PERFORMANCE ANALYSIS OF METAHEURISTIC OPTIMIZERS ON THE ESTIMATION

OF THE INTERFACIAL HEAT TRANSFER COEFFICIENT IN ALUMINUM–SILICON

PERMANENT MOLD CASTING

Gianfranco de M. Stieven
Laboratory of Nano and Microfluidics and Microsystems, Mechanical Engineering Department, Universidade Federal do
Rio de Janeiro, Avenida Athos da Silveira Ramos, 274, Rio de Janeiro 21945-970, RJ, Brazil

Daniele S. Stieven
Mechanical Engineering Department, Universidade Federal do Pará, Av. Augusto Correa, 1, Belém 66075-900, PA, Brazil

Edilma P. Oliveira and Erb F. Lins

Faculty of Mechanical Engineering, Universidade Federal Rural de Pernambuco, Rua. Cento e Sessenta e Três, 300, Cabo
de Santo Agostinho 54518-430, PE, Brazil

Copyright  2023 American Foundry Society

https://doi.org/10.1007/s40962-023-01010-z

Abstract

The advancement of modeling and optimization research, law was used for the IHTC model. The numerical dis-
especially in metallurgical applications, usually aims at cretization of the problem is considered the Finite Volume
casting products with fewer macroscopic defects and Method approach. Statistical results from the Markov
superior mechanical resistance. Based on that, any casting Chain Monte Carlo method were considered a reference
industry with this objective, that requires the thermal for qualitative and quantitative performance analysis. The
optimization of the solidification process, which mainly Moth-Flame Optimizer (MFO) presented a probability of
controls the mechanical behavior of castings, must obtain a success of 85%, outperforming the other methods in all
non-trivial and relevant answer: which optimizer should parametric and non-parametric metrics. Further research
one use to get a quick and satisfactory answer to my is encouraged to verify the influence and effectiveness of
casting manufacturing process? This paper proposes a the intrinsic search to the MFO method to propose more
performance analysis of ten metaheuristic optimization efficient optimizers for this physical problem.
algorithms applied to an inverse estimation of the Inter-
facial Heat Transfer Coefficient (IHTC) coupled with the Keywords: optimization, Bayesian approach,
solidification phenomenon. For this purpose, a unidirec- nature-inspired algorithms, solidification
tional permanent mold casting was simulated considering
the Al–7wt%Si alloy. A multi-parameter temporal power

Introduction estimation using numerical routines derived from differ-

ential operators, Bayesian methodologies, or even nature
Many problems in science deal with parameter uncertainty. observations techniques. In the last case, studies involving
As a result, predictions of physical phenomena may lose the so-called Nature-Inspired Optimization Algorithms
accuracy, and scientific conclusions may be compromised. (NIOA) and Metaheuristic Optimization Algorithms
Usually, parameter estimation is done by optimization, (MOA) are usual. These methods are based on local and
where some objective functions should be minimized. In global random search and adaptative weighting
this context, several strategies are proposed: parameter mechanisms.

From the knowledge of the existence of numerous opti-

mization methods to reduce the uncertainty of parameters
Received: 27 September 2022 / Accepted: 24 February 2023

International Journal of Metalcasting

in the most diverse areas, the researcher must define which In another work9 the performance of two deterministic
method to use, regardless of his knowledge of the most methods, named Compass Search (CS) and NEWUOA, and
diverse types of optimizers. Based on that, by No Free two heuristic ones, named GA and PSO, was compared in
Lunch (NFL) theorems,1,2 originally established for search the adjustment of the metal casting simulation models,
and for optimization,3,4 it is known that no optimization considering two, three, six, and ten variables. For that, a
algorithm is better than others overall possible functions or commercial Finite Element Method (FEM) software Pro-
real-world problems. However, as reminded in the past,5 CAST 2018 was used. The authors concluded that only CS
despite some incorrect ideas about NFL, the original and PSO have succeeded in reaching the error target. They
statement does not ignore the fact that there is some also commented that as the ‘‘true’’ value of each variable is
algorithm A that can outperform B, once A could be spe- unknown, it is difficult to evaluate if any of the algorithms
cialized to the set of problems analyzed in the case. In this have a better performance in assigning more reliable values
sense, NFL encourages researchers to identify this spe- to the variables. This comment can be related either to an
cialization.6 One can do that by creating a tailored-hand experimental response that guides the optimization or the
algorithm to the problem to achieve better performance or, existence of a numerical reference such that it is possible to
instead, identifying among all developed algorithms one guarantee statistical reliability.
that is already specialized or is somehow better for a par-
ticular problem subset. A comparison of well-known heuristic approaches coupled
with a solidification model, applied to the secondary
In this context, there is very little information available cooling zone in continuous steel casting, in terms of
about the effectiveness of metaheuristic methods in specific computational efficiency, robustness, and accuracy was
applications of Mechanical and Metallurgical Engineering. presented before.10 For that, was selected five optimizers:
One of these is the transient heat and mass transfer coupled SA, DE, PSO, the Firefly Algorithm (FA), and the
with phase transformation, that is, the phenomenon of Teaching-Learning-Based Optimization (TLBO) method.
solidification, especially on permanent metallic mold Based on the best median value and convergence rate, the
castings.7 In this process, the Interfacial Heat Transfer authors concluded that the DE algorithm was the best-
Coefficient (IHTC) represents one of the most important performing algorithm for the considered control problem.
thermal coefficients since it predominantly controls the The performance of the SA, the FA, and the TLBO algo-
heat transfer of the metallic system, then directly influences rithms was evaluated as good, which makes them also
several thermal parameters, such as Cooling Rate (TR ) and suitable for the solution of this kind of control problem.
Tip Growth Rate (VL ), which can be related to metallur- The PSO method was the worst algorithm, presenting many
gical characteristics and mechanical properties. Some non-converged runs.
researchers, in addition to using metaheuristic algorithms
to solve inverse heat transfer problems in metallic solidi- Based on the exposed works, determining a metaheuristic
fication, have also proposed to evaluate the response of algorithm that can satisfactorily optimize the thermal
more than one optimizer. Among them, some aimed only at parameters in solidification with good reliability represents
verifying the response of more than one method to guar- an advance for thermal and metallurgical research. For this,
antee the accuracy of the estimated output. However, some a methodology capable of evaluating several metaheuristic
studies also aimed to assess which of the optimizers was methods concomitantly, using a statistically reliable answer
the most competent. as a reference, and considering parametric and non-para-
metric metrics, is relevant and utile to define which opti-
A performance comparison of DE, PSO, and an Improved mizer outperforms others on a specific physical problem.
Differential Evolution (IDE) algorithm on the identification
of heat transfer coefficient parameters of a SAE1007 steel So, in this paper, ten metaheuristic optimization methods
billet in continuous casting was done before.8 The FDM will be analyzed to determine which of them, in this par-
was used for the discretization of a 2D heat transfer ticular physical problem, returns the best temperature
equation. By the author’s conclusion, the PSO algorithm prediction. For this purposed, these optimizers were cho-
had a fast convergent speed, but it was easy to trap itself sen: Particle Swarm Optimization (PSO),11 Differential
into a local minimum. DE optimizer maintains the diversity Evolution (DE),12 Bat Algorithm (BA),13 Flower Pollina-
of the population. However, IDE presented better results in tion Algorithm (FPA),14 Grey Wolf Optimizer (GWO),15
comparison with DE. Even though these works presented a Moth-Flame Optimization (MFO),16 Sine Cosine Algo-
relevant result for solidification and metal casting, other rithm (SCA),17 Whale Optimization Algorithm (WOA),18
important considerations for performance analysis were not Dragonfly Algorithm (DA)19 and Harris Hawks Opti-
used for the studied metaheuristic optimizers, such as mization (HHO).20 These methods were selected based on
repeatability of the simulations, comparative analysis of their scientific relevance and applicability, once these
standard deviation, the average number of iterations, methods have already been applied by several researchers
probability of success, and non-parametric tests. in the most varied areas of science and their performances
are notoriously recognized.

International Journal of Metalcasting

As there is no defined analytical response to this engi- based alloys, the aluminum–silicon (Al–Si) system has
neering problem, it is necessary to obtain a reliable sta- considerable prestige and is widely used in the automotive
tistical response to evaluate the performance of the industry, especially on engines,27 due to their attractive
optimizers. In this sense, Markov Chain Monte Carlo strength-to-weight ratio as well as superior casting char-
(MCMC) method21,22 will be used to determine the sta- acteristics.28 Regarding the production volume, Al–Si
tistical consistency of the results provided by the opti- alloys account for 80 to 90% of the aluminum castings
mization algorithms under study. produced commercially.29 Based on this author, there is an
optimum range of silicon content, depending on the casting
Therefore, the contributions of this work are: To present a process: 5 to 7wt%Si for slow cooling-rate processes, such
Posterior Probability Distribution (PPD) of the IHTC as casting in plaster, investment, and sand molds; 7 to
parameters using MCMC, considering the Al–7wt%Si 9wt%Si for permanent molds; and 8 to 12wt%Si for die
alloy directionally fabricated in a permanent metallic mold, casting.
and; To analyze, based on MCMC’s PPD, the performance
of ten metaheuristic algorithms on the inverse optimization Therefore, given the industrial and academic relevance of
of the IHTC in directional permanent mold casting, con- this non-ferrous system and its usage ranges, this paper
sidering parametric metrics (sample mean, percentiles, chose Al–7wt%Si (also registered by the Aluminum
kurtosis, error, iterations, and probability of success) and a Association under the number A444.0) as the studied alloy.
non-parametric test, named Friedman Ranking Test. This alloy, especially on permanent mold casting, exhibits
excellent properties such as resistance to hot cracks, cor-
The results obtained via Markov Chain Monte Carlo will rosion resistance, good fluidity, shrinkage tendency,
be considered a reference to determine which method castability, and weldability.30
stands for the most appropriate inference. It is worth
pointing out that this analysis will not indicate which of the It is important to highlight that this type of study has direct
methods mentioned above will always be superior to the application in the reality of simulation, manufacturing, and
others in all applications involving the solidification phe- design of castings. In every manufacturing process, thermal
nomenon. This conclusion cannot exist using the approach process control is essential as it impacts the mechanical
proposed for this work. properties and performance of components. As shown in
previous works31,32 the search for IHTC control by man-
The contribution of this work is to clarify which of the ufacturing characteristics such as pressure, surface rough-
aforementioned methods has the greatest capacity to opti- ness, and casting thickness is essential for the quality of the
mize the IHTC in this phenomenon, for this manufacturing casting process.
process. Based on the ability of the most qualified opti-
mizer, another researcher will feel more comfort- In direct applications,33 made predictions about casting
able choosing this algorithm to optimize his specific defects, specifically hot tears, in magnesium alloys in
problem, related to the phenomenon under analysis, permanent mold casting. For this, the authors had to esti-
knowing that this optimizer is more likely to deliver better mate the IHTC by trial and error to simulate the behavior
results than the others analyzed in this work. of this casting product. An algorithm specialized in casting
optimization would reduce the estimated IHTC error and
Thus, the generalization of the proposition of the most thus provide greater confidence in the prediction of casting
suitable optimizer for similar problems (i.e. boundary defects.34 presented a study on the performance of riser
conditions and different geometries) can be considered, sleeves for steel castings. For this, the IHTC had to be
especially about metallic solidification in local interfacial estimated from inverse problems, in which a satisfactory
equilibrium,23 since the optimization surface will be optimization method could be applied to estimate, with a
topologically similar. high degree of reliability, this thermal parameter.35 studied
the dimensional control of turbine blades in investment
Based on this proposition, the selection of the metallic casting. For this, it was necessary to implement an IHTC as
system in this work does not directly influence the main a temporal power function, which was raised by.36 In this,
objective, which resides in the determination of an ade- there are three parameters that, when properly optimized,
quate optimizer for the physical problem in question. result in the real performance of the casting system. The
However, an aluminum-based alloy was the material cho- quality of optimization directly impacts the quality of the
sen to be analyzed based on its industrial relevance. This prediction of turbine blades, an important mechanical
type of alloy has notorious importance, since, to improve component that requires maximum operational quality.37
vehicle fuel efficiency by lower consumption, lightweight worked on an investigation of the IHTC during the casting
aluminum alloys have been replacing other materials for of an aluminum alloy and a magnesium alloy into steel and
use in automotive integrated systems and components24 graphite molds. The last two works estimated based on
such as engine blocks and cylinder heads,25 and also on algorithms that do not use a metaheuristic optimization
aerospace and marine applications.26 Among aluminum- strategy, that is, it performs a global and local search and

International Journal of Metalcasting

were not properly tested against other methods to verify Once the metal/mold interface and the ingot mold are
that this method is capable of providing satisfactory opti- prepared, they must be fixed together in the solidification
mization results by the level of demand of the manufac- device. Then, type K thermocouples must be dovetailed to
turing process. the holes on the side of the ingot mold to collect the
temperature distribution during the process. After com-
In this context, this work presents itself as an aid for the pleting these operational steps, the experiment is started.
satisfactory determination of the IHTC function. The def- Initially, the temperature of the muffle furnace (which
inition of an algorithm that is capable of optimizing an contains the crucible) is defined, which must be equal to
IHTC function for a more reliable simulation of a real part, the pouring temperature (Tv ). When the material is in the
predicting defects, and improving component design is furnace long enough such that its temperature equals Tv ,
extremely relevant for the industry, with direct application the crucible is removed and the liquid metal is poured into
in the development of better castings. the ingot mold. Simultaneously, the resistors of the direc-
tional furnace are connected to keep the material at the
pouring temperature. This control is performed with the aid
Experimental Procedure of thermocouples immersed in the liquid.

