame work used for the modeling of
input variables . The decision tree structure used in tree based ML algorithm is shown
in Fig. 15. The predictive performance of the above-mentioned ML algorithms was
assessed by using R2 , M SE (mean absolute error), and M AE (mean absolute error).
Those can be defined as follows:

18
sl. no. s f d Ra F CWL CWW FW
1 40 0.04 0.2 0.45 84 0.25 0.036 0.044
2 40 0.06 0.2 0.5 98.4 0.3 0.053 0.044
3 40 0.08 0.3 0.5 110 0.41 0.075 0.057
4 40 0.1 0.3 0.65 120.77 0.45 0.089 0.060
5 40 0.12 0.4 0.74 149.99 0.56 0.10 0.068
6 40 0.14 0.4 0.80 155.63 0.57 0.12 0.071
7 40 0.16 0.5 0.89 202.99 0.699 0.159 0.070
8 50 0.04 0.2 0.38 72.85 0.23 0.032 0.043
9 50 0.06 0.2 0.43 87.55 0.29 0.048 0.043
10 50 0.08 0.3 0.51 95.65 0.377 0.07 0.055
11 50 0.1 0.3 0.58 107.52 0.42 0.084 0.058
12 50 0.12 0.4 0.67 137.15 0.52 0.10 0.065
13 50 0.14 0.4 0.74 141.91 0.54 0.122 0.069
14 50 0.16 0.5 0.83 189.38 0.66 0.15 0.067
15 55 0.04 0.3 0.35 54.99 0.23 0.034 0.05
16 55 0.06 0.3 0.40 73.15 0.30 0.05 0.05
17 55 0.08 0.2 0.46 96.99 0.32 0.06 0.044
18 55 0.1 0.2 0.53 106.08 0.36 0.07 0.05
19 55 0.12 0.5 0.68 174 0.62 0.12 0.06
20 55 0.14 0.5 0.74 180.88 0.64 0.14 0.06
21 55 0.16 0.4 0.76 137 0.54 0.13 0.07
22 60 0.04 0.3 0.33 50.9 0.23 0.03 0.05
23 60 0.06 0.3 0.38 69.43 0.3 0.05 0.05
24 60 0.08 0.2 0.45 94.4 0.32 0.06 0.05
25 60 0.1 0.2 0.51 103.59 0.35 0.07 0.05
26 60 0.12 0.5 0.51 170.04 0.61 0.12 0.06
27 60 0.14 0.5 0.51 175.62 0.63 0.13 0.06
28 60 0.16 0.4 0.51 132.85 0.53 0.13 0.07
29 70 0.04 0.4 0.51 57 0.28 0.04 0.06
30 70 0.06 0.4 0.51 80.18 0.35 0.05 0.06
31 70 0.08 0.5 0.51 136 0.52 0.09 0.06
32 70 0.01 0.5 0.51 154.53 0.57 0.1 0.06
33 70 0.12 0.2 0.51 108 0.36 0.06 0.06
34 70 0.14 0.2 0.51 108.46 0.37 0.07 0.06
35 70 0.16 0.3 0.51 109 0.45 0.1 0.07
36 80 0.04 0.4 0.51 64.48 0.31 0.03 0.06
37 80 0.06 0.4 0.51 88.01 0.38 0.04 0.06
38 80 0.08 0.5 0.51 137.31 0.53 0.08 0.06
39 80 0.1 0.5 0.51 155.03 0.59 0.09 0.06
40 80 0.12 0.2 0.51 114.61 0.38 0.05 0.06
41 80 0.14 0.2 0.51 115.55 0.39 0.06 0.07
42 80 0.16 0.3 0.51 119 0.48 0.09 0.08
43 90 0.04 0.5 0.51 96 0.42 0.04 0.07
44 90 0.06 0.5 0.51 118.56 0.49 0.05 0.07
45 90 0.08 0.4 0.51 101.07 0.42 0.04 0.07
47 90 0.06 0.4 0.51 133 0.5 0.05 0.07
48 90 0.14 0.3 0.51 134.31 0.52 0.06 0.08
49 90 0.16 0.2 0.51 127 0.46 0.08 0.08

Table 2.: Dataset used in ML model development

19
 Figure 14.: The machine learning framework used in the modeling of machining
parameters.

(φi − φ̂i )2
P
2
R = 1 − Pi 2
(2)
i (φi − φ̄i )

N
1 X
M SE = (φi − φ̂i )2 (3)
N
i=1

N
1 X
M AE = |φi − φ̂i |. (4)
N
i=1

Here |.| represents the absolute value.

