You are on page 1of 16

Chemical Engineering Research and Design 194 (2023) 163–178

Available online at www.sciencedirect.com

Chemical Engineering Research and Design

journal homepage: www.elsevier.com/locate/cherd

Phase stability analysis and phase equilibrium


calculations in reactive and nonreactive systems
]]
]]]]]]
]]

using new hybrids of Pelican and Gorilla troops


algorithms

John O. Bamikole a, Caleb Narasigadu a,
a
Chemical Engineering Department, Faculty of Engineering and the Built Environment, University of
Johannesburg, Doornfontein 2028, South Africa

a r t i c l e i n f o a b s t r a c t

Article history: Phase stability and equilibrium calculations are very significant in the process industries,
Received 17 December 2022 ranging from process simulation, equipment design, operation and design. These pro­
Received in revised form 13 March blems involve several systems of complex and nonlinear equations, which are mathe­
2023 matically challenging. An alternative solution is formulating the problems into a global
Accepted 19 April 2023 optimisation problem. A robust and reliable optimisation technique is required to solve
Available online 21 April 2023 these global optimisation problems. In this study, two recently developed stochastic
global optimisation algorithms, POA and GTO were applied to phase stability and equili­
Keywords: brium problems. The need to enhance their performance necessitated the development of
Phase stability two modified algorithms, PGOA1 and PGOA2, which are the hybridisation of POA and GTO.
Equilibrium All the algorithms’ abilities were explored for solving computationally challenging and
Stochastic optimisation multidimensional phase stability and equilibrium problems, and their performances were
Pelican algorithm compared. The implementation of POA and GTO was good in solving most problems,
Gorilla though GTO slightly outperformed POA. The hybridisation of both to develop PGOA1 and
Hybrid PGOA2 was worthwhile as both outperformed the POA and GTO. PGOA2 is more robust
and reliable in solving phase stability and equilibrium, and its performance superseded
the performances of the other algorithms.
© 2023 The Authors. Published by Elsevier Ltd on behalf of Institution of Chemical
Engineers.

condition at which phases coexist in a system with no major


1. Introduction
change in the composition of all the phases in the system.
Phase stability and equilibrium problems can be solved
Phase stability and equilibrium are two interrelated funda­
through a system of nonlinear equations. These equations
mental concepts that play a significant role in the design,
are often generated from the mass balance, equilibrium laws,
simulation, operation and optimisation of separation units in
reactions, and correlation equations for physical properties
the process industries. In most cases, phase stability analysis
such as the Antoine equation, activity coefficient models,
is a precursor to phase equilibrium. Phase stability of any
equations of state and others. The analytical solution of all
chemical system is the ability of the system to remain in a
these problems is quite challenging. A viable option is to
single phase at a given condition of temperature, pressure
solve them numerically by formulating the problems into an
and composition. While phase equilibrium has to do with the
optimisation problem and minimising the objective func­
tions (Khalil et al., 2016). The objective function of phase

stability is the tangent plane distance function (TPDF), while
Corresponding author.
that of phase equilibrium is the Gibbs free energy.
E-mail address: cnarasigadu@uj.ac.za (C. Narasigadu).
https://doi.org/10.1016/j.cherd.2023.04.042
0263-8762/© 2023 The Authors. Published by Elsevier Ltd on behalf of Institution of Chemical Engineers.
164 Chemical Engineering Research and Design 194 (2023) 163–178

tendency to converge to the global minimum (Khalil


Nomenclature
et al., 2016).
c number of components Stochastic algorithms have been applied to phase stability
CS cuckoo search and equilibrium problems, and include; genetic algorithms
D Dimension and simulated annealing (Rangaiah, 2001), particle swarm
EoS Equation of state optimisation (Bonilla-Petriciolet and Segovia-hernández,
g reduced Gibbs free energy 2010), differential evolution and unified bare-bones particle
G Gibbs free energy swarm optimisation (Zhang et al., 2011), genetic algorithm
GTO Gorilla troops optimiser and differential evolution with tabu list (Bonilla-Petriciolet
HB Honey badger et al., 2011), Covariance Matrix Adaptation Evolution
HH Horse herd optimisation algorithm Strategy, Shuffled Complex Evolution and Firefly Algorithms
IDE integrated differential evolution algorithm (Fateen et al., 2012; Fateen and Bonilla-Petriciolet, 2015),
Keq equilibrium constant improved ant colony (Fernández-Vargas et al., 2013), Herd
KH Krill herd algorithm algorithm and the modified Lévy-Flight Krill Herd algorithm
LKH Lévy-flight krill herd algorithm (Moodley et al., 2015), Charged System Search Algorithms
LLE liquid liquid equilibria (Elnabawy et al., 2014), Hybrid of Monkey and Krill herd al­
LB Lower parameter bound gorithm (Khalil et al., 2015), Intelligent Firefly Algorithm,
n number of moles Cuckoo Search, Artificial Bee Algorithm and Bat Algorithm
NFE number of function evaluations (Fateen and Bonilla-Petriciolet, 2015), Gravitational Search,
NP Number of population Monkey and Krill Herd Swarm Algorithms (A. M. E. Khalil
PE phase equilibria et al., 2016), Pathfinder, Honey Badger, Red Fox and Horse
PGOA Pelican-Gorilla optimisation algorithms Herd (Bamikole and Narasigadu, 2023) and others.
POA Pelican optimisation algorithm The widespread use of stochastic algorithms to solve
PS Phase stability phase stability and equilibrium problems has shown that
rPE reactive phase equilibria different algorithms perform differently for various pro­
SC stopping criteria blems; this corroborates the no-free launch theory, which
TPDF Tangent plane distance function says that it is quite impossible for a given algorithm to suf­
UB Upper parameters bound ficiently solve all problems (Wolpert and Macready, 1997).
VLE Vapour liquid equilibria Therefore it is a worthwhile venture to check the perfor­
yi component i for the trial phase mances of newly developed algorithms, improve existing
zi mole fraction of component i in the ones and check their performances in solving phase stability
εi fraction of component i and phase equilibrium problems. This work aims to apply
μi Chemical potential of component i two new algorithms: the Pelican (Trojovský and Dehghani,
µio Chemical potentials for pure component i 2022) and Gorilla (Abdollahzadeh et al., 2021) stochastic al­
ˆi component i fugacity coefficient gorithms to phase stability and equilibrium, and compare
i pure fugacity coefficient for component i their performances and develop new hybrids of both algo­
γi component i activity coefficient rithms for enhanced performance.
π number of phases

Subscripts and superscripts


2. The description of the stochastic
T Total
optimisation algorithms
ref reference components
The formulation of an optimisation problem is to minimise
i component i
an objective function F(x) that is dependent on the D vari­
j Phase j
ables, which are known as decision variables (x1,x2,x3,…xD)
that are bounded between the minimum limit
1 2 3 D
(xmin , x min , x min …x min ) and a maximum limit
Optimisation methods are generally classified into two 1 2 3 D
(xmax , x max , x max …x max ).
classes; deterministic and stochastic algorithms. Phase sta­
This study examined two new hybrid and metaheuristic
bility and equilibrium issues have seen increased use of de­
algorithms to tackle phase stability and equilibrium pro­
terministic algorithms. McDonald and Floudas (1995) solved
blems. The application of these algorithms to phase stability
stability problems for NRTL and UNIQUAC activity models
and equilibriums are of interest because it has not been re­
using global optimisation and the bound and branch algo­
ported in the literature. An overview of the algorithms is
rithms. Several stationary points for TPDF were obtained
discussed in the following section.
using the Newton homotopy-continuation method (Sun and
Seider, 1995; Hua et al., 1998a,1998b; Jalali et al., 2008), In­
terval analysis ( Tessier et al., 2000), continuation method 2.1. Pelican optimisation algorithms
(Wasylkiewicz et al., 2013).
Despite the popularity of the deterministic method to The pelican optimisation algorithm (POA) is a stochastic
phase stability and equilibrium problems, it was observed optimisation algorithm that was recently proposed by
that the higher the dimension of the problems, the higher the Trojovský and Dehghani (2022) to mimic the hunting strategy
computational effort exponentially (Nichita et al., 2002, of the pelican. The algorithm is modelled using the pelican’s
Wakeham and Stateva, 2004). Stochastic algorithms are ea­ hunting strategy in two stages; the exploration stage, in
sier to implement without assumption and have a higher which the pelican moves towards its prey and the exploita­
tion stage, winging on the water’s surface. The mathematical
Chemical Engineering Research and Design 194 (2023) 163–178 165

model of the exploration phase of the pelican algorithms is L = C *l (7)


shown in Eq. (1): where l is used to enhance the leadership of the silverback,
and it has values within the range of − 1 and 1.
xit + rand *(P I . xit ), Fp < Fi
xit + 1 = H = Z *xit
xit + rand *(xit P ), else (8)
(1)

where xit + 1 is the location of ith pelican member at an itera­ Z = [ C, C ] (9)