In this Section, the experimental procedure that originated Thus, when the internal temperature of the liquid is ade-
the thermal profiles that will be used in this work will be quate, the resistors of the directional device are turned off
presented. The experimental description to be mentioned and the refrigerant flow is activated, cooling the system. The
here was not performed by the authors of this work. cooling process must not be interrupted until the temperature
Despite that, for the reader’s full comprehension of this of the ingot is at room temperature, finishing the procedure.
paper, a brief elucidation of the main experimental steps After the experiment, the thermal profiles are stored so that
will be made. For further details, please consult.38,39 they serve as likelihood information in an inverse approach.
This approach will be presented in more detail in the section
To obtain the thermal distribution in a permanent mold about Markov Chain Monte Carlo. The thermophysical
experiment using a vertical upward unidirectional device, it is properties of the alloy in question, as well as the operational
important to first determine the necessary mass of each solidification parameters, can be observed in Table 1.
chemical element based on the ingot mold volume, the
specific mass of each element, and the solute percentage. In It is not possible to identify in the experimental work the
this work, the Al–7wt%Si alloy was chosen. Subsequently, thickness of the alumina layer, nor the frequency of data
these chemical elements must be accurately weighed. After acquisition. For this work, the positions of the
this stage, the directional solidification equipment is prepared.

First, the metal/mold interface must be polished to reduce Table 1. Thermophysical Properties of Al–7wt%Si and
roughness. After polishing, the metal/mold interface, as Experimental Parameters38,39
well as the ingot mold, are coated with alumina to facilitate Properties/parameters Symbol/units Al– SAE 310
demolding. For this experiment, an SAE 310 stainless steel 7wt%Si steel
interface with a 4 mm thickness was used. In Figure 1 is
presented an illustration of the solidification device under Thermal conductivity ks ½W m1 k1  90 33
analysis. (solid)
kl ½W m1 K1  90 –
(liquid)
Specific heat cs ½J kg1 K1  963 528
(solid)
cl ½J kg1 K1  963 –
(liquid)
Specific mass qs ½kg=m3  1080 8020
(solid)
ql ½kg m3  1080 –
(liquid)
Latent heat of fusion L½J kg1  397440 –
Solidus temperature Ts ½ C 557 –
Liquidus Tliq ½ C 610 –
temperature
Pouring temperature Tv ½ C 612 –
Figure 1. Illustration of an upward unidirectional solid-
ification device.

International Journal of Metalcasting

thermocouples at 4 mm, 8 mm, and 12 mm away from the where L is the latent heat coefficient and fs the local solid
metal/mold interface were considered. fraction. To estimate fs in the mushy zone, Scheil’s rule
was used (which is well used as segregation model ahead
phase-transformation interface), given by Eqn. 7,
Mathematical and Numerical Descriptions   1
for the Forward Problem Tf  T k0 1
fs ¼ 1  Eqn: 7
Tf  Tliq
In this section, an introduction is done to present the heat
where Tf represents the melting temperature of the pure
problem in analysis. We consider a numerical model based
solvent, Tliq the liquidus temperature of the metal alloy and
on the Finite Volume Method. In this technique, the
k0 the partition coefficient. The derivative ofots in Eqn. 6 may
domain is divided into a set of interconnected volumes,
be computed as a function of oT ot using a pseudo specific
and, afterward, the conservation law is applied to each
heat,
volume. More details of this particular mathematical and
numerical approach can be seen in another work.40 ofs
Eqn: 8
cp ¼ cm  L
oT
To model the phenomenon of solidification consider a where cm is the specific heat at mushy zone, which is given
point x in an arbitrary volume in 3D space V 2 R3 . For any by a simple mixing law as,
time t, the energy conservation equation in differential can
be written as, cm ¼ fs cs þ ð1  fs Þcl Eqn: 9
where cs and cl are the specific heats of solid and liquid
oT
qðTÞcðTÞ  r  ðkðTÞrTÞ ¼ Q_ Eqn: 1 zones, respectively. Thermal conductivity and density are
ot given by,
where T  Tðx; tÞ represents temperature, x is a spatial
k ¼fs ks þ ð1  fs Þkl Eqn: 10
vector, t is time, qðTÞ represents the material density, c(T)
the specific heat and k(T) is the thermal conductivity. These q ¼fs qs þ ð1  fs Þql Eqn: 11
variables are temperature dependent. Supposing an upward
where ks and kl are thermal conductivity of solid and liquid
solidification device, the boundary conditions for the
zones, as well as qs and ql are density of solid and liquid
mathematical approach are: the lateral walls of the mold
zones, respectively.
are thermal insulated, and the heat is extracted only from
the lower boundary so that solidification occurs only in the All thermophysical properties from the alloy are consid-
vertically upward direction. The advective effect was ered constant. The experimental data related to the thermal
neglected once it is weak in upward solidification.41 So this profile of Al–7wt%Si alloy, as well as thermophysical
problem is applied can be applied to one dimensional properties considered in this paper, can be found in
region, x  x 2 ½0; L, L is domain length, and Eqn. 1 can experimental work.39
be set as
  In the present work, the heat transfer coefficient U0 , shown
oT o oT in Eqn. 3,4,5, is obtained using an inverse method. This
qðTÞcðTÞ  kðTÞ ¼ Q_ Eqn: 2
ot ox ox coefficient is represented in this paper using the parametric
The initial and boundary conditions to this problem can be model expressed in Eqn. 12. This IHTC was set once it is
set as one of the most applied non-constant model of interface
heat transfer coefficient (IHTC),40,42–44 extensively studied
Tðx; t ¼ 0Þ ¼T0 ðxÞ Eqn: 3 and referenced in various metal systems (Al–Si,45 Bi–Sb,46

oT  Sn–Pb47 etc) and in various types of experiments. In this
k  ¼U0 ðTðx ¼ 0; tÞ  T1 Þ Eqn: 4 equation, t representstime and t0 a referential time (t0 = 1
ox x¼0;t
 s) and A Wm2 K1 and B ½ constants. In this paper,
oT  UL ¼ 50 Wm2 K1 .
k  ¼UL ðTðx ¼ L; tÞ  T1 Þ Eqn: 5
ox x¼L;t
 B
T0 ðxÞ is the initial temperature distribution, UL and U0 are t
U 0 ¼ hi ¼ A Eqn: 12
overall heat transfer coefficient between ingot and vicinity, t0
which is at temperature T1 . In Eqn. 2, Q_ is the energy
The mathematical modeling shown in this section and the
generated or consumed due to the energy sources. Here, the
use of the parametric IHTC model expressed in Eqn. 12
contribution of the internal energy Q_ is only by latent heat,
does not restrict the ability of the optimizer to be selected
which is defined as,
to optimize other geometries and parametric IHTC models
ofs that may better meet the needs of other researchers. The
Q_ ¼ qL Eqn: 6
ot generality of the work is based on the use of a simple

International Journal of Metalcasting

geometry and an IHTC model widely used in the literature.
6. Moth-Flame Optimization: The MFO algorithm was
It is assumed that the validity of the work extends to other
conceived by Mirjalili16, which was inspired by
parametric models of IHTC since the optimization
the navigation method of moths in nature called
topology should not degenerate with this change. transverse orientation;
It is important to highlight that there is no single model of 7. Sine Cosine Algorithm: The SCA is a population-
IHTC in the literature that is a consensus. As previously based optimizer17 that uses a mathematical
commented, most works use temporal parametric models to model based on sine and cosine functions for
represent this thermal coefficient. Among those that do not solving optimization problems. This method is
use temporal models, the constant IHTC is the majority. For easy to implement because it uses only two
example, in a study48 for the solidification of an ingot search functions, which are changeable accord-
weighing more than 10 tons. In this, as well as in this work, ing to a random number;
they considered the IHTC model as an effective heat transfer 8. Whale Optimization Algorithm: The WOA was
coefficient, since the determination of each contact resis- developed by Mirjalili and Lewis.18 This opti-
tance is onerous. For this, the authors used, in most of the mizer mimics the social behavior of humpback
geometry, piecewise constant time IHTCs. Likewise, a whales;
study49 employed, in several regions of the casting geometry, 9. Dragonfly Algorithm: The DA was created by
piecewise constant time IHTCs to represent the complex Mirjalili.19 This method is inspired on the static
thermal interaction between the casting and the mold. and dynamic swarming behaviors of dragonflies
in nature;
Finally, it is also possible to model the IHTC using ther- 10. Harris Hawks Optimization: The HHO algorithm
mophysical properties, such as thermal conductivity, or was developed by Heidari et al.,20 which were
even thermodynamic properties, such as pressure and inspired by the behavior and chasing style of
temperature. In this way the IHTC model has a direct Harris hawks in nature called ‘‘surprise
dependence of material properties and process temperature. pounce’’.
A discussion about this topic is beyond the scope of this
paper. For further information please refer to previous
Markov Chain Monte Carlo
works.7,31,50,51
In this section, the Markov Chain Monte Carlo (MCMC)
Metaheuristic Optimization Algorithms method is presented, which will be used as a reference for
comparing the results of the metaheuristic methods pre-
This section presents a brief description of the ten meta- sented in the previous section. The MCMC result is
heuristic methods applied in this paper. A more detailed obtained within a Bayesian inference approach. The
description for each methods, including their algorithms, is Bayesian statistics approach allows the coupling of all
shown in Appendix A: available information to reduce the uncertainty on infer-
ence or decision-making problems.
1. Particle Swarm Optimization: The PSO algorithm
was developed by Kennedy and Eberhart11 As stated54 the statistical inversion approach is based on
which was inspired by the observation of natural the following principles:
clusters, such as migratory behavior of fish and
birds;52 1. All variables in the model are modeled as random
2. Differential Evolution: The DE algorithm is a vector- variables;
based algorithm created by Storn and Price12,53 to 2. The randomness describes the degree of infor-
minimize nonlinear and non-differentiable con- mation concerning their realizations;
tinuous space functions. This method consists in 3. The degree of information concerning these
mutation, crossover and selection; values is coded in probability distributions;
3. Bat Algorithm: The BA algorithm was created by 4. The solution of the inverse problem is the
Yang.13 This method is inspired by bat’s behav- posterior probability distribution.
ior, especially by their echolocation ability; The inverse problem solution is reformulated in the form of
4. Flower Pollination Algorithm: The FPA was created statistical inference from the posterior probability density
by Yang14, based on the behavior of the flow PðhjXÞ, which is the model for the conditional probability
pollination process of flowering plants; distribution of the unknown parameters h given the
5. Grey Wolf Optimizer: The GWO was conceived by measurements X. The measurement model (function) that
Mirjalili et al.15, which were inspired on hunting incorporates the measurement error model and related
technique and the social hierarchy of grey uncertainties is called likelihood PðXjhÞ, that is, the
wolves; conditional probability of the measurements given the

International Journal of Metalcasting

unknown parameters. The model for the unknowns that Definition of Search Parameters for MCMC
reflects all the uncertainty of the parameters without any and MOAs
extra information given by the measurements is called the
prior model PðhÞ.55 In Mathematical Description section, the mathematical
Based on Bayes rule, the probability density function of h modeling and numerical discretization used to obtain the
given X can be written as in Eqn. 13, direct model for the solidification phenomenon in question
were presented. After, in Algorithms’ section, the meta-
PðXjhÞ PðhÞ heuristic optimization algorithms were presented. These
PðhjX Þ ¼ Eqn: 13 optimizers will be evaluated based on the statistical
PðXÞ
R response given by the MCMC method, which was
where PðX Þ ¼ H PðXjhÞ PðhÞdh is a normalizing factor.56 addressed in previous section. In this section, the operating
However, once PðX Þ is a constant, this information may be conditions for each metaheuristic method and for the
temporarily disregarded in order to obtain a posterior MCMC method will be presented.
probability density function as Eqn. 14,
For performance analysis of the metaheuristic algorithms,
PðhjX Þ / PðXjhÞ PðhÞ Eqn: 14
we used a similar approach that has been used before59,60
Assuming that the parameters h have a prior Gaussian and set the following parameters:
distribution with mean lh and covariance matrix V, PðhÞ
can be written as in Eqn. 15,57 • Search range of parameter
Nvar 12
A : ½0; 10000 Wm2 K1 ;
PðhÞ ¼ ð2pÞ jVj 2
• Search range of parameter B : ½0:5; 0:005;
 