9.1. Polynomial regression (PR)

A functional relationship between the features and targets in the data may be de-
veloped using arbitrary approaches. Algorithms such as deep neural networks and
ensemble approaches lead to functional relationships that are very flexible. But these
functions are often too complicated for human analysis a posteriori. This limits the
interpretability of the ML model. For human interpretable models, we utilize polyno-
mial regression. Polynomial regression represents the target as a linear combination
of basis functions of the input features (Ostertagová, 2012). These basis functions
can be arbitrarily convoluted based on the complexity of the problem. Based on ex-
perimentation with the dataset, we use the second-order polynomial basis function.
The mathematical expression for the second-order polynomial regression model can

20
Figure 15.: The structure of a decision tree as used in tree based ML algorithms (RF,
AB and GB) J. Panda & Warrior (2022).

(a) (b)

Figure 16.: Hyperparameter optimization of random forest, for a) d=10 and for b)
n=5.

21
be written as follows:
k
X k
X XX
y = β0 + βi xi + βii x2i + βij xi xj +  (5)
i=1 i=1 i j

The β0 term represents the intercept and captures constant baseline effects that
are independent of the input features. The βi and βii terms capture the linear and
quadratic relationships between the input features and the target value. The βij term
represents the interactions amongst the input features. Thus the coefficients of this
polynomial closure have clear human interpretability. Training of this model invokes
gradient descent to find the appropriate coefficients for the dataset.

9.2. Random Forests (RF)

Random Forests are an ensemble learning algorithm where a meta-model is devel-
oped using a group (or ensemble) of decision trees (J. Panda & Warrior, 2022; Chung,
Mishra, Perakis, & Ihme, 2021). Decision Trees (also referred to as Classification And
Regression Trees) are a simpler modeling algorithm. Each decision tree represents a
set of inequalities that leads to a partitioning of the feature space. Different subsets
of this partitioning are assigned specific values of the target. A simple decision tree is
shown in figure 15. While decision trees are simple to use, they lead to overfitting. To
ameliorate this, an ensemble of decorrelated trees can be trained on the same data.
Here the decorrelation is produced by bootstrapping the dataset and feature bagging.
The final prediction of the Random Forest model is via aggregation of the predictions
of the individual decision trees (Chung et al., 2020). Due to their flexibility and ac-
curacy, Random Forests have been extensively used in physics and physical problems,
such as renewable energy generation (Serras et al., 2019), turbulence modeling (Heyse
et al., 2021), Meteorological prediction (Loken et al., 2019), combustion simulations
(Chung, Mishra, & Ihme, 2021), etc. The important hyperparameters of the random
forest are the number of decision trees (n) and maximum depth (d). Hyperparameter
optimization is an important task in any ML development. In this article, a manual
approach was followed, in optimizing the n and d. The graphs of R2 variation in terms
of n and d are presented in figure 16. For figure 16 a and b, d, n are taken as 10
and 5 respectively. From the analysis of hyperparameters, the best hyperparameters
(n,d) for modeling the Ra was found to be (5,10). The hyperparameter optimization is
shown in Fig. 16. For GB and AB, we also have considered the same hyperparameters.

9.3. Gradient Boosting (GB)

Similar to random forests the basic building block of gradient boosted trees are the
decision trees (J. Panda & Warrior, 2022). However while random forests create ran-
dom samples for training based on Bootstrap aggregation (or Bagging), other ensem-
ble learning approaches utilize Boosting. In Boosting-based approaches, several weak
learners are trained sequentially. The training of each subsequent tree is affected by
the errors of the prior trees, where a higher weight is placed on samples where prior
trees have been inaccurate. The boosting process is slower in comparison to the bag-
ging method since in boosting the trees are added sequentially. The trees in boosting
focus on errors from the previous one, which makes the algorithm self-correcting and
accurate in the asymptotic limit.

22
9.4. Adaptive Boosting (AB)
The Adaboost regressor is a meta-estimator that starts by fitting a regressor on the
actual dataset and then fits extra copies of the regressor on the same dataset but the
weights of the instances are adjusted according to the error of the current prediction
(Margineantu & Dietterich, 1997). The core principle of AdaBoost is to fit a series of
weak learners on a repeatedly modified version of the data. The predictions of all of
the weak learners are then combined through a weighted majority vote to produce a
final prediction.