tion time t + 1, P is the location of the prey, I is a set of At the exploitation stage, the silverback takes the lead,
random numbers that range from 0 to 1. Fp is the value of the and other members of the troops follow. As time goes on,
objective function of the prey’s location. some members can decide not to follow but compete for the
At every stage of the iteration, it is necessary to check the headship of the colony and female gorillas with the silver­
fitness of the newly calculated members’ positions and up­ back. These characteristics are factored in the mathematical
date accordingly. This is done using Eq. (2): formulation of the exploitation stage, and this is done by
assigning a specific value to C. If C ≥ W, the silverback leads,
xit + 1, Fit + 1 < Fit
xit + 1 = but if C < W, the female competition is enhanced, given that
xit else (2) W must be selected before the optimisation process
(Abdollahzadeh et al., 2021).
The second phase of POA involves modelling the process
The following set of equations was modelled to factor in
in which the pelican spreads its wings on the water’s surface
the lead of the silverback and competition with other
to attract more fish. This process is modelled mathematically
members of the troops.
using Eq. (3):
xit + 1 = L *M *(xit t
xsilverback ) + xit (10)
t
xit + 1 = xit + R. 1 *(rand 1)*xit
T (3) 1
N g g
1
Eqs. (1) and (3) are the fundamental and peculiar equa­ M= xit
N i=1 (11)
tions for the pelican algorithms, the first set of the popula­
tion is randomly generated within the lower and upper limit g = 2L (12)
of the decision variables. At every point, there is a need to
check new members’ fitness before updating them, and it The competition for the adult female is modelled with
terminates as soon as the stopping criteria are satisfied. The Eq. (13):
pelican algorithm has no unique parameters that must be t
xit + 1 = xsilverback t
(xsilverback *Q xit *Q )* *E (13)
declared before or during the optimisation process. Details of
the pelican algorithms can be found in the work of Trojovský Q = r5 1 (14)
and Dehghani ( 2022).
N1, rand 0.5
E=
N2, rand < 0.5 (15)
2.2. Artificial Gorilla troop optimizer
where is a parameter that is assigned a value before the
The artificial Gorilla troop optimiser (GTO) was modelled to optimisation process, and r5 is a random number that ranges
mimic the Gorilla troop’s social intelligence (Abdollahzadeh between 0 and 1.
et al., 2021). The Gorilla troop walk in groups, and in the
application of their pattern to solving optimisation problems, 2.3. The proposed hybrid algorithms
each of the member in the group is considered as a solution,
but the best of all the solutions at a particular time is called The evaluation of the performances of both POA and GTO on
the silverback. The algorithm is modelled using three me­ phase stability and equilibrium problems has necessitated
chanisms for the exploration phase: movement to an un­ the need to enhance the performance of both algorithms by
known location, transit to a well-defined location, and the developing new algorithm, which will be a hybrid of both
group movements of other gorillas to the known location. A algorithms. Hybridisation aims to improve the performance
selection criteria P is imposed to determine the condition for of algorithms by using the comparative strength of both al­
switching within the mechanisms. If rand < P, the first me­ gorithms to compensate for their weaknesses. Each algo­
chanism is activated, if rand < 0.5, it switches to the known rithm has its unique characteristic and property, which can
location, and if rand ≥ 0.5, the other gorillas move to the favour the exploration or exploitation phases.
known location, as seen in Eq. (4) (Abdollahzadeh et al., 2021). The POA seems to be more effective in solving phase
problems, but findings in this research showed that it could
(UB LB )*r1 + LB , rand < P
sometimes be trapped in the local minimum, depending on
xit + 1 = (r2 C )*xrt L *B , rand 0.5
the nature of the problem and the stopping criteria. On the
xit L *(L *(xit xrt ) + r3*(xit xrt ), rand < 0.5 (4) other hand, the GTO has a fast convergence rate, but the
probability of locating the global minimum of some problems
where xit + 1 is the Gorilla individual position at t + i iterations,
is low.
r1, r2 and r3 are random parameters which range from 0 to 1.
These new algorithms combine the POA’s exploration
The P value should be chosen between 0 and 1. LB and UB are
strength with the GTO’s exploitative strength. A good ex­
the lower and upper limits, respectively. xrt is a randomly
plorative technique ensures that the search space is well
selected member of the gorilla troop at time t.
combed for possible solutions, while a robust exploitative
t process will intensify the explored space for the best solu­
C = F* 1
T (5) tion. The proposed algorithms ensure a dynamic balance and

F = cos(2*r4 ) + 1 (6)
166 Chemical Engineering Research and Design 194 (2023) 163–178

dimension and the constraints for phase stability and phase


equilibrium for nonreactive and reactive problems, are given
in the following sections.

3.1. Phase stability

A chemical system’s phase stability determines whether it


will separate into various phases under a specific condition.
It is often a precursor to phase equilibrium calculations. The
objective function employed for solving phase stability issues
is frequently the tangent plane distance function (TPDF). It is
the vertical distance between the tangent of the Gibbs energy
surface at the feed composition and any other location on the
Gibbs energy surface. Anywhere else on the Gibbs free energy
surface where the tangent contacts, indicate that the system
is unstable at that composition (Michelsen,1982). The pur­
pose of solving PS problems is to identify the objective
function’s global minimum values. The system is unstable if
this value is negative; otherwise, it is stable. Eq. (18) is the
TPDF formulation (Michelsen,1982; Sun and Seider, 1995;
Fig. 1 – Pseudo-Code for PGOA. Harding and Floudas, 2000).
c
TPDF = yi (µi |y µi |zi )
i (18)
stability between the extensive and intensive phases. The i

proposed algorithms integrate the explorative ability of the where c is the number of components in the system, z and y
POA, as given in Eq. (1), with the exploitative power of GTO, are the feed mole fractions and the mole fractions at the split
as given in Eqs. (10) and (13), which were modified to Eqs. 16 point, µi | y and µi |zi are the chemical potentials at the split
i
and 17 , and were used in place of Eqs. (10) and (13) for better and feed composition, respectively.
performance. The pseudo-code and flowchart are shown in The chemical potential is expressed in terms of the fu­
Figs. 1 and 2, respectively. gacity coefficient and activity coefficient in Eq. (19):
Two sets of hybrid algorithms were developed and stu­
xi ˆ i
died; the first is called PGOA1, while the second is called µi = µio + RTln = µio + RTln(xi i )
PGOA2. In both modifications, the explorative phase adapts i (19)
the exploration phase of POA, which is given in Eq. (1). The Where µio ,
ˆ i , i , i are the chemical potential of pure
exploitative phase is a modification of that of GTO. component, fugacity coefficient, the pure fugacity coefficient,
In PGOA1, in the exploitative stage, if C ≥ W, the silverback and the activity coefficients for component i respectively.
leads, but if C < W, the female competition is enhanced, and The fugacity and activity coefficient can be evaluated via the
this is done using Eq. (16). equation of state and activity models, respectively.
xit + 1 Eq. (20) is the reformulation of the TPDF given in Eq. (18).

t
xbest t
[xbest *(2*rand 1) xit *(2*rand 1)]* µi µio xi ˆ i
= ln = ln(xi i )
= RT i (20)
*g , C W
t t
xbest [xbest *(2*rand 1) xit *(2*rand 1)]* *E , else The mole fractions, x, are often the decision variables
(16) used to minimise the problem when solving the TPDF. A new
parameter can be introduced into the formulation for ex­
The objective’s fitness will determine each member’s
pressing the mole fractions of component i and a percentage
status in Eq. (16), and the update will be done, while the best
of the previous feed composition. Another variable that can
solution will be saved as the global minimum.
be utilised is the number of moles (n).
In PGOA2, a modified exploitative phase is given in Eq.
Stability problems can be solved by globally minimising
(17), and it replaces Eq. (16) of PGOA1.
Eq. (18) subject to the constraint of mole balance as shown in
xit + 1 Eq. (21).

( )*(2*rand
t c
xit + 0.2* 1 tmax
1)*xit , C W yi = 10 yi 1for i = 1, …c
= (21)
t t i= 1
xbest [xbest *(2*rand 1) xit *(2*rand 1)]* *E , else
With a new decision variable i , the problem’s constraint
(17)
formulation given in Eq. (21) can be changed to an un­
Eqs. (16) and (17) are followed by a fitness estimation and constraint formulation.
update to get the global minimum values.
niy = i zi nT (22)

3. Formulation of PS, PE and rPE problems niy


yi = c
i = 1 niy (23)
The description of the global optimisation problems, in­
cluding the objective functions, decision variables, This reformulation can be applied to more phases.
Chemical Engineering Research and Design 194 (2023) 163–178 167

Fig. 2 – The flowchart for PGOA.

3.2. Phase Equilibrium a constraint to an unconstraint optimisation problem. i, j is


the fraction of component i in the phase j.
Phase equilibrium calculation involves knowing the compo­ ni, j
sitions in all the phases in an unstable system. A phase i, j = k=
k =j nik (26)
equilibrium problem can be formulated into an optimisation
problem by minimising the Gibbs free energy of all phases.
The reformulation by using i, j as a decision variable re­
Phase equilibrium reduced Gibbs free energy formulation duces the complexity of the optimisation problems and
is given as (Rangaiah, 2001; Srinivas and Rangaiah, computational time without violating the mass balance
2007a,2007b; Fernández-Vargas et al., 2013): constraint.
c c The number of moles ni ,j of any component i in any phase
xi ˆ i
g= nij ln(xi i ) = nij ln j can be related using Eq. (27):
j i j i i (24)
K =j 1
Eq. (24) is subjected to mass balance constraint, which is ni ,j = i .j zi nT nik
k =1 (27)
given as:
With the inclusion of Eq. (27), the direct relationship of ni ,j
ni, j = zi nF , 0 ni, j nF to the initial feed composition is established.
j= 1 (25)
For the last phase , Eq. (27) can be written as:
where π is the number of phases involved and nF is the total K= 1
number of moles for the feed. The introduction of i, j [0,1], ni , = zi nT nik
k =1 (28)
which is given in Eq. (26) can be used to convert Eq. (25) from
168 Chemical Engineering Research and Design 194 (2023) 163–178

3.3. Phase equilibrium calculations for chemically

(Bonilla-Petriciolet et al., 2006)

(Bonilla-Petriciolet et al., 2006)


(McDonald and Floudas, 1997)
(McDonald and Floudas, 1995)
reacting systems

(Harding and Floudas, 2000)


(Sun & Seider, 1995; Balogh

(Ammar and Renon, 1987)


The reactive phase equilibrium is the determination of

References

(Nichita et al., 2002)


equilibrium compositions in the presence of a reaction. The
inclusion of reaction kinetics makes the problem more
complicated than the nonreactive system. The objective

et al., 2003)
function for a reactive system with multiple components,
phases and reactions r, phase equilibrium is given by Eq. (29)
(Srinivas and Rangaiah, 2007a,2007b; Bonilla-Petriciolet et al.,
2011; Bhargava et al., 2013).

