Computers and Chemical Engineering 180 (2024) 108501

Contents lists available at ScienceDirect

Computers and Chemical Engineering

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

A hybrid deterministic-stochastic algorithm for the optimal design of

process flowsheets with ordered discrete decisions
David A. Liñán a, Gabriel Contreras-Zarazúa b, c, Eduardo Sánchez-Ramírez b, Juan
Gabriel Segovia-Hernández b, Luis A. Ricardez-Sandoval a, *
a
Department of Chemical Engineering, University of Waterloo N2L 3G1, Canada
b
Departamento de Ingeniería Química, Universidad de Guanajuato GTO 36050, Mexico
c
Centro de Innovación Aplicada en Tecnologías Competitivas (CIATEC) GTO 37545, Mexico

A R T I C L E I N F O A B S T R A C T

Keywords: This work presents a hybrid stochastic-deterministic algorithm for optimal design of process flowsheets, i.e.,
Discrete-steepest descent algorithm finding the optimal design variables and operating conditions of multiple interconnected units using rigorous
Differential evolution with tabu list phenomenological chemical engineering models. Unlike previous studies that propose hybrid deterministic and
Hybrid deterministic-stochastic optimization
stochastic algorithms in sequential and nested arrangements, the present work proposes a parallel configuration
Distillation flowsheet
to perform the hybridization. The proposed hybrid algorithm combines a stochastic method (SM) with the
Optimization
deterministic Discrete-Steepest Descent Algorithm with Variable Bounding (DSDA-VB). The SM and DSDA-VB
strategies interact in parallel by exchanging new feasible solutions identified by the SM and improved search
bounds determined by the DSDA-VB. The proposed method is illustrated using a thermally coupled system and a
sequence of reactive, extractive, and traditional distillation columns. The results indicate that the proposed al­
gorithm outperforms the traditional Differential Evolution with Tabu List (DETL) algorithm, showing faster and
improved convergence.

1. Introduction literature (Aliaga-Vicente et al., 2017; Ma et al., 2021; Ma and Li, 2022;
Zhang et al., 2018). This work is focused on flowsheet design optimi­
Flowsheet optimization is typically regarded as a two-stage process zation problems, which are regarded as Mixed-Integer Nonlinear Pro­
involving the optimal synthesis problem and the optimal design prob­ gramming (MINLP) problems. Formulating and initializing MINLP
lem. The former (i.e., flowsheet synthesis) determines the type of op­ flowsheet design problems may require modeling expertise, and con­
erations and how they are interconnected to achieve a desired structing accurate mechanistic models for systems with complex ther­
production target from raw materials and energy inputs, e.g., see Cre­ modynamic equilibrium and hydrodynamic relationships can be
maschi (2015), Gooty et al. (2022), Göttl et al. (2022), Ryu et al. (2020), challenging and time-consuming. Typically, these problems are
and Tula et al. (2017). The latter (i.e., flowsheet design) aims to deter­ non-convex and challenging to solve due to the nonlinearity of non-ideal
mine the optimal operating conditions and other operation/equipment thermodynamic models, mass, and energy balances in chemical pro­
related variables (e.g., discrete and continuous design variables) of the cesses; the inherent non-convexity of discrete variables, and the non­
synthetized process with a fixed flowsheet, e.g., see Contreras-Zarazúa linearities that arise from the interactions of discrete and continuous
et al. (2019), Gómez et al. (2006), Hong et al. (2019), Jia et al. (2023), variables. Hence, optimizing the design of rigorous flowsheet models
Peccini et al. (2023). The flowsheet synthesis problem often considers can turn into a computationally expensive and time-consuming process.
shortcut or simplified models that allow the screening of several flow­ In this work, a hybrid algorithm that parallelizes stochastic and deter­
sheet configurations, which are later evaluated and optimized in the ministic optimization methods is proposed to exploit the advantages of
flowsheet design stage using rigorous nonlinear models (Ramapriya stochastic and deterministic techniques and to reduce the computational
et al., 2018). Alternatively, the simultaneous synthesis and design effort required to optimize the design of chemical process flowsheets
problem using rigorous nonlinear models has also been studied in the using rigorous process models. Chemical engineering software with

* Corresponding author.
E-mail address: [email protected] (L.A. Ricardez-Sandoval).

https://doi.org/10.1016/j.compchemeng.2023.108501
Received 27 May 2023; Received in revised form 12 October 2023; Accepted 1 November 2023
Available online 4 November 2023
0098-1354/© 2023 Elsevier Ltd. All rights reserved.
D.A. Liñán et al. Computers and Chemical Engineering 180 (2024) 108501

Fig. 1. Hybrid arrangements for optimal flowsheet design aided with simulation software. A: sequential hybrid, B: nested hybrid, and C: parallel hybrid. DOF means
degrees of freedom.

