You are on page 1of 32

A Novel MOALO-MODA Ensemble Approach for

Multi-objective Optimization of Machining


Parameters for Metal Matrix Composites
Kanak Kalita
Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology
Vikas Kumar
Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology
Shankar Chakraborty (  s_chakraborty00@yahoo.co.in )
Jadavpur University

Research Article

Keywords: MMC, Optimization, Process parameter, MOALO, MODA, Ensemble

Posted Date: August 2nd, 2022

DOI: https://doi.org/10.21203/rs.3.rs-1896308/v1

License:   This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License
A Novel MOALO-MODA Ensemble Approach for Multi-objective Optimization of Machining
Parameters for Metal Matrix Composites
Kanak Kalita1, Vikas Kumar2, Shankar Chakraborty3*
1
Department of Mechanical Engineering, Vel Tech Rangarajan Dr. Sagunthala R&D Institute of
Science and Technology, Avadi, India
2
Department of Automobile Engineering, Vel Tech Rangarajan Dr. Sagunthala R&D Institute of
Science and Technology, Avadi, India
3
Department of Production Engineering, Jadavpur University, Kolkata, India
Communicating author’s email: s_chakraborty00@yahoo.co.in
*

Abstract
Machining of metal matrix composites (MMCs) has now become an essential task in shaping them to
their final usable forms for various industrial applications. These machining operations may range
from drilling of simple through holes to cutting and shaping the composites into complex shapes.
Determination of the optimal process parameters during their machining presents a combinatorial
optimization problem, which may turn out to be more complex due to involvement of various material
dependent parameters of the MMCs, like particle size, reinforcement percentage etc. The importance
of optimization increases manifold due to conflicting settings of different process parameters for
attaining the desired quality characteristics. In this paper, two nature-inspired optimization algorithms,
i.e. multi-objective antlion optimization (MOALO) and multi-objective dragonfly algorithm (MODA)
are applied for multi-objective optimization of various process parameters during machining of
MMCs. To mitigate uncertainties in global optimality of the predicted Pareto optimal fronts arising
due to stochastic nature of the metaheuristics, an ensemble approach integrating MOALO and MODA
techniques is proposed here. It is demonstrated with the help of two case studies that MOALO-
MODA ensemble is far superior to its individual counterparts and thus, can be employed for
development of robust and reliable Pareto optimal fronts.
Keywords: MMC, Optimization, Process parameter, MOALO, MODA, Ensemble
1. Introduction
Composites are usually made of two or more materials having different physical and chemical
properties. In these materials, combination of reinforcement to matrix provides superior properties to
those of the individual materials. Due to their several advantageous features, like light weight, higher
strength and stiffness, excellent fatigue resistance, higher thermal conductivity, increased product life
span etc., they have now become an integral part of every industry, like automobile, aerospace, civil
infrastructure, marine, electrical appliances etc. To overcome the shortage of raw materials and fulfil
the day-to-day diversified engineering requirements, they have turned out to be the promising
materials of recent interest. Metal matrix composites (MMCs) are composed of metallic or non-
metallic reinforcements embedded in a metal matrix. The matrix is usually a lighter metal (such as
aluminium, titanium or magnesium) which acts as a support for the reinforcement (like carbon fibers,

1
boron carbide, silicon carbide or ceramics). The reinforcement endows the composite with higher
strength and better wear resistance properties. However, it also adds machinability problems, like
higher tool wear and surface roughness. The machining of MMCs is thus complicated than that of the
conventional metals because of the complex properties of the reinforcement materials. It has been
noticed that when ceramics are used as reinforcement, they provide powder-like discontinuous chips
during the machining operation causing third body abrasion between the tool and the material,
resulting in poor surface finish and higher tool wear [1].
Any machining operation of MMCs is largely influenced by various process parameters, which
may be material specific (like weight % of reinforcement, particle size etc.) or machining operation
specific (like spindle speed, feed rate, depth of cut etc.). For example, in drilling which is the most
common operation for industrial applications of MMCs, there are many parameters, like drill speed,
feed rate, temperature, tool type, particle fraction etc. affecting the drilled surface quality and tool
wear rate. Thus, the most influential parameters affecting efficiency of the machining operation and
quality of the final products must be carefully determined. Basavarajappa et al. [2] identified that
during drilling of MMCs, the major factor influencing drilling force had been feed rate rather than
cutting speed. It is also reported in the literature that surface roughness of the drilled holes on MMCs
would increase with increase in feed rate and decrease with increase in cutting speed [3, 4]. But, many
other researchers are against these observations [5, 6]. However, most of the researchers have agreed
that feed rate would be the dominant parameter over cutting speed with respect to achieved surface
roughness. Cutting tool material also plays an important role in influencing the drilled surface quality.
For example, carbide coated tools would provide better drilled surface quality than uncoated cutting
tools [2] and HSS would not show any promising result on drilled surface quality of MMCs. The HSS
tools would often undergo excessive tool wear, resulting in rough surface during MMC drilling [7].
Presence of reinforcement has a great influence on the machining characteristics as it affects
the machinability property of composites. Reinforcement volume fraction hinders chip formation due
to higher plastic shear. During machining of MMCs, the hard ceramic particles when come in contact
with the tool, start forming built-up edge (BUE). These small particles form a sandwich layer between
the tool and the material, and the BUE acts as cutting tool causing generation of rough surface with
higher tool wear. Machinability of MMCs can be effectively controlled while varying the
reinforcement shape, size, properties of reinforcement, manufacturing method etc. [8]. Determination
of the optimal values of different machining parameters, like feed rate, depth of cut and speed helps in
attaining better machinability without changing the properties of the considered materials. As surface
quality of the machined composites plays a pivotal role in product performance, the concerned
manufacturers should strictly attempt to adhere to its target value. Typically, surface roughness is
considered as one of the main factors influencing surface integrity of the machined components [9]
According to Manna and Bhattacharayya [10], high speed, low depth of cut and low feed rate would

2
provide better surface finish during machining of Al/SiC composite. Similar optimal process
parameter combinations have also been reported by the other past researchers.
A significant number of studies has relied on the applications of various multi-criteria
decision making (MCDM) tools for determination of the optimal intermixes of different process
parameters during machining of MMCs. Sidhu and Yazdani [11] employed desirability function and
lexicographic goal programming to identify the optimal machining condition for having minimum
residual stress, tool wear rate (TWR) and maximum material removal rate (MMR) during electrical
discharge machining (EDM) of SiC/A359 MMC. Tamiloli et al. [12] integrated fuzzy logic with
technique for order of preference by similarity to ideal solution (TOPSIS) for optimization of the end
milling parameters while machining Al-SiC-fly ash MMC. It was observed that feed rate, speed and
depth of cut would significantly affect surface roughness and cutting force during the machining
operation. Patel and Pradhan [13] investigated the effects of various EDM process parameters on
MRR, TWR, surface roughness and radial overcut during machining of Al7075/SiC/WS2 composite.
While machining AlSiC MMC, Satpathy et al. [14] proposed the application of TOPSIS to find out
the optimal levels of various EDM process parameters. Kumar et al. [15] adopted additive ratio
assessment (ARAS) method for optimality evaluation of EDM process parameters during machining
of AA7050 composite.
Despite their significant advantages, like extremely low computational cost and ease of
implementation, MCDM techniques have a major drawback that they rarely can find out the global
optimal solution. They work best as a suitable alternative selection approach. On the other hand,
several researchers have solved the machining process optimization problems treating them as single
objective optimization (SOO) problems where only one response can be optimized. Kumar et al. [16]
applied genetic algorithm (GA) for optimizing the wear behaviour of Al2618 alloy reinforced with
Si3N4, AlN and ZrB2 materials. Different tribological properties, like wear rate, wear resistance,
specific wear rate and coefficient of friction were considered to examine the wear behaviour of
composite materials. Radhika et al. [17] examined the wear behaviour of aluminium hybrid MMC and
optimized the process parameters using GA. It was noticed that minimum wear rate for alumina
composite could be achieved at 5 wt% reinforcement. Rao et al. [18] applied GA and Taguchi
methodology to optimize the drilling parameters for AlSiC MMC. Bongale et al. [19] attempted to
minimize wear on Al/SiCnp/E-glass fiber MMC while adopting Taguchi methodology and GA
technique. Adithya et al. [20] applied grey wolf optimizer and K-nearest neighbour for optimizing the
process parameters of stir squeeze casting of AA2219 reinforced MMC.
Although SOO is quite relevant for some applications where the main focus is on a particular
response, in majority of the industrial applications, simultaneous achievement of multiple responses is
often required. Thus, multi-objective optimization (MOO) needs special emphasis. However, many
researchers have adopted formulations, like weighted sum approach [21, 22] to transform an MOO
problem into a SOO problem. This is an efficient way to reduce the complexity and computational