1  T 1   Eqn: 15 • Maximum number of iterations on each optimiza-
exp  h  lh V h  l h tion: 100;
2
• Number of particles: 20;
where Nvar is the number of parameters. • Stop criterion: 10 iterations resulting on the same
As in Eqn. 15, assuming that the measurement errors are best parameter particle;
Gaussian, with zero mean, the covariance matrix W is • Number of independent runs: 40.
known, so the fact that the errors are additive and inde- Table 2 contains the parameters intrinsic to each meta-
pendent of the vector of parameters h, the likelihood heuristic algorithm considered in this research.
function can be expressed as,57 The metrics to be analyzed in this contribution, for meta-
Nvar 1
heuristic methods, will be: Sample mean, standard devia-
PðXjhÞ ¼ð2pÞ 2 jWj2 tion and kurtosis of the optimal points, the average number
  of iterations, convergence, error analysis, and probability
1
exp  ½X  YðhÞT W 1 ½X  YðhÞ of success (probability of a method provides a solution
2
vector that is in the posterior probability density of the
Eqn: 16 MCMC method). The objective function F ðhÞ used in this
where YðhÞ is the numerical result of the direct model paper is the sum of the standard deviation between the
proposed on section about Mathematical Description. simulated curves obtained by the direct model presented on
Mathematical section and experimental thermal data of all
The Markov Chain Monte Carlo method (MCMC) bases
thermocouples. So, considering X the measurement data
itself on the fact that the probability density function built
and Y ðhÞ the simulated curves, the fitness function is
by the Markov Chain process converges on the actual
expressed as in Eqn. 17,
distribution as the sample size increases.58 This method
estimates the referred posterior probability density using (P )0:5
Nmeas
sampling and rejection techniques. One of the most used i¼1 ½Xi  Yi ðhÞ2
strategies for implementing the Monte Carlo technique F ð hÞ ¼ Eqn: 17
Nmeas
with Markov chains is the Metropolis-Hasting algorithm.
The Algorithm on Appendix A, expresses the iterative where Nmeas stands for the number of experimental
process of the MCMC method considered in this work. On measurements.
Figure 2 is shown an illustration of this process. Regarding the MCMC method, which will be the reference
method in this paper, the parameters were considered as follows:
This method will be used to infer the PPD of the interface
parameters of the temporal model exposed in Eqn. 12. • Reference starting point for parameter A: 6430
After obtaining the statistical metrics for both parameters, Wm2 K1 ;40
these metrics will be considered as a reference to evaluate • Reference starting point for parameter
the optimization performance of the metaheuristic B : 0:153;40
methods.

International Journal of Metalcasting

Table 2. Parameter Values Used on the Metaheuristic
Algorithms

Method Parameters References

Particle Swarm Acceleration constants 11

Optimization aPSO and bPSO ¼ 2
Inertia constant hPSO = 1
Differential Differential weights 12
1 2
Evolution FDE ¼ FDE ¼ 0:5
Crossover parameter
CrDE ¼ 0:8
Bat Algorithm Loudness uBA ¼ 0:25 13
Pulse rate rBA ¼ 0:5
Frequency fBA 2 ½0; 2
Constants cBA and aBA =
0.9
Scaling factor rBA = 1
Figure 2. Illustration of Metropolis-Hasting algorithm.
Flower Pollination Switch probability 14
Algorithm pFPA ¼ 0:8
Lévy flight parameter Results and Discussion
kFPA ¼ 1:5
Scaling factor cFPA ¼ 0:1 In this section, the results from the MCMC method will be
Grey Wolf – 15
exposed, as well as the results of the metaheuristic methods
Optimizer and the comparison between them.
Moth-Flame Spiral shape parameter 16
Optimization bMFO ¼ 1
Results via Markov Chain Monte Carlo
Sine Cosine Weighting factor aSCA ¼ 2 17
Algorithm
The posterior probability distribution obtained by MCMC
Whale Spiral shape parameter 18 of parameters A and B can be seen in Figure 4a. The
Optimization bWOA ¼ 1
267000 states are represented, in this figure, by 20 bins,
Algorithm
based on Doane’s formula.61 Presenting a positive excess-
Dragonfly Lévy flight parameter 19
kurtosis and skewness of 0.03 each, these histograms are
Algorithm kDA ¼ 1:5
leptokurtic and approximately symmetrical.
Harris Hawks Lévy flight parameter 20
Optimization kHHO ¼ 1:5
Based on Figure 4a, the parameter value that minimizes the
model error with experimental data, i.e., has the highest
posterior probability value, is between 6200 and
6400 Wm2 K1 for parameter A and between 0:15 and
• Number of Markov Chains analyzed: 7;
0:14 for parameter B. The expected value, such as per-
• Markov Chain starting points: 40%, 60%, 80%,
centile of 95%, 5% and standard deviation of the Markov
100%, 120%, 140% and 160% of the reference
Chain states are show in Table 3.
starting point for each parameter under analysis
(Figure 3a);
In Figure 4b is shown the numerical thermal profile
• Number of states for each Markov Chain: 40000;
resulted by the expected value of A and B obtained by
• Search Step (wSS ): 0.005;
MCMC. The small distance between the experimental and
• Standard deviation considered for experimental
numerical curves corroborates the accuracy of the values
measurements: 5 K.
raised for the heat exchange interface parameters. Applying
After removing the burn-in samples, as can be seen in the expected value obtained by MCMC in the simulation,
Figure 3a, the Markov chains were concatenated and the standard deviation between the experimental and sim-
analyzed to obtain the expected value, standard deviation, ulated thermal profile for the thermocouples in 4 mm, 8
percentiles 95% and 5% of each parameter under analysis. mm, and 12 mm were, respectively, 4.05 K, 3.72 K, and
The concatenation of Markov chains, resulting in 267000 2.82 K.
states, can be seen in Figure 3b.

International Journal of Metalcasting

 (a) (b)
12000
Parameter A [W/m2 K]

10000

8000

6000

4000

2000
0 2000 4000 6000 8000 10000
Number of States
-0.05
Parameter B [-]

-0.1

-0.15

-0.2

-0.25

-0.3
0 2000 4000 6000 8000 10000
Number of States

Figure 3. Markov chains (a) initial states: Burn-in regions and; (b) concatenation for the two parameters.

(a) (b)
0.2 0.2 1000
Thermocouple 4 mm
Thermocouple 8 mm
Thermocouple 12 mm
MCMC Expected Value
900
Posterior Probability Distribution

0.15 0.15
Temperature [K]

800

0.1 0.1

700

0.05 0.05
600

0 0 500
5800 6000 6200 6400 6600 6800 -0.17 -0.16 -0.15 -0.14 -0.13 -0.12 0 20 40 60 80 100
Parameter A [W/m2 K] Parameter B [-] Time [s]

Figure 4. (a) Histograms obtained by MCMC and; (b) Experimental38 and simulated thermal profiles of Al–7wt%Si
solidification.

Table 3. Expected Value, Standard Deviation, 95% and 5% Percentile Values of the Parameters Obtained by MCMC

Parameter Expected value Standard deviation MCMC percentile 5% MCMC percentile 95%
1
A [W m2 K ] 6301 91 6126 6476
B [-] -0.147 0.004 -0.156 -0.139

Performance Analysis of Metaheuristic Friedman Ranking Test, Iterations, Probabilities and Con-
Algorithms vergence, and Overall Performance.

In this section, a performance analysis of all metaheuristic An evaluation table of all the information present in this
algorithms shown is presented. To show the results more section can be seen in Table 4. The quantitative data were
coherently, this section presents four subsections: Sample represented qualitatively based on parameters derived both
Mean, Percentiles and Kurtosis, Error Analysis and from the information obtained by the MCMC method and
the performance of each metaheuristic algorithm. In this

International Journal of Metalcasting

table, ‘‘Relative Performance’’ is a qualitative comparison uncertainty range obtained by the MCMC method. It is
of the results (sample mean, kurtosis, and standard devia- observed that in a range from 0 to 10000 Wm2 K1 , the
tion) between the metaheuristic methods and MCMC. results of all methods are consistent, indicating that any
‘‘Convergence’’ represents how the method was able to optimizer out of ten can retrieve the value of this param-
search the best minimum on the domain without being eter. However, from the chart:
trapped on a local one, taking into account the average
number of iterations and the probability of success. • BA, FPA, and WOA presented a sample mean
‘‘Average Error’’ is a quantitative evaluation of the average outside the uncertainty range of the MCMC
error of the optimized values of each method under anal- method. Based on kurtosis, the optimizers do not
ysis, considering expected value, standard deviation, min- recover exactly and precisely the expected value
imum and maximum values. ‘‘Overall Performance’’ of reference. Even though the uncertainty range of
indicates a subjective evaluation of the authors about the each method covers part or all of the MCMC
results available during this research, taking into account uncertainty range, the probability that these
all quantitative metrics analyzed. At the end of the table, methods will optimize the parameter to the exact
the ‘‘Friedman Test’’ shows the results of the non-para- range is quite small;
metric Friedman Ranking Test for all ten metaheuristic • PSO, GWO, SCA, and HHO presented a sample
optimizers on analysis. mean within the MCMC uncertainty range, but
exposed a wide range compared to the other
methods, covering a region of approximately
Sample Mean, Percentiles and Kurtosis 2000 Wm2 K1 ;
• Among the analyzed methods, DE, MFO, and DA
In Figure 5, it is shown the results related to the studied stand out. The sample mean of these methods is
optimizers. In this, the methods on the horizontal axis of quite consistent with the reference range and the
the charts are in chronological sequence with respect to standard deviation comprises a small range of
their publication. The horizontal lines represent the results values. It is noteworthy the ability of MFO to
obtained by MCMC. The circles represent the sample mean recover, in a small uncertainty range and very
of the optimal values of each method, along with the high kurtosis, the optimal value of parameter A.
standard deviation bars above and below that point. The In Figure 5b is expressed the same quantitative parameters
second graphic shows the kurtosis of each method, in of statistical analysis as shown in Figure 5a, but relative to
which the yellow horizontal line on the blue bars indicates parameter B.
the value 3, that guides the argument about tail size of the As expressed on parameter A analysis, at different levels,
distribution presented in each method. all methods were able to estimate the value of parameter B,
considering the MCMC uncertainty range. In the interval
It is clear, based on the uncertainty range of each method, from -0.005 to -0.5, the results of all optimizers are con-
that all methods were able, at different coherence levels, to sistent. However, the methods had difficulty in optimizing
estimate the value of parameter A, considering the

(a) (b)
MCMC Expected Value MCMC Percentile 95% MCMC Percentile 5% MCMC Expected Value MCMC Percentile 95% MCMC Percentile 5%

8000 -0.08

7500 -0.1
Parameter A [W/m2 K]

Parameter B [-]

7000 -0.12

6500 -0.14

6000 -0.16

5500 -0.18

5000 -0.2

4500 -0.22
PSO DE BA FPA GWO MFO SCA WOA DA HHO PSO DE BA FPA GWO MFO SCA WOA DA HHO

Kurtosis Reference = 3 Kurtosis Reference = 3

25 25
20 20
Kurtosis

Kurtosis

15 15
10 10
5 5
0 0
PSO DE BA FPA GWO MFO SCA WOA DA HHO PSO DE BA FPA GWO MFO SCA WOA DA HHO
Metaheuristic Methods Metaheuristic Methods

Figure 5. Uncertainty range and kurtosis of the metaheuristic methods and MCMC, for parameter (a) A and (b) B.

International Journal of Metalcasting

 Table 4. Comparative Performance Analysis of the Metaheuristic Optimizers

Methods Relative performancea Convergenceb Average errorc Overall performanced Friedman ranke

PSO Good Good Low Good 2

DE Excellent Good Medium Good 5
BA Regular Premature High Regular 9
FPA Regular Premature High Regular 10
GWO Good Premature Low Regular 3
MFO Excellent Satisfactory Low Excellent 1
SCA Regular Good Medium Regular 7
WOA Regular Premature Low Regular 4
DA Good Good Medium Good 6
HHO Good Good High Regular 8
In this work:
a
‘ Relative Performance’’ is classified as ‘ Regular’’, ‘ Good’’ or ‘ Excellent’’.
b
‘ Convergence’’ is classified as ‘ Premature’’, ‘ Good’’ or ‘ Satisfactory’’.
c
‘ Average Error’’ is classified as ‘ Low’’, ‘ Medium’’ or ‘ High’’.
d
‘ Overall Performance’’ is classified as ‘ Regular’’, ‘ Good’’ or ‘ Excellent’’.
e
‘ Friedman Rank’’ goes from ‘ 1’’ to ‘ 10’’, where ‘ 1’’ represents the best and ‘ 10’’ the worst.

parameter B, which can be seen in the figure above. From Table 5 is shown the abovementioned metrics for each
the chart it is possible to visualize that: metaheuristic method analyzed.