simulation and Nonlinear Programming (NLP) optimization capabilities optimization with discrete variables fixed is used for each solution
is considered for modeling purposes. candidate in an inner loop. Although these memetic algorithms have
The available optimization strategies to solve the MINLP flowsheet demonstrated improvements in solution quality and computational
design problem are generally classified into deterministic and stochastic performance, they fail to guarantee local optimality for discrete DOF in
(Segovia-Hernández et al., 2015). Both deterministic and stochastic the inner loop. To address this issue, Kruber et al. (2021) and Skibor­
optimization techniques exhibit limitations when solving flowsheet owski et al. (2015) used GAMS to investigate the inclusion of both
optimization problems. Local optimization strategies highly depend on discrete and continuous DOF when evaluating solution candidates in the
the initialization of continuous and discrete decisions, due to “zero flow” inner loop. To avoid the rigorous solution of an MINLP for each solution
numerical issues (Liñán and Ricardez-Sandoval, 2023). Also, finding a candidate, Skiborowski et al. (2015) recommend a continuous refor­
global solution to non-convex problems is not guaranteed with local mulation of discrete variables such as feed locations and number of
solvers (Kronqvist et al., 2019). Conversely, traditional global deter­ stages, which approximates the solution of the MINLP as a series of
ministic MINLP methods may be computationally prohibitive when NLPs.
dealing with large-scale flowsheet design problems that consider Considering a simulation software to define the flowsheet design
continuous and discrete degrees of freedom (DOF) and rigorous optimization problem exhibits a few attractive features. For instance,
nonlinear models for multiple processing units (Franke, 2017; Kruber they do not require the explicit formulation of the model equations and
et al., 2021). Furthermore, the performance of both local and global they allow to modify the underlying property methods, e.g., test
deterministic solvers may be impacted by the level of detail of the different thermodynamic packages. Also, models embedded in chemical
process models and the formulation selected to represent the in­ engineering software can be applied to a wide range of flowsheets with
teractions between continuous and discrete variables, e.g., Big-M, multiple unit operations and degrees of freedom (DOF) while preserving
convex hull, etc (Liñán et al., 2020). In contrast, stochastic methods the robustness of the models, thanks to their readily available unit
that use randomized search strategies and exploration techniques that operation, thermodynamic, and physical property models. Nevertheless,
are not tied to specific model assumptions or simplifications. This makes a practical limitation of a simulation tool is the lack of integration be­
stochastic methods especially suitable to consider the simulation soft­ tween mathematical programming techniques and process simulators.
ware as a black box. Despite their advantages, black box stochastic ap­ Chemical engineering simulation software are typically able to perform
proaches have other limitations such as a large number of function local NLP optimizations; nonetheless, most simulation tools do not have
evaluations, which may result in slow convergence, or refined param­ MINLP optimization capabilities (Franke, 2017; Hernández-Pérez et al.,
eter tuning to achieve optimal performance (Costa and Bagajewicz, 2020; Javaloyes-Antón et al., 2022). This hinders the application of
2019). To partially address these issues, previous works have suggested hybrid algorithms that rely on MINLP optimization to flowsheet design
different arrangements of hybrid deterministic-stochastic algorithms optimization with simulation software. For instance, the nested hybrid
that counterbalance the benefits and drawbacks of each strategy. The algorithm proposed by Skiborowski et al. (2015) relies on the contin­
hybrid strategies that are currently available for optimal flowsheet uous reformulation of a MINLP problem; however, traditional simula­
design aided with simulation software are shown in Fig. 1A and B, as tion tools do not allow the continuous reformulation of discrete
explained below. decisions within their corresponding NLP optimizer. To the authors’
The simplest hybrid deterministic-stochastic technique consists of knowledge, a hybrid deterministic-stochastic algorithm that can guar­
applying these techniques sequentially (Fig. 1A), where stochastic antee local optimality for both discrete and continuous DOF is not
optimization is used first, and the best solution found is further refined currently available to optimize the process design using chemical engi­
using a deterministic optimization solver, see e.g., Srinivas and Ran­ neering simulation software.
gaiah (2006), Munawar and Gudi (2005), Staudt and Soares (2009), This work aims to propose a new hybrid stochastic-deterministic
Chia et al. (2021), and Herrera Velázquez et al. (2022). Given that there algorithm to optimize the design of chemical process flowsheets
is no iterative interaction between the deterministic and the stochastic involving a mixture of continuous and ordered discrete decisions1,
step, the performance of neither of the algorithms improves. Moreover, which is a common characteristic in chemical engineering applications.
this sequential methodology would fail at providing local optimality Examples of ordered discrete decisions include selecting the number of
guarantees in flowsheet optimization problems implemented within a stages in multi-effect evaporator sequences (Hong et al., 2019), number
commercial simulator without MINLP capabilities. More advanced of reactors connected in series or parallel (Zhang et al., 2018), or the
strategies hybridize stochastic and deterministic strategies in a nested number of trays and location of interconnecting streams in separation
fashion (Fig. 1B), by keeping the stochastic method in an outer loop and
the deterministic strategy in an inner loop, see e.g., Urselmann et al.
(2011a, 2011b, 2016) , Zhou et al. (2017), Gómez et al. (2006), and 1
We refer to ordered discrete decisions instead of integer decisions, given
Holtbruegge et al. (2015). Most of those works implement a memetic
that, depending on the problem formulation, a reformulation step may be
algorithm, i.e., an evolutionary algorithm that optimizes the discrete (or needed to identify these ordered structures (Liñán and Ricardez-Sandoval,
a mix of discrete and continuous) DOF in an outer loop, while NLP 2023).