10. Germinal center algorithm for optimization of the machining process

parameters

The algorithm for the Germinal center optimization (GCO) technique is discussed
here (Nguyen, 2020; Villaseñor et al., 2018). It is a unique multivariate continuous
optimization technique mainly inspired by the germinal center reaction. The germinal
centers are formed by the accumulation of (B cells) lymphocytes B and other unsus-
ceptible cells. These cells are bounded by some sluggish B cells and these sluggish B
cells form when any type of infection persists. The antibody was developed to get the
best harmony in germinal centers when B cells undergo a diversified and aggressive
or ruthless process. The germinal center reaction is shown in figure 17. In germinal
centers, there are two zones, i.e. dark and light. The procedures adopted for two zones
are repeatedly influenced by the GCO technique which is mentioned in Algorithm 1
as described in the article. A flow chart of the GCO is presented in the figure 17. 17
and discussed in following sub sections:

10.1. Process of dark zone

The enhancement of B cells in the dark zone is mainly due to clonal expansion. And
when new B cells formed through clonal expansion, some of them transformed through
a process called hyper-mutation. The multiplicity of the B cells in our work is confirmed
by the above two processes called clonal expansion and hyper-mutation. The solution
for the multivariate optimization problem through germinal search is represented by B
cells and Ab. Three providences a B shell can face in the dark zone like cell generation,
cell death, and cell mutation. Algorithms 2 and 3 are selected for this work.

10.2. Process of light zone

The B cells are engaged in phagocytizing Ag in the light zone. The B cell incorporates
the Ag and degrades in the proteins culture. When B cells are in the proteinous
elements to a cell of T, the life signal was inhabited in the B cell. In our experimental
analysis, the honor for the life signal L was the light zone. The amount of reward was
decided based upon the performance of the solution. The requisite algorithm 4 was
stated in the paper. In general, the population distribution of the B cell is done through
the competition in the germinal center optimization algorithm. The procedure adopted
in this optimization technique is quite important. Because it’s different from DE,
where uniform distribution is adopted. In uniform distribution, all the particles will
get the chance for mutation. In the PSO technique, the mutation process is influenced
by the leader’s behavior. In the GCO technique, the method was assumed by the

23
Figure 17.: The flow chart of germinal center algorithm reproduced from Villaseñor
et al. (2018).

population distribution. After that, the algorithm works like DE and PSO. When all
the solutions will be having the same performance, the GCO technique is imprecise
with DE. Similarly, when there will be one critical solution then the GCO technique
is imprecise with PSO. By providing such methodologies, the exploration-exploitation
can be balanced that changes its actions through the time span. The delayed leadership
quality is also there. And this phenomenon occurred, when one normal particle will
defeat the best particle. But its impact will be slowly reduced through no of iterations.

11. Results and Discussion

In this section, the predictive capability of four ML models (polynomial regression,

random forest, gradient boosting, and adaptive boosting) is discussed and compared
based on their prediction accuracy. We have used three performance parameters for
the development and comparison of the predictive capability of the ML models, those
are R2 , M SE and M AE as defined in section 9. In table 3 and 4 the R2 , M SE and
M AE values of different ML models are [Link] 3 and 4 correspond to the
train-test dataset D1 and D2 respectively. The R2 values of all prediction of all the
five parameters are found to be more than 0.95 for RF and PR for train-test dataset
D1 . However, for the train-test dataset D2 the R2 values are greater than 0.95 only for
PR. The R2 values of results predicted through the other two ML algorithms e.g. ADA
and GB is much lesser than 0.95, which signifies that those models are not suitable for
predicting the machining parameters. Similar to R2 , M SE and M AE of predictions
are small for PR. The actual vs predicted values of the predicted ML variables are
presented in figures 18 and 19.

24
11.1. ML based prediction of responses:
For modeling the responses of Ra, F , CWL , CWW , F W in terms of speed (s), feed (f),
and depth of cut (d), we have used the equations generated by polynomial regression
(PR) using the dataset D1, since PR can provide interpretable equations, those can
be used in the development of the complex objective function, that is required to be
used in the optimization of the process parameters. The equations learned through
PR, are written as follows:

Ra = 0.51 − 0.05s + 0.19f + 0.03d + 0.09f s + 0.0005sd−

0.005f d − 0.004s2 + 0.006f 2 + 0.016d2 ,
F = 106.28 − 0.39s + 33.69f + 25.38d + 21.95f s + 2.91sd−
4.22f d − 22s2 + 14.9f 2 + 21.38d2 ,
CL = 0.43 + 0.007s + 0.12f + 0.11d + 0.06f s + 0.008sd−
(6)
0.0003f d − 0.06s2 + 0.03f 2 + 0.02d2 ,
CW = 0.07 − 0.02s + 0.04f + 0.02d − 0.004f s − 0.002sd+
0.004f d − 0.002s2 + 0.002f 2 + 0.006d2 ,
F W = 0.06 + 0.004s + 0.008f + 0.004d + 0.006f s − 0.001sd−
0.002f d + 0.004s2 − 0.003f 2 − 0.006d2

R2 MSE MAE
PR 0.9916 0.0002 0.0119
RF 0.9826 0.0003 0.0184
Ra

ADA 0.9397 0.0013 0.0276

GB 0.5645 0.0099 0.0813
PR 0.9979 2.85 1.17
RF 0.9638 49.58 5.84
ADA 0.9011 135.69 11.08
F

GB 0.6676 455.98 17.95

PR 0.9949 0.0001 0.0077
RF 0.9886 0.0002 0.0123
CWL

ADA 0.9613 0.0007 0.0264

GB 0.6531 0.0068 0.0721
PR 0.9981 2.9E − 06 0.0014
RF 0.9769 0.00003 0.0053
CWW

ADA 0.9342 0.0001 0.0083

GB 0.6381 0.0005 0.0194
PR 0.9885 1.2E − 06 0.0008
RF 0.9865 1.4E − 6 0.0011
FW

ADA 0.9519 5.1E − 6 0.0021

GB 0.5274 4.9E − 6 0.0058

Table 3.: ML performance parameters for the train-test dataset D1

25
 (a) (b)

(c) (d)

(e)

Figure 18.: Actual vs predicted values of the machining parameters for train-test
dataset D1 . The predictions were made using ML models. square symbols PR, cross
symbols RF, plus symbols AB and circles correspond to GB.

26
 (a) (b)

(c) (d)

(e)

Figure 19.: Actual vs predicted values of the machining parameters for train-test
dataset D2 . The predictions were made using ML models. square symbols PR, cross
symbols RF, plus symbols AB and circles correspond to GB.

27
 R2 MSE MAE
PR 0.9974 000000 0.0055
RF 0.9002 0.0015 0.0314

Ra
ADA 0.8933 0.0016 0.0361
GB 0.5260 0.0074 0.0751
PR 0.9956 4.6432 1.3173
RF 0.9457 57.7313 6.3032
ADA 0.8521 157.4010 10.4949

F
GB 0.6094 415.6819 17.7077
PR 0.9947 9.94e − 05 0.0086
RF 0.9258 0.0014 0.0330
CWL ADA 0.9474 0.0010 0.0266
GB 0.6467 0.0067 0.0704
PR 0.9948 7.78e − 06 0.002
RF 0.9769 0.00003 0.0053
CWW

ADA 0.9342 0.0001 0.0083

GB 0.6381 0.0005 0.0213
PR 0.9950 5.27e − 07 0.0005
RF 0.9033 1.02e − 05 0.0023
FW

ADA 0.8737 1.34e − 05 0.0027

GB 0.5495 4.79e − 05 0.0057

Table 4.: ML performance parameters for the train-test dataset D2

11.2. Multi-objective optimization:

Although we have developed the predictive machine learning models using different
ensemble methods, those are characterized by complex non-linear functions, so it is
not possible to use traditional optimization methods in conjunction with the ML
methods. We have developed input-output correlation by using polynomial regression
and the correlation was used to establish the objective function to be used in the
multi-objective optimization. The objective the multi-objective optimization here is
to minimize the F , Ra, CWL , CWW and F W . The following objective function as
shown in equation 7 has been proposed for the above-mentioned purpose:

Ra F CWL CWW FW
COF (s, f, d) = w1 + w2 + w3 + w4 + w5 (7)
Ram Fm CWLm CWW m F Wm

where w1 , w2 , w3 , w4 and w5 weight values for the output parameters. Here, we have
taken all the weight values as same, i.e. w1 = w2 = w3 = w4 = w5 = 1/5. The
minimization of the above complex objective function was performed based on con-
straints of machining operation. Three operation constraints were used to develop the
complex objective function. The three constraints are the lower and upper limits of
the experimental parameters. The constraints are as follows: 40 ≤ s ≤ 90, 0.04 ≤ f ≤
0.16 and 0.2 ≤ d ≤ 0.5. There are different meta-heuristic methods such as genetic
algorithm, differential evolution, particle swarm optimization, and simulated anneal-
ing available by which the above-mentioned problem can be solved. In this work, to
solve the optimization problem a recently developed algorithm, named Germinal Cen-
ter Optimization was used. The detailed theory of the Germinal center algorithm is

28
available in (Villaseñor et al., 2018). The optimal combination of input parameters
(speed, feed, and depth of cut) for minimizing the outputs are found to be 60, 0.04,
0.2.