• BA, FPA, SCA, and DA presented a sample mean Considering the E½F ðG Þ, it is possible to comment that:
out of the uncertainty range of MCMC method.
• PSO, GWO, WOA, and HHO presented a sample • MFO, PSO, GWO, and WOA, in this order,
mean inside or, at least, almost within MCMC presented very good results, outperforming the
uncertainty range, but exposed a very wide range other methods in these metrics. The data show
compared to other methods; that, on average, the optimization of these meth-
• DE and MFO stand out. The consistency of the ods was quite satisfactory, reaching low error
sample mean in relation to the reference range and values and presenting a very low E½F ðG Þ,
the small standard deviation corroborate the sometimes being less than the error obtained by
excellent result of these methods. DE did not the expected value of MCMC method;
present high kurtosis value, but its performance is
satisfactory. It is noteworthy the performance of
MFO, which presented an almost exact sample
mean and reaches a kurtosis value above 20. This
Table 5. Expected Value, Standard Deviation, Maximum
indicates that MFO, besides being accurate, is and Minimum Values of Error for the Metaheuristic
extremely precise. Algorithms Under Analysis

Methods E½F ðG Þ rðF ðG ÞÞ maxðF ðG ÞÞ minðF ðG ÞÞ

Error Analysis and Friedman Ranking Test
PSO 10.47 0.34 12.07 10.31
In this subsection, the optimization errors of the meta- DE 11.19 0.79 13.40 10.34
heuristic algorithms under study will be analyzed, as well BA 25.76 18.05 78.89 10.66
as the non-parametric Friedman Ranking Test. Consider by FPA 24.02 16.15 84.36 10.80
notation, as exposed in previous Section, that F ðhÞ is the
GWO 10.52 0.15 10.91 10.32
fitness (objective) function and G is the vector that cor-
MFO 10.34 0.05 10.60 10.31
responds to the optimized values of each method. F ðG Þ
stands for the error of each optimized parameter vector. SCA 11.75 1.31 14.75 10.41
Four metrics will be considered in terms of error: Expected WOA 10.71 0.61 13.19 10.32
value ðE½F ðG ÞÞ, standard deviation ðrðF ðG ÞÞÞ and DA 11.44 1.55 17.14 10.39
maximum ðmaxðF ðG ÞÞÞ and minimum ðminðF ðG ÞÞÞ HHO 12.26 3.26 29.41 10.42
values. First, the error analysis will be presented. In

International Journal of Metalcasting

 Table 6. Friedman Ranking Test for All Ten Metaheuristic Optimizers in Analysis Based on 40 Optimization Runs

Runs PSO DE BA FPA GWO MFO SCA HHO DA HHO

1 1 8 6 10 3 2 9 4 7 5
2 1 5 6 8 3 2 9 4 10 7
3 3 8 5 10 1 4 7 2 9 6
4 6 3 10 9 5 2 8 1 7 4
5 1 7 10 9 3 2 8 5 4 6
6 2 3 10 9 4 1 8 5 6 7
7 2 6 8 10 5 1 9 3 4 7
8 2 6 5 10 4 1 8 3 7 9
9 4 3 7 10 6 1 8 2 9 5
10 3 7 9 10 5 1 4 1 8 6
11 8 5 10 7 3 1 2 4 9 6
12 1 4 7 10 5 3 8 2 9 6
13 2 4 10 9 6 1 5 8 3 7
14 3 7 10 9 5 1 8 2 6 4
15 2 7 9 10 5 1 8 3 4 6
16 1 6 10 9 3 2 5 4 7 8
17 2 6 10 9 3 1 8 4 5 7
18 2 5 8 9 4 1 10 3 6 7
19 4 9 6 10 2 1 8 7 3 5
20 2 7 9 10 5 1 3 4 6 8
21 1 5 9 10 7 2 8 6 3 4
22 1 6 10 9 3 2 4 7 8 5
23 4 6 8 9 3 1 7 2 5 10
24 4 3 9 10 5 1 6 2 7 8
25 5 7 10 9 3 1 2 4 6 8
26 2 6 8 10 3 1 5 4 9 7
27 2 5 10 6 4 1 3 8 9 7
28 2 6 10 8 4 1 5 3 7 9
29 1 4 9 10 3 2 7 8 5 6
30 2 7 10 9 4 1 6 5 3 8
31 1 9 8 10 2 6 3 5 4 7
32 2 9 10 6 3 1 8 5 4 7
33 1 5 9 10 6 2 7 3 4 8
34 2 6 9 10 4 1 8 5 3 7
35 4 9 5 10 2 1 6 3 7 8
36 2 9 10 8 4 1 6 7 5 3
37 1 6 10 9 5 2 7 4 3 8
38 3 5 10 9 2 1 8 6 4 7
39 2 3 10 9 7 1 4 6 8 5
40 2 6 10 9 3 1 7 4 5 8
Sum 96 238 349 367 157 60 260 168 238 266
Rank 2 5 9 10 3 1 7 4 5 8

International Journal of Metalcasting

 Average Number of Iterations For the analysis of the Friedman Ranking Test, Table 6
40
shows for each simulation the ascending performance
30 ranking. The results from this test are intelligible and
20
straightforward. For each independent optimization, each
optimizer will present an optimal value. In this way, the
10 optimizers are ranked by performance with each opti-
mization. The best optimizer receives the number 1, the
0
PSO DE BA FPA GWO MFO SCA WOA DA HHO second-best the number 2, and so on. At the end of the
1 evaluation, all optimizer points are added. Whoever gets
Probabibity of Success

0.8 the fewest points is considered the best optimizer for the
problem in question. In this rank, the test statistic Q ¼ 245
0.6
and pvalue  0:001.
0.4

0.2 It can be seen that the ranking above coincides almost

0 entirely with the error analysis considering E½F ðG Þ. The
PSO DE BA FPA GWO MFO SCA WOA DA HHO optimizers MFO and PSO, in that order, were the two best
Metaheuristic Methods
algorithms for this problem. Next, the GWO and WOA
Figure 6. Average number of iterations and probability methods were also suitable, even though they had a much
of success for each optimization method. lower capacity than the first two. After, DE and DA
methods showed considerable optimization capacity, but
• DE showed good result, showing low E½F ðG Þ. they did not present such satisfactory results. In the end, the
However, it was almost 10% higher than the best SCA, HHO, BA, and FPA methods, in this order, showed
result of this parameter (MFO), reducing its the worst optimization capabilities in this problem.
relative efficiency;
• DA, SCA, HHO, FPA, and BA, in that order, It is worth pointing out that Friedman Ranking Test only
presented unsatisfactory results in these metrics. considers the final value of the optimization, disregarding
In comparison with better performance methods, data average variability. That is why the overall perfor-
these presented very high E½F ðG Þ, being 1.1 to mance presented in Table 4 presents a slight divergence
2.5 times greater than the lowest one (MFO). from the ranking shown on Table 6. Even though the
optimizers GWO and WOA were the third and fourth-best
Based on rðF ðG ÞÞ, maxðF ðG ÞÞ and minðF ðG ÞÞ, it is
methods, respectively, they were considered regular
possible to argue that:
behavior by the authors. This assessment was mainly
attributed to the low probability of success added to the
• MFO and GWO presented very small maximum
average number of average iterations shown by the meth-
intervals, indicating high precision. The two
ods, which will be shown in the next section.
methods remained in the error range from 10 to
11, providing the best results compared to all
In addition, it is worth pointing out that one method that
algorithms. The small standard deviation value
has the best performance compared to other methods does
presented by MFO is highlighted which, corrob-
not necessarily classify it as a satisfactory method. Thus, a
orated by the high kurtosis presented before,
method may be classified as the best, but its result is reg-
indicates that this method is the most accurate and
ular. In this work, a method is considered satisfactory
precise of all;
based on its performance in comparison with MCMC
• PSO and WOA showed good results such as small
posterior probability distribution.
maximum interval and standard deviation. How-
ever, the maximum value presented by the two
methods is at least 10 % higher than the lowest
Iterations, Probabilities and Convergence
maximum values found (MFO);
• DE, SCA, DA, HHO, BA, and FPA did not
present satisfactory results. In addition to high Figure 6 shows the average number of iterations and
standard deviation, their maximum value is 1.25 probability of success of each method until the stopping
to 8 times greater than the best maximum value criterion is achieved. As stated previously, the stopping
(MFO). criterion considered in this paper is the repetition of the
same parameter vector for ten iterations. It is worth
It is important to note that, due to the results of pointing out that, for all MOAs presented in this paper, an
minðF ðG ÞÞ, it is observed that at least in one optimiza- iteration calls the same number of times the direct model.
tion, all methods were able to satisfactorily minimize the
parameters of the IHTC under study. It is possible to note from results that:

International Journal of Metalcasting

 • BA, FPA, GWO, and WOA presented premature method did not show premature convergence, showing a
convergence, based on low/medium average num- good probability of success. However, the error analysis
ber of iterations and low probability of success (\ did not obtain satisfactory results like other methods. In
50%); general, the algorithm showed acceptable convergence and
• PSO, DE, SCA, DA, and HHO presented good results. Thus, DE represents a good method to optimize this
convergence given the probability of success kind of physical problem, presented as the fifth-best opti-
between 50 and 70%. Even with similar conver- mizer based on Friedman Ranking Test.
gence quality, these models presented a very
different average of iterations. PSO showed a high Bat Algorithm (BA) showed regular relative performance
number of iterations, followed by DE and, finally, in both parameters under analysis. The uncertainty interval
SCA, DA, and HHO, which iterated little. There is produced by the method was quite inconsistent with the
great potential for the latter methods to optimize reference. This may be due to a very premature conver-
this physical problem. However, they present gence, based on the average number of iterations, which
some difficulty in the proceeding, probably has a value very close to the minimum number of accept-
because they are stuck in local minimums and able iterations, producing a very low probability of success.
are unable to move from them. The performance This is reflected in the error analysis, which presented
of DA is highlighted, which presented a small awful metrics. Therefore, BA has not behaved as an ade-
uncertainty interval for the two parameters under quate method for this type of inverse estimate, in this case,
analysis and presented the second-highest proba- showing regular overall performance. By Friedman Rank-
bility of success. ing Test, BA is the second-worst method.
• Finally, MFO presented a satisfactory conver-
gence, showing an average of iterations three Flower Pollination Algorithm (FPA) also presented a reg-
times higher than the stopping criterion. This ular relative performance in both parameters under analy-
indicates that the search space was better sis. The uncertainty range presented was very inconsistent
explored. This corroborates the quantitative based on MCMC’s posterior probability distribution. FPA,
results. Even though the method required a higher like BA, suffered very premature convergence, showing a
average number of iterations, its results were very low number of iterations and probability of success. The
satisfactory, presenting a probability of success of error presented by this method was not satisfactory, pre-
85%. senting a very high expected value. That said, FPA had a
regular overall performance, being the worst optimizer
based on Friedman Ranking Test.
Overall Performance
Grey Wolf Optimizer (GWO) showed good relative per-
From the previous analysis, it is possible to conclude the formance and its uncertainty interval was consistent with
parameter estimation in each metaheuristic method studied. the reference, but the average number of iterations of this
One should keep in mind that the reference values used in method was median and the probability of success does not
this study are relative to MCMC results presented in pre- exceed 50%, indicating premature convergence, even with
vious section. good error metrics. Based on that, GWO presented a reg-
ular overall performance, being the third-best algorithm by
Particle Swarm Optimization (PSO) presented good rela- Friedman Ranking Test.
tive performance. The sample means of parameters A and
B were into the uncertainty range of reference and showed Moth-Flame Optimization (MFO) exposed excellent rela-
one of the highest excess-kurtosis for both parameters. PSO tive performance. The uncertainty range presented was
presented a high average number of iterations and a good quite consistent with the reference, both due to the prox-
probability of success, resulting in a good convergence. imity of the sample mean to the reference average and the
PSO presented the lowest minimum value of all methods small standard deviation presented. This fact is accentuated
and very good results, such as low expected value and by the kurtosis of the distribution of the optimized
standard deviation. Based on the above, this method has a parameters, higher than all the other methods. The average
good overall performance in optimizing IHTC parameters number of iterations of this method was high. However, it
in this physical problem, being the second-best on Fried- presented very satisfactory convergence, presenting a
man Ranking Test. probability of success of 85%. In error analysis, MFO stood
out. All metric values shown by this algorithm were the
Differential Evolution (DE) exposed excellent relative best ones, managing to be able to estimate as well, or even
performance in both parameters under analysis. The better, than the reference method itself. MFO presented the
uncertainty range produced by DE was consistent with the shortest maximum interval among all methods, also
reference, indicating accuracy and precision. The average standing out for the small standard deviation. So, MFO is
number of iterations was reasonable, indicating that the the best optimizer analyzed, presenting a satisfactory

International Journal of Metalcasting

overall performance and the best position on Friedman considered in this paper may change when optimizing the
Ranking Test. same physical problem, but with smoother experimental data.