3
intensiveness of an optimization problem. However, among other challenges, specifying suitable
weights to the responses a priori during the optimization phase is a major hurdle. The MOO generally
predicts a set of optimal solutions known as non-dominated solutions which together form the Pareto
optimal front. Such MOO involves two or more solutions that need to simultaneously maximized or
minimized. Selvakumar et al. [23] applied non-sorting genetic algorithm-II (NSGA-II) to optimize the
corresponding machining parameters that would minimize flank wear loss and maximize MRR of Al-
4%Cu-7.5%SiC MMC. Singh et al. [24] applied NSGA-II to optimize MRR and TWR during
machining of MMCs. Kumar et al. [25] also employed NSGA-II to maximize MRR and minimize
surface roughness while machining Al6061 silver-coated copper MMCs. Daniel at al. [26] maximized
MRR and minimized tool temperature during drilling operations on MMCs. Ikhlas et al. [27]
compared the predictive performance of artificial neural network and response surface methodology
for metamodeling of the machining operations on MMCs. The NSGA-II technique was also deployed
to optimize surface roughness and cutting speed during machining of AISI 4140 hardened steel with
mixed ceramic MMC. It was proved that NSGA-II would be more efficient than desirability function
approach.
It can be revealed from the literature review of MOO applications in machining of MMCs that
NSGA-II is the most commonly used algorithm. In this paper, two recently developed nature-inspired
optimization algorithms, i.e. multi-objective antlion optimization (MOALO) and multi-objective
dragonfly algorithm (MODA) are employed for multi-objective optimization of process parameters
during machining of two different MMCs. The MOALO and MODA are then robustified while using
an ensemble approach. A combined MOALO-MODA ensemble is finally proposed for the Pareto
optimal front generation. The rest of the paper is organized as follows: Section 2 outlines the
proposed ensemble approach of Pareto optimization. Discussions regarding MOALO, MODA and
non-dominated sorting (NDS) are presented in Sections 2.1, 2.2 and 2.3 respectively. Two case
studies on machining of MMCs from the existent literature are solved using the proposed ensemble
approach and presented in Section 3. Section 3.1 demonstrates the effectiveness of this approach in
predicting 2D Pareto optimal fronts for MOO during turning of MMCs. Similarly, in Section 3.2, 3D
Pareto optimal fronts for MOO during drilling of MMCs are presented. Conclusions based on this
comprehensive study are drawn in Section 4.
2. Methodology
In this paper, MOALO and MODA are applied for MOO of machining process parameters for
MMCs. Taking inspiration from ensemble machine learning (ML) models, where a series of weak
learners are considered to build a strong ML system, an ensemble approach to Pareto optimal front
generation is proposed here. It is well known that to account for stochastic nature of metaheuristic
algorithms, multiple independent trials (n) of the algorithm must be carried out. Generally, the best
Pareto optimal front with respect to spread and generational distance among n trials is reported.
However, there may be other Pareto optimal fronts from those n trials that may have better non-

4
dominated solutions at certain sections of the Pareto optimal front. To exploit this, an ensemble
approach can be a suitable solution. A NDS algorithm based on NSGA-II is applied for secondary
sorting in MOALO and MODA ensembles. Similarly, a tertiary NDS is also carried out in MOALO-
MODA ensemble. It should be noted here that the primary NDS is carried out within each
independent trial of MOALO and MODA.
The flowchart of the proposed methodology is illustrated in Figure 1. The optimization process
starts with specifying the objective function, design variables and their respective bounds. Then n
independent optimization trials are carried out by both MOALO and MODA individually. The
corresponding Pareto optimal fronts are predicted by MOALO and MODA in each of the n trials. The
n Pareto optimal fronts predicted by MOALO are collated together and a NDS is carried out. The
resulting Pareto optimal front is hereafter referred to as MOALO ensemble. Similarly, the Pareto
optimal front obtained after NDS of n Pareto optimal fronts generated by MODA is referred to as
MODA ensemble. The MOALO ensemble and MODA ensemble may be categorized as homogenous
ensembles as the individual n Pareto optimal fronts that contribute to the final output are obtained
using the same type of algorithm. Next, the ensemble Pareto optimal fronts of MOALO and MODA
are collated together and a tertiary NDS is carried out to obtain MOALO-MODA ensemble Pareto
optimal front. This ensemble is heterogeneous as it contains non-dominated solutions predicted by
two different algorithms.
2.1 Multi-objective antlion optimization
Antlion optimizer (ALO) is a popular nature-inspired population-based metaheuristic algorithm
proposed by Mirjalili [28] in 2015. The MOALO, the multi-objective version of ALO, was developed
in 2017 by Mirjalili et al. [29]. It is based on the hunting tactics of an insect called the antlion.
Antlions make traps for catching their prey by digging a cone-shaped pit in sand with their jaws. After
making the pit, the antlion hides itself at the bottom of the cone and waits for the prey (primarily ants)
to fall into the trap. Once the prey enters the pit, the antlion tries to consume it. However, the prey
tries to evade the antlion and escape the pit by crawling up. At the same time, the antlion starts tossing
sand from the center of the hole towards the edges so that the prey may slide back. Once the prey gets
slided to the center of the pit, it is unable to come out and antlion consumes it. To simulate the
random walk of ants (prey), the following equation is used.
X(t) = [0, cs(2r(t1) – 1),...,cs(2r(tm) – 1) (1)
where cs calculates the cumulative sum, t shows the step of random walk and m is the maximum
number of iterations. To keep the random walk within the boundaries of the search space and letting
the ant from overshooting the bounds, the following equation is considered for normalization.

( X it  ai )  (dit  cit )
X it   cit (2)
(bi  ai )

5
where ai and bi are the minimum and maximum of random walks of ith variable, and cit and d it are the
minimum and maximum of random walks of ith variable at tth iteration. The ALO simulates the
entrapment of ants in antlion’s pits by changing the random walks around antlion pits.

cit  Atj  c t (3)

d it  Atj  d t (4)

where ct and dt are the minimum and maximum of all variables at tth iteration, cit and d it are the

minimum and maximum of all variables for ith ant, and Atj is the position of jth antlion at tth iteration.

For simulating the sliding of prey towards the antlion, the boundaries of the random walks are
adaptively decreased.
ct
ct  (5)
I

dt
dt  (6)
I
t
I 1 10w (7)
T
where t and T are the current and maximum number of iterations, and w is the parameter responsible
for adjusting the accuracy of the exploitation. Based on the recommendations of Mirjalili et al. [29], in
this paper, w = 2, 3, 4, 5 and 6, for (t/T) > 0.1, 0.5, 0.75, 0.9 and 0.95 are respectively considered.
Once the antlion catches the prey and consumes it, it reconstructs the pit.
Atj  antit if f (antit )  f ( Atj ) (8)

where antit is the position of ith ant at tth iteration.


Finally, elitism is introduced in ALO, wherein the fittest antlion so far is recorded. The elite antlion
influences the random walks of all the ants. Any ant performing random walk around an antlion which
is selected by roulette wheel mechanism and the elite antlion can be mathematically modelled as
below:
( RAt  REt )
Antit  (9)
2
where RAt and REt are the random walks at tth iteration around the roulette wheel selected and elite
antlions respectively. An archive is used in MOALO to store the Pareto optimal solutions. To improve
the Pareto optimal solution distribution and diversity in the archive, niching is adopted. The Pareto
optimal solutions from the least populated neighbourhoods are chosen with the following probability:
c
Pi  (10)
Ni
where c is a constant (greater than 1) and Ni is the number of solutions near ith solution.