-0.032466

-0.015767

-0.003932

-0.000002

-0.002688

-1.486205

-2.05E-05
-0.29454
GT = g ln K eqN 1nref , j

Global optimum

PS
j =1 (29)

Where g is given in Eq. (24), K eq


is the equilibrium constant, N

-0.35296
-0.54779

-0.01989

-1.18365

-0.83878

-0.76977

-1.12118
is a square matrix of the stoichiometric coefficients of re­

-0.0202
PE
ference components in r reactions. nref , j is a matrix with a
single column that relates other components with the re­
ference components. The extent of reaction and reduction in

SRK EoS with classical mixing rules (Hua et al.,

SRK EoS with classical mixing rules (Hua et al.,


NRTL model (Heidemann and Mandhane, 1973)
decision variables are usually done with the introduction of
the reference components and transformed variables (Ung
and Doherty, 1995).

NRTL model(Bender and Block, 1975)

SRK EoS with classical mixing rules

SRK EoS with classical mixing rules

SRK EoS with classical mixing rules

SRK EoS with classical mixing rules


Thermodynamic models
In reactive systems, the rate of production and reaction
must be accounted for, and this is not similar to phase
equilibria without reactions in which the total mole is con­
stant. The composition of a component in a given phase is
the difference in the number of moles of the component in
the feed with that of other phases. But, in reactive systems,
the rate of consumption and reaction must be accounted for.
1998a,1998b)

1998a,1998b)
The stoichiometric coefficients can effectively be used for
accounting for the changes in composition due to reactions,
as shown in Eq. (30).
n 1
ni = zi nT vi N 1 (n nref , ) (ni , j vi N 1nref , j )
ref ,
nF = (0.401, 0.293, 0.199, 0.0707, 0.0363) at 390 K and 5583 kP

j =1

nF = (0.6436, 0.0752, 0.0474, 0.0412, 0.0297, 0.0138, 0.0303,


(30)
nF = (0.29989, 0.20006,0.50005) at 298 K and 101.325 kPa

nF = (0.7212, 0.09205,0.04455, 0.03123, 0.01273, 0.01361,

nF = (0.614, 0.10259,0.04985, 0.008989, 0.02116, 0.00722,


The value of ni the number of moles of component i in
0.01187,0.01435, 0.16998) at 314 K and 2010.288 kPa
the last phase π, is significant in formulating the optimisa­
tion problem for phase equilibrium. It is the difference be­
nF = (0.30, 0.10, 0.60) at 270 K and 7600 kPa

nF = (0.9813, 0.0187) at 190 K and 4053 kPa

tween the components in all the other phases and the feed
0.07215, 0.01248) at 353 K and 38,500 kPa
Feed Conditions

nF =(0.5, 0.5) at 298 K and 101.3250kPa

compositions. During optimisation, it can have a negative


value, which is an infeasible region of the search environ­
0.0371, 0.0415, 0.0402) at 435.35 K

ment. A penalty is included in the objective function to cor­


rect this anomaly, as shown in Eq. (31).

Fobj if ni, j > 0 for i = 1, 2…C and j = 1, 2…


f= otherwise
Table 1 – Details of studied PS and PE problems.

Fobj + Penalty
(31)
where Penalty is the penalty function given in Eq. (32)
(Srinivas and Rangaiah, 2007a,2007b).
nunf
Penalty = 10. |ni |
C1 +C2 +C3 + iC4 + C4 + iC5 + C5

i =1 (32)
n-Butyl acetate + H2O (LLE)

C1 +C2 +C3 + C4 + C5 + C6

C1 +C2 +C3 + C4 + C5 + C6

where nunf is the number of ni


Toluene + H2O + aniline

that have negative va­


lues, ni < 0 .
C2 +C3 + C4 + C5 + C6

+ C7 + C8 + C9 + C10
+ C6 + C7 −16 + C17+
System

+ C7 −16 + C17+

4. Implementation of the methods


N2 + C1 + C2

C1 + H2S

The studied algorithms were used to solve several phase


(VLE)

(VLE)

(VLE)

stability, nonreactive phase equilibrium, and reactive phase


equilibrium benchmark problems, as shown in Tables 1 and
2. In using the algorithms, the objective function of PS pro­
No

1
2
3

blems to be minimised is the TPDF given in Eq. (18); the


Table 2 – Detailed of phase equilibrium of chemically reacting systems.
No System Feed Condition Model Global Reference
minimum

1 Ethanol + acetic acid ↔ ethyl acetate + water (LLE) nF = (0.5, 0.5, 0.0, 0.0) NRTL model and ideal gas − 2.058120 (Peng Lee et al., 1999)
at 355 K and Keq= 18.670951
101.325 kPa
2 Isobutene + methanol ↔ methyl tert-butyl ether with n- F = (0.30, 0.30, 0.0, 0.40) Wilson model and ideal gas. − 1.434267 (McDonald and Floudas, 1997)
butane as an inert component at 373.15 K and Keq= 15.6465
(LLE) 101.325 kPa
Wilson model and ideal gas.
Grxs◦/R =
− 4205.05 + 10.0982 T − 0.2667
3 2-Methyl-1-butene + 2-methyl-2-butene + 2 *methanol ↔ (0.354, 0.183, Keq= 36.6646 − 1.226367 (Ung and Doherty, 1995) (Vianna Neto and
2 *tert-amyl methyl ether 0.463, 0.0) at 355 K Oliveira, 2017)
(reactive with liquid–liquid equilibrium) and 151.95 kPa
4 Acetic acid + n-butanol ↔ water + n-butyl acetate(LLE) = (0.3, 0.4, 0.3, 0.0) UNIQUAC model and ideal gas − 1.106300 (Wasylkiewicz and Ung, 2000)
at 298.15 K and Keq= 10.0674 (Vianna Neto and Oliveira, 2017)
101.325 kPa
5 (reactive with liquid–liquid equilibrium) (0.6, 0.4, 0.0) Margules solution model − 0.144508 (Bonilla-Petriciolet et al., 2008)
Keq= 0.9825 (Vianna Neto and Oliveira, 2017)
6 2-Methyl-1-butene + 2-methyl-2- nF = (0.1, 0.15, 0.7, Wilson model and ideal gas − 0.872577 (Bonilla-Petriciolet et al., 2008) (Vianna Neto
butene + 2 *methanol ↔ 2 *tert-amyl methyl ether 0.0, 0.05) at 335 K Keq= 36.6646 and Oliveira, 2017)
with n-pentane as an inert component and 151.9875 kPa
(reactive with liquid–liquid equilibrium)
7 A1 + A2 ↔ A3(LLE) nF = (0.52, 0.48, 0.0) Margules solution model. Keq − 0.653756 (Wasylkiewicz and Ung, 2000) (Bonilla-
Chemical Engineering Research and Design 194 (2023) 163–178

at 323.15 K and = 3.5 Petriciolet et al., 2011)


101.325 kPa
8 A1 + A2↔ A3 + A4 (0.048, 0.5, 0.452, NRTL model 0.311918 (Wasylkiewicz and Ung, 2000) (Bonilla-
0.0) at 360 K and Keq= 4.0 Petriciolet et al., 2011)
101.325 kPa
169
170 Chemical Engineering Research and Design 194 (2023) 163–178