2
D.A. Liñán et al. Computers and Chemical Engineering 180 (2024) 108501

sequences (Ma et al., 2021). While previous works hybridize determin­ 2. Problem statement
istic and stochastic algorithms in sequential and nested arrangements
(Fig. 1A and B), the present work proposes a parallel configuration to This work considers the optimization problem presented in Eq. (1),
perform the hybridization, i.e., the stochastic algorithm runs in one which involves the minimization of an objective function (f), e.g.,
processor whereas the deterministic strategy runs in an alternate pro­ minimize capital and operating costs. Vector of variables x ∈ Rnx . rep­
cessor (Fig. 1C). The proposed hybrid algorithm combines a stochastic resents the continuous DOF of the problem, e.g., output flow specifica­
method (SM) with a deterministic optimization algorithm that is based tions, equipment sizes, energy requirements, among others. Variables in
on the previously developed deterministic Discrete-Steepest Descent y ∈ Zny . are the discrete DOF of the problem, i.e., a given combination of
Algorithm (DSDA) (Liñán et al., 2021, 2020). A new version of the y completely specifies structural decisions of the flowsheet such as
traditional DSDA that integrates this method with a variable bounding number of sequential/parallel units, feed locations, location of inter­
(VB) strategy that iteratively generates tighter search bounds and avoids connecting flows, etc. There are other continuous variables that are
convergence failures is presented herein. A key feature of this deter­ handled internally by the simulator. Among these, variables in vector
ministic algorithm is its capability to optimize discrete DOF within z ∈ Rnz . are the outputs of interest among the internally managed vari­
flowsheet simulation software. This capability eliminates the need to ables, e.g., output flows and compositions that are not specified by the
rely solely on a SM for optimizing discrete decisions. Rather than user.
focusing on the comparison of different hybrid strategies, this work
minf (x, y, z)
highlights the performance of the proposed parallel-hybrid strategy x,y,z
s.t. (1)
against a pure SM. The utilization of pure SMs, specifically the Differ­
xL ≤ x ≤ xU , yL ≤ y ≤ yU , g(x, y, z) ≤ 0, [x, y] ∈ Ω
ential Evolution with Tabu List (DETL), has undergone extensive testing
in solving highly nonlinear problems characterized by a large number of Different types of constraints are considered, including bounds on
equations, which include features such as discontinuous and the decision variables (xL ≤ x ≤ xU , yL ≤ y ≤ yU ), constraints with a
multi-objective functions (Contreras-Zarazúa et al., 2022; known mathematical expression (g(x, y, z) ≤ 0) and the constraints
González-Navarrete et al., 2022; Rangaiah, 2008; Rangaiah et al., 2020; handled internally by black-box simulations ([x, y] ∈ Ω). The upper (su­
Romero-García et al., 2022; Sánchez-Ramírez et al., 2022; Wang et al., perscript U) and lower (superscript L) bounds for continuous and
2020; Wang and Rangaiah, 2017). SMs have proven to be efficient at discrete DOF are denoted by xL ,xU and yL , yU respectively. Given that one
finding good quality solutions, thanks to their convergence proofs to a of the key features of the proposed hybrid algorithm is the iterative
global optimum in infinite time (Liberti and Kucherenko, 2005; Zie­ refinement of these bounds, their values will be updated throughout the
linski and Laur, 2008). Nonetheless, the computational effort required iterative solution of optimization problems. Thus, bounds xL,0 , xU,0 , yL,0
by these methods is often significant and is still an open challenge. This and yU,0 will be referred to as the user-defined bounds of the problem,
bottleneck has prompted the exploration of a parallel-hybrid method­ which impose limits over the iteratively refined algorithmic bounds as
ology, conceived with the objective of tackling the existing computa­ follows: yL,0 ≤ yL ≤ y ≤ yU ≤ yU,0 and xL,0 ≤ xL ≤ x ≤ xU ≤ xU,0 . Vector of
tional time limitations. By reducing computational time, the proposed constraints g includes those design requirements that are specified by
parallel-hybrid approach aims to make SMs more accessible and appli­ the user, e.g., product compositions. Other constraints are handled
cable in real-world scenarios. The key novelties of the proposed hybrid internally by the simulator, and these are represented through set Ω.
algorithm are summarized as follows: These constraints include mass and energy balances, equilibrium re­
lationships and thermodynamic models, reaction rates, etc. Set Ω is
• The proposed hybrid algorithm is the first methodology that executes therefore defined as those combinations in x and y that guarantee
a deterministic algorithm and a stochastic algorithm in parallel, convergence of all the constraints within the simulator. Throughout this
which is a feature that has not been exploited by the available hybrid work, it is assumed that objective function f is single-objective function
strategies for optimal design of process flowsheets that makes use of that is bounded from below, and that the effective domain of f is
chemical process simulators. This ultimately leads to solutions that nonempty. Also, functions f and g are assumed to be twice differentiable
satisfy local optimality for both discrete and continuous DOF, i.e., with respect to the continuous variables. Concerning the structure of the
concepts from discrete convex analysis (Murota, 2003) are used to DOF, the only requirement is that variables y must be ordered, which is a
verify the local optimality of discrete DOF, while the local optimality coon feature in process flowsheets such as sequences of intensified
requirement of continuous variables is assessed by an NLP solver. distillation systems, reactor networks, multi-effect evaporator se­
These optimality criteria are not guaranteed by the available meth­ quences, among others.
odologies for flowsheet design available in the literature. The solution strategy developed in this study aims to handle prob­
• The DSDA-VB is developed as a deterministic MINLP algorithm in­ lems involving multiple interconnected units in a flowsheet, with high
tegrated within the proposed hybrid method. This results in the first dimensionality in the discrete (y) and continuous (x) DOF. This work
application of a DSDA-based algorithm to optimize problems emphasizes the application of the proposed strategy to process optimi­
involving flowsheet simulators. The proposed DSDA-VB strategy al­ zation with flowsheet simulators using rigorous process models, thanks
lows to iteratively update the search region explored by the SM. The to the flexibility process simulators offer to include multiple operation
VB strategy also provides adequate variable bounds when the NLP units in a simulation, i.e., their readily available unit operations, non-
solver used by DSDA-VB is prone to convergence failure, which is a ideal thermodynamic and physical property models allow to simulate
common situation in process flowsheet simulation software. processes with multiple interconnected unit operations. Nevertheless,
the hybrid algorithm proposed in this work could also be applied in the
This study is organized as follows: Section 2 presents a general context of optimization modeling languages, e.g., GAMS or Pyomo,
description of the optimization problem addressed in this work. Section although these would require the users to build their own first-principles
3 describes the working principle of the SM and DSDA algorithms. This process models.
section also introduces the proposed hybrid algorithm and the DSDA-VB
strategy. Section 4 illustrates the main features of the proposed method 3. Mathematical framework
using a thermally coupled system and a sequence of reactive, extractive,
and traditional distillation columns as case studies. Concluding remarks The approach proposed in this work to solve problem (1) is a hybrid
and future areas of research are outlined in Section 5. deterministic-stochastic algorithm that combines an SM with the herein
proposed DSDA-VB strategy. An overview of stochastic and