6
To accommodate new Pareto optimal solutions in the archive, in case if it is full, some solutions from
the most populated regions are removed with the following probability:
Ni
Pi  (11)
c
Further, for multi-objective optimization, Eq. (8) is modified as follows:
Atj  antit if f (antit )  f ( Atj ) (12)

Similarly, Eq. (9) is also modified for selection of antlions and elite. A roulette wheel mechanism and
Eq. (9) are finally considered for selecting a Pareto optimal solution from the archive.
2.2 Multi-objective dragonfly algorithm
The MODA was introduced by Mirjalili [30] in early 2015. This is also a nature-inspired
population-based metaheuristic algorithm based on the behavioural patterns of dragonflies, mainly the
hunting technique for food and swarming behaviour to save themselves from the enemy. Thus, it
draws inspiration from the static and dynamic behaviours of the dragonflies. In static behaviour, they
form small swarms in which they fly over a small area for hunting of other flying preys, like
mosquitoes and butterflies. In dynamic behaviour, they form a very large swarm in which they
migrate long distances in a particular direction. These static and dynamic swarming behaviours of the
dragonflies are akin to exploration and exploitation respectively. Movement of the dragonflies is
governed by five main principles, i.e. separation, alignment, cohesion, attraction and distraction.
Separation is the mechanism by which dragonflies avoid collision with the neighbouring dragonflies.
It is mathematically modelled as:


N
Si   j 1
(Xi  X j ) (13)

where Si is the separation of ith dragonfly, Xi and Xj are the current dragonfly and jth neighbouring
dragonfly respectively, and N is the number of dragonflies in the neighbourhood. Alignment is the
mechanism by which dragonflies match their velocity to the neighbouring individuals, which can be
modelled as below:


N
j 1
Vj
A
i (14)
N
where Ai is the alignment of ith dragonfly and Vj is the velocity of jth neighbouring dragonfly. Cohesion
is the mechanism by which dragonflies are attracted towards the center of mass of the neighbourhood.
It is mathematically modelled as follows:


N
j 1
Xj
C 
i  Xi (15)
N
where Ci is the cohesion of ith dragonfly. Attraction is the mechanism by which the dragonflies are
attracted towards a food source. It can be mathematically expressed as below:
Fi = X+ – Xi (16)

7
where Fi is the food source of ith dragonfly and X+ is the location of the food source. Distraction is the
mechanism by which the dragonflies move away from an enemy. It is mathematically modelled as:
Ei = X– – Xi (17)
where Ei is the enemy of ith dragonfly and X– is the location of the enemy.
Based on these five behavioural patterns of the artificial dragonfly, step vector (∆X) and position
vector (X), the following relation is modelled to update its position.
X it 1  sSi  aAi  cCi  fFi  eEi   wX it (18)
where s a, c, f, e and w are the separation weight, alignment weight, cohesion weight, food factor,
enemy factor and inertia weight respectively, and t is the iteration counter. Next, the position vectors
are updated using the following relation:
X it 1  X it  X it 1 (19)
In absence of any neighbourhood solution, dragonflies perform a Levy flight [31] to update their
positions.
X it 1  X it  Levy(d )  X it (20)
where d is the dimension of the position vector. The Levy flights are computed as shown below:
r1  
Levy( x)  0.01 1
(21)
r2 

where r1 and r2 are random numbers between [0,1], β is a constant and σ is defined as follows:
1
 
 (1   )  sin   
  
   2  
  1 
 (22)
  1     
      2  
2

  2  
where Γ(x) = (x – 1)!
An archive is used in MODA to store the Pareto optimal solutions. The position updating of
search agents is identical to that of single objective DA, but the food sources are selected from the
archive. To find out the least populated area of the Pareto optimal front, the search space is
segmented. To achieve this, the best and the worst objective functions of the Pareto optimal solutions
found so far are identified. A hyper-sphere is defined to cover all these solutions and the hyper-
spheres are then divided into equal sub-hyper-spheres in each iteration. The Pareto optimal solutions
from the least populated neighbourhoods are chosen with the probability defined in Eq. (10). For
selecting enemies from the archive, the worst (most populated) hyper-sphere is identified so that the
artificial dragonflies are discouraged from searching around the non-promising crowded areas. Thus,
some worst solutions from the most populated regions are removed with the probability defined in Eq.
(11).
2.3 Non-dominated sorting

8
The NDS carried out in this paper during the secondary and tertiary sorting is the same as in
NSGA-II [32]. For secondary sorting, the Pareto optimal fronts obtained in n independent trials of
MOALO (or MODA) are collated together and the number of solutions dominating each solution in
the collated set is found out. Any solution B is said to be dominated by solution A if any the following
three conditions is true [33-35]:
a) Both A and B are feasible, and A is better than B in at least one objective and equal to B in
other objectives. But solutions A and B are non-dominated to each other if A is better than B
in at least one objective and B is better than A in at least one other objective.
b) Solution A is feasible but solution B is infeasible.
c) Both solutions A and B are infeasible, but A has a smaller number of violated constraints
compared to B. If both solutions have the same number of violated constraints, then A has
lesser total absolute constraint violation than B.
The number of solutions (nd) dominating each solution in the collated set is calculated. Any solution
with nd = 0 is assigned rank one. Here, nd = 0 means that they are non-dominated solutions and are not
dominated by any other solution in the collated set. Next, these solutions assigned rank as one are
excluded from the collated set and the nd = 0 in the reduced collated set are assigned rank two. This
ranking procedure is iteratively carried out until all the solutions in the collated set are ranked. All the
solutions in the collated set are then sorted according to their Pareto ranks assigned in the previous
steps and the Pareto rank of Nth solution is found out (where N is the a priori set maximum number of
non-dominated solutions required for the ensemble Pareto optimal front). This rank is denoted as ndc.
For all the solutions with Pareto rank = ndc, the crowding distance in the objective function space is
calculated. All the solutions are sorted according to their Pareto ranks, and then sorted with Pareto
rank = ndc in the order of decreasing crowding distance. The first N solutions form the ensemble
Pareto optimal set.
2.4 Metrics for Pareto optimal front comparison
The Pareto optimal fronts are assessed using four different metrics, i.e. generational distance
(GD), convergence metric (CM), spread (SP) and inverse generational distance (IGD). The GD
evaluates the closeness of non-dominated solutions to the known or true Pareto-optimal front [36]. In
this paper, to establish the superiority of MOALO-MODA ensemble, it is treated as the true Pareto
optimal front. The GDs of MOALO ensemble and MODA ensemble are then compared with respect
to MOALO-MODA ensemble. The value of GD is calculated using the following equation:


NP 2
d
i 1 i
GD  (23)
NP
where NP is the number of solutions in MOALO ensemble Pareto optimal front (or MODA ensemble
Pareto optimal front) and di is the Euclidean distance of each of these solutions to the nearest solution

9
in the true Pareto optimal front. Similarly, CM too evaluates the closeness of non-dominated solutions
to the true Pareto-optimal front [37].


NP
d
i 1 i
CM  (24)
NP
On the other hand, SP measures the distribution of non-dominated solutions. It was originally
propounded by Deb et al. [38] and later extended by Zhou et al. [37] as shown below:

 
M NP
m 1
(em , S )  i 1
i  
SP  (25)

M
m 1
(em , S )  NP.

i  min Si  S j (j = 1,2,...,NP) (except i = j)


2

where M is the number of objective functions, em are the M extreme solutions on the true Pareto
optimal front, ∆(em,S) is the Euclidean distance between the extreme solution on the true Pareto
optimal front and extreme non-dominated solution on MOALO ensemble Pareto optimal front (or
MODA ensemble Pareto optimal front) based on mth objective function, ∆i is the Euclidean distance of
ith non-dominated solution (Si) to its nearest non-dominated solution (Sj) on MOALO ensemble Pareto
optimal front (or MODA ensemble Pareto optimal front) and  is the mean of ∆i for all non-
dominated solutions. The value of IGD [39] is calculated based on the following expression:


NT
d
i 1 i
IGD  (26)
NT
where NT is the number of solutions in the true Pareto optimal front and di is the Euclidean distance
of ith non-dominated solution on the true Pareto optimal front to its nearest solution in MOALO
ensemble Pareto optimal front (or MODA ensemble Pareto optimal front). Further, for ease of
comparison, the non-dominated solutions (Si) predicted by the three ensembles are normalized ( Ŝ i )
using Eq. (27) and the Pareto optimal front comparison metrics presented in Eqs. (23)-(26) are
calculated again in their normalized forms.