12. Concluding remarks

After analyzing the experimental results, following conclusions are obtained. a) A

columnar free dense structure coating was observed using scalable pulsed power plasma
process, b) With 40, 50, 55 and 60m/min speed levels, cutting force was found to be
diminished, whereas, when machining was conducted through 70, 80 and 90m/min sud-
den increment in cutting force was noticed. Mainly two wear mechanisms are observed
i.e., abrasion and adhesion. Feed influenced the tool-chip contact length significantly.
Both chip sticking and sliding was noticed on the tool rake face. Speed was observed
to be the critical parameter for flank wear. Both morphological aspect and surface
roughness was highly influenced by depth of cut and tool wear. Self-induced vibration
observed at higher depth of cut.
The experimental dataset was used for modeling the various machining parameters
such as surface roughness, cutting force, flank wear, crater wear length, and crater
wear width were modeled using machine learning algorithms (polynomial regression,
random forest, gradient boost, and ADA boost). The predictive capability of four ML
models was assessed in the modeling of the machining parameters. The performance
parameters assessed are R2 , M SE, and M AE of the predictions of the machining
parameters. It was noticed that for both random forest and polynomial regression the
R2 value is greater than 0.9, which signifies the fact that the learned correlation func-
tion maps the input and output parameters more accurately. However, ADA boost
and gradient boost predictions are poor in comparison to RF and PR. In contrast to
RF, PR predictions are comparatively better and the surrogate model learned through
PR can be easily extracted and implemented in the formulation of the complex objec-
tive function. We recommend the use of PR for predictive modeling of the machining
parameters.
Hence we have used the PR learned surrogate models of different machining param-
eters in the development of the complex objective function that is required to be used
in the germinal center-based optimization of the input parameters for the minimiza-
tion of the output parameters. The optimal machining process parameters are found
to be 60, 0.04, and 0.2 (speed, feed, and depth of cut respectively) that minimize the
output parameters.
In the future course of work, in place of speed, feed and depth of cut, tools of various
shapes can be used for ML-based modeling and optimization in a machining process.
The input features for various tools will be processed through shape parametrization
using convolutional neural networks.

References

Akhil, V., Raghav, G., Arunachalam, N., & Srinivas, D. (2020). Image data-based
surface texture characterization and prediction using machine learning approaches
for additive manufacturing. Journal of Computing and Information Science in En-
gineering, 20 (2), 021010.