Sine Cosine Algorithm (SCA) presented regular relative

performance. Its parameter uncertainty interval did not stand Conclusions
out when compared with the reference. The average number
of iterations of this method was low but exposed a proba- This paper presented a qualitative and quantitative perfor-
bility of success of 62.5%, showing good convergence. mance analysis of ten nature-inspired metaheuristic algo-
When observing the performance through error analysis, the rithms to verify which metaheuristic method excels in
algorithm presents median results, not as satisfactory as optimizing the Interfacial Heat Transfer Coefficient (IHTC)
other analyzed methods. Based on the above, SCA presented parameters in a unidirectional permanent mold casting pro-
itself as a regular method for this type of application. By cess for Al–7wt%Si alloy. For this purpose, was selected the
Friedman Ranking Test, SCA was the seventh-best one. optimizers: Particle Swarm Optimization (PSO), Differen-
tial Evolution (DE), Bat Algorithm (BA), Flower Pollination
Whale Optimization Algorithm (WOA) showed regular Algorithm (FPA), Grey Wolf Optimizer (GWO), Moth-
relative performance. The uncertainty range presented was Flame Optimization (MFO), Sine Cosine Algorithm (SCA),
quite bad compared to the reference. The standard devia- Whale Optimization Algorithm (WOA), Dragonfly Algo-
tion of the estimation of the parameters was one of the rithm (DA) and Harris Hawks Optimization (HHO).
highest of all methods. The average number of iterations of
this method was relatively high and presented a low Among the results obtained from the metaheuristic methods,
probability of success, indicating premature convergence. all of them would be able to reasonably optimize parameters
By the error analysis, this method was satisfactory, pre- A and B. However, based on qualitative and quantitative
senting low average error in several metrics. However, metrics, BA, FPA, GWO, SCA, WOA, and HHO did not
once this method presented out-range A and B parameter present significant results. In general, these methods showed
values (based on MCMC’s posterior probability distribu- premature convergence, low correlation of the optimized
tion) 80% of the times, WOA presented itself as a regular values with the reference uncertainty range, high error val-
method in this specific case. By the non-parametric rank- ues related to the optimized parameters, and low probability
ing, WOA was considered the fourth-best one. of success. Thus, these methods are not the most suitable for
optimizing parameters in this specific problem.
Dragonfly Algorithm (DA) exposed good relative perfor-
mance. The uncertainty interval presented by the method, On the other hand, PSO, DE, and DA showed reasonably
in both parameters under analysis, partially contained the good results, with good convergence (based on the average
reference interval, with a low standard deviation. The number of iterations and probability of success) and
average number of iterations of this method was small but acceptable error metrics. They presented significantly
resulted in a probability of success of 65%, which is a superior results concerning the abovementioned methods.
relatively good convergence. Given these characteristics, However, they did not provide a satisfactory error level.
DA presents good results in general, being classified at the Thus, these methods show relevant estimates for this type
end as a good method for this kind of problem, being the of physical problem, but they are not the most suitable.
sixth-best optimizer based on Friedman Ranking Test.
Unlike the others, MFO exposed promising results. It
Harris Hawks Optimization (HHO) presented good relative presented the best qualitative and quantitative results,
performance. The uncertainty range exposed by the method showing an above-average performance. In general, the
was regular, presenting a sample mean within the reference MFO sample mean was very close to the expected value of
range but a very high standard deviation. The average reference, its standard deviation was small compared to the
number of iterations of this method was small but resulted in others, and its kurtosis was high. The MFO convergence
a good probability of success. Presenting high error metrics, was satisfactory, with a reasonable average number of
based on other methods’ behavior, HHO presented itself as a iterations and a probability of success of 85%. Its error
regular method in this optimization application, considered level was satisfactory, classifying itself as an excellent
the third-worst optimizer based on Friedman Ranking Test. method for this type of problem and application.

It is worth pointing out that some metaheuristic methods As MFO outperformed the other methods in different
studied had difficulty in optimizing the parameters contained qualitative and quantitative parameters, it has total poten-
in the IHTC because the experimental data considered in this tial to optimize the IHTC parameters A and B coupled in
study are not smooth, presenting disturbances that create the solidification phenomenon applied in the permanent
several local minima near the optimal point. This causes mold casting without loss of accuracy and precision. It is
premature convergence by creating dense local minimum worth pointing out that the selected values for the opti-
zones. Therefore, the behavior of the metaheuristic methods mizer’s intrinsic parameters highly influence their

International Journal of Metalcasting

performances. Further research based on these parameters Data Availability
is encouraged to expand the applicability and efficiency of
these algorithms in engineering and metallurgy applica- The experimental results necessary to reproduce all
tions, such as the casting problem studied in this work. analyses are provided in the paper.

Conflict of interest The authors declare that they have no conflict

Finally, the MFO method can be of great value when it is of interest.
necessary to identify the most suitable prior for feeding
Bayesian techniques, such as Approximate Bayesian Appendix A Algorithms Detailed Description
Computation (ABC) and MCMC. Due to the low compu-
tational cost of this type of optimizer and the great Particle Swarm Optimization
numerical exploration power, relevant estimates can be
raised when there is no good knowledge about the In this algorithm, consider X ¼ xti and V ¼ vti , respectively,
parameters under study. the position and velocity for particle i at generation or
iteration t. The particles xt and g represent the best
Acknowledgements position on the iteration t and the best ones obtained so far
on the simulation, respectively. The velocity upgrade is
The authors would like to thank the CAPES and determined by Eqn. A1,where hPSO represents the inertia
FAPESPA for financial support by research grant constant, aPSO and bPSO the acceleration constants and u1
88881.159113/2017-1. Also, the authors are thankful and u2 random numbers uniformly distributed, such that
to CNPq and PROPESP - UFPA. uh U ½0; 1. The initial velocity of the particle can be zero,
that is, v0i ¼ 0. The new particle position is updated as
Author Contributions G. de M.Stieven contributed to Investigation, Eqn. A2.
Methodology, Programming, Validation, Formal Analysis, Writing:
Original Draft, Review and Editing. D.S. Stieven contributed to    
Investigation, Conceptualization, Methodology, Validation, Formal vtþ1
i ¼hPSO vti þ aPSO u1 g  xti þ bPSO u2 xt  xti
Analysis Writing: Review and Editing. E.P. Oliveira contributed to Eqn: A1
Investigation, Conceptualization, Formal Analysis, Supervision,
Writing: Original Draft. E.F. Lins contributed to Investigation, Con- xtþ1
i ¼xti þ vtþ1
i Eqn: A2
ceptualization, Formal Analysis, Supervision, Writing: Original Draft,
Review. In Algorithm 1 a pseudo-code of this method is presented.
The mathematical explanation of the equations below was
Funding taken from reference52.
No funding was received to assist with the preparation
of this manuscript. Differential Evolution

First, for a d-dimensional problem with d parameters, a

population of n solution vectors are initially generated. So,

Algorithm 1 Pseudo-code of PSO algorithm52

Objective function F (X) , X = xti
Initialize locations X and velocity V of npart , where npart is the number of
particles
Find g∗ for min[F(X)] (at t = 0)
while (criterion) do
for each search agent do
Generate new velocity vit+1 using Eq. A1
Calculate new locations using Eq. A2
Evaluate objective functions at new locations xt+1
i
Find the current best for each particle xt∗
end for
end for
Find the current global best g∗
end while
end while
Output the final results xt∗ and g∗

International Journal of Metalcasting

for i ¼ 1; 2; . . .; n, generations are given by 1. All bats use echolocation to sense distance, and
xti ¼ xt1;i ; xt2;i ; . . .; xtd;i . Consider X ¼ xti and V ¼ vti , they also ‘‘know’’ the difference between food/
respectively, the position and velocity for particle i at prey and background barriers;
generation or iteration t. A velocity vti , generated by ran- 2. Bats fly randomly with velocity vti at position xti .
domly selected xti positions, here established as xtki , is used They can automatically adjust the frequency of
for mutation. In this paper, V is created by a DE/Rand/2/ their emitted pulses and adjust the rate of pulse
Bin scheme52, which includes five xtki positions, upgrading emission rBA 2 ½0; 1, depending on the proximity
1 of their target;
mutation, as can be seen in Eqn. A3. In this equation, FDE
2 h
and FDE are differential weights, such that FDE 2 ½0; 2. 3. Although the loudness can vary in many ways,
we assume that the loudness uBAi varies from a
large (positive) uBA0 to a minimum value uBAmin .
vtþ1
i ¼ xtk1 þ FDE
1
xtk2  xtk3 þ FDE
2
xtk4  xtk5 Eqn: A3
Defining the rules of how position X and velocities V are
After mutation, crossover is used to select randomly if vti or updated, Eqs. A6, A7 and A8 express the update proce-
xti will be used in the next step of optimization. That dure. In these expressions, fBA 2 ½fBAmin ; fBAmax  is the
decision is based on a crossover parameter CrDE 2 ½0; 1, as frequency value, uh U ½0; 1 is an uniformly distribution
expressed in Eqn. A4. In this equation, u1 represents a random number and xt is the current best global location or
uniformly distributed random number, such that solution.
uh U ½0; 1. fBAi ¼fBAmin þ u1 ðfBAmax  fBAmin Þ Eqn: A6
 t
vi ; if u1 CrDE vtþ1 ¼vti þ fBAi xti  xt Eqn: A7
utþ1
i ¼ t
Eqn: A4 i
xi ; otherwise
xtþ1
i ¼xti þ vtþ1
i Eqn: A8
At last, in selection, the position utþ1
and
i xti
will be
compared in terms of the direct problem output for each After this computation, an uniformly distributed random
vector ðF ðhÞÞ, as in Eqn. A5. number uh is generated. If it is greater than the pulse rate
( rBA , xtþ1
i suffers a perturbation as in Eqn. A9, where
tþ1 utþ1
i ; if F utþ1
i F xti ; uBAmean is the average loudness of all the bats at this time
xi ¼ t
Eqn: A5 step, nh N ð0; 1Þ is a normally distributed random
xi ; otherwise:
number and rBA is a scaling factor.
This process continues until convergence. The principal
xtþ1
i ¼ xti þ rBA n1 uBAmean Eqn: A9
steps to implement DE can be seen on Algorithm 2.
After this procedure, an uh is selected again. Two
Bat Algorithm conditions need to be confirmed: if uh is lower than the
loudness uBAi and the fitness F xti is lower than the fitness
Based on its developer52, the algorithm follows the F xt , the new point is accept as the new solution and the
approximate rules: pulse rate and loudness suffer alterations, as expressed in
Eqn. A10. Is assumed that uBA0 ¼ 1, uBAmin ¼ 0, cBA and
aBA are constants and kint represents iteration. For better
understanding on Algorithm 3 is summarized all the steps
to code this method.

Algorithm 2 Pseudo-code of DE algorithm52

Initialize the population X with randomly generated solutions
1 2
Set the weights FDE and FDE ∈ [0, 2] and crossover probability CrDE ∈ [0, 1]
while (criterion) do
for each search agent do
Select randomly xtk1 , xtk2 , xtk3 , xtk4 and xtk5
Generate velocity V by Eq. A3
Generate an uniformly distributed random number u1 ∼ U [0, 1]
Select and update the solution by Eq. A4 and A5
end for
end for
end while
end while
Post-process and output the best solution found

International Journal of Metalcasting

 Mathematically, rule 1 and 3 can be represented as in
Eqn. A11, where X ¼ xti is the pollen (particle) i at
iteration t, xt is the current best solution found among all
solutions, cFPA is a scaling factor, LFPA is a step-size
utþ1 t
BAi ¼ aBA uBAi ;
tþ1
rBA ¼ u0BAi ½1  expðcBA kint Þ
i
parameter, which can be observed in Eqn. A12, and kFPA is
Eqn: A10 a step-size constant. In this equation, C represents the
Gamma Function.