S  S min
T
Sˆi  T i (27)
S max  S min
T

T T
where S min and S max are the lower and upper extremum non-dominated solutions on the true Pareto
optimal front.
3. Case studies
3.1 Case study1: Turning of aluminium MMC
To demonstrate the efficacy of the proposed ensemble approach, a couple of case studies on
machining of MMCs are considered here. Prakash et al. [40] employed an L27 orthogonal array to
design a set of experiments during turning of rock dust reinforced aluminium MMC. The effects of
reinforcement weight % (W), particle size (PS) (μm), turning speed (S) (rpm), feed rate (F) (mm/rev)

10
and depth of cut (DC) (mm) on MRR (g/min) and Ra (average surface roughness) (μm) were studied.
During those experiments, each of the process parameters was varied at three different levels. The
experimental values of the considered process parameters and responses are presented in
supplementary Table A1. Prakash et al. [40] employed grey relation analysis (GRA) to find out the
optimal combination of the process parameters.
It is well documented in the literature that MCDM techniques are often unable to locate the
global optimal solution, which leads to the deployment of metaheuristic algorithms for the said
purpose. In this case study, the objective is to simultaneously maximize MRR and minimize Ra.
However, due to conflicting nature of the parametric settings needed to individually maximize MRR
and minimize Ra, instead of a unique solution, a set of non-dominated solutions called the Pareto
optimal front is preferred. The proposed MOALO ensemble, MODA ensemble and MOALO-MODA
ensemble are subsequently applied to determine the best Pareto optimal front. However, for
deployment in the optimization scenario, the information (regarding dependency of the responses on
the process parameters) derived from the experimental data in Table A1 must be expressed in the
form of metamodels.
For this purpose, polynomial regression is used to express MRR and Ra as functions of the
considered process parameters, i.e. W, PS, S, F and DC. Initially, a full quadratic metamodel is
developed and then step-wise elimination method is employed to remove the statistically insignificant
terms from the metamodels. These insignificant terms in the model are identified based on their
corresponding p-values (> 0.1) from analysis of variance (ANOVA) results. Thus, the following
reduced quadratic metamodels are derived for MRR and Ra.
MRR = 24.64739 + 0.18106×PS – 0.00844×W – 0.02999×S – 143.52735×F – 64.41538×DC –
0.50416×PS×DC + 0.154025×S×F + 0.07927×S×DC + 382.25×F×DC (28)
Ra = 1.35025 + 0.52190×PS + 0.03555×W – 0.00025×S – 63.67138×F + 7.34895×DC – 0.00046×PS×S
+ 1.20777×PS×F– 0.07916×PS×DC + 0.06027×S×F – 0.00450×S×DC– 0.005611×PS2 (29)
The corresponding ANOVA results for the reduced quadratic MRR and Ra metamodels are
presented in Table 1. It can be noticed from this table that for MRR metamodel except for PS and W,
the p-values for all other terms are less than 0.1, indicating them as statistically significant. It should
be noted that these first-order terms (PS and W in MRR and W in Ra) with p-values greater than 0.1
are not eliminated to maintain hierarchical integrity of the metamodels. The responses predicted by
the above-developed metamodels are compared with the experimental observations in Figure 2. It is
observed from Figure 2 that all the data points lie on or just near the identity (diagonal) line indicating
extremely high predictive power of the metamodels. To further examine these metamodels, their
externally studentized residuals are plotted as normal probability plots with 95% confidence interval
(CI) bands in Figure 3. For MRR, all the residuals lie within the CI bands indicating a normal
distribution of the residuals. On the other hand, for Ra, only one residual is noticed to be marginally
outside the lower CI band. The goodness of fit of the metamodels is measured using R2, adjusted R2

11
and predicted R2 values, and are reported in Table 2. There are negligible differences between R2 and
adjusted R2 values for both the responses, which indicate that no statistically insignificant terms are
included in the metamodels. Higher values of predicted R2 (which is the leave-one-out cross
validation error metric) further confirm the accuracy of the developed metamodels.
[Insert Figure 1 about here]
[Insert Figure 2 about here]
[Insert Figure 3 about here]
[Insert Table 1 about here]
[Insert Table 2 about here]
These metamodels are now deployed for optimization of the turning process parameters while
machining the considered aluminium MMCs. All the subsequent optimization tasks are performed on
an ASUS ROG Strix G531GT-AL030T series Windows system with Intel(R) CoreTM i7 9th Gen, CPU
@ 2.59GHz, L2 Cache Size 1.5MB and 8GB RAM with 4GB NVIDIA GeForce GTX1650. Five
independent trials of MOALO and MODA are carried out, and the resulting Pareto optimal fronts are
exhibited in Figure 4. A lot of variability is noticed in the five different MOALO Pareto optimal
fronts (Figure 4(a)). It is unveiled from the box plots in Figure 4(a) that the spread of the Pareto
optimal fronts is inconsistent across various trials. The average number of non-dominated solutions in
MOALO Pareto optimal fronts is 436 with a standard deviation of 94. A comparatively low (average)
number of solutions as compared to the maximum archive size of 500 (for primary NDS) and
relatively high standard deviation among various trials indicate that the results of MOALO are too
stochastic. The best Pareto optimal front with respect to spread is obtained in trial 5 of MOALO. The
average computational time for MOALO is 107 s with a standard deviation of 49 s. The curvature
noticed at the lower side of some of the Pareto optimal fronts suggests that special care should be
taken in choosing the optimal solutions from the lower Ra region. The Pareto optimal fronts generated
by five different trials of MODA (Figure 4(b)) are very similar to each other. Small variation is
noticed only in the lower end (lower Ra region) of the Pareto optimal fronts of MODA. The average
and standard deviation of the number of non-dominated solutions in MODA Pareto optimal fronts are
492 and 8 respectively. Each MODA run takes on an average 107 s (± 56 s) to complete. Thus, so far
it can be concluded that MODA is more robust as compared to MOALO though both of them have
similar computational time requirements. However, the largest range (i.e. difference between two
extremum values of the Pareto optimal front) among the 10 trials is observed in case of MOALO trial.
[Insert Figure 4 about here]
The ensemble MOALO Pareto optimal front (generated using the five MOALO trials) and the
ensemble MODA Pareto optimal front (generated based on the five MODA trials) are presented in
Figure 5(a) and 5(b) respectively. It is worthwhile to mention here that a sufficiently high maximum
archive size (5000) is considered in this paper during the secondary and tertiary NDS stages. This
ensures that only the dominated solutions in the collated set are eliminated. It can be revealed from

12
Figure 5 that 747 and 2217 non-dominated solutions are present in MOALO and MODA ensembles
respectively. However, MOALO seems to have better non-dominated solutions at lower end of the
Pareto optimal front as compared to MODA. Thus, it becomes clear that despite high robustness of
MODA, the utility of MOALO in navigating certain complex regions of the objective function cannot
be ignored. This also justifies the need for heterogeneous ensemble.
[Insert Figure 5 about here]
Thus, the two homogenous ensembles discussed above are further passed through NDS for
tertiary sorting that would eventually lead to the development of MOALO-MODA ensemble Pareto
optimal front, presented in Figure 6. The MOALO-MODA ensemble is made up of 2387 non-
dominated solutions out of which only 174 (~7.3%) are contributed by MOALO. Thus, as also
observed from this figure, MODA accounts for majority of the non-dominated solutions in the
heterogenous ensemble, but MOALO contributes some of the most important solutions towards the
lower end of the Pareto optimal front. This is further highlighted in Figure 7 using parallel plots for
the ensembles. The solutions represented by red and blue colours vividly contrast the superiority of
MODA in the middle and upper portions of the Pareto optimal fronts while MOALO solutions
dominate the lower portion of the Pareto optimal fronts. It is also important to point out here that the
secondary and tertiary NDS cost less than 5s each. As discussed earlier, for calculation of various
metrics, like GD, CM, SP and IGD, MOALO-MODA ensemble is considered as the true Pareto
optimal front. The MOALO and MODA ensembles are quantitatively compared with the true Pareto
optimal front using these metrics in Table 3. It is noticed that MODA has better performance as
compared to MOALO.
[Insert Figure 6 about here]
[Insert Figure 7 about here]
[Insert Table 3 about here]
It should also be stated here that any Pareto optimal front contains a set of optimal non-
dominated solutions, and the superiority of one cannot be arbitrary established over the others. Thus,
each solution of the Pareto optimal solution is a compromise solution between two or more conflicting
objectives. On the other hand, industrial problems generally require a single solution adhering to
certain pre-specified needs. Perhaps, this may be the reason of much popularity of MCDM techniques
in such applications, where objective or criteria weights (i.e. importance quotients) are implied a
priori. Nevertheless, it is hypothesised here and subsequently shown that MCDM application to
Pareto optimal front can lead to much better solutions as opposed to MCDM-based level averaging
approach widely prevalent in the literature.
Table 4 shows the comparison of the derived solutions with those of Prakash et al. [40]. As
stated earlier, GRA technique was employed by Prakash et al. [40], which would not require objective
weights to be explicitly described, but would calculate their values based on the decision matrix.
Thus, in this paper, GRA is also employed to determine the optimal solutions from the ensemble