decision variables are the mole fractions which are the po­ 1
j ( )= size { problem : ij }
sition of members in the algorithms. The objective functions Np (36)
for PE and rPE are given in Eqs. (24) and (29), respectively.
j ( ) is the cumulative probability of a solver j to optimise
MATLAB software was used for coding, implementing,
all the problems under consideration within the set value of
and computing all the problems, thermodynamic models
, while size is the number of problems with ij . The plot of
and algorithms. The algorithms’ ability to solve all problems
the j ( ) versus is known as the performance profile. The
was carried out using two major terminating criteria.
higher the pp, the better the performance of the solver (Dolan
Implementing the first stopping criteria (SC1) terminates the
& More, 2002). As soon the stochastic optimisation is over,
optimisation process when the maximum number of itera­
the final values of the decision variable for the best minimum
tions is reached. While in the second stopping criteria (SC2),
of an iteration becomes the initial values for the local opti­
the program terminates, if there is no significant change in
misation. Sequential quadratic program (SQP) was the local
the value of the objective function after a declared number
optimiser used in this study.
(SCmax) of successive iterations. The algorithms that were
HP EliteBook 840 G4 computer, Processor Intel(R) Core(TM)
considered are population-based, a fixed population (NP)
i5–7300 U CPU @ 2.60 GHz, 2712 Mhz, 2 Core(s), 4 Logical
= 10D was used for all the simulations, where D is the
Processor(s) and 16 GB was used for all computations.
number of decision variables. The study considered 24 pro­
blems with different degrees of computational complexity, 8
for each set of problems, as shown in Tables 1 and 2. Each 5. Results and discussion
problem was solved 100 times with a different set of random
populations, and the performance of each algorithm was 5.1. The performance of all the algorithms on PS problems
evaluated based on their success rate in obtaining the global
minimum within the two stopping criteria. The number of The ability of all the stochastic algorithms with and without
objective functions evaluation (NFE) for each problem and a local intensifier to solve phase stability problems of dif­
algorithms using SC2 was also considered. A run is suc­ ferent dimensions and complexity, as given in Table 1, was
cessful if the tolerance i.e. the difference between its best explored using the two stopping criteria described in
global value and the known global minimum, is less than the Section 4.
set tolerance limits. PS5 and PS8 were assigned a 10−6, while PS1 is a phase stability problem of a dimension of 2, and it
others were assigned a tolerance of 10−5. The Global success has low computational resolution (Fonseca-Pérez et al.,
rate (GSR) was also used to compare the performances of all 2021), the result without a local optimiser shows that at the
algorithms for all the problems. The GSR is shown in Eq. (33): 10th iteration, the SR of POA, PGOA2, GTO and PGOA1 are 22
np
SRi %, 74 %, 84 % and 100 %, respectively. From 50 iterations and
GSR = above, all the stochastic algorithms attained an SR of 100.
i=1 np (33)
The average SR of all the algorithms for SC1 is 84.4, 94.8, 96.8
Where SRi is the success rate for problem i and np is the total and 100 for POA, PGOA2, GTO and PGOA1, and PGOA1 has the
number of problems. best average SR. The effects of stopping criteria SC2, using
The performance profile (PP) by Dolan and Mor (2002) was SCmax of 10, 20 and 50, were explored and shown in Table 3.
also used in comparing the performances of all the algo­ At SCmax = 10, all the algorithms except PGOA1 have an SR of
rithms. There is a need for a performance matrix in evalu­ 100 %. GTO has the smallest average value of NFE, 46, while
ating the PP, a performance matrix contains the performance GPOA2, POA, and PGOA1 have average NFE of 54, 58 and 65,
measure of interest, which could be the computation time, respectively. Including a local optimiser improved the SR to
NFE or objective function deviation from the global value. In 100 % for all cases in SC1 and SC2. These results show that all
this study, the element of the performance matrix i, j was the algorithms are reliable in the accurate prediction of the
chosen as a deviation between the mean values of the cal­ point of instability and composition of the components of
culated global minimum fij obtained in solving problem i and PS1 system.
solver j and the known global minimum value f j* as shown in PS2 has a medium computational complexity (Fonseca-
Eq. (34). For a total number of solver Ns and problems Np, i ϵ Pérez et al., 2021), as the number of iterations increases, the
Np and j ϵ Ns SR of all the algorithms increases. At 50 iterations, GTO and
i, j = fij f j* (34) PGOA1 have an SR of 98 %, while POA and PGOA2 have an SR
of 100 %. For SC2, all the algorithms have an average SR of
The smallest value of i, j of all solvers for a given problem 100 % except for PGOA1, with an average SR of 96 %, and SR
serves as the basis for comparison with the performance of
value of 90 % at SCmax = 10. The presence of an intensifier
all other solvers. As shown in Eq. (35).
equates the SR of all the algorithms for all cases in SC1 and
The solver with the smallest value of i, j for a particular
SC2. The accuracy of the algorithms in locating the global
problem was used as the basis of comparison as shown in
minimum of PS2 problems is an indicator that the predicted
Eqs. (35) and (36).
compositions at the point of instability are correct.
i* = min( i , j ), 1 j Ns (35) PS3 is a low computational resolution problem (Fonseca-
Pérez et al., 2021), using SC2 for stochastic algorithms
i, j
ij = i Np, j Ns, without a local optimiser shows that PGOA1 has the highest
i* (36)
SR of 93 % at the 10th iteration. POA, GTO, PGOA2 and PGOA1
The performance ratio ij is the ratio of i, j all solvers for have average SR of 85.4 %, 96.8 %, 96.8 % and 98.6 %, re­
the particular problem to the best solver with i* for the spectively. In SC2, PGOA1 has an average SR of 96.7 %, which
problem. The solver with the best performance will have a is the lowest of all the algorithms, while the other algorithms
value of 1. The fraction of all problems Np in which a solver j have an average value of 100 %. Including local optimisation
with ij , is given as j as seen in Eq. (36). has little effect on SC2 but is more effective in SC1.
Chemical Engineering Research and Design 194 (2023) 163–178 171

Table 3 – SR and NFE for PS Problems with SC2.


D Scmax POA GTO PGOA1 PGOA2

SR NFE SR NFE SR NFE SR NFE

10 100 3588 100 2810 95 3358 100 3392


PS1 2 25 100 5751 100 4240 100 6815 100 5174
50 100 8172 100 6794 100 9219 100 7741
10 87 5815 98 3586 70 5143 94 5452
PS2 3 25 99 8794 100 5129 100 10,323 100 7415
50 100 11,490 99 7622 100 12,944 100 10,217
10 100 4969 100 2955 90 5642 100 4308
PS3 3 25 100 6957 100 4441 100 8686 100 6023
50 100 9386 100 6994 100 11,021 100 8473
10 97 2908 98 2246 99 2767 98 2885
PS4 2 25 100 5568 100 3724 100 4515 100 4963
50 100 8605 100 6232 100 7093 100 7364
10 24 6105 100 5921 65 11,544 75 8823
PS5 5 25 81 17,718 98 7268 100 21,431 96 14,822
50 95 32,114 100 10,151 100 22,258 95 18,223
10 7 4314 92 12,289 0 970,050 62 9234
PS6 8 25 22 13,793 100 16,674 0 200,000 99 22,210
50 40 38,537 100 19,136 0 200,000 100 26,568
10 100 1200 100 1200 100 1200 100 1200
PS7 9 25 100 2700 100 2700 100 2700 100 2700
50 100 5200 100 5200 100 5200 100 5200
10 1 8320 72 42,177 17 63,821 19 18,243
PS8 10 25 4 18,293 85 59,032 30 152,670 90 104,442
50 8 44,084 74 64,097 39 143,398 99 116,093

Problem PS4 has similar computational complexity to PS1, classified by Fonseca-Pérez et al. ( 2021). The implementation
in solving PS4 with the algorithms, using SC1 without a local of the first stopping criteria showed that at the 10th iteration,
optimiser shows average values of 98.4 %, 99.8 %, 99.8 % and all algorithms failed to locate the global minimum, while at
100 % for POA, GTO, PGOA2 and PGOA1, respectively. For SC2 the 50th iteration, PGOA1 had the highest SR of 14 %. As the
in Table 3, all the algorithms have an average SR of 99 %. The iteration increases, the SR of POA, GTO, and POA2 increases
intensification by local optimisation improved the SR of all while that of POA1 decreases. The average SR in SC1 for
the algorithms. PGOA2, GTO, POA and PGOA1 are 50 %, 39 %, 30.5 % and 30 %,
Problem PS5 has a medium computational resolution respectively. In implementing SC2, as shown in Table 3, the
(Fonseca-Pérez et al., 2021), at the 10th iteration, all the al­ SR of POA was very poor in all the cases, while GTO had a
gorithms failed to solve the problem, but as the SC1 in­ relatively constant SR of 74 %, 85 % and 74 % for SCmax of 10,
creased, there was an improvement in the SR of POA, while 25 and 50, respectively. The SR of PGOA1 was too low, with an
the SR of GTO, PGOA1 and PGOA2 were fairly constant. The average of 28.6 %. At SCmax = 10, the SR of PGOA2 was 19 %,
same trend was experienced in the implementation of SC2, but it became 99 % at SCmax = 50. With local optimisation, the
as seen in Table 3, as SCmax increased. In both SC1 and SC2, SR of the algorithms were significantly changed, though the
POA has the lowest SR. SR of PGOA1 trailed the others. There was a sharp contrast in
PS6 is computationally demanding with high resolution the SR of POA for SC1 and SC2 for PS8, the SR of the former
(Fonseca-Pérez et al., 2021), using SC1, the average SR of was far better than that of the latter; this anomaly affects the
PGOA1, POA, GTO and PGOA2 are 7.4 %, 62.4 %, 69 % and 77 %, overall GSR of POA for all the PS problems in SC2.
respectively. The result of implementing SC2 in Table 3 Considering the GSR of all PS problems in SC1 without the
shows that GTO has the highest SR of 97. 3 % followed by 87 local optimiser, as seen in Fig. 3a, the GSR increases as the
% of PGOA2, while POA and PGOA1 performed poorly. In­ iterations increase. At the 10th iteration, PGOA1 had a GSR of
cluding a local optimiser improved the performances of POA, 51.8 %, which was the highest among all the algorithms. The
GTO and PGOA2, and it did not affect the performance of trend changed at 50 iterations; the GSR of PGOA2 and GTO
PGOA1. This shows that the inclusion of local optimisation were higher than that of PGOA1. At the 100th iteration, the
may not be necessary for PGOA1. To correctly predict the GSR of all the algorithms was better than that of PGAO1. This
point of instability and the composition of various compo­ trend was maintained till the 1000th iteration, with PGOA2
nents in this system, GTO and PGOA2 can be used with an having the highest GSR of 98.3 %. Fig. 3, plot b shows that
average NFE of 200 iterations. with the intensification by a local optimiser, the reliability of
PS7 has a high computational resolution, and despite its all the algorithms increases irrespective of the number of
high computational demand, all algorithms have fast con­ iterations. The GSR of all the algorithms was greatly en­
vergence with high success rates for all the considered hanced and looked alike except for PGOA1, which trailed the
stopping criteria. This shows that all the algorithms can ac­ others.
curately predict the split point for this system, and the need The result of the implementation of SC2 are shown in
for a local optimiser is not necessary. Table 3; as SC2 increases, the GSR and overall NFE increase.
PS8 is computationally challenging and belongs to the At SCmax= 50, the GSR of PGAO1, POA, GTO and PGOA2 are 80
group with the high computational resolution, as recently
172 Chemical Engineering Research and Design 194 (2023) 163–178

In PS problems, different algorithms performed differ­


ently for various problems with varying criteria of termina­
tion. The GSR of all the algorithms in SC2 was slightly better
with fewer NFE, except for POA, which was affected by the SR
of PS8 in SC2. The implication of the wrong global solution
for PS stability problems will lead to wrong phase split and
wrong compositions at the point of instability. PGOA2 has the
best performance and reliability for PS problems.