3
D.A. Liñán et al. Computers and Chemical Engineering 180 (2024) 108501

deterministic optimization using the DSDA is provided in sections 3.1 DSDA outperforms traditional MINLP and GDP optimization algorithms
followed by the proposed hybrid algorithm presented in Section 3.2. when optimizing systems involving binary or Boolean variabs defined
over ordered discrete sets (Liñán et al., 2021, 2020; Liñán and Ricar­
3.1. Overview of stochastic methods and the DSDA dez-Sandoval, 2023). Note that several discrete analogues of convexity
such as integral-convexity, M♮-convexity and L♮-convexity have been
Stochastic methods (SMs) aim to find the optimal solution of a studied in the literature to detect when a local optimum according to
problem by iteratively generating and evaluating a set of candidate so­ Definition 1 is a global solution (e.g., see Chen and Li (2021); Murota
lutions based on random sampling. These algorithms use a probabilistic and Tamura (2023), and Zhang et al. (2022)); however, function f ∗ (y).
approach to explore the search space and escape local optima, rather does not have an analytical representation because it is obtained via NLP
than relying on deterministic approaches that often converge to a local optimization, i.e., the convexity properties of f ∗ (y) are unknown a priori.
solution. The operation of stochastic optimization algorithms typically For this reason, only local optimality can be guaranteed in this work.
involves four steps: initialization, generation, evaluation, and selection. Some of the limitations of stochastic and deterministic methods can
These steps are repeated until a satisfactory solution is found or a be overcome by combining them into a single optimization framework
stopping criterion is met. The initialization step involves generating an when solving flowsheet design problems. Typically, exceeding a
initial set of candidate solutions randomly. The generation step involves resource limit is considered as the stopping criterion for SMs, e.g., stop
creating new candidate solutions by applying random perturbations or once a maximum number of objective function evaluations is attained
mutations to the current population. The evaluation step involves (Fmax). However, selecting the appropriate resource limit (e.g., Fmax
calculating the objective value of each candidate solution using the value) requires trial-and-error tests, given that the computational per­
objective function and constraints defined for the problem. The objec­ formance depends on the specific characteristics of each optimization
tive value is used to assess the quality of the solution and compare it to problem, e.g., number of variables, non-linearities, and variable’s
other candidate solutions. The selection step involves choosing the best bounds. Additionally, the stopping criterion may differ in different runs
candidate solutions based on some criteria, such as the highest objective of the same problem due to the stochastic nature of SM algorithms
value, lowest cost, or shortest distance. Once the selection step is (Zielinski and Laur, 2008). A deterministic optimization algorithm such
completed, the algorithm repeats the generation, evaluation, and se­ as DSDA can help to alleviate these issues since it can be used to improve
lection steps until a satisfactory solution is found or a stopping criterion and assess the optimality characteristics of the current best solutions
is met, e.g., the number of iterations, the quality of the best solution identified by the SM. Hence, if the current SM solution is close to a
found, or a predefined threshold (Deb, 2009; Spall, 2003). locally optimal solution, then the execution of the SM can stop. Never­
The DSDA is a deterministic decomposition algorithm originally theless, the DSDA exhibits other limitations, such as the initialization
developed for integer programming (Murota, 2003), and recently requirement (i.e., a feasible combination of discrete decisions), or the
extended to MINLP optimization problems involving continuous and exponential growth of discrete neighbors when using a ℓ∞ -norm to
ordered discrete decisions such as those emerging in vector y (Liñán define the vicinity of a discrete point (Liñán and Ricardez-Sandoval,
et al., 2020). To initialize the DSDA, the user must provide a feasible 2023)The DSDA may benefit from SMs, which use randomness and
initialization for the variables in y. At each iteration, the DSDA fixes the probabilistic techniques to search for feasible and optimal solutions.
values of discrete decisions at their current trial value (denoted by y∗ ) in These methods do not guarantee finding optimal solutions, but they can
Problem (1) and solves for the remaining continuous variables x and z be efficient in exploring large search spaces and finding acceptable
using NLP optimization to return an objective function value f ∗ (y∗ ) = feasible solutions. In the case of constrained optimization problems,
f(x∗ , y∗ , z∗ ) with optimized continuous variables x∗ an z∗ . The NLP stochastic optimization methods typically use a penalty function
optimization problem that results from fixing discrete decisions in approach, where the objective function is modified to include a term
Problem (1) is referred to as a subproblem. Discrete variables are then that penalizes infeasible solutions. This penalty term increases as the
fixed at every discrete point within a neighborhood of y∗ denoted as solution violates constraints, and the objective function is minimized
NΘ (y∗ ), where Θ indicates the neighborhood type. For example, this while keeping the penalty as low as possible. Another approach used in
neighborhood typically considers those discrete points whose ℓ∞ -norm stochastic optimization for constrained problems is the use of constraint
(infinity norm) with respect to y∗ is less or equal to l, and it is denoted by handling techniques, such as the use of repair mechanisms to transform
N∞ (y∗ ). To reduce the computational effort, subsets of N∞ (y∗ ) such as infeasible into feasible solutions. These mechanisms can modify the
the N2 -neighborhood (N2 (y∗ )) with 2ny + 1 points within an ℓ2 -norm of solution to satisfy the constraints, but they must be designed carefully to
y∗ are often considered in practice. Each neighbor y ∈ NΘ (y∗ ). results in avoid introducing new infeasibilities. Concerning the size of the neigh­
an NLP subproblem that returns a new objective function value f ∗ (y) = borhoods, subsets of N∞ (y∗ ) such as NΘ (y∗ ) = N2 (y∗ ). have been
f(x∗ (y), y, z∗ (y)), where notation x∗ = x∗ (y) and z∗ = z∗ (y) represents the considered in previous DSDA works to decrease computational costs, at
optimized values of x and z with discrete variables y fixed in Problem the expense of scarifying local optimality within the larger N∞ (y∗ )
(1). Given that discrete variables in y are ordered, concepts from discrete neighborhood. To alleviate this issue, both the SM and the DSDA can
convex analysis can be used to define local optimality, as stated in perform the neighborhood exploration, by exploring NΘ (y∗ )⊂N∞ (y∗ )
Definition 1. neighborhoods with DSDA and heuristically using the SM to find
candidate solutions within the larger N∞ (y∗ ) neighborhood.
Definition 1. Consider function f ∗ (y) = f(x∗ (y), y, z∗ (y)), where func­
One problem that affects both the DSDA and SMs is the adequate
tion f ∗ : Zny ↦R ∪ {+∞} takes the optimal objective function of the
selection of bounds to perform the search (i.e., xL,0 , xU,0 , yL,0 and yU,0 ).
subproblem obtained after fixing y in Problem (1) for feasible values of y
Bounds on optimization variables limit the search space of SMs, which
and +∞ otherwise. A point y∗ is a local minimum of f ∗ . if f ∗ (y∗ ) ≤ f ∗ (y) for
can have a significant impact on the optimization process. While tight
every y that belongs to NΘ (y∗ ) (Murota, 2003).
bounds can lead to premature convergence and limit the exploration of
Based on Definition 1, if f(y∗ ) takes the minimum value within its the search space, thereby leang to suboptimal solutions, overly loose
neighborhood, then y∗ is declared as locally optimal and the DSDA al­ bounds can lead to a broader search space, thus allowing the identifi­
gorithm stops; otherwise, the combination of discrete decisions that cation of attractive solutions, at the cost of lengthy (or even prohibitive)
minimize the objective function is used to define a steepest-descent di­ computational times. Selecting bounds that balance the trade-off be­
rection. NLP problems with discrete variables fixed are then sequentially tween exploration and exploitation of the search space is thus key to find
solved along the steepest-descent direction until no improvement is attractive feasible solutions in acceptable turnaround times. If bounds
found. Then, a new iteration of the algorithm begins with a new are not properly selected for SMs, slow convergence, algorithm stagna­
neighborhood exploration. Previous works have demonstrated that the tion or lack of convergence may arise. In the case of the DSDA, an