13
Pareto optimal fronts. It can be noticed that the solutions by MOALO ensemble are 14% and 99%
better for MRR and Ra respectively as compared to the process parameters reported in [40] and
calculated in this paper using the developed metamodels. Similarly, MODA ensemble solutions are
found to be 51% and 10% better for MRR and Ra respectively. The prediction of MOALO-MODA
ensemble is the same as that of MOALO ensemble. Since MOALO-MODA ensemble is entirely made
up of the solutions from both MOALO and MODA ensembles, it is quite likely that the same solution
as reported by one of the homogenous ensembles may be achieved.
[Insert Table 4 about here]
To further show the effectiveness of the proposed ensemble approach in generating tailor-
made solutions based on different end requirements, some additional example solutions are also
reported in Table 4. A priori weights of 25%, 50% and 75% are assigned to MRR in three different
scenarios. Technique for order of preference by similarity to ideal solution (TOPSIS) is applied here
to determine the optimal solutions from the Pareto optimal fronts. Description of TOPSIS is beyond
the scope of this paper and may be found in Wang et al. [35], Singh et al. [41], and Kalita et al. [42].
In all these scenarios, the current solutions almost outperform the results reported in [40]. It is also
interesting to observe that at lower MRR weights (i.e. 25% and 50%), the TOPSIS-based optimal
MOALO-MODA Pareto solutions are those which are drawn from MOALO ensemble. This further
indicates that despite MODA’s overall dominance in generating better non-dominated solutions,
MOALO can also navigate certain portions of the objective function space better than MODA.
3.2 Case study 2: Drilling of boron carbide reinforced MMCs
In this case study, drilling operation of boron carbide reinforced MMC is considered. Based on
Taguchi’s L27 orthogonal array, Taşkesen and Kütükde [43] conducted 27 experiments to determine
thrust force (Fa) (N), drilling torque (Mc) (N cm) and surface roughness (Ra) (μm) while treating B4C
particle weight % (W), feed rate (F) (mm/rev), spindle speed (S) (rev/min) and drill material (D) as
the input drilling parameters. The experimental values of those process parameters and the measured
responses are presented in supplementary Table A2. During the experimental runs, Taşkesen and
Kütükde [43] considered three different levels for each of the process parameters and finally applied
GRA technique to search out the optimal parametric combination.
In order to carry out multi-objective optimization for simultaneous minimization of Fa, Mc and
Ra using the proposed ensemble technique, the following metamodels for the three responses are
subsequently developed.
Fa = –57 + 115.2×W + 182×F + 0.061×S – 0.018×D – 0.193×W2 + 2833×F2 + 0.000011×S2 –
0.000013D2 – 64.5×W×F – 0.02266×W×S – 0.00486×W×D + 0.110×F×S + 0.288×F×D +
0.000039×S×D (30)
Mc = 766 – 3.1×W + 1109×F – 0.454×S – 0.018×D – 0.033×W + 452×F + 0.000087×S2 +
2 2

0.000010D2 + 7.8×W×F + 0.00281×W×S + 0.00597×W×D – 0.278×F×S – 0.280×F×D –


0.000016×S×D (31)

14
Ra = 43.2 + 0.60×W – 97.8×F – 0.0248×S – 0.01360×D – 0.0056×W2 + 102×F2 + 0.000007×S2 +
0.000003D2 + 0.19×W×F – 0.000264×W×S – 0.000012×W×D+ 0.0188×F×S+ 0.0091×F×D –
0.000001×S×D (32)
Since the focus of this paper is on development of a novel ensemble approach for multi-
objective optimization and for the sake of compactness of this paper, the steps and results regarding
ANOVA and other visual analyses of the metamodels are not presented here. These metamodels are
now employed as objective functions to predict 3D Pareto optimal fronts.
At first, multi-objective optimization is carried for simultaneous minimization of Fa, Mc and Ra
using MOALO and MODA. Five independent trials of MOALO and MODA are carried out and the
corresponding results are reported in the form of violin plots in Figure 8. The three-dimensional (due
to three responses) Pareto optimal fronts are easier to compare for each dimension, algorithm-wise
using these violin plots. From Figure 8(a), it can be noticed that the 5th trial of MOALO is
significantly different than the rest for Fa values. On the other hand, MODA has similar distribution
for all the trials. It is interesting to observe from Figure 8(a) that in the all the trials of MOALO,
ample number of non-dominated solutions exist in the upper value region of Fa, which is almost
inexistent in MODA trials. Similar observation is also noticed from Figure 8(b) for Mc. The stark
differences between the non-dominated solutions of MOALO and MODA are perhaps most
prominent in Figure 8(c). From the distribution of the Pareto optimal fronts across the three
dimensions for each trial, it is also become evident that while a certain trial may have better values in
a particular dimension, other trials may have better solutions in some other dimension. This justifies
the current strategy of collating multiple trials of the algorithm and obtaining an ensemble Pareto
optimal front that would amalgamate the critical information of multiple trials. It should be noted here
that MODA (average 72s, standard deviation 15s) is marginally less computationally intensive as
compared to MOALO (average 87s, standard deviation 14.7s). Further, in both MOALO and MODA,
for all the trials, the maximum archive size of 500 non-dominated solutions is reached.
[Insert Figure 8 about here]
Figure 9 exhibits the homogenous ensemble 3D Pareto optimal fronts derived using MOALO
and MODA. There are 1386 and 1885 non-dominated solutions in MOALO ensemble and MODA
ensemble respectively after secondary NDS. Figure 9 shows that MOALO has better solution density
in higher Mc and lower Ra regions, while the cluster of solutions by MODA is much better than
MOALO in the mid Ra region. These contrasts in the solutions of MOALO ensemble and MODA
ensemble in 3D objective space clearly indicate that there is sufficient scope for improvement of
results for both the algorithms. Thus, tertiary NDS is adopted by collating MOALO ensemble and
MODA ensemble solutions to obtain MOALO-MODA ensemble, presented in Figure 10. Among
2868 non-dominated solutions in MOALO-MODA ensemble, 1124 (i.e. ~39%) solutions are from
MOALO ensemble. This is a much better equitable distribution of non-dominated solutions than that
observed in the previous 2D Pareto optimal case study. The parallel plots in Figure 11 depict the