29
Bricher, D., & Müller, A. (2020). A supervised machine learning approach for intelli-
gent process automation in container logistics. Journal of Computing and Informa-
tion Science in Engineering, 20 (3).
Bustillo, A., Pimenov, D. Y., Mia, M., & Kaplonek, W. (2021). Machine-learning
for automatic prediction of flatness deviation considering the wear of the face mill
teeth. Journal of Intelligent Manufacturing, 32 (3), 895–912.
Caggiano, A., Rimpault, X., Teti, R., Balazinski, M., Chatelain, J.-F., & Nele, L.
(2018). Machine learning approach based on fractal analysis for optimal tool life
exploitation in cfrp composite drilling for aeronautical assembly. CIRP Annals,
67 (1), 483–486.
Carrino, S., Guerne, J., Dreyer, J., Ghorbel, H., Schorderet, A., & Montavon, R.
(2020). Machining quality prediction using acoustic sensors and machine learning.
In Multidisciplinary digital publishing institute proceedings (Vol. 63, p. 31).
Cheng, M., Jiao, L., Shi, X., Wang, X., Yan, P., & Li, Y. (2020). An intelligent
prediction model of the tool wear based on machine learning in turning high strength
steel. Proceedings of the Institution of Mechanical Engineers, Part B: Journal of
Engineering Manufacture, 234 (13), 1580–1597.
Chiu, H.-W., & Lee, C.-H. (2017). Prediction of machining accuracy and surface
quality for cnc machine tools using data driven approach. Advances in Engineering
Software, 114 , 246–257.
Cho, S., Asfour, S., Onar, A., & Kaundinya, N. (2005). Tool breakage detection
using support vector machine learning in a milling process. International Journal
of Machine Tools and Manufacture, 45 (3), 241–249.
Chung, W. T., Mishra, A. A., & Ihme, M. (2021). Interpretable data-driven methods
for subgrid-scale closure in les for transcritical lox/gch4 combustion. arXiv preprint
arXiv:2103.06397 .
Chung, W. T., Mishra, A. A., Perakis, N., & Ihme, M. (2020). Random forests for
accelerating turbulent combustion simulations. In Proceedings of the third workshop
on machine learning and the physical sciences (neurips 2020), vancouver, bc, canada
(Vol. 11).
Chung, W. T., Mishra, A. A., Perakis, N., & Ihme, M. (2021). Data-assisted combus-
tion simulations with dynamic submodel assignment using random forests. Com-
bustion and Flame, 227 , 172–185.
Cica, D., Sredanovic, B., Tesic, S., & Kramar, D. (2020). Predictive modeling of turn-
ing operations under different cooling/lubricating conditions for sustainable manu-
facturing with machine learning techniques. Applied Computing and Informatics.
Elsheikh, A. H., Muthuramalingam, T., Shanmugan, S., Ibrahim, A. M. M., Ramesh,
B., Khoshaim, A. B., . . . Sathyamurthy, R. (2021). Fine-tuned artificial intelligence
model using pigeon optimizer for prediction of residual stresses during turning of
inconel 718. Journal of Materials Research and Technology, 15 , 3622–3634.
Fang, N., Pai, P. S., & Edwards, N. (2016). Neural network modeling and predic-
tion of surface roughness in machining aluminum alloys. Journal of Computer and
Communications, 4 (5), 1–9.
Gouarir, A., Martı́nez-Arellano, G., Terrazas, G., Benardos, P., & Ratchev, S. (2018).
In-process tool wear prediction system based on machine learning techniques and
force analysis. Procedia CIRP , 77 , 501–504.
Guo, Y., Yan, P., Wu, D., Zhou, H., Shi, Y., & Yi, R. (2021). Analysis method
for factors influencing gear hobbing quality based on density peak clustering and
improved multi-objective differential evolution algorithm. International Journal of
Computer Integrated Manufacturing, 34 (4), 385–406.

30
Heyse, J. F., Mishra, A. A., & Iaccarino, G. (2021). Estimating rans model uncertainty
using machine learning. Journal of the Global Power and Propulsion Society. Special
Issue: Data-Driven Modelling and High-Fidelity Simulations, 1–14.
Jumare, A. I., Abou-El-Hossein, K., Abdulkadir, L. N., & Liman, M. M. (2019).
Predictive modeling and multiobjective optimization of diamond turning process of
single-crystal silicon using rsm and desirability function approach. The International
Journal of Advanced Manufacturing Technology, 103 (9), 4205–4220.
Jurkovic, Z., Cukor, G., Brezocnik, M., & Brajkovic, T. (2018). A comparison of
machine learning methods for cutting parameters prediction in high speed turning
process. Journal of Intelligent Manufacturing, 29 (8), 1683–1693.
Kahya, M., Ozbayoglu, M., & Unver, H. O. (2021). Precision and energy-efficient
ball-end milling of ti6al4v turbine blades using particle swarm optimization. Inter-
national Journal of Computer Integrated Manufacturing, 34 (2), 110–133.
Khoshaim, A. B., Moustafa, E. B., Bafakeeh, O. T., & Elsheikh, A. H. (2021). An
optimized multilayer perceptrons model using grey wolf optimizer to predict me-
chanical and microstructural properties of friction stir processed aluminum alloy
reinforced by nanoparticles. Coatings, 11 (12), 1476.
Kothuru, A., Nooka, S. P., & Liu, R. (2018). Application of audible sound signals
for tool wear monitoring using machine learning techniques in end milling. The
International Journal of Advanced Manufacturing Technology, 95 (9), 3797–3808.
Krishnakumar, P., Rameshkumar, K., & Ramachandran, K. (2018). Machine learning
based tool condition classification using acoustic emission and vibration data in
high speed milling process using wavelet features. Intelligent Decision Technologies,
12 (2), 265–282.
Liu, R., Kothuru, A., & Zhang, S. (2020). Calibration-based tool condition monitoring
for repetitive machining operations. Journal of Manufacturing Systems, 54 , 285–
293.
Loken, E. D., Clark, A. J., McGovern, A., Flora, M., & Knopfmeier, K. (2019).
Postprocessing next-day ensemble probabilistic precipitation forecasts using random
forests. Weather and Forecasting, 34 (6), 2017–2044.
Madhusudana, C., Budati, S., Gangadhar, N., Kumar, H., & Narendranath, S. (2016).
Fault diagnosis studies of face milling cutter using machine learning approach. Jour-
nal of Low Frequency Noise, Vibration and Active Control , 35 (2), 128–138.
Margineantu, D. D., & Dietterich, T. G. (1997). Pruning adaptive boosting. In Icml
(Vol. 97, pp. 211–218).
McLeay, T., Turner, M. S., & Worden, K. (2020). A novel approach to machin-
ing process fault detection using unsupervised learning. Proceedings of the In-
stitution of Mechanical Engineers, Part B: Journal of Engineering Manufacture,
0954405420937556.
Mishra, A. A., Mukhopadhaya, J., Alonso, J., & Iaccarino, G. (2020). Design explo-
ration and optimization under uncertainty. Physics of Fluids, 32 (8), 085106.
Mohanraj, T., Yerchuru, J., Krishnan, H., Aravind, R. N., & Yameni, R. (2021). De-
velopment of tool condition monitoring system in end milling process using wavelet
features and hoelder’s exponent with machine learning algorithms. Measurement,
173 , 108671.
Nagargoje, A., Kankar, P. K., Jain, P. K., & Tandon, P. (2021). Performance eval-
uation of the data clustering techniques and cluster validity indices for efficient
toolpath development for incremental sheet forming. Journal of Computing and
Information Science in Engineering, 21 (3), 031001.
Nain, S. S., Garg, D., & Kumar, S. (2018). Evaluation and analysis of cutting speed,