5.2. The performance of all the algorithms on PE problems

The studied PE problems have different degrees of compu­


tational difficulties, which are similar to the PS problems.
The global solutions predict the composition of all the equi­
librium phases at phase split.
For PE1, the performances of the stochastic algorithms
alone show that in SC1, the SR increases as the iteration in­
creases. The rate increase of the SR with iterations was faster
in GTO, PGOA1 and PGOA2 than in POA. The average SR for
all the algorithms are 35 %, 68.2 %, 71.8 % and 76 % for POA,
GTO, PGOA2 and PGOA1, respectively. The inclusion of local
Fig. 3 – The GSR versus the number of Iteration for PS optimisation in SC1 brought a slight increase in the SR of the
Problems with SC1 (a) Stochastic Algorithms only (b) algorithms, with an average SR of 52.2 %, 79.4 %, 85.6 % and
88.8 % for POA, GTO, PGOA2 and PGOA1, respectively. The
Stochastic and local optimisation.
implementation of SC2 without a local optimiser was not
favourable to all the algorithms, as seen in Table 4; they all
%, 80.3 %, 96.6 % and 99.3 %, respectively. Intensification with have an average SR of 43 % except for POA, with an average
local optimiser increases the GSR slightly. SR of 33 %. Local optimisation in SC2 brought a slight in­
For the performance profile (PP) of all the algorithms for crease in the SR of the algorithms.
all the PS problems, The PP of PGOA2 was higher and better For problem PE2, in implementing SC1, without a local
than others, closely followed by that of GTO, while the PP of optimiser, the success rate of GTO was lower than that of
PGOA1 and POA are similar. PGOA1 has the fastest con­ others; the average SR were 49 %, 79 %, 81.8 % and 84.6 % for
vergence profile to the global minimum for most PS pro­ GTO, POA, PGOA1 and PGOA2, respectively. The presence of a
blems. This was closely followed by the convergence profile local optimiser for SC1 improved the performances of all the
of PGOA2, while the convergence profile of GTO was slightly algorithms. The result of the SC2 for all the algorithms is
better than that of POA. shown in Table 4, the average SR of all algorithms in SC2

Table 4 – SR and NFE of all algorithms using SC2 For PE Problems.


Scmax POA GTO PGOA1 PGOA2

D SR NFE SR NFE SR NFE SR NFE

10 31 3172 32 3318 14 2614 29 3071


PE1 2 25 35 6188 48 6303 37 9702 45 7539
50 33 9412 55 9141 78 21,878 56 12,333
10 87 6534 55 3373 74 16,473 95 6959
PE2 3 25 100 12,540 56 4962 98 32,577 100 9079
50 100 16,290 49 7589 96 55,225 100 11,922
10 97 6178 98 3457 87 7174 100 6066
PE3 3 25 99 10,090 100 4996 100 11,554 100 8019
50 100 13,876 98 7536 100 14,194 100 10,823
10 100 3898 100 2932 99 4182 100 3702
PE4 2 25 100 6247 100 4280 100 8184 100 5612
50 100 9017 100 7001 100 11,295 100 8222
10 80 7499 75 6661 60 11,161 94 10,672
PE5 5 25 96 17,543 79 9542 100 25,656 100 22,690
50 100 34,556 68 13,601 100 28,856 100 33,271
10 1 4630 73 18,468 24 7346 24 11,440
PE6 8 25 1 11,753 79 22,398 96 31,500 85 73,149
50 3 33,911 81 26,691 100 36,117 98 94,556
10 0 4590 99 14,818 47 9663 38 8069
PE7 9 25 0 11,942 98 16,373 99 26,607 93 26,166
50 1 31,556 99 20,367 100 30,763 99 32,318
10 1 5702 100 31,918 32 10,024 42 11,397
PE8 10 25 2 13,820 100 46,871 100 53,203 91 72,353
50 6 35,603 100 89,478 100 61,472 100 193,195
Chemical Engineering Research and Design 194 (2023) 163–178 173

without a local optimiser are 53.3 %, 89.5 %, 95.7 % and 98.3 %


for GTO, PGOA1, POA and PGOA2, respectively. GTO has the
lowest NFE, which was followed by PGOA2. The best algo­
rithm for PE2 using SC2 without a local optimiser is POA2
because it has the highest SR and low NFE. With the inclu­
sion of local optimisation, the success rate of all the algo­
rithms improved and are similar.
In solving PE3, the average SR of all algorithms using SC1
without a local optimiser are 84.2 %, 91.6 %, 95.4 % and 96.4 %
of POA, PGOA1, PGOA2 and GTO, respectively. These SRs
were high and similar, though POA is slightly lower than the
rest. Including local optimisation brought a slight improve­
ment in the SR of all the algorithms. Implementing SC2
without local optimisation, as seen in Table 4, the SR of all
the algorithms were higher than that of SC1.
The solution of problem PE4 for all cases in SC1 and SC2
shows that all the algorithms performed well with high SR,
which are pretty similar.
In problem PE5, as SC1 increases, the SR of all the algo­
rithms increases, though the SR of GTO were slightly lower
than the rest. The average SR is 63.8 %, 74.8 %, 77.8 % and 80
% for GTO, POA, PGOA1 and PGAO2, respectively. The im­
plementation of SC2 with a local optimiser brought about Fig. 4 – The GSR versus the number of Iteration for PE
average SR of 74 %, 92 %, 86.7 % and 98 % for GTO, PGOA1, Problems with SC1 (a) Stochastic Algorithms only (b)
POA and PGOA2, respectively. Stochastic and local optimisation.
For problem PE6, in implementing SC1 without a local
optimiser, the SR of POA was very poor for the set of SC1 until The performance profile of PGOA2 was far better than the
the 500th iteration, when it became 88 %. The SR of GTO has rest, and it was faster and higher. The PP of POA was better
a slight increase, with the highest value of 77 % at 500 than that of GTO, while PGOA1 has the lowest performance
iterations, while the SR of PGOA1 and PGOA2 were better profile. In most problems, PGOA1 and PGOA2 convergence
than the others. The implementation of SC2, as seen in profiles are similar and faster than that of POA and GTO.
Table 4, was not favourable to POA; the SR were very poor These results also showed that to predict the points of
and far below the SR of SC1. The SR for all the other algo­ equilibrium and the compositions of all the components in
rithms were better than that of POA. Including a local opti­ all the phases for PE systems, PGOA2 has the highest relia­
miser improved the SR in all cases; this improvement was bility and performance with an average NFE of 50 and 100
more pronounced in SC2 of POA, with a wide difference in iterations for SC1 stopping criterion with and without local
the SR, it is most likely that the convergence rate of POA was optimization, respectively.
slow, and it was trapped in a local minimum with proximity
to the global minimum, and this led to the predictions of
compositions that are slightly different from the equilibrium 5.3. The performance of all the algorithms on rPE
compositions of the global solutions. The performances of problems
the algorithms for PE7 and PE8 in all the cases are similar to
that of PE6. rPE problems are computationally demanding; the studied
The plot of the overall performance of all the algorithms problem has different degrees of computational complexity.
for all PE problems in case SC1 is shown in Fig. 4a, at SC1 = 10, The presence of a chemical reaction increases the com­
the GSR of GTO, PGOA1 and PGO2 are quite similar, while plexity because the reaction’s extent must be accounted for,
that of POA was lower. At SC1 = 50, PGOA1 has the highest thereby increasing the problems’ dimensionality.
GSR, and PGOA2 is closely followed; the same trend was For problem rPE1, in implementing SC1 without a local
experienced when SC1 = 100. At SC1 = 500, the GSR of PGOA2 optimiser, there was an increase in all the SR of all the al­
was slightly higher than that of PGOA1, and this trend was gorithms; the average SR of POA, GTO, PGOA1 and PGOA2 are
maintained at the 1000th iteration. At 1000 iterations, the 58.4 %, 60.2 %, 78.4 % and 78.4 % respectively. The im­
GSR of POA, GTO, PGOA1, and PGO2 are 87 %, 88 %, 97.8 % and plementation of SC2 shown in Table 5 was far better than
99.5 %, respectively. Including local optimisation improved SC1, with an average SR of 92 %, 76 %, 100 % and 100 % for
the GSR for all the iterations, as seen in Fig. 4b. The result for POA, GTO, PGOA1 and PGOA2, respectively. Including local
the implementation of SC2 without a local optimiser is optimisation improved the performance of all the algorithms
shown in Table 4; at SC2 = 10, GTO has the highest GSR, at for SC1 and that of POA and GTO.
SC2 = 25, PGOA1 has the highest GSR, and PGAO2 closely For rPE2, implementing the SC1, at the 10th iteration, all
followed, while the GSR of GTO trails that of PGOA1 and the algorithms failed to locate the global minimum. As the
PGOA2, and POA has the worst GSR. The GSR of all algorithms number of iterations increased, there were improvements in
at SC2 = 50 was similar to that of SC2 = 25. The NFE of SC2 can the SR of all the algorithms. The average SR for all the algo­
be seen in Table 4; at SC2 = 10, POA has the lowest NFE, rithms are 52.8 %, 60 %, 76.2 % and 78.4 % for GTO, POA,
though the GSR value was also the lowest; therefore, its PGOA1 and PGOA2, respectively. The implementation of SC2
performance does not supersede that of others. with local optimisation was favourable to all the algorithms,
the SR were far better than those in SC1 and with fewer NFE.
174 Chemical Engineering Research and Design 194 (2023) 163–178

Table 5 – SR and Total NFE of algorithms for rPE problems.