4
D.A. Liñán et al. Computers and Chemical Engineering 180 (2024) 108501

Fig. 2. Proposed deterministic-stochastic algorithm. Black rectangles at the top and the bottom represent the beginning and the end of the parallelization,
respectively. Dashed lines represent the exchange of information.

5
D.A. Liñán et al. Computers and Chemical Engineering 180 (2024) 108501

inadequate selection of bounds for discrete variables (i.e., yL,0 and yU,0) DSDA-VB (Steps S4 and S2.2). This procedure is repeated until the
does not represent a major concern since the algorithm is successively stopping criterion in Step D6 is satisfied, meaning that both the deter­
fixing discrete variables and discarding infeasible neighbors. Neverthe­ ministic (Step D7) and stochastic (Step S5) algorithms are terminated.
less, bounds for continuous variables (i.e., xL,0 and xU,0) must be care­ According to this stopping criterion, the parallel hybrid strategy stops
fully selected for the NLP solver used within DSDA, whose performance when the relative difference between fBk and fSk is below an ϵ-tolerance.
may be affected by these bounds. For instance, the NLP optimization Note that iterations k are controlled by the deterministic algorithm i.e.,
performed by process flowsheet simulators such as Aspen Plus typically Steps D2-D6. Iterations are managed by the deterministic steps because
implements Sequential Quadratic Programming (SQP), which requires these define when the SM must be reinitialized (see Step D5 and red
good initial estimations and suitable bounds for the continuous DOF. dashed lines in Fig. 2), and when both deterministic and stochastic steps
Overly restricted variable bounds may result in optimal solutions lying must stop (see Steps D6 and D7 and green dashed lines in Fig. 2). Fig. S1
at their lower/upper bounds, while loose bounds may result in lack of in the supplementary material provides in-depth (high level) informa­
convergence of the SQP algorithm. tion about the algorithm presented in Fig. 2, including information
related to solver interactions, variables and constraints considered at
3.2. Proposed hybrid deterministic-stochastic algorithm each step, and data exchange aspects.