15
distribution of the Pareto optimal solutions in the three ensembles for different drilling parameters and
responses. These plots further help in understanding the solution pattern of the two different
algorithms used in this paper. The quantitative comparison of the two homogeneous ensembles with
the true Pareto optimal front is presented in Table 5, which proves that MODA has moderately better
performance than MOALO.
[Insert Figure 9 about here]
[Insert Figure 10 about here]
[Insert Figure 11 about here]
[Insert Table 5 about here]
The optimal solutions reported in [43] are compared with the current solutions in Table 6.
Minor improvements are noticed in Fa for all the three ensembles, whereas, for Mc, the current results
are poorer than [43]. However, the current Ra solutions are at least 54% better than the literature [43].
It should again be stated here that the solutions selected using GRA (or any other MCDM technique)
from the ensemble Pareto optimal fronts are strictly dependent on the objective weights. Thus, the
ensemble Pareto optimal fronts present the decision maker (or end user) with a plethora of
opportunities to choose from. For sake of brevity, unlike in previous case study, no additional
solutions for different weights are shown in Table 6.
[Insert Table 6 about here]
4. Conclusions
Multiple independent trials of metaheuristic optimization are usually carried out to account for
the stochastic nature of these algorithms. The best Pareto optimal front with respect to spread and/or
generational distance is then selected as a suitable solution to the optimization problem. However, the
selected Pareto optimal front may have some sub-optimal solutions in certain regions of the objective
function space. Thus, in this paper, a novel ensemble approach to multi-objective optimization is
proposed while considering MOALO and MODA algorithms. It relies on multiple re-runs of the
metaheuristic optimization and then combines the non-dominated solutions from these independent
runs to form an ensemble Pareto optimal front. This is performed by collating the individual Pareto
optimal fronts generated by the same algorithm and carrying out a secondary NDS to obtain the
ensemble MOALO or ensemble MODA. These homogenous Pareto optimal fronts are further
enhanced by collating the previous two ensembles and carrying out a tertiary NDS to obtain the
heterogeneous MOALO-MODA ensemble. Based on two comprehensive case studies for multi-
objective optimization of machining of MMCs, it is revealed that MODA is more robust and faster
than MOALO. The non-dominated solution density of MODA is also observed to be better. However,
MOALO is able to navigate certain sections of the objective function to locate effective non-
dominated solutions that MODA totally ignored. Nevertheless, MOALO solutions are accounted for
only 7.3% and 39% of the total solutions in 2D and 3D MOALO-MODA ensemble Pareto fronts
obtained in case studies 1 and 2 respectively. The run times of secondary and tertiary sorting are only

16
a fraction of the total computational cost of the metaheuristic simulations. Thus, it is proposed that at
practically negligible additional computational cost, better Pareto optimal fronts can be obtained by
the proposed ensemble method. The limitation of this paper is in sticking to only two metaheuristics
and that too from the same class of nature-inspired algorithms. However, in principle, this
methodology can be extended to include Pareto optimal solutions from several other optimization
algorithms. Enhancement of the Pareto optimal solutions by the proposed ensemble methodology
while encompassing algorithms from different classes, like metaphor-less, gradient-based etc. may
perhaps lead to more interesting results.
Author Declarations
Funding: No external financial support (fund) has been required.
Competing interests: There is no conflict of interest among the authors. All the authors have gone
through the manuscript and agreed to submit this paper.
Ethics approval/declarations: No ethics have been violated.
Consent to participate: All the authors have given their consents to participate.
Consent for publication: All the authors have given their permission to publish.
Availability of data and material/ Data availability: The data is shown in Supplementary.
Code availability: All the related codes are developed indigenously.
Authors’ contributions: KK: Conceptualization, coding and writing, VK: Data collection, coding,
analysis, SC: Technical writing and editing
References
1. Übeylı M, Acir A, Serdar Karakaş M, Ögel B (2008) Effect of feed rate on tool wear in milling
of Al-4% Cu/B4Cp composite. Materials and Manufacturing Processes 23: 865-870
2. Basavarajappa S, Chandramohan G, Davim JP, Prabu M, Mukund K, Ashwin M, Prasanna
Kumar M (2008) Drilling of hybrid aluminium matrix composites. International Journal of
Advanced Manufacturing Technology 35: 1244-1250
3. Rubi CS, Prakash JU, Čep R, Elangovan M (2022) Optimization of process variables in the
drilling of LM6/B4C composites through grey relational analysis. Materials 15(14): 4860
4. Davim JP, Monteiro Baptista A (2001) Cutting force, tool wear and surface finish in drilling
metal matrix composites. Proceedings of the Institution of Mechanical Engineers, Part E:
Journal of Process Mechanical Engineering 215: 177-183
5. Tosun G, Muratoglu M (2004) The drilling of an Al/SiCp metal-matrix composites. Part I:
microstructure. Composites Science and Technology 64: 299-308
6. Ahamed AR, Asokan P, Aravindan S, Prakash MK (2010) Drilling of hybrid Al-5% SiC-5%
B4Cp metal matrix composites. International Journal of Advanced Manufacturing Technology
49: 871-877
7. Tosun G, Muratoglu M (2004) The drilling of Al/SiCp metal-matrix composites. Part II:
workpiece surface integrity. Composites Science and Technology 64: 1413-1418

17
8. Ozben T, Kilickap E, Cakır O (2008) Investigation of mechanical and machinability properties
of SiC particle reinforced Al-MMC. Journal of Materials Processing Technology 198: 220-225
9. Outeiro JC (2014) Surface integrity predictions and optimisation of machining conditions in the
turning of AISI H13 tool steel. International Journal of Machining and Machinability of
Materials 15: 122-134
10. Manna A, Bhattacharayya B (2003) A study on machinability of Al/SiC-MMC. Journal of
Materials Processing Technology 140: 711-716
11. Sidhu SS, Yazdani M (2018) Comparative analysis of MCDM techniques for EDM of
SiC/A359 composite. Arabian Journal for Science and Engineering 43: 1093-1102
12. Tamiloli N, Venkatesan J, Murali G, Kodali SP, Kumar TS, Arunkumar MP (2019)
Optimization of end milling on Al-SiC-fly ash metal matrix composite using Topsis and fuzzy
logic. SN Applied Sciences 1: 1-15
13. Patel RK, Pradhan MK (2019) Process parameters optimization of electrical discharge
machining of Al7075/SiC/WS2 by using MCDM. In: Advances in Computational Methods in
Manufacturing, Springer 329-336
14. Satpathy A, Tripathy S, Senapati NP, Brahma MK (2017) Optimization of EDM process
parameters for AlSiC-20% SiC reinforced metal matrix composite with multi response using
TOPSIS. Materials Today: Proceedings 4: 3043-3052
15. Kumar A, Hussain SAI, Rai RN (2019) Optimization by AHP-ARAS of EDM process
parameters on machining AA7050-10% B4C composite. In: Advances in Industrial and
Production Engineering, Springer 285-296
16. Kumar NM, Kumaran SS, Kumaraswamidhas LA (2016) Wear behaviour of Al 2618 alloy
reinforced with Si3N4, AlN and ZrB2 in situ composites at elevated temperatures. Alexandria
Engineering Journal 55: 19-36
17. Radhika N, Vijaykarthik KT, Shivaram P (2015) Adhesive wear behaviour of aluminium
hybrid metal matrix composites using genetic algorithm. Journal of Engineering Science and
Technology 10: 258-268
18. Rao CS, Rao YS, Marimuthu P, Jeyapaul R, Chakravarthy NK, Murugesan P (2021)
Optimisation of drilling parameters of metal matrix composites using genetic algorithm in the
Taguchi method. In: IOP Conference Series: Materials Science and Engineering 1126: 012035
19. Bongale AM, Kumar S, Sachit TS, Jadhav P (2018) Wear rate optimization of Al/SiCnp/e-glass
fibre hybrid metal matrix composites using Taguchi method and genetic algorithm and
development of wear model using artificial neural networks. Materials Research Express 5:
035005
20. Adithiya T, Chandramohan D, Sathish T (2020) Optimal prediction of process parameters by
gwo-knn in stirring-squeeze casting of AA2219 reinforced metal matrix composites. Materials
Today: Proceedings 21: 1000-1007