D Scmax POA GTO PGOA1 PGOA2

SR NFE NFE SR NFE SR NFE

6D 79 24,609 72 10,283 100 23,616 100 22,849


rPE1 5 12D 97 54,570 81 13,467 100 27,430 100 27,430
24D 100 101,520 75 19,525 100 32,047 100 33,135
6D 68 21,021 62 9761 100 33,810 100 22,859
rPE2 5 12D 90 63,131 58 13,662 99 36,329 100 26,435
24D 97 112,337 59 19,108 100 43,193 100 34,047
6D 100 20,222 99 12,259 100 19,444 100 24,085
rPE3 5 12D 100 41,927 100 14,156 100 22,734 100 29,609
24D 100 74,049 100 20,940 100 29,556 100 34,539
6D 81 49,178 31 17,081 18 157,285 93 28,839
rPE4 5 12D 95 73,920 47 26,372 18 160,185 95 34,766
24D 92 106,009 52 34,856 13 173,166 96 41,556
6D 100 14,482 97 6495 100 15,957 100 11,371
rPE5 4 12D 99 26,104 98 8689 100 17,974 100 14,262
24D 100 40,554 99 13,483 100 23,047 100 19,376
6D 38 31,667 16 9273 100 33,931 100 27,663
rPE6 5 12D 72 87,190 25 14,210 100 39,294 100 34,036
24D 81 149,441 20 20,668 100 44,985 100 38,869
6D 70 19,461 71 8963 78 24,119 92 17,149
rPE7 4 12D 88 36,529 66 11,269 74 27,698 92 19,232
24D 97 66,258 63 16,060 75 31,887 94 24,854
6D 87 20,116 61 9051 100 21,242 91 17,972
rPE8 5 12D 96 51,426 71 13,286 100 24,934 96 21,388
24D 100 95,032 74 20,105 100 32,808 97 28,224

Including local optimisation in both SC1 and SC2 sig­ was greatly enhanced for all algorithms. The implementation
nificantly increased the SR of GTO. of SC2 without a local optimiser is shown in Table 5. While
For rPE3, the performances of all the algorithms in all the SR for POA1 and POA2 were better and higher than that of
cases are pretty similar to the performance in problem rPE2. SC1, there is no significant improvement in the SR of POA
However, the performance of GTO was better in rPE3 than and GTO.
in rPE2. For rPE7, the result of the implementation of SC1 without
rPE4 is the most computationally challenging problem out a local optimiser shows that the average SR of all the algo­
of all the reactive problems considered because it has a rithms 48.3 %, 50.8 %, 58 % and 67.8 % for GTO, POA, PGOA1
prominent local minimum point value of − 1.10326 (Bamikole and PGO2, respectively. The inclusion of a local optimiser
and Narasigadu, 2023). Most of the algorithms get trapped using SC1 had little effect on the SR of all the algorithms, it
against the global maximum point of − 1.10630 shown in improved the average SR to 60 %, 87.6 %, 72.4 % and 80.8 %,
Table 2. In the implementation of SC1 without a local opti­ for GTO, POA, PGOA1 and PGO2, respectively. This improve­
miser, the average SR of all the algorithms are 7.4 %, 36.6 %, ment was better enhanced for POA than the other algo­
42.8 % and 62 % for GTO, PGAO1, and POA2, respectively. The rithms. The effect of implementing SC2 without a local
inclusion of a local optimiser for SC1 improved the average optimiser can be seen in Table 5; the SR of the algorithms are
SR to 48 %, 51.4 %, 78 % and 82 % for GTO, PGOA1, and PGOA2, 66.7 %, 85 %, 75.7 % and 92.7 % for GTO, POA, PGOA1 and
respectively. The result of implementing SC2 without a local PGO2, respectively. The inclusion of a local optimiser to SC2
optimiser is shown in Table 5; the result shows that the led to average SR values of 66.7 %, 93.6 %, 77.3 % and 93 %;
average SR of GTO, PGOA1 and PGOA2 is 43 %, 16 %, 89 % and these effects were very negligible in GTO, PGOA1 and PGOA2
94.7 %, respectively. The SR of GTO was better in the im­ but more pronounced in POA.
plementation of SC2 than in SC1, while in the POA1, it was In rPE8, the implementation of SC1 without including a
otherwise. The inclusion of local optimisation improves the local optimiser, the average SR was 65.2 %, 77.4 %, 80.2 % and
SR of all algorithms in SC2, but the SR of PGOA1 was not 87.4 % for the GTO, PGOA2, POA and POA and PGOA1, re­
affected. spectively. The effect of SC2, without a local optimiser,
In rPE5, the SR of the algorithms was high in all the cases; yielded SR averages of 68.7 %, 94.6 %, 94.3 % and 100 % for the
the average SR in the implementation of SC1 without a local GTO, PGOA2, POA and PGOA1, respectively. While the SR of
optimiser are 81.6 %, 83.8 %, 89.2 % and 92.5 % for POA, GTO, SC2 was more favourable for POA, PGOA1 and PGOA2 than
PGOA1 and PGOA2 respectively. The performance in SC2, as that of SC1 but GTO had a slight improvement. Including
seen in Table 5 was slightly better than SC1 for all the algo­ local intensifiers in SC1 and SC2 increased and equated the
rithms. SR of all the algorithms to 100 %.
For rPE6, in the implementation of the SC1 without a local The overall performance of all the algorithms for all rPE
optimiser, the performance of POA, PGOA1 and PGOA2 look problems shows that as the number of algorithms increases,
alike, while that GTO was poor. The average SR of all the the GSR of all the algorithms increases though at different
algorithms are 15.4 %, 46.8 %, 70.4 %, and 77.6 % for GTO, rates, as seen in plot of GSR against the number of iterations
POA, PGOA1 and PGOA2, respectively. When a local optimiser in Fig. 5a. The rate increase of POA was the most steady as
was included to intensify the result of SC1, the performance
Chemical Engineering Research and Design 194 (2023) 163–178 175

Fig. 5 – The GSR versus the number of iterations for rPE Fig. 6 – The GSR(plot a) and NFE(plot b) of stochastic
Problems with SC1 (a) Stochastic Algorithms only (b) algorithms for all problems with SC2.
Stochastic and local optimisation.
5.4. The performance of all the algorithms for all
the iteration increased. At the 10th iteration, the GSR was 2 % problems
but increased to the highest GSR of 97.6 % at the 1000th
iteration. At the same time, PGOA2 had a GSR value of 21 % at The overall performances of all the algorithms were com­
the 10th iteration and had a sudden leap to 79.1 % at the 50th pared for all 24 problems. The comparison was based on the
iteration, an additional GSR value of 12 % was added at the SC2, and the plot of the GSR against the different values of
100th iteration. Beyond the 500th iteration, it maintained a SCmax is shown in Fig. 6a. From the plot, it can be seen that in
GSR average of 97.5 %. GTO and PGOA1 have a similar rate of the first set of SCmax, the GSR of POA, PGOA1, GTO, and
increases in GSR as the iteration increases. Beyond the 10th PGOA2 are 64 %, 69.5 %, 79.2 % and 81 %, respectively; PGOA2
iteration, the GSR value was slightly maintained as the has the highest GSR, which GTO closely followed. As the
iteration progressed, though the performance of POA1 was SCmax progressed, the previous trend in the GSR of the al­
better than GTO. The effect of the local optimisation on SC1 gorithms changed; though PGOA2 still retained the highest
can be seen in Fig. 5b; all the algorithms have similar per­ GSR, the GSR of PGOA1 was slightly better than that of GTO,
formance, though the GSR of GTO was slightly lower than the while the GSR of POA was the lowest in all the iterations. The
rest. It was also seen that the effect of the number of itera­ effect of intensification of a local optimiser on the GSR of all
tions was negligible. the algorithms led to a similar performance in all cases,
The result of the implementation of SC2 without local though PGOA2 still retained the highest GSR.
optimisation is shown in Table 5. The GSR was fairly con­ The plot of NFE is shown in Fig. 6b, it shows that GTO has
stant as SCmax increased from 6D to 24D for most of the al­ the lowest NFE, which means that it is computationally less
gorithms except for POA, in which the increase of GSR was expensive. The NFE of the PGOA2 is slightly higher than that
more significant from SCmax of 6D to 12D. The inclusion of a of GTO, while the NFE of PGOA1 and POA are higher than that
local optimiser in SC2 yielded a GSR average of 86.7 %, 92 %, of PGOA2 and POA.
98.5 % and 97.5 % for PGOA1, GTO, PGOA2 and PAO, respec­ Without intensification, the PP of PGOA was far above that
tively. of others, followed by that of POA, while GTO and PGOA1
The performance profile POA and PGOA2 are similar and have similar PP. Including a local intensifier improves the PP
better than that of GTO and PGOA1, though the PP of GTO of all the algorithms and reduces the difference between the
was slightly higher than that of PGOA1. The convergence rate PP of PGOA2 and others, though PGOA2 still has the
of PGOA2 is the fastest and the most consistent for most of highest PP.
the rPE problems. The accuracy in the obtaining the global The combination of the GSR, NFE and PP shows that
solutions for rPE systems, will in turn yield the correction PGOA2 has the best performance for all problems; though the
composition of all components in the all the phases at NFE of GTO was the smallest, its GSR was lower than that of
equilibrium, the extent of reaction and the reaction equili­ PGOA2, which makes the performance of PGOA2 more robust
brium constant. and reliable than the other algorithms.
176 Chemical Engineering Research and Design 194 (2023) 163–178