The hybrid algorithm presented in this section addresses the limi­ 3.2.1. Convergence analysis and summary
tations described above for SMs and the DSDA by iteratively refining The proposed hybrid method enforces requirements over the sto­
variable bounds within a parallel-hybrid algorithm. In this work, the chastic and deterministic algorithms to successfully converge to a local
variable’s bounds considered by the SM are updated by the deterministic optimum. In general, the hybrid stochastic-deterministic strategy
algorithm. That is, iteratively refined variable bounds (i.e., modified introduced above is expected to converge and close the gap within an
algorithmic bounds xL,M , xU,M , yL,M and yU,M ) are used to iteratively up­ ε-tolerance (ε≥0) if the three conditions detailed in Theorem 1 are
date the bounds on the decision variables (xL , xU , yL and yU ), instead of satisfied (a detailed proof this theorem is available in section S2 of the
using fixed user-defined bounds (i.e., xL = xL,0 , xU = xU,0 , yL = yL,0 and supplementary material).
yU = yU,0 ) throughout the execution of the SM. For continuous variables Theorem 1. Assume that: 1) the first run of the SM in Step S1 (iteration
(x), this is achieved by embedding the DSDA within a variable bounding k=0) can identify at least two different feasible solutions which are
(VB) algorithm that iteratively refines bounds xL,M and xU,M based on the added to FS; 2) the DSDA-VB in Step D4 can eventually find at least one
convergence status of the NLP optimizer, i.e., the DSDA-VB algorithm. locally optimal objective function (fDk ) using feasible information in FS,
For discrete variables (y), the N∞ neighborhood of each locally optimal and 3) the objective function in optimization problem (1) is bounded, i.
solution found by DSDA-VB defines refined bounds yL,M and yU,M. e., the objective function cannot be improved indefinitely without
Parallelizing the SM and the DSDA-VB offers different advantages. violating the constraints, and the SM in Steps S2.2 and S3 eventually
First, both the SM and DSDA-VB are being executed simultaneously, converges towards the global optimum, subject to the current selection
resulting in faster convergence compared to processing the SM and of bounds (xL,M , xU,M , yL,M , yU,M ). Then, the proposed hybrid stochastic-
DSDA-VB optimization steps sequentially. Also, the proposed paralleli­ deterministic algorithm will eventually terminate in a local optimum
zation allows discrete and continuous DOF to be optimized by both the for any user-defined ε-tolerance greater or equal to 0.
deterministic and stochastic techniques simultaneously, which leads to
solutions that satisfy local optimality requirements for both discrete and The hybrid strategy discussed above has the potential to improve the
continuous DOF. This feature would not be possible using the available performance of both the SM and the DSDA when solving flowsheet
hybrid stochastic-deterministic strategies for flowsheet optimization design problems. First, the hybrid algorithm in Fig. 2 proposes a parallel
with process simulators. The parallel implementation presented in this execution of a deterministic and a stochastic algorithm that not only
work also offers advantages with respect to the independent execution of ensures local optimality for continuous variables, but also guarantees
the DSDA and an SM, e.g., improvements in neighborhood exploration local optimality for discrete decisions within N2 -neighbors. To the au­
compared to the traditional DSDA, and improved convergence of the SM thors’ knowledge, these optimality guarantees cannot be provided by
in situations where the traditional SMs may stagnate. other hybrid algorithms available in the literature, see e.g., Gómez et al.
Based on the above, the SM and DSDA-VB algorithms benefit from (2006), Staudt and Soares (2009), Zhou et al. (2017), Herrera Velázquez
each other by considering their parallel execution in an iterative et al. (2022). This parallel execution strategy also addresses some of the
manner. As shown in Fig. 2, the algorithm starts with an initialization limitations that SMs or deterministic algorithms such as the DSDA
that consists of sequential Steps S1, S2.1 and D1. Then, the parallel al­ exhibit when executed separately. For instance, the inability of DSDA to
gorithm begins by simultaneously executing Step S3 of a stochastic al­ handle infeasible initializations is addressed by retrieving feasible so­
gorithm that iteratively executes the SM (see Section 3.2.2), and Step D2 lutions from the SM through set FS , and the prohibitive evaluation of
of a deterministic algorithm that iteratively executes the DSDA-VB (see N∞ -neighborhoods for problems with many discrete decisions is
Section 3.2.3). The main iterations of the parallel-hybrid algorithm are addressed by delegating the N∞ -neighborhood exploration to the SM.
denoted by superscript k. The overall idea of the proposed algorithm is Also, the slow convergence of the SM with respect deterministic stra­
to initialize and start the execution of the SM (Steps S1 and S2.1) and tegies is addressed by continuously generating updated search bounds
have the SM running, while the deterministic algorithm is constantly with the DSDA-VB method, and by defining a stopping criterion that
retrieving a database consisting of all feasible solutions (FS ) detected by incorporates stochastic and deterministic objective function values. In
the SM. The deterministic algorithm waits until enough feasible solu­ summary, the hybrid method takes advantage of the SM to identify
tions are available in FS . Then the DSDA-VB is initialized (Step D1) by feasible solutions at early stages of the execution (iteration k = 0) and
assigning an objective function value of +∞ to the best objective func­ explore a neighborhood of the current solution candidate (iterations
tion identified thus far by the DSDA-VB (fB0 ← + ∞). The DSDA-VB uses k ≥ 1). The SM heuristically seeks the global optimum of the problem
within the neighborhood it is exploring, but it does not provide opti­
the solution with the best objective function value from FS (i.e., fSk ) as
mality guarantees. The deterministic DSDA-VB strategy, on the other
initialization (Step D2), and then performs a local optimization to
hand, uses information from the SM to update search limits and, under a
generate modified bounds xL,M , xU,M , yL,M and yU,M (Steps D3 and D4),
deterministic approach, DSDA-VB satisfies the local optimality condi­
which are returned to the SM when there is an improvement in fBk (see
tions in Definition 1 for the highly nonlinear model in Problem 1. This is
Step D5 and purple dashed lines in Fig. 2). If fBk improved in Step D5, then expected to result in shorter computational times, as illustrated in
the SM search is re-started using the modified bounds obtained by