18
21. Bazgan C, Ruzika S, Thielen C, Vanderpooten D (2022) The power of the weighted sum
scalarization for approximating multiobjective optimization problems. Theory of Computing
Systems 66(1): 395-415
22. Chagas JB, Wagner M (2022) A weighted-sum method for solving the bi-objective traveling
thief problem. Computers & Operations Research 138: 105560
23. Selvakumar V, Muruganandam S, Tamizharasan T, Senthilkumar N (2016) Machinability
evaluation of Al-4% Cu-7.5% SiC metal matrix composite by Taguchi-Grey relational analysis
and NSGA-II. Sādhanā 41: 1219-1234
24. Singh B, Kumar S, Kumar J (2017) Multi-objective optimization in electrical discharge
machining of 6061 Al/SiCp using RSM and NSGA-II. Key Engineering Materials 748: 207-
211
25. Kumar KP, Vinay PV, Chintada AK, Bhimuni V (2016) Optimization of machining parameters
using micro NSGA-II for Al6061-silver coated copper metal matrix composite. In: Proceedings
of 6th International and 27th All India Manufacturing Technology, Design and Research
Conference, India 1667-1671
26. Daniel SAA, Kumar R, Vijay Ananth S, Pugazhenthi R (2020) Multi-objective optimization of
drilling of Al5059-SiC-2% MoS2 composites using NSGA-II. Materials Today: Proceedings
22: 853-857
27. Ikhlas M, Athmane YM, Hamza B, Ahmed K, Mohamed E (2018) Prediction of surface
roughness and cutting forces using RSM, ANN, and NSGA-II in finish turning of AISI 4140
hardened steel with mixed ceramic tool. International Journal of Advanced Manufacturing
Technology 97: 1931-1949
28. Mirjalili S (2015) The ant lion optimizer. Advances in Engineering Software 83: 80-98
29. Mirjalili S, Jangir P, Saremi S (2017) Multi-objective ant lion optimizer: a multi-objective
optimization algorithm for solving engineering problems. Applied Intelligence 46: 79-95
30. Mirjalili S (2016) Dragonfly algorithm: a new meta-heuristic optimization technique for
solving single-objective, discrete, and multi-objective problems. Neural Computing and
Applications 27: 1053-1073
31. Yang X-S (2010) Nature-inspired Metaheuristic Algorithms. Luniver Press.
32. Deb K, Pratap A, Agarwal S, Meyarivan TAMT (2002) A fast and elitist multiobjective genetic
algorithm: NSGA-II. IEEE Transactions on Evolutionary Computation 6: 182-197
33. Sharma S, Rangaiah GP, Maréchal F (2017) Multi-objective optimization programs and their
application to amine absorption process design for natural gas sweetening. In: Multi-objective
Optimization: Techniques and Application in Chemical Engineering, World Scientific 533-560
34. Wang Z, Rangaiah GP (2017) Application and analysis of methods for selecting an optimal
solution from the Pareto-optimal front obtained by multiobjective optimization. Industrial &
Engineering Chemistry Research 56: 560-574

19
35. Wang Z, Parhi SS, Rangaiah GP, Jana AK (2020) Analysis of weighting and selection methods
for Pareto-optimal solutions of multiobjective optimization in chemical engineering
applications. Industrial & Engineering Chemistry Research 59: 14850-14867
36. Van Veldhuizen DA, Lamont GB (1998) Evolutionary computation and convergence to a
Pareto front. In: Late Breaking Papers at the Genetic Programming 1998 Conference,
Wisconsin, USA
37. Zhou A, Jin Y, Zhang Q, Sendhoff B, Tsang E (2006) Combining model-based and genetics-
based offspring generation for multi-objective optimization using a convergence criterion. In:
Proceedings of IEEE International Conference on Evolutionary Computation 892-899
38. Deb K (2011) Multi-objective optimisation using evolutionary algorithms: an introduction. In:
Multi-objective Evolutionary Optimisation for Product Design and Manufacturing, Springer 3-
34
39. Wang Y, Wang W, Ahmad I, Tag-Eldin E (2022) Multi-objective quantum-inspired seagull
optimization algorithm. Electronics 11(12): 1834
40. Prakash KS, Gopal PM, Karthik S (2020) Multi-objective optimization using Taguchi based
grey relational analysis in turning of rock dust reinforced Aluminum MMC. Measurement 157:
107664
41. Singh A, Ghadai RK, Kalita K, Chatterjee P, Pamučar D (2020) EDM process parameter
optimization for efficient machining of Inconel-718. Facta Universitatis, Series: Mechanical
Engineering 18: 473-490
42. Kalita K, Pal S, Haldar S, Chakraborty S (2022) A hybrid TOPSIS-PR-GWO approach for
multi-objective process parameter optimization. Process Integration and Optimization for
Sustainability. https://doi.org/10.1007/s41660-022-00256-0
43. Taşkesen A, Kütükde K (2014) Experimental investigation and multi-objective analysis on
drilling of boron carbide reinforced metal matrix composites using grey relational analysis.
Measurement 47: 321-330

20
Table 1 ANOVA results for reduced quadratic MRR and Ra metamodels
MRR Ra
Source
SS df F-value p-value SS df F-value p-value
Model 11999.33 9 903.81 < 0.0001 132.75 11 138.02 0.0000
PS 0.48 1 0.32 0.5769 2.45 1 28.06 0.0001
W 0.03 1 0.02 0.8845 0.08 1 0.93 0.3503
S 1925.31 1 1305.16 < 0.0001 0.27 1 3.08 0.0999
F 4406.73 1 2987.31 < 0.0001 14.65 1 167.57 0.0000
DC 4341.88 1 2943.35 < 0.0001 1.54 1 17.58 0.0008
PS×S - - - - 3.23 1 36.96 0.0000
PS×F - - - - 2.19 1 25.02 0.0002
PS×DC 12.20 1 8.27 0.0105 0.30 1 3.44 0.0834
S×F 179.93 1 121.97 < 0.0001 7.35 1 84.03 0.0000
S×DC 305.02 1 206.77 < 0.0001 0.99 1 11.28 0.0043
F×DC 701.35 1 475.45 < 0.0001 - - - -
PS2 - - - - 1.08 1 12.35 0.0031
Residual 25.08 17 - - 1.31 15 - -
Cor Total 12024.41 26 - - 134.06 26 - -
Table 2 Predictive performance of the metamodels for case study 1
Metric MRR Ra
R² 0.9979 0.9902
Adjusted R² 0.9968 0.9830
Predicted R² 0.9915 0.9669

Table 3 Comparison of MOALO and MODA ensembles with true Pareto (MOALO-MODA
ensemble) in case study 1
MOALO MODA
Metric
Original Normalized Original Normalized
GD 0.025010 0.000489 0.000000 0.000000
CM 0.005054 0.000161 0.040941 0.000838
SP 0.066600 0.003498 0.000000 0.000000
IGD 1.362931 1.278484 0.834328 0.833726

21
Table 4 Comparison of the derived solutions with literature for case study 1
Optimal parameters Optimal responses
Weight for
Method % %
MRR PS, W, S, F, DC MRR
improvement
Ra
improvement
34.041 1.566
GRA [40] 20, 10, 1273, 0.1, 0.6 - -
(41.54)2 (1.3)
MOALO ensemble 30, 11.89, 1273, 0.12, 0.6 38.82 14% 0.020 99%
GRA
MODA ensemble 10, 5, 643, 0.3, 0.6 51.50 51% 1.414 10%
MOALO-MODA
30, 11.89, 1273, 0.12, 0.6 38.82 14% 0.020 99%
ensemble
25% 30, 11.89, 1273, 0.12, 0.6 38.82 14% 0.020 99%
MOALO ensemble
50% 10, 5.23, 637, 0.3, 0.6 51.11 50% 1.359 13%
(TOPSIS)
75% 10, 5, 1118, 0.3, 0.6 81.80 140% 6.417 -310%3
25% 30, 8.58, 1272, 0.13, 0.6 41.96 23% 0.458 71%
MODA ensemble
50% 10, 5, 643, 0.3, 0.6 51.50 51% 1.414 10%
(TOPSIS)
75% 10, 5, 1200, 0.3, 0.6 87.06 156% 7.286 -365%
25% 30, 11.89, 1273, 0.12, 0.6 38.82 14% 0.020 99%
MOALO-MODA
50% 10, 5.23, 637, 0.3, 0.6 51.11 50% 1.359 13%
ensemble (TOPSIS)
75% 10, 5, 1200, 0.3, 0.6 87.06 156% 7.286 -365%
1Forunbiased comparison, MRR and Ra values by the current metamodels are used.
2Values within the parentheses are the experimental values reported in [40].
3Negative values indicate that the current results are poorer than [40].