Algorithm 3 Pseudo-code of BA52

Initialize X and V
Initialize fBAi , pulse rate rBAi and loudness ϕBAi
while (criterion) do
Generate new solutions by adjusting frequency
Update velocities and locations/solutions by Eqs. A6, A7 and A8
if u1 > rBAi then
Select a solution among the best solutions
Generate a local solution around the selected best solution by
Eq. A9
end if
end if
Generate a new solution by flying randomly
if u2 < ϕBAi and F(xti ) < F(xt∗ ) then
Accept the new solution
Increase rBAi and reduce ϕBAi by Eq. A10
end if
end if
Rank the bats and find the current best xt∗
end while
end while

Flower Pollination Algorithm

Flower Pollination Algorithm (FPA) was created by14

based on the behavior of the flow pollination process of
flowering plants. As written by the FPA developer52, this xtþ1
i ¼xti þ cFPA LFPA ðkFPA Þ xt  xti Eqn: A11
method can be explained by four rules: kFPA CðkFPA Þsin pkFPA
1
2
LFPA ðkFPA Þ ¼
p ðsFPA Þ1þkFPA
1. Biotic and cross-pollination can be considered
processes of global pollination, and pollen-carry- Eqn: A12
ing pollinators move in a way that obeys Lévy The step-size sFPA can be expressed by Eqn. A13, which
flights; relates kFPA and two Gaussian distributions:
2. For local pollination, abiotic pollination and self- ng N 0; r2FPA and n1 N ð0; 1Þ. The variance r2FPA can
pollination are used; be calculated by Eqn. A14.
3. Pollinators such as insects can develop flower
ng
constancy, which is equivalent to a reproduction sFPA ¼ 1 Eqn: A13
probability that is proportional to the similarity of jn1 jkFPA
two flowers involved; " #k 1
2 Cð1 þ kFPA Þ sin pk2FPA FPA
4. The interaction or switching of local and global rFPA ¼ k 1
Eqn: A14
pollination can be controlled by a switch prob- kFPA C 1þk2FPA 2 FPA2
ability pFPA 2 ½0; 1, slightly biased toward local
For local pollination, rules 2 and 3 can be mathematically
pollination.
expressed as in Eqn. A15, where xtj and xtk are pollen from

International Journal of Metalcasting

different flower of the same plant species. In this equation, As the analogy suggests, because we do not know the
u1 U ½0; 1. solution to the problem, that is, the position of the prey, the
new position suggested by the formulation of this method
xtþ1
i ¼ xti þ u1 xtj  xtk Eqn: A15 is obtained from the position of the aw , bw and dw wolves.
Replacing xtp by xti , it is suggested that the location of the
For better understanding of this method, Algorithm 4 prey (optimal point) is close to the location of the wolves
explains by a pseudo-code its numerical implementation. (points with the best position until the current iteration).
Equations A19, A20 and A21 expose this behavior, where
Grey Wolf Optimizer xtI , xtII and xtIII represent, respectively, the best, second and
third best positions, guided by the wolves position, that is,
It is required a piece of information of how to express xtaw , xtbw and xtdw . The new position is determined by the
mathematically the behavior of these animals in order to arithmetic mean of the best positions constructed by the

Algorithm 4 Pseudo-code of FPA52

Objective min or max F (X), X = xti
Initialize a population of npart flower/pollen gametes with random solutions
Find the best solution x∗i in the initial population
Define a switch probability pFPA ∈ [0, 1]
while (criterion) do
for each search agent do
if u ∼ U [0, 1] < pFPA then
Global pollination via Eq. A11
else
Local pollination via Eq. A15
end if
end if
Evaluate new solutions
If new solutions are better, update them in the population
end for
end for
Find the current best solution xt∗
end while
end while
Output the best solution found

understand this method. Comparing the social hierarchy of location of the dominant wolves (best positions previously
wolves and optimization language, the alpha wolf ðaw Þ is obtained), as expressed in Eqn. A22.
considered the fittest solution. The second and third best
xtI ¼xtaw  AGWO jCGWO xtaw  xti j Eqn: A19
solutions are named beta ðbw Þ and delta ðdw Þ. The rest of
the solutions is named omega ðxw Þ. In this method, aw , bw xtII ¼xtbw  AGWO jCGWO xtbw  xti j Eqn: A20
and dw wolves are the hunters and xw wolves follow them. xtIII ¼xtdw  AGWO jCGWO xtdw  xti j Eqn: A21

The mathematical approach for encircling prey are xt þ xtII þ xtIII

xtþ1
i ¼ I Eqn: A22
expressed by Eqn. A16, where AGWO (in Eqn. A17) and 3
CGWO (in Eqn. A18) are coefficients, xtp is the position of On Algorithm 5 it is possible to observe a pseudo-code
the prey, xti is the position of a grey wolf, aint is a linearly containing the programming procedure of this method
decreasing parameter that goes from 2 to 0 based on the based on the formulations presented above.
number of iterations and u1 and u2 are uniformly dis-
tributed random numbers.

xti ¼ xtp  AGWO jCGWO xtp  xti j Eqn: A16

AGWO ¼ 2aint u1  aint Eqn: A17
CGWO ¼ 2u2 Eqn: A18

International Journal of Metalcasting

 Algorithm 5 Pseudo-code of GWO algorithm15
Initialize the grey wolf population X
Initialize aGWO , AGWO and CGWO
Calculate the fitness of each search agent
Identify xtαw , xtβw and xtδw (first, second and third best solutions)
while (criterion) do
for each search agent do
Update the position of the current search agent by Eq. A22
end for
end for
Update aint , AGWO and CGWO
Calculate the fitness of all search agent
Update xtαw , xtβw and xtδw
t=t+1
end while
end while

Moth-Flame Optimization indicates the distance between moth and flame

(Eqn. A25), bMFO is a constant for defining the shape of
Consider M ¼ mti and F ¼ fit the moth and flame positions, the logarithmic spiral, and tint is a random number such that
respectively, in which i and t are, respectively, the particle tint 2 ½1; 1.
and iteration index. OM ¼ omti and OF ¼ ofit represents
S mti ; fit ¼dit expðbMFO tint Þcosð2ptint Þ þ fit Eqn: A24
the fitness values of the moth and flame particles, such that
OM ¼ F ðMÞ and OF ¼ F ðFÞ. As exposed in16, it is worth dit ¼jfit  mti j Eqn: A25
pointing out that moths and flames are both solutions. The
In order to balance exploration and exploitation during the
difference between them is the way we treat and update
iterative process, it was observed that the number of flames
them in each iteration. The moths are actual search agents
should be reduced in order to favor exploitation in the
that move around the search space, whereas flames are the
search for more promising solutions. Thus, a formulation
best position of moths that obtains so far. In other words, F
was proposed with the objective of reducing the number of
is a sorted vector obtained by the best particles in M during
search positions in a given iteration number, which can be
the entire simulation. Flames, in this method, are consid-
seen in Eqn. A26,
ered as flags that are dropped by moths when searching the 
search space. Equation (A23) represents the modification Nmax  1
fn ¼ Nmax  t Eqn: A26
of the moth position with respect to the geometry ðS ðhÞÞ Tint
used to search the best position.
where t is the current number of iteration and Nmax and Tint
represent the maximum number of flames and iterations,
mtþ1 ¼S mti ; fit Eqn: A23
i respectively. In Eqn. A26, bhe represents rounding to the
In the original paper a logarithmic spiral was chosen, nearest integer, such that bhe ¼ bh þ 0:5c. A pseudo-
which formulation can be seen in Eqn. A24, where dit code containing the programming procedure of this method
based can be seen on Algorithm 6.

International Journal of Metalcasting

 Algorithm 6 Pseudo-code of MFO algorithm16
Update fn by Eq. A26
Create M and OM
if iteration == 1 then
F = sort(M)
OF = sort(OM)
else
F = sort(Mt−1 , Mt )
OF = sort(Mt−1 , Mt )
end if
end if
while (criterion) do
for each search agent do
Calculate dti using Eq. A25 with respect to the corresponding moth
Update mti using Eqs. A23 and A24 with respect to the corresponding
moth
end for
end for
end while
end while

Sine Cosine Algorithm Whale Optimization Algorithm

Consider X ¼ xti the i-th position of the current solution at Following Grey Wolf Optimizer, consider that the best
t-th iteration. In Eqn. A27 the random exchange mecha- candidate solution is represented by Eqn. A29. In this
nism between the two search equations is shown. In this equation, AWOA (Eqn. A30) and CWOA (Eqn. A31) are
equation, riter is a parameter given by Eqn. A28, rU1 and rU2 coefficients, aint is linearly decreased parameter from 2 to 0
are random numbers such that rU1 ¼ 2pu1 and rU2 ¼ 2u2 , in over the course of iterations, u1 and u2 is an uniformly
which u1 and u2 are uniformly distributed random numbers distributed random number, such that uh U ½0; 1, g is the
ðuh U ½0; 1Þ and g is the best position obtained so far. best solution obtained so far and X ¼ xti is the particle
position.
(
xtþ1
i ¼ xti þ rint sin rU1 jrU2 g  xti j; if u3 \0:5
xtþ1
i ¼ xtþ1 ¼g  AWOA jCWOA g  xti j Eqn: A29
xtþ1
i ¼ xti þ rint cos rU1 jrU2 g  xti j; if u3 0:5 i

Eqn: A27 AWOA ¼2aint u1  aint Eqn: A30

CWOA ¼2u2 Eqn: A31
In Eqn. A28, t is the current iteration, kmax is the maximum
number of iterations and aSCA is a constant. For exploitation, Eqn. A32 is used to perform a spiral
aSCA updating position, which mimics the helix-shaped
rint ¼ aSCA  t Eqn: A28 movement of humpback whales, very similar to the one
kmax
presented section about Moth-Flame Optimization,
An implementation of this methods is shown in
Algorithm 7. xtþ1
i ¼ dit expðbWOA tint Þcosð2ptint Þ þ g Eqn: A32

where dit ¼ jg  xti j indicates the distance of the i-th

position to the best solution obtained so far, bWOA is a

Algorithm 7 Pseudo-code of SCA17

Initialize a set of search agents X
while (criterion) do
Evaluate each of the search agents by the objective function
Update the position of search agents using Eq. A27 and Eq. A28
end while
end while

International Journal of Metalcasting

constant related to the shape of the spiral and tint is a Dragonfly Algorithm
random number in ½1; 1. Two mechanisms are presented
simultaneously by humpback whales: swim around the As shown on the original paper, based on62, the behavior of
prey within a shrinking circle and along a spiral-shaped swarms follows three primitive principles:
path. In the original paper is assumed probability of 50%
for occurrence of both models, as expressed in Eqn. A33. 1. Separation: Static collision avoidance of the
 individuals from other individuals in the
g  AWOA jCWOA g  xti j; if u3 \0:5
xtþ1
i ¼ t
neighbourhood;
di expðbWOA tint Þcosð2ptint Þ þ g ; if u3 0:5 2. Alignment: Velocity matching of individuals to
Eqn: A33 that of other individuals in neighbourhood;
3. Cohesion: Tendency of individuals towards the
According to the behavior of AWOA , the prey search
center of the mass of the neighbourhood.
equations are altered in order to favor exploitation, as seen
in Eqn. A34. This search modification can be observed on So, the principal goal for any swarm is survival, so all of
Algorithm 8. the individuals should be attracted towards food sources
and distracted outward enemies. Mathematically, separa-
xtþ1
i ¼ xtk1  AWOA jCWOA xtk1  xti j Eqn: A34
tion is calculated by Eqn. A35.
where xtki is a random position chosen from the current X
Nnb

population. SDA
i ¼  xti  xtj Eqn: A35
j¼1

Algorithm 8 Pseudo-code of WOA18

Initialize the whales population X
Calculate the fitness of each search agent
g∗ = the best search agent
while (criterion) do
for earch search agent do
Update aint , AWOA , CWOA , tint and u3
if u3 < 0.5 then
if |AWOA | < 1 then
Update the position of the current search agent by the
Eq. A29
else if |AWOA | ≥ 1 then
Update the position of the current search agent by the
Eq. A34
end if
end if
else if u3 ≥ 0.5 then
Update the position of the current search by the Eq. A32
end if
end if
end for
end for
Check if any search agent goes beyond the search space and amend it
Calculate the fitness of each search agent
Update g∗ if there is a better solution
t=t+1
end while
end while
Return g∗

International Journal of Metalcasting

where xti is the position of the current individual, xtj is the
position j-th neighboring individual and Nnb is the number
of neighboring individuals. Alignment is calculated as in
Eqn. A36, where vtj represents the velocity of j-th
Dxtþ1
i ¼ sDA SDA
i þ aDA ADA
i þ cDA CiDA þ fDA FiDA
neighboring individual.
PNnb t þeDA EiDA wDA Dxti
DA j¼1 vj Eqn: A36
Ai ¼ Eqn: A40
Nnb
xtþ1
i ¼xti þ Dxtþ1
i Eqn: A41
Cohesion is calculated by Eqn. A37, where xti is the
position of the current individual and xtj shows the position One mechanism to enhance the stochastic behavior and
j-th neighboring individual. exploration of the particles is by adding a random walk by
PNnb t Lévy flight where there is no neighboring solutions. So, the
j¼1 xj position of dragonflies can be updated by Eqn. A42. The
CiDA ¼  xti Eqn: A37
Nnb Lévy flight can be calculated by Eqn. A43, where u1 and u2
are uniformly distributed random numbers in [0, 1], kDA is
Attraction towards a food source FiDA is calculated as
a constant and rDA is calculated by Eqn. A44, where C
Eqn. A38, where xþi shows the position of the food source.
represents the Gamma Function.
Distraction outwards an enemy EiDA is calculated as
Eqn. A39, where x i shows the position of the enemy. xtþ1
i ¼xti þ LDA ðkDA Þxti Eqn: A42
u1 rDA
FiDA ¼xþ
i  xti Eqn: A38 LDA ðkDA Þ ¼0:01 1 Eqn: A43
ju2 jkDA
EiDA ¼x
i  xti Eqn: A39
" #k 1
The step Dxtþ1 shows the movement direction of the Cð1 þ kDA Þ sin pk2DA DA
i rDA ¼ b1
Eqn: A44
dragonflies, which is defined in Eqn. A40, where sDA kDA C 1þk2 DA 22
shows the separation weight, SDA i indicates the separation
of the i-th individual, aDA is the alignment weight, ADA is A better understanding of the code building process can be
i
the alignment of i-th individual, cDA indicates the cohesion taken from the pseudo-code expressed on Algorithm 9.
weight, CiDA is the cohesion of the i-th individual, fDA is the
food factor, FiDA is the food source of the i-th individual,
eDA is the enemy factor, EiDA is the position of enemy of the
i-th individual and wDA is the inertia weight. After that, the
position vector are updated as Eqn. A41.