6
D.A. Liñán et al. Computers and Chemical Engineering 180 (2024) 108501

Section 4. It is also acknowledged that the proposed hybrid algorithm saturation, moving away from an optimal solution, or moving into re­
has limitations. The proposed hybrid strategy still needs improvements gions with large function values, large derivative values, or singularities.
to handle multi-objective, discontinuous, and highly multimodal Despite their relevance, variable bounds typically remain fixed, and the
objective functions. Also, updating the search bounds of the SM using task of defining appropriate bounds is typically delegated to the user. To
DSDA-VB help the algorithm to converge faster. However, the bounds alleviate this issue, the DSDA-VB algorithm has been developed as an
obtained using the proposed VB strategy only guarantee feasibility and optimization method that iteratively executes the DSDA and a VB
local optimality; hence, any set of bounds returned by DSDA-VB may strategy that recomputes bounds for continuous variables (xL,M , xU,M ).
miss the global optimum. Detailed descriptions of the steps involved in The VB strategy includes three distinctive features. First, it helps to
the proposed parallel-hybrid algorithm are discussed below. circumvent potential convergence issues of the NLP solver. Second, it
allows to keep the continuous variables returned by DSDA-VB (i.e., xk )
3.2.2. Steps followed by the stochastic algorithm away from their algorithmic bounds xL,M and xU,M ; otherwise, a solution
Steps S1 and S2.1: In Step S1, the stochastic algorithm begins by may be erroneously declared as locally optimal. Third, the VB strategy
setting up Problem (1) aided with a process simulator for flowsheet helps to indirectly accelerate the SM, whose convergence rate may
calculations and initializing the algorithm using: 1) the user-defined benefit from exploring a subset of the original search space that is known
variable bounds (i.e., xL,0 , xU,0 , yL,0 and yU,0 .) for Problem (1), 2) a to contain at least one locally optimal solution. Note that bounds for
value of False is selected for control flow parameter stopS ., which will be discrete decisions do not need to be iteratively recalculated within
latter used to define when to execute or reinitialize the SM, and 3) DSDA-VB, because the DSDA can consider the user-defined bounds yL,0
known feasible DOF combinations for set FS , e.g., FS0 = ∅, indicating that and yU,0 whenever it is executed, without affecting its performance.
no feasible solution is known for the problem. After that, an iterative Instead, updated bounds of discrete decisions (yL,M , yU,M ) are defined as
procedure starts with the first execution of the SM in Step S2.1. upper and lower bounds of a discrete neighborhood of the local opti­
Step S3 and S4: According to Step S3, the execution of the SM keeps mum identified by DSDA-VB. Note that these updated bounds of discrete
updating set FS with new feasible solutions until the deterministic DOF define a N∞ -neighborhood consisting of 3ny combinations of
strategy determines that the execution of the SM must be stopped, i.e., discrete DOF, which are considered by the SM whenever it is reini­
until stopS takes a value of True. As described below in Step D5, stopping tialized. In contrast, each execution of the DSDA within DSDA-VB con­
the SM occurs whenever the deterministic algorithm identifies an siders the smaller N2-neighborhood of size 2ny + 1, given that the
improvement in the best locally optimal solution found. Once stopS = evaluation of N∞ -neighborhoods with DSDA may be computationally
True, the current SM run is stopped as indicated in Step S4. prohibitive. This means that the SM is allowed to explore many discrete
Steps S4 and S5: As indicated in Step S4, stopS is set back to False and DOF combinations that may improve the objective function, but that
variable bounds in Problem (1) are updated using the modified bounds may be ignored by the DSDA-VB. In summary, the DSDA-VB executed in
returned by the deterministic algorithm, denoted by purple dashed lines Step D4 of Fig. 2 is an algorithm that not only optimizes the discrete and
in Fig. 2 (xL,M , xU,M , yL,M and yU,M .). One of the purposes of these new continuous DOF of the problem locally, but also returns new bounds for
bounds is to guide the SM towards the best locally optimal solution these variables as part of the solution.
identified by the deterministic strategy thus far, e.g., by keeping yL,M and The DSDA-VB returns an objective function (fDk ) and the locally
yU,M within a N∞ -neighborhood of the best local solution. Before re- optimal solution found ([xk , yk ]), as well as new variable bounds
starting the execution of the SM using these new bounds, a global (xL,M , xU,M , yL,M and yU,M ) that are tighter than those originally defined by
stopping criterion is verified in Step S5. This global stopping criterion
the user, i.e., xL,0 ≤ xL,M , xU,M ≤ xU,0 , yL,0 ≤ yL,M and yU,M ≤ yU,0 . On the
uses control flow parameter StopD, which stops the execution of both
one hand, continuous variables bounds (xL,M , xU,M ) guarantee that the
algorithms based on the best objective functions found by both the
continuous variables reported by DSDA-VB ([xk , yk ]) are not at their
deterministic and stochastic algorithms. Flow parameters StopD and
bounds and that NLP convergence issues are avoided, e.g., discontinu­
Stops (red and green dashed lines in Fig. 2) are controlled by the
ities in the constraints that cause convergence errors when bounds are
deterministic algorithm (Steps D1-D7), which is explained next.
inappropriately selected are avoided. On the other hand, updated
3.2.3. Steps followed by the deterministic algorithm discrete variable bounds (yL,M , yU,M ) are restricted within an N∞ -neigh­
Step D1: The deterministic strategy is constantly retrieving the most borhood of the locally optimal solution yk returned by DSDA-VB as
updated version of set FS from either Step S2.1 or Step S3 (see Steps D1 indicated in Eqs. (2A) and (2B), e.g., for a problem with a single discrete
and D2). In Step D1, the deterministic strategy must await until at least DOF, the improved bounds for optimal solution yk = 3 are yL,M =
two different feasible solutions are available in FS, which is a require­ max(yL,0 , 2) and yU,M = min(yU,0 ,4). The best objective function found by
ment to initialize the DSDA-VB strategy. These two feasible solutions the deterministic algorithm (fBk ) is updated after the execution of the
from the stochastic technique are required to provide a feasible initial­ DSDA-VB in Step D4 as fBk = min (fDk̃ ). A detailed diagram and
ization ([xI , yI ]) to Step D3 and an initialization for continuous variables k̃∈{1,…,k}