Table 6 – Comparison of the studied algorithms with other methods in the literature.
PS PE rPE All Problems

GSR NFE GSR NFE GSR NFE GSR NFE

IDE 87 20,821 99.8 8629 89.4 19,648 92.1 16,366


CS 81 27,167 99 35,208 90 12,161 90.0 24,845
KH 89.5 1406 99 3543 90 3062 92.8 2670
LKH 81 614 98 2506 89 3122 89.3 2081
HB 85.9 211 96.9 254 84.4 299 89.1 255
HH 92.9 235 86.3 338 77.9 364 85.7 312
POA 80.4 197 55.4 230 95.9 932 77.2 453
GTO 96.6 158 81.3 227 67.8 206 81.9 197
PGOA1 79.9 514 96.8 325 86 513 87.6 451
PGOA2 99.3 250 94.1 496 97.3 354 96.9 367

5.5. The comparison with the performance of other comparison, the result of stochastic optimisation without the
algorithms reported in the literature inclusion of local optimisation in this study were used for
comparison, also the result of stochastic algorithms is sol­
It is essential to compare the performance of the applied and ving the problem without local optimisation indicate the
proposed algorithms to other methods available in the lit­ reliability of the algorithms in find the solutions ( Khalil et al.,
erature. Several algorithms have been used to solve phase 2016). The comparison for the studied algorithms with other
stability and equilibrium problems. In order to have a fair methods in the literature is shown using the SCmax = 50
basis for comparison, the selected algorithms were chosen stopping criteria is shown in Table 6.
because; the same set of problems was solved, the same kind The results of the comparison for PS problems are shown
of problem formulation was used, the same set of stopping in Table 6. From the result of the ten algorithms presented,
criteria were also used, and the comparison was done based the GSR of PGOA2 was the highest, and this was followed by
on the GSR after solving each problem 100 times with a dif­ that of GTO. The NFE of GTO was the lowest, though the NFE
ferent set of random initialisation. The NFE was also a factor of HB and POA were lower than that of PGOA2, the GSR of the
that was used for comparison, NFE is a good yardstick for latter was relatively higher. This shows that GTO and PGOA2
comparison of computational efficiency, and it is in­ are very reliable in solving PS problems with low computa­
dependent of the computer and computation software tional cost and will favourably compete with most methods
(Fateen et al., 2012). without local intensification.
The bases above were used in selecting the following al­ The performance of all the algorithms for the PE problems
gorithms for comparison with the studied algorithms in this can be seen in Table 6. The GSR of PGOA2 with local opti­
work. In the work by Zhang et al. (2011), four algorithms, misation is 99.9 %, which was similar to the GSR of IDE, and
namely; unified bare-bones particle swarm optimisation, surpasses that of the other algorithms. The NFE of PGOA2 is
integrated differential evolution (IDE) and IDE without tabu lower than that of IDE and other algorithms with similar GSR,
list and radius, were studied and used to solve the same set and this makes PGOA2 to be computationally less expensive
of phase equilibrium and stability problems. Their finding than the others for solving PE problems.
shows that IDE has the highest SR and the lowest NFE, and it The comparison of the algorithms for solving rPE in
was termed the best. In the application of the Covariance Table 6 shows that PGOA2 has the highest GSR with a small
Matrix Adaptation Evolution Strategy (CMAES), Shuffled NFE; therefore, it is highly robust and more computationally
Complex Evolution (SCE) and Firefly Algorithms (FA) for sol­ less expensive than most of the algorithms in solving rPE
ving the same set of problems, the results were compared problems.
with the performance of IDE, and IDE performed better than The comparison of the performance of the algorithms for
CMAES, SCE and FA. (Fateen et al., 2012). The performance of all problems is shown in Table 6. The GSR of the algorithms
IDE surpasses several other algorithms and therefore it was shows that the overall performance of PGOA2 surpasses that
chosen for comparison with the algorithms in this study. of most algorithms in solving all the problems. It is also
In the study by Khalil et al. (2016), Gravitational Search, computationally less expensive, reliable, and robust in opti­
Monkey, and Krill Herd Swarm Algorithms were studied for mising phase stability and equilibrium problems.
the same set of problems, all the algorithms were compared
with the cuckoo search (CS) algorithms. The result shows 6. Conclusion
that CS was better than the other algorithms. In a recent
study by Bamikole and Narasigadu (2023), four different al­ In this study, two recently developed stochastic global opti­
gorithms, Pathfinder, Honey Badger, Red Fox and Horse Herd, misation algorithms, POA and GTO, were studied to solve
were examined, the findings showed that honey badger and multidimensional and challenging phase stability, phase and
horse herd algorithms were better than the others. chemical equilibrium problems. To improve the perfor­
Therefore, in concise but detailed comparison, integrated mances of these algorithms, two new hybrid algorithms,
differential evolution (IDE), cuckoo search (CS), Krill Herd namely PGOA1 and PGOA2 were developed in this study.
(KH), Lévy-Flight Krill Herd (LKH), Honey Badger (HB) and These newly developed hybrid algorithms combine the ex­
horse herd (HH) were used for comparison with the present plorative ability of POA and the exploitative ability of GTO for
study. The result presented for IDE, and CS include local improved performance. The effect of different stopping cri­
optimisation, while others do not. To have a fair basis for teria, the number of iterations and local optimiser inclusion
Chemical Engineering Research and Design 194 (2023) 163–178 177