Algorithm 9 Pseudo-code of DA19

Initialize the dragonflies population xti
Initialize step Δxti
while (criterion) do
Calculate the objective values of all dragonflies
Update the food source and enemy
Update wDA , sDA , aDA , cDA , fDA and eDA
Calculate SiDA , ADA DA
i , Ci , Fi
DA
and EiDA using Eqs. A35 to A39
Update neighbouring radius
if a dragonfly has at least one neightbouring dragonfly then
Update velocity using Eq. A40
Update position using Eq. A41
else
Update position using Eq. A42
end if
end if
Check and correct the new positions based on the boundaries of variables
end while
end while

International Journal of Metalcasting

Harris Hawks Optimization in iteration t and J HHO ¼ 2  ð1  u6 Þ represents the random
jump strength of the rabbit throughout the escaping
In this method, as expressed in the original paper, the procedure. The J HHO value changes randomly in each
Harris hawks are the candidate solutions and the best iteration to simulate the nature of rabbit motions.
candidate solution in each step is considered as the inten-
xtþ1
i ¼Dxti  EHHO jJ HHO xt  xti j Eqn: A48
ded prey or nearly the optimum. The Harris’ hawks perch
randomly on some locations and wait to detect prey based Dxti ¼jxt  xti j Eqn: A49
on two strategies, resulting in an equal chance for each
When u6 0:5 and jEHHO j\0:5, Eqn. A50 is used. This
perching strategy. They perch based on the positions of
mechanism is called ‘‘hard besiege’’.
other family members and the rabbit, which is modeled in
Eqn. A45, where xtþ1 i is the position of hawks in the next xtþ1
i ¼ xt  EHHO jDxti j Eqn: A50
iteration, xt is the position of rabbit, that is, the best
solution in the iteration t, xti is the current position vector of For ‘‘soft besiege with progressive rapid dives’’, that is
hawks, u1 , u2 , u3 , u4 and u5 are uniformly distributed u6 \0:5 and jEHHO j 0:5, Eqs. (A51), (A52), (A53), (A54)
random numbers ðuh U ð0; 1ÞÞ, LB and UB show the and (A55) are used. In these equations LHHO is the Lévy
upper and lower bounds of the variable, xtki is a randomly flight function. For Lévy flight computation, two Gaussian
selected hawk from the current population and xtmean is the distributions are used, ng N ð0; rHHO Þ and n1 N ð0; 1Þ,
average position of the current population of hawks. In kHHO is a default constant and C represents the Gamma
Eqn. A46, npart denotes the total number of hawks. function.

8 yti ¼xt  EHHO jJ HHO xt  xti j Eqn: A51

< xtk1 u1 jxtk1 2u2 xti j; if u5 0:5 zti ¼yti þ u7 LHHO ðkHHO Þ Eqn: A52
xtþ1
i ¼
: xtki xtmean u3 ½LBþu4 ðUBLBÞ; if u5 \0:5 ng rHHO
LHHO ðkHHO Þ ¼0:01 1 Eqn: A53
jn1 jkHHO
Eqn: A45 " #k 1
npart
1 X Cð1 þ kHHO Þ sin pkHHO
2
HHO

xtmean ¼ xti Eqn: A46 rHHO ¼ k 1

Eqn: A54
npart i¼1
kHHO C 1þk2HHO 2 HHO2
(
To transit from exploration to exploitation the parameter yti ; if F yti \F xti
xtþ1
i ¼ Eqn: A55
EHHO is applied, which can be seen in Eqn. A47, where zti ; if F zti \F xti
EHHO indicates the escaping energy of the prey, kmax is the
maximum number of iterations and E0HHO 2 ½1; 1 is the At last, for ‘‘hard besiege with progressive rapid dives’’,
initial state of its energy. From the calculation of EHHO the that is u6 \0:5 and jEHHO j\0:5, Eqn. A56 and Eqn. A57
algorithm will alternate the mechanism of search. If can be used to calculate the next position by Eqn. A58.
jEHHO j 1, exploration phase take account. If jEHHO j\1, yti ¼xt  EHHO jJ HHO xt  xtmean j Eqn: A56
exploitation phase is considered. Besides that parameter, a
random number u6 is used to determine if the prey has or zti ¼yti
þ u7 LHHO ðkHHO Þ Eqn: A57
(
has not successfully escaped. Numerically, this control yt ; if F yti \F xti
besieges with progressive rapid dives, that is, more xtþ1
i ¼ ti Eqn: A58
zi ; if F zti \F xti
rigorous exploitation of the method.
  A more detailed view of the iterative process considered in
HHO HHO t
E ¼ 2E0 1 Eqn: A47 this method can be seen on Algorithm 10.
kmax
If u6 0:5 and jEHHO j 0:5, Eqs. (A48) and (A49) will be
used for ‘‘soft besiege’’, where Dxti is the difference
between the position of the rabbit and the current location

International Journal of Metalcasting

 Algorithm 10 Pseudo-code of HHO algorithm20
Initialize the random population X
while (criterion) do
Calculate the fitness values of hawks
Set xt∗ as the location of rabbit (best location)
for earch search agent do
Update the initial energy E0HHO and jump strength J HHO
Update the E HHO using Eq. A47
if (|E HHO | ≥ 1) then
Update location using Eq. A45
end if
end if
if (|E HHO | < 1) then
if (u6 ≥ 0.5 and |E HHO | ≥ 0.5 ) then
Update using Eq. A48
else if (u6 ≥ 0.5 and |E HHO | < 0.5 ) then
Update using Eq. A50
else if (u6 <0.5 and |E HHO | ≥ 0.5 ) then
Update using Eq. A55
else if (u6 <0.5 and |E HHO | < 0.5 ) then
Update using Eq. A58
end if
end if
end if
end if
end for
end for
end while
end while
Return xt∗

Markov Chain Monte Carlo Algorithm unit standard deviation, such that nh N ð0; 1Þ. After we
calculate the probability of acceptance aðh; h Þ as stated in
This algorithm is used to decide which values to accept or Eqn. A60 (based on the symmetry of the proposed
discard.We begin by selecting a candidate h from a den- distributions, as previously exposed).
sity distribution qðh ; ht1 Þ, which gives the probability of " #
changing from ht1 to h , (in this paper, the proposal dis-  Pðh jX Þ
aðh; h Þ ¼ min 1; Eqn: A60
tribution qðh ; ht1 Þ is a random walk and is symmetric to P ht1 jX
q ht1 jh ), such that,
Then, aðh; h Þ is compared with a random number ur such
that ur U ð0; 1Þ. If ur aðh; h Þ, h is accepted to be ht .
h ¼ ht1 ½1 þ wSS n1  Eqn: A59
Otherwise, ht ¼ ht1 . The Algorithm 11 expresses the
where wSS is the search step and n1 is random number iterative process of the MCMC method considered in this
sampled from a normal distribution with zero mean and work.

International Journal of Metalcasting

 Algorithm 11 Pseudo-code of MCMC based on Refs. 54 , 63 and 64.
Initialize the Markov Chain
while (criterion) do
Select a candidate θ ∗ by Eq. A59
Calculate the probability of acceptance α (θ, θ ∗ ) by Eq. A60
Generate a random number ur such that ru ∼ U (0, 1)
if ur ≤ α (θ, θ∗ ) then
θt = θ∗
else
θ t = θ t−1
end if
end if
end while
end while

REFERENCES 9. E. Anglada, A. Meléndez, A. Obregón, E. Villanueva,

I. Garmendia, Performance of optimization algorithms
1. X.S. Yang, Mathematical analysis of nature-inspired in the model fitting of the multi-scale numerical
algorithms, in Nature-Inspired Algorithms and simulation of ductile iron solidification. Metals 10(8),
Applied Optimization. ed. by X.S. Yang (Springer 1071 (2020). https://doi.org/10.3390/met10081071
International Publishing, Cham, Switzerland, 2018), 10. M. Brezina, T. Mauder, L. Klimes, J. Stetina, Com-
pp.1–25 parison of optimization-regulation algorithms for
2. T. Joyce, J.M. Herrmann, A review of no free lunch secondary cooling in continuous steel casting. Metals
theorems, and their implications for metaheuristic 11(2), 237 (2021). https://doi.org/10.3390/
optimisation, in Nature-Inspired Algorithms and met11020237
Applied Optimization. ed. by X.S. Yang (Springer 11. J. Kennedy, R. Eberhart, Particle swarm optimization,
International Publishing, Cham, 2018), pp.27–51 in Proceedings of ICNN’95 - International Conference
3. D.H. Wolpert, W.G. Macready, et al., No free lunch on Neural Networks. vol. 4. (IEEE, Perth, WA, IEEE
theorems for search. Technical Report SFI-TR-95-02- Service Center, Piscataway NJ, 1995, pp. 1942–1948
010, Santa Fe Institute (1995) 12. R. Storn, K. Price, Differential evolution—a simple
4. D.H. Wolpert, W.G. Macready, No free lunch theo- and efficient heuristic for global optimization over
rems for optimization. IEEE Trans Evolut Comput. continuous spaces. J. Global Optim. 11(4), 341–359
1(1), 67–82 (1997). https://doi.org/10.1109/4235. (1997). https://doi.org/10.1023/a:1008202821328
585893 13. X.S. Yang, A new metaheuristic bat-inspired algo-
5. J. McDermott, When and why metaheuristics
rithm, in Nature Inspired Cooperative Strategies for
researchers can ignore ‘‘no free lunch’’ theorems.
Optimization (NICSO 2010). ed. by V. Fischer, S.
SN Comput. Sci. 1(1), 1–18 (2020). https://doi.org/10.
Hickinbotham, J.R. González, D.A. Pelta, C. Cruz, G.
1007/s42979-020-0063-3
Terrazas et al. (Springer, Berlin, 2010), pp.65–74
6. X.S. Yang, Free lunch or no free lunch: that is not just
14. X.S. Yang, Flower pollination algorithm for global
a question? Int. J. Artif. Intell. T. 21(03), 1240010
optimization, in Unconventional Computation and
(2012). https://doi.org/10.1142/s0218213012400106
7. G. de M Stieven, D. dos R Soares, E.P. Oliveira, E.F. Natural Computation, 11th International Conference,
Lins, Interfacial heat transfer coefficient in unidirec- UCNC 2012. Orle´ans, France (Springer, Berlin,
tional permanent mold casting: modeling and inverse 2012), pp. 240–249
estimation. Int. J. Heat Mass Trans. 166, 120765 15. S. Mirjalili, S.M. Mirjalili, A. Lewis, Grey wolf
(2021). https://doi.org/10.1016/j.ijheatmasstransfer. optimizer. Adv. Eng. Softw. 69, 46–61 (2014). https://
2020.120765 doi.org/10.1016/j.advengsoft.2013.12.007
8. Y. Yu, X. Luo, Identification of heat transfer coeffi- 16. S. Mirjalili, Moth-flame optimization algorithm: a
cients of steel billet in continuous casting by weight novel nature-inspired heuristic paradigm. Knowl.-
least square and improved difference evolution Based Syst. 89, 228–249 (2015). https://doi.org/10.
method. Appl. Therm. Eng. 114, 36–43 (2017). https:// 1016/j.knosys.2015.07.006
doi.org/10.1016/j.applthermaleng.2016.11.173 17. S. Mirjalili, SCA: a sine cosine algorithm for solving
optimization problems. Knowl.-Based Syst. 96,