bounds ([xL,S , xU,S ]) for the execution of DSDA-VB in Step D4. The description of the DSDA-VB algorithm is provided in the supplementary
deterministic strategy is then initialized in Step D1 by assigning a value material (Section S1, Fig. S2).
of +∞ to the best locally optimal objective function known fB0 , where ( ( ( )))
yL,M = max yL,0 , min N∞ yk (2A)
subscript B stands for “best”. Having fB0 = +∞ means that no locally
optimal solution has been detected yet. Then, the iteration count k. is set ( ( ( )))
yU,M = min yU,0 , max N∞ yk (2B)
at k = 1 and the iterations of the parallel hybrid algorithm start.
Steps D2 and D3: In Step D2, the solution with the best objective Steps D5, D6 and D7: The remaining steps (Steps D5, D6 and D7) of the
function value is retrieved from FS and assigned to fSk , i.e., fSk is the best deterministic algorithm shown in Fig. 2 decide when the stochastic al­
solution reported by the stochastic algorithm at iteration k. The DSDA- gorithm must be reinitialized, and when both algorithms must be
VB algorithm is then initialized at Step D3 using the DOF associated to stopped through control flow parameters stopS and stopD , respectively.
Step D5 decides if the SM must be stopped and reinitialized, based on an
fSk (i.e., [xI , yI ]), and the maximum ([xU,S , yU,S ]) and minimum ([xL,S , yL,S ])
objective function improvement criterion that uses a small tolerance
DOF values in FS.
δ>0. If the best objective function found by the deterministic strategy in
Step D4 (The DSDA-VB): Selecting adequate variable bounds for
deterministic NLP optimization is critical to avoid execution errors and the current iteration (fBk ) improved with respect to the best objective
improve solver performance, e.g., good bounds m avoid variable found in previous iterations (fBk− 1 ), then control flow parameter stopS is

7
D.A. Liñán et al. Computers and Chemical Engineering 180 (2024) 108501

Fig. 4. Thermally coupled system that separates a mixture of isobutane, n-

butane and n-pentane.

run an implementation of the DETL code that was developed in previous

works (Sharma et al., 2012; Srinivas and Rangaiah, 2007).
The rigorous Mass-Equilibrium-Summation-entHalpy (MESH) model
with non-ideal equilibrium relationships available in Aspen plus was
considered. Vector of constraints g in Problem (1) was handled by DETL
using the penalty function method, with a high penalty factor of 1 ∗ 1010
that is likely to avoid infeasible solutions. The traditional DETL method
was used as a benchmark to compare the performance of the proposed
Fig. 3. Hybrid algorithm implementation and software considered in this work. hybrid algorithm. To perform a fair comparison, the tuning parameters
of DETL were kept fixed for all runs of the hybrid and traditional DETL
set to True indicating that the current SM run in the stochastic strategy method, respectively. That is, both algorithms were executed under the
must be stopped and reinitialized (see Steps S3 and S4 of the stochastic same conditions. We adopted standard tuning parameters available in
algorithm, which use the updated information from the red and purple the literature for DETL (Srinivas and Rangaiah, 2016; Storn, 1996):
dashed lines in Fig. 2). This allows the stochastic algorithm to update the population size NP = 10 ∗ (nx + ny ), crossover rate CR=0.8, mutation
search bounds of the SM and guide it towards the solution with objective factor F=0.7, tabu list size TLS = NP/2, and tabu radius TR = 1 ∗ 10− 6 .
fBk . If the stochastic technique does not require to be reinitialized, then For the DSDA exploration, N2-neighborhoods were considered.
the global stopping criterion of both algorithms is evaluated in Step D6. The objective function considered for both case studies is shown in
This global stopping criterion verifies if the relative difference between Eq. (3A), which was obtained using the estimation method presented by
the best solutions found by the stochastic (fSk ) and deterministic (fBk ) Douglas (Douglas, 1988) and cost data available from previous works
algorithms is below an ϵ-tolerance. If the global stopping criterion is (Bernal et al., 2018; Ciric and Gu, 1994; Gómez et al., 2006). This
satisfied, the stochastic technique h thus converged near the best local function represents the addition of the annualized investment cost
solution found by the deterministic algorithm, hence both algorithms (Investmentc .) and the operating costs associated with utilities
stop (Step D7). Otherwise