31
 wire wear ratio, and dimensional deviation of wire electric discharge machining of
super alloy udimet-l605 using support vector machine and grey relational analysis.
Advances in Manufacturing, 6 (2), 225–246.
Nguyen, T. (2020). A collection of the state-of-the-art meta-heuristics algorithms
in python: Mealpy. Zenodo. Retrieved from [Link]
.3711948 doi:
Oleaga, I., Pardo, C., Zulaika, J. J., & Bustillo, A. (2018). A machine-learning based
solution for chatter prediction in heavy-duty milling machines. Measurement, 128 ,
34–44.
Ostertagová, E. (2012). Modelling using polynomial regression. Procedia Engineering,
48 , 500–506.
O’Donnell, J., & Yoon, H.-S. (2020). Determination of time-to-failure for automotive
system components using machine learning. Journal of Computing and Information
Science in Engineering, 20 (6), 061003.
Panda, J., & Warrior, H. (2022). Evaluation of machine learning algorithms for
predictive reynolds stress transport modeling. Acta Mechanica Sinica, -.
Panda, J. P., & Warrior, H. V. (2021). Modeling the pressure strain correla-
tion in turbulent flows using deep neural networks. Proceedings of the Institu-
tion of Mechanical Engineers, Part C: Journal of Mechanical Engineering Science,
09544062211042920.
Parwal, V., & Rout, B. (2021). Machine learning based approach for process supervi-
sion to predict tool wear during machining. Procedia CIRP , 98 , 133–138.
Patange, A. D., & Jegadeeshwaran, R. (2021). A machine learning approach for
vibration-based multipoint tool insert health prediction on vertical machining centre
(vmc). Measurement, 173 , 108649.
Peng, B., Bergs, T., Schraknepper, D., Klocke, F., & Döbbeler, B. (2019). A hybrid
approach using machine learning to predict the cutting forces under consideration
of the tool wear. Procedia Cirp, 82 , 302–307.
Rašovi&39;c, N. (2021). Recommended layer thickness to the powder-based addi-
tive manufacturing using multi-attribute decision support. International Journal of
Computer Integrated Manufacturing, 34 (5), 455–469.
Ruiz, E., Ferreño, D., Cuartas, M., L&39;opez, A., Arroyo, V., & Guti&39;errez-
Solana, F. (2020). Machine learning algorithms for the prediction of the strength of
steel rods: an example of data-driven manufacturing in steelmaking. International
Journal of Computer Integrated Manufacturing, 33 (9), 880–894.
Saidi, R., Fathallah, B. B., Mabrouki, T., Belhadi, S., & Yallese, M. A. (2019). Mod-
eling and optimization of the turning parameters of cobalt alloy (stellite 6) based on
rsm and desirability function. The International Journal of Advanced Manufacturing
Technology, 100 (9), 2945–2968.
Sanchez, J. A., Conde, A., Arriandiaga, A., Wang, J., & Plaza, S. (2018). Unex-
pected event prediction in wire electrical discharge machining using deep learning
techniques. Materials, 11 (7), 1100.
Serras, P., Ibarra-Berastegi, G., Sáenz, J., & Ulazia, A. (2019). Combining ran-
dom forests and physics-based models to forecast the electricity generated by ocean
waves: A case study of the mutriku wave farm. Ocean Engineering, 189 , 106314.
Shastri, A., Nargundkar, A., Kulkarni, A. J., & Benedicenti, L. (2021). Optimization
of process parameters for turning of titanium alloy (grade ii) in mql environment
using multi-ci algorithm. SN Applied Sciences, 3 (2), 1–12.
Shen, Y., Yang, F., Habibullah, M. S., Ahmed, J., Das, A. K., Zhou, Y., & Ho, C. L.
(2020). Predicting tool wear size across multi-cutting conditions using advanced