Table 5 Comparison of MOALO and MODA ensembles with true Pareto (MOALO-MODA
ensemble) in case study 2
MOALO MODA
Metric
Original Normalized Original Normalized
GD 0.078404 0.000139 0.041417 0.000099
CM 0.055312 0.000261 0.056035 0.000126
SP 0.508263 0.001292 0.291016 0.000806
IGD 1.481294 1.447995 1.384211 1.199136
Table 6 Comparison of current solutions with literature for case study 2
Optimal
Optimal responses
parameters
Method
% % %
W, F, S, D Fa Mc Ra
improvement improvement improvement
10, 0.1, 2500, 805.671 272.552 2.722
GRA [43] - - -
2800 (804.3)2 (249.3) (2.36)
MOALO
10, 0.1, 2275,
ensemble 803.83 0.23% 290.802 -7%3 1.254 54%
2800
(GRA)
MODA
10, 0.1, 2247,
ensemble 803.68 0.25% 293.734 -8% 1.118 59%
2800
(GRA)
MOALO- 10, 0.1, 2253,
803.71 0.24% 293.101 -8% 1.146 58%
MODA 2800

22
ensemble
(GRA)
1Forunbiased comparison, Fa, Mc and Ra values by the current metamodels are used.
2Values within the parentheses are the experimental values reported in [43].
3Negative values indicate that the current results are poorer than [43].

Figure 1 Flowchart of the proposed three-level non-dominated sorting methodology

Figure 2 Experimental versus predicted plots for (a) MRR and (b) Ra

23
Figure 3 Normal probability plots of externally studentized residuals for (a) MRR and (b) Ra

Figure 4 Pareto optimal fronts for five independent trials (a) MOALO and (b) MODA [Each colour
represents an independent trial. The vertical box plots show distribution of Ra in the Pareto front. The
horizontal box plots are for MRR]

24
Figure 5 Ensemble Pareto optimal fronts computed using five independent trials of (a) MOALO and
(b) MODA

Figure 6 Ensemble Pareto front computed after tertiary NDS of ensemble MOALO and MODA
Pareto fronts [The inset images show the portions of the homogenous ensembles that make up the
heterogenous ensemble]

25
Figure 7 Parallel plots showing the process parameters and responses for the ensemble Pareto optimal
front computed using (a) MOALO (b) MODA and (c) MOALO-MODA

Figure 8 Violin plots showing spread of the responses computed for five independent trials using
MOALO and MODA for (a) Fa, (b) Mc and (c) Ra

26
Figure 9 Ensemble 3D Pareto optimal fronts computed using five independent trials of (a) MOALO
and (b) MODA

Figure 10 MOALO-MODA ensemble Pareto front computed after tertiary NDS of ensemble
MOALO and MODA Pareto optimal fronts

27
Figure 11 Parallel plots showing the process parameters and responses for the ensembled Pareto
optimal fronts computed using (a) MOALO (b) MODA and (c) MOALO-MODA

28
Supplementary
Table A1 Experimental data for case study 1 [40]
Exp. No. PS (µm) W (%) S (rpm) F (mm/rev) DC (mm) MRR (g/min) Ra (µm)
1 10 5 637 0.1 0.2 5.12 2.42
2 10 5 637 0.1 0.4 10.67 3.28
3 10 5 637 0.1 0.6 15.58 4.02
4 10 10 955 0.2 0.2 15.77 3.49
5 10 10 955 0.2 0.4 31.12 3.88
6 10 10 955 0.2 0.6 46.74 4.06
7 10 15 1273 0.3 0.2 32.11 7.88
8 10 15 1273 0.3 0.4 62.5 8.42
9 10 15 1273 0.3 0.6 92.44 8.34
10 20 5 955 0.3 0.2 24.22 6.49
11 20 5 955 0.3 0.4 46.87 6.97
12 20 5 955 0.3 0.6 70.73 6.53
13 20 10 1273 0.1 0.2 10.38 1.31
14 20 10 1273 0.1 0.4 21.43 1.66
15 20 10 1273 0.1 0.6 32.01 1.71
16 20 15 637 0.2 0.2 11.21 4.19
17 20 15 637 0.2 0.4 21.23 5.07
18 20 15 637 0.2 0.6 31.21 5.65
19 30 5 1273 0.2 0.2 20.43 3.74
20 30 5 1273 0.2 0.4 41.87 3.7
21 30 5 1273 0.2 0.6 61.98 3.78
22 30 10 637 0.3 0.2 15.11 6.89
23 30 10 637 0.3 0.4 31.2 7.32
24 30 10 637 0.3 0.6 47.66 8.17
25 30 15 955 0.1 0.2 7.73 2.77
26 30 15 955 0.1 0.4 14.99 1.97
27 30 15 955 0.1 0.6 23.29 2.18

29
Table A2 Experimental data for case study 2 [43]
Exp. No. W (%) F (mm/rev) S (rpm) D (Hv) Fa (N) Mc (N cm) Ra (µm)
1 10 0.1 1500 800 843.2 293 9.49
2 10 0.1 2000 1500 834.3 300.9 1.04
3 10 0.1 2500 2500 804.3 249.3 2.36
4 10 0.2 1500 1500 966.8 478.4 0.41
5 10 0.2 2000 2500 924.1 344 0.42
6 10 0.2 2500 800 880.7 294 5.03
7 10 0.3 1500 2500 1113 434.7 0.45
8 10 0.3 2000 800 1048 446.6 2.39
9 10 0.3 2500 1500 1096 291.3 12.87
10 15 0.1 1500 1500 1168 500.3 2.8
11 15 0.1 2000 2500 1062 398.5 0.73
12 15 0.1 2500 800 973.3 309.1 11.54
13 15 0.2 1500 2500 1244 496.7 0.33
14 15 0.2 2000 800 1132 370.6 8.32
15 15 0.2 2500 1500 1089 340 1.34
16 15 0.3 1500 800 1313 578.3 5.86
17 15 0.3 2000 1500 1310 328 0.51
18 15 0.3 2500 2500 1218 388.5 0.66
19 25 0.1 1500 2500 1631 627.1 0.57
20 25 0.1 2000 800 1655 434.1 4.44
21 25 0.1 2500 1500 1374 415.4 0.45
22 25 0.2 1500 800 1826 496.2 6.05
23 25 0.2 2000 1500 1477 492.8 0.37
24 25 0.2 2500 2500 1379 547.8 0.38
25 25 0.3 1500 1500 1834 674.1 0.92
26 25 0.3 2000 2500 1629 641 0.35
27 25 0.3 2500 800 1441 492.2 6.25

List of Tables

Table 1 ANOVA results for reduced quadratic MRR and Ra metamodels


Table 2 Predictive performance of the metamodels for case study 1
Table 3 Comparison of MOALO and MODA ensembles with true Pareto (MOALO-MODA
ensemble) in case study 1
Table 4 Comparison of the derived solutions with literature for case study 1
Table 5 Comparison of MOALO and MODA ensembles with true Pareto (MOALO-MODA
ensemble) in case study 2
Table 6 Comparison of current solutions with literature for case study 2
Table A1 Experimental data for case study 1 [40]
Table A2 Experimental data for case study 2 [43]

30
List of Figures
Figure 1 Flowchart of the proposed three-level non-dominated sorting methodology
Figure 2 Experimental versus predicted plots for (a) MRR and (b) Ra
Figure 3 Normal probability plots of externally studentized residuals for (a) MRR and (b) Ra
Figure 4 Pareto optimal fronts for five independent trials (a) MOALO and (b) MODA [Each colour
represents an independent trial. The vertical box plots show distribution of Ra in the Pareto front. The
horizontal box plots are for MRR]
Figure 5 Ensemble Pareto optimal fronts computed using five independent trials of (a) MOALO and
(b) MODA
Figure 6 Ensemble Pareto front computed after tertiary NDS of ensemble MOALO and MODA
Pareto fronts [The inset images show the portions of the homogenous ensembles that make up the
heterogenous ensemble]
Figure 7 Parallel plots showing the process parameters and responses for the ensemble Pareto optimal
front computed using (a) MOALO (b) MODA and (c) MOALO-MODA
Figure 8 Violin plots showing spread of the responses computed for five independent trials using
MOALO and MODA for (a) Fa, (b) Mc and (c) Ra
Figure 9 Ensemble 3D Pareto optimal fronts computed using five independent trials of (a) MOALO
and (b) MODA
Figure 10 MOALO-MODA ensemble Pareto front computed after tertiary NDS of ensemble
MOALO and MODA Pareto optimal fronts
Figure 11 Parallel plots showing the process parameters and responses for the ensembled Pareto
optimal fronts computed using (a) MOALO (b) MODA and (c) MOALO-MODA

31

You might also like