International Journal of Metalcasting

 120–133 (2016). https://doi.org/10.1016/j.knosys. Anderson, J. Weritz, J.G. Kaufman (ASM Interna-
2015.12.022 tional, Materials Park, OH, 2018)
18. S. Mirjalili, A. Lewis, The whale optimization algo- 30. G. Sigworth, Aluminum, Casting Alloys, Casting
rithm. Adv. Eng. Softw. 95, 51–67 (2016). https://doi. Processes, in Aluminum Science and Technology, vol.
org/10.1016/j.advengsoft.2016.01.008 2A, ed. by K. Anderson, J. Weritz, J.G. Kaufman
19. S. Mirjalili, Dragonfly algorithm: a new meta-heuris- (ASM International, Materials Park, OH, USA, 2018),
tic optimization technique for solving single-objec- pp.119–142
tive, discrete, and multi-objective problems. Neural 31. A. Hamasaiid, M. Dargusch, T. Loulou, G. Dour, A
Comput. Appl. 27(4), 1053–1073 (2016). https://doi. predictive model for the thermal contact resistance at
org/10.1007/s00521-015-1920-1 liquid-solid interfaces: analytical developments and
20. A.A. Heidari, S. Mirjalili, H. Faris, I. Aljarah, M. validation. Int. J. Therm. Sci. 50(8), 1445–1459
Mafarja, H. Chen, Harris hawks optimization: algo- (2011)
rithm and applications. Future Gener Comput. Syst. 32. A. Hamasaiid, M. Dargusch, G. Dour, The impact of
97, 849–872 (2019). https://doi.org/10.1016/j.future. the casting thickness on the interfacial heat transfer
2019.02.028 and solidification of the casting during permanent
21. K.G.P. Nunes, I.V.J. Davila, D. Arnold, C.H.R. mold casting of an A356 alloy. J. Manuf. Process. 47,
Moura, D.C. Estumano, L.A. Féris, Kinetics and 229–237 (2019)
thermodynamic study of laponite application in caf- 33. M. Pokorny, C. Monroe, C. Beckermann, L. Bichler,
feine removal by adsorption. Environ. Process. 9(3), C. Ravindran, Prediction of hot tear formation in a
1–17 (2022). https://doi.org/10.1007/s40710-022- magnesium alloy permanent mold casting. Int.
00598-4 J. Metalcast. 2(4), 41–53 (2008)
22. B.M. Viegas, E.M. Magalhães, H.R.B. Orlande, D.C. 34. T.J. Williams, R.A. Hardin, C. Beckermann, Ther-
Estumano, E.N. Macêdo, Experimental study and mophysical properties and performance of riser
mathematical modelling of red mud leaching: appli- sleeves for steel castings. Int. J. Metalcast. 10(4),
cation of Bayesian techniques. Int. J. Environ. Sci. 535–555 (2016)
Technol. (2022). https://doi.org/10.1007/s13762-022- 35. F. Qiu, K. Bu, J.H. Song, G.L. Tian, X.D. Zhang,
04346-x Dimensional control of nickel-based single crystal
23. D.M. Stefanescu, in ed. by A. Tomsett Science and turbine blade investment casting by process control
Engineering of Casting Solidification (Springer, optimization. Int. J. Metalcast. 12(3), 469–479 (2018)
Cham, 2015) 36. Y. Dong, K. Bu, Y. Dou, D. Zhang, Determination of
24. S.C. Johnson, C.D. Clark, J.S. Alvarez, Development interfacial heat-transfer coefficient during investment-
and analysis of Al7075 alloy materials using press and casting process of single-crystal blades. J. Mater.
sinter processing, in Metals Materials Series. ed. by T. Process. Technol. 211(12), 2123–2131 (2011)
Minerals (Springer, Cham, 2020), pp.233–240 37. V. Bazhenov, Y.V. Tselovalnik, A. Koltygin, V.
25. J. Stroh, A. Piche, D. Sediako, A. Lombardi, G. Belov, Investigation of the interfacial heat transfer
Byczynski, The effects of solidification cooling rates coefficient at the metal-mold interface during casting
on the mechanical properties of an aluminum inline-6 of an A356 aluminum alloy and AZ81 magnesium
engine block, in Light Metals 2019 (Springer, 2019), alloy into steel and graphite molds. Int. J. Metalcast.
pp. 505–512 15(2), 625–637 (2021)
26. A. Kordijazi, S.K. Behera, O. Akbarzadeh, M. Povolo, 38. M. Peres, C. Siqueira, A. Garcia, Macrostructural and
P. Rohatgi, A statistical analysis to study the effect of microstructural development in Al–Si alloys direc-
silicon content, surface roughness, droplet size and tionally solidified under unsteady-state conditions.
elapsed time on wettability of hypoeutectic cast J. Alloys Compd. 381(1–2), 168–181 (2004)
aluminum-silicon alloys, in Metals Mater. Ser. ed. by 39. M.D. Peres, Desenvolvimento da macroestrutura e da
T. Minerals (Springer International Publishing, Cham, microestrutura na solidificacao unidirecional transito-
Switzerland, 2020), pp.185–193 ria de ligas Al–Si [Doctoral Thesis]. Federal Univer-
27. S. Chen, K. Liu, X.G. Chen, Effect of mo on elevated- sity of Campinas (2005)
temperature low-cycle fatigue behavior of Al-Si 356 40. E.P. Oliveira, G. de M. Stieven, E.F. Lins, J.R.P. Vaz,
cast alloy, in The Minerals, Metals & Materials Series An inverse approach for the interfacial heat transfer
(Springer, 2020), pp. 261–266 parameters in alloys solidification. Appl. Therm. Eng.
28. J. Rakhmonov, G. Timelli, F. Bonollo, L. Arnberg, 155, 365–372 (2019). https://doi.org/10.1016/j.
Influence of grain refiner addition on the precipitation applthermaleng.2019.03.084
of Fe-rich phases in secondary AlSi7Cu3Mg alloys. 41. H. Wang, M.S. Hamed, S. Shankar, Interaction
Int. J. Metalca St. 11(2), 294–304 (2017). https://doi. between primary dendrite arm spacing and velocity of
org/10.1007/s40962-016-0076-9 fluid flow during solidification of Al–Si binary alloys.
29. G. Sigworth, Aluminum science and technology, in J. Mater. Sci. 53(13), 9771–9789 (2018). https://doi.
Aluminum Science and Technology, vol. 2A, ed. by K. org/10.1007/s10853-018-2239-y

International Journal of Metalcasting

42. N.A. El-Mahallawy, A.M. Assar, Effect of melt 54. E. Somersalo, J. Kaipio, Statistical and computational
superheat on heat transfer coefficient for aluminium inverse problems. Appl. Math. Sci. (2005). https://doi.
solidifying against copper chill. J. Mater. Sci. 26(7), org/10.1007/b138659
1729–1733 (1991). https://doi.org/10.1007/ 55. D.C. Knupp, H.M. da Fonseca, C.P. Naveira-Cotta,
bf00543594 H.R.B. Orlando, R.M. Cotta, O. Fudym, Thermal
43. L. Zeng, W. Zhang, Y. Ji, Y. Huang, J. Li, Improving characterization of nonhomogeneous media, in Meth-
cooling rate during solidification by eliminating the ods and Applications. ed. by S. Foundations (Springer,
metal-mold interfacial gap. Metall. Mater. Trans. A. Cham, 2014), pp.119–139
46(7), 2819–2822 (2015). https://doi.org/10.1007/ 56. Z. Chen et al., Bayesian filtering: From Kalman filters
s11661-015-2945-x to particle filters, and beyond. Statistics 182(1), 1–69
44. P.S. Vishweshwara, N. Gnanasekaran, M. Arun, (2003)
Inverse estimation of interfacial heat transfer coeffi- 57. J.V. Beck, K.J. Arnold, Parameter Estimation in
cient during the solidification of Sn-5wt%Pb alloy Engineering and Science. Wiley Series in Probability
using evolutionary algorithm, in Lecture Notes in and Mathematical Statistics (Wiley, New York, 1977)
Mechanical Engineering (Springer Singapore, Cham, 58. J. Lee, W. Sung, J.H. Choi, Metamodel for efficient
2019), pp. 227–237 estimation of capacity-fade uncertainty in Li-Ion
45. Y. Natsume, Y. Oka, J. Ogawa, M. Ohno, Estimation batteries for electric vehicles. Energies 8(6),
of time-dependent heat transfer coefficient in unidi- 5538–5554 (2015). https://doi.org/10.3390/en8065538
rectional casting using a numerical model coupled 59. A.E. Ezugwu, O.J. Adeleke, A.A. Akinyelu, S. Viriri,
with solidification analysis and data assimilation. Int. A conceptual comparison of several metaheuristic
J. Heat Mass Transf. 150, 119222 (2020) algorithms on continuous optimisation problems.
46. T. Soares, C. Cruz, M. Xavier, R.V. Reyes, F. Bertelli, Neural Comput. Appl. 32(10), 6207–6251 (2019).
A. Garcia et al., Interfacial heat transfer and https://doi.org/10.1007/s00521-019-04132-w
microstructural analyses of a Bi-5% Sb lead-free alloy 60. S. Bejinariu, H. Costin, A comparison of some nature-
solidified against Cu, Ni and low-C steel substrates. inspired optimization metaheuristics applied in
J. Alloys. Compd. 860, 158553 (2021) biomedical image registration. Method Inform Med.
47. P. Vishweshwara, N. Gnanasekaran, M. Arun, Inverse 57(05/06), 280–286 (2018). https://doi.org/10.1055/s-
estimation of interfacial heat transfer coefficient 0038-1673693
during the solidification of Sn-5wt% Pb alloy using 61. D.P. Doane, Aesthetic frequency classifications. Am
evolutionary algorithm, in Advances in Materials and Stat. 30(4), 181–183 (1976)
Metallurgy: Select Proceedings of ICEMMM 2018 62. C.W. Reynolds, Flocks, herds and schools: A dis-
(Springer, 2019), pp. 227–237 tributed behavioral model, in Proceedings of the 14th
48. M. Wu, A. Ludwig, A. Kharicha, A four phase model Annual Conference on Computer Graphics and
for the macrosegregation and shrinkage cavity during Interactive Techniques (1987), pp. 25–34
solidification of steel ingot. Appl. Math. Model. 41, 63. J.M. Costa, C.P. Naveira-Cotta, Estimation of kinetic
102–120 (2017) coefficients in micro-reactors for biodiesel synthesis:
49. H. Zhang, X. Liu, D. Ma, M. Song, A. Ludwig, A. Bayesian inference with reduced mass transfer model.
Kharicha et al., Digital twin for directional solidifi- Chem. Eng. Res. Des. 141, 550–565 (2019). https://
cation of a single-crystal turbine blade. Acta Mater. doi.org/10.1016/j.cherd.2018.11.023
244, 118579 (2023) 64. C.H.R. Moura, B.M. Viegas, M. Tavares, E.N.
50. L. Chen, Y. Wang, L. Peng, P. Fu, H. Jiang, Study on Macêdo, D.C. Estumano, J.N.N. Quaresma, Parameter
the interfacial heat transfer coefficient between estimation in population balance through Bayesian
AZ91D magnesium alloy and silica sand. Exp. technique markov chain Monte Carlo. J. Appl. Com-
Thermal Fluid Sci. 54, 196–203 (2014) put. Mech. 7(2), 890–901 (2021). https://doi.org/10.
51. R. Dou, T. Ge, X. Liu, Z. Wen, Effects of contact 22055/JACM.2021.35741.2725
pressure, interface temperature, and surface roughness
on thermal contact conductance between stainless
Publisher’s Note Springer Nature remains neutral with
steel surfaces under atmosphere condition. Int. J. Heat
regard to jurisdictional claims in published maps and
Mass Transf. 94, 156–163 (2016)
institutional affiliations.
52. X.S. Yang, Nature-Inspired Optimization Algorithms
(Elsevier, London, 2014)
Springer Nature or its licensor (e.g. a society or other
53. R. Storn, On the usage of differential evolution for
partner) holds exclusive rights to this article under a
function optimization, in Proceedings of North
publishing agreement with the author(s) or other rightsh-
American Fuzzy Information Processing. IEEE.
older(s); author self-archiving of the accepted manuscript
Berkeley, California, USA by M.H. Smith, M.A. Lee,
version of this article is solely governed by the terms of
J. Keller, J. Yen, eds. (IEEE, 1996), pp. 519–523
such publishing agreement and applicable law.

International Journal of Metalcasting