32
 machine learning techniques. Journal of Intelligent Manufacturing, 1–14.
Shukla, S. K., & Priyadarshini, A. (2019). Application of machine learning techniques
for multi objective optimization of response variables in wire cut electro discharge
machining operation. In Materials science forum (Vol. 969, pp. 800–806).
Surleraux, A., Lepert, R., Pernot, J.-P., Kerfriden, P., & Bigot, S. (2020). Machine
learning-based reverse modeling approach for rapid tool shape optimization in die-
sinking micro electro discharge machining. Journal of Computing and Information
Science in Engineering, 20 (3).
Tian, C., Zhou, G., Lu, F., Chen, Z., & Zou, L. (2020). An integrated multi-objective
optimization approach to determine the optimal feature processing sequence and
cutting parameters for carbon emissions savings of cnc machining. International
Journal of Computer Integrated Manufacturing, 33 (6), 609–625.
Ulas, M., Aydur, O., Gurgenc, T., & Ozel, C. (2020). Surface roughness prediction
of machined aluminum alloy with wire electrical discharge machining by different
machine learning algorithms. Journal of Materials Research and Technology, 9 (6),
12512–12524.
Villaseñor, C., Rios, J. D., Arana-Daniel, N., Alanis, A. Y., Lopez-Franco, C., &
Hernandez-Vargas, E. A. (2018). Germinal center optimization applied to neu-
ral inverse optimal control for an all-terrain tracked robot. Applied Sciences, 8 (1),
31.
Wang, J., Li, Y., Zhao, R., & Gao, R. X. (2020). Physics guided neural network for
machining tool wear prediction. Journal of Manufacturing Systems, 57 , 298–310.
Wang, J., Sánchez, J., Iturrioz, J., & Ayesta, I. (2019). Artificial intelligence for
advanced non-conventional machining processes. Procedia Manufacturing, 41 , 453–
459.
Wang, P., Liu, Z., Gao, R. X., & Guo, Y. (2019). Heterogeneous data-driven hybrid
machine learning for tool condition prognosis. CIRP Annals, 68 (1), 455–458.
Wu, D., Jennings, C., Terpenny, J., Gao, R. X., & Kumara, S. (2017). A comparative
study on machine learning algorithms for smart manufacturing: tool wear prediction
using random forests. Journal of Manufacturing Science and Engineering, 139 (7).
Zhang, Q., Wang, Z., Wang, B., Ohsawa, Y., & Hayashi, T. (2020). Feature extraction
of laser machining data by using deep multi-task learning. Information, 11 (8), 378.
Zhang, Z., Shi, J., Yu, T., Santomauro, A., Gordon, A., Gou, J., & Wu, D. (2020).
Predicting flexural strength of additively manufactured continuous carbon fiber-
reinforced polymer composites using machine learning. Journal of Computing and
Information Science in Engineering, 20 (6), 061015.
Zhao, G., Li, C., Lv, Z., Cheng, X., & Zheng, G. (2020). Specific energy consumption
prediction model of cnc machine tools based on tool wear. International Journal of
Computer Integrated Manufacturing, 33 (2), 159–168.

33