were explored. Most of the algorithms studied require a energy minimization using simulated annealing for two-
reasonable number of iterations for optimum performance phase equilibrium calculations in reactive systems. Chem.
Biochem. Eng. Quart. 22, 285–298.
and therefore caution should be taken in implementing the
Bonilla-Petriciolet, A., Rangaiah, G.P., Segovia-Hernández, J.G.,
stopping criteria. Their performances vary with the nature of
2011. Constrained and unconstrained Gibbs free energy
the problems. Though the performance of GTO and POA is minimisation in reactive systems using genetic algorithm and
good, considering the GSR of all the problems, GTO slightly differential evolution with tabu list. Fluid Phase Equilib. 300,
outperformed POA. The SC2 was more favourable for the 120–134. https://doi.org/10.1016/j.fluid.2010.10.024
performance of GTO than POA, and the inclusion of a local Bonilla-Petriciolet, A., Segovia-hernández, J.G., 2010. A com­
optimiser (Sequential Quadratic Program) improved the parative study of particle swarm optimisation and its variants
for phase stability and equilibrium calculations in multi­
performances of both algorithms. The hybrid algorithms,
component reactive and non-reactive systems. Fluid Phase
PGOA1 and PGOA2 were better than that of GTO and POA for Equilib. 289, 110–121. https://doi.org/10.1016/j.fluid.2009.11.
all the problems combined. Though PGOA1 seems to have an 008
overall performance that is slightly above that of GTO and Bonilla-Petriciolet, A., Vázquez-Román, R., Iglesias-Silva, G.A.,
POA, the performance seems inconsistent and can be more Hall, K.R., 2006. Performance of stochastic global optimisation
computationally expensive than both. For all cases and per­ methods in the calculation of phase stability analyses for
formance indices, PGOA2 is more robust and reliable, and it nonreactive and reactive mixtures. Ind. Eng. Chem. Res. 45,
4764–4772. https://doi.org/10.1021/ie051081g
outperforms all the other algorithms in locating the points of
Dolan, E.D., Mor, J.J., 2002. Benchmarking Optimisation Software
instability, and the composition of all components in all the with Performanc. pp. 201–213.
phases at Equilibruim. Comparison with other studied algo­ Elnabawy, A.O., Fateen, S.E.K., Bonilla-Petriciolet, A., 2014. Phase
rithms in the literature shows that PGOA2 has a higher stability analysis and phase equilibrium calculations in re­
probability than most algorithms in locating the global active and non-reactive systems using Charged System
minimum with very little deviation from the known global Search algorithms. Ind. Eng. Chem. Res. 53, 2382–2395.
Fateen, S.-E.K., Bonilla-Petriciolet, A., Rangaiah, G.P., 2012.
minimum value, equilibrium compositions and it is less
Evaluation of covariance matrix adaptation evolution
computationally expensive than most algorithms.
strategy, shuffled complex evolution and firefly algorithms for
PGOA2 is highly reliable without the presence of a local phase stability, phase equilibrium and chemical equilibrium
intensifier. The performance of PGOA2 without a local opti­ problems. Chem. Eng. Res. Des. 90, 2051–2071.
miser surpasses that of several algorithms with local in­ Fateen, S.E.K., Bonilla-Petriciolet, A., 2015. On the performance of
tensification. Therefore PGOA2 is highly reliable, less swarm intelligence optimisation algorithms for phase stabi­
computationally expensive, and does not necessarily need lity and liquid-liquid and vapor-liquid equilibrium calcula­
tions. Period. Polytech. Chem. Eng. 59, 186–200. https://doi.org/
the enhancement of a local optimiser. Its implementation is
10.3311/PPch.7636
quite easy, and its application can be extended to other Fernández-Vargas, J.A., Bonilla-Petriciolet, A., Segovia-
global optimisation and multidimensional problems, as such Hernández, J.G., 2013. An improved ant colony optimisation
equipment design, chemical kinetic modelling, parameter method and its application for the thermodynamic modeling
evaluation for thermodynamic models and other forms of of phase equilibrium. Fluid Phase Equilib. 353, 121–131.
global optimisation problems. https://doi.org/10.1016/j.fluid.2013.06.002
Fonseca-Pérez, R.M., Bonilla-Petriciolet, A., Tapia-Picazo, J.C., Jaime-
Leal, J.E., 2021. A reconsideration on the resolution of phase sta­
Declaration of Competing Interest bility analysis using stochastic global optimisation methods:
Proposal of a reliable set of benchmark problems. Fluid Phase
The authors declare that they have no known competing fi­ Equilib. 548. https://doi.org/10.1016/j.fluid.2021.113180
Harding, S.T., Floudas, C.A., 2000. Phase stability with cubic
nancial interests or personal relationships that could have
equations of state: global optimisation approach. AIChE J. 46.
appeared to influence the work reported in this paper.
Heidemann, R., Mandhane, J., 1973. Some properties of the NRTL
equation in correlating liquid–liquid equilibrium data. Chem.
References Eng. Sci. 28, 1213–1221.
Hua, J.Z., Brennecke, J.F., Stadtherr, M.A., 1998a. Enhanced in­
terval analysis for phase stability: cubic equations of state
Abdollahzadeh, B., Soleimanian Gharehchopogh, F., Mirjalili, S.,
models. Ind. Eng. Chem. Res. 37, 1519–1527.
2021. Artificial gorilla troops optimiser: a new nature-inspired
Hua, James Z., Brennecke, J.F., Stadtherr, M.A., 1998b. Reliable
metaheuristic algorithm for global optimisation problems. Int.
computation of phase stability using interval analysis: cubic
J. Intell. Syst. 36, 5887–5958. https://doi.org/10.1002/int.22535
equation of state models. Comput. Chem. Eng. 22, 1207–1214.
Ammar, M., Renon, H., 1987. The isothermal flash problem: new
https://doi.org/10.1016/S0098-1354(98)00024-6
methods for phase split calculations. AIChE J. 33, 926–939.
Jalali, F., Seader, J.D., Khaleghi, S., 2008. Global solution approaches in
Balogh, J., Csendes, T., Stateva, R.P., 2003. Application of a sto­
equilibrium and stability analysis using homotopy continuation
chastic method to the solution of the phase stability problem:
in the complex domain. Comput. Chem. Eng. 32, 2333–2345.
cubic equations of state. Fluid Phase Equilib. 212, 257–267.
https://doi.org/10.1016/j.compchemeng.2007.12.001
https://doi.org/10.1016/S0378-3812(03)00266-8
Khalil, A.M.E., Fateen, S.E.K., Bonilla-Petriciolet, A., 2016.
Bamikole, J.O., Narasigadu, C., 2023. Application of Pathfinder,
Gravitational search, monkey, and krill herd swarm algo­
Honey Badger, Red Fox and Horse Herd algorithms to phase
rithms for phase stability, phase equilibrium, and chemical
equilibria and stability problems. Fluid Phase Equilib. 566,
equilibrium problems. Chem. Eng. Commun. 203, 389–406.
1–17.
https://doi.org/10.1080/00986445.2015.1004666
Bender, E., Block, U., 1975. Thermodynamische Berechnung der
Khalil, A.M.E., Fateen, S.E.K., Bonilla-Petriciolet, A., 2016.
Flüssig-Flüssig-Extraktion. Verfahrenstechnik 9, 106–111.
Gravitational search, monkey, and krill herd swarm algo­
Bhargava, V., Fateen, S.E.K.K., Bonilla-Petriciolet, A., 2013. Cuckoo
rithms for phase stability, phase equilibrium, and chemical
Search: a new nature-inspired optimisation method for phase
equilibrium problems. Chem. Eng. Commun. 203, 389–406.
equilibrium calculations. Fluid Phase Equilib. 337, 191–200.
https://doi.org/10.1080/00986445.2015.1004666
https://doi.org/10.1016/j.fluid.2012.09.018
Bonilla-Petriciolet, A., Bravo-Sánchez, U., Castillo-borja, F.,
Frausto-hernández, S., Segovia-Hernández, J., 2008. Gibbs
178 Chemical Engineering Research and Design 194 (2023) 163–178

Khalil, A.M.E., Fateen, S.E.K., Bonilla-Petriciolet, A., 2015. MAKHA- Sun, A.C., Seider, W., 1995. Homotopy-continuation method for
A new hybrid swarm intelligence global optimisation algo­ stability analysis in the global minimisation of the Gibbs free
rithm. Algorithms 8, 336–365. https://doi.org/10.3390/a8020336 energy. Fluid Phase Equilib. 103, 213–249.
McDonald, C.M., Floudas, C.A., 1997. GLOPEQ: a new computa­ Tessier, S.R., Brennecke, J.F., Stadtherr, M.A., 2000. Reliable phase
tional tool for the phase and chemical equilibrium problem. stability analysis for excess Gibbs energy models. Chem. Eng. Sci.
Comput. Chem. Eng. 21, 1–23. https://doi.org/10.1016/0098- 55, 1785–1796. https://doi.org/10.1016/S0009-2509(99)00442-X
1354(95)00250-2 Trojovský, P., Dehghani, M., 2022. Pelican optimization algorithm:
McDonald, C.M., Floudas, C.A., 1995. Global optimisation for the a novel nature-inspired algorithm for engineering applica­
phase stability problem. Am. Inst. Chem. Eng. J. 41, 1798–1814. tions. Sensors 22. https://doi.org/10.3390/s22030855
Michelsen, M.L., 1982. The isothermal flash problem. Part I. Ung, S., Doherty, M.F., 1995. Vapor-liquid phase equilibrium in
Stability. Fluid Phase Equilib. 9, 1–19. https://doi.org/10.1016/ systems with multiple chemical reactions. Chem. Eng. Sci. 50,
0378-3812(82)85001-2 23–48. https://doi.org/10.1016/0009-2509(94)00180-Y
Moodley, K., Rarey, J., Ramjugernath, D., 2015. Application of the Vianna Neto, M.R., Oliveira, E.D., 2017. Chemical and phase
bio-inspired Krill Herd optimisation technique to phase equilibrium calculations by Gibbs energy minimisation using
equilibrium calculations. Comput. Chem. Eng. 74, 75–88. deterministic methods based on globally convergent branch
https://doi.org/10.1016/j.compchemeng.2014.12.008 and bound algorithms. Fluid Phase Equilib. 447, 95–106.
Nichita, D.V., Gomez, S., Luna, E., 2002. Phase stability analysis https://doi.org/10.1016/j.fluid.2017.05.020
with cubic equations of state by using a global optimisation Wakeham, W.A., Stateva, R.P., 2004. Numerical solution of the
method. Fluid Phase Equilib. 194–197, 411–437. https://doi.org/ isothermal, isobaric phase equilibrium problem. Rev. Chem.
10.1016/S0378-3812(01)00779-8 Eng. 20, 1–56.
Peng Lee, Y., Pandu Rangaiah, G., Luus, R., 1999. Phase and che­ Wasylkiewicz, S.K., Li, Y.K., Satyro, M.A., Wasylkiewicz, M.J.,
mical equilibrium calculations by direct search optimisation. 2013. Application of a global optimisation algorithm to phase
Comput. Chem. Eng. 23, 1183–1191. https://doi.org/10.1016/ stability and liquid-liquid equilibrium calculations. Fluid
S0098-1354(99)00283-5 Phase Equilib. 358, 304–318. https://doi.org/10.1016/j.fluid.
Rangaiah, G.P., 2001. Evaluation of genetic algorithms and si­ 2013.08.030
mulated annealing for phase equilibrium and stability pro­ Wasylkiewicz, S.K., Ung, S., 2000. Global phase stability analysis
blems. Fluid Phase Equilib. 187–188, 83–109. https://doi.org/10. for heterogeneous reactive mixtures and calculation of re­
1016/S0378-3812(01)00528-3 active liquidliquid and vapor-liquid-liquid equilibria. Fluid
Srinivas, M., Rangaiah, G., 2007a. A study of differential evolution Phase Equilib. 175, 253–272.
and tabu search for benchmark, phase equilibrium and phase Wolpert, D.H., Macready, W.G., 1997. No free lunch theorems for
stability problems. Comput. Chem. Eng. 31, 760–2007. optimisation. IEEE Trans. Evol. Comput. 1, 67–82.
Srinivas, Mekapati, Rangaiah, G.P., 2007b. Differential evolution Zhang, H., Fernández-Vargas, J.A., Rangaiah, G.P., Bonilla-
with tabu list for global optimization and its application to Petriciolet, A., Segovia-Hernández, J.G., 2011. Evaluation of
phase equilibrium and parameter estimation problems. Ind. integrated differential evolution and unified bare-bones par­
Eng. Chem. Res. 3410–3421. ticle swarm optimisation for phase equilibrium and stability
problems. Fluid Phase Equilib. 310, 129–141. https://doi.org/10.
1016/j.fluid.2011.08.002

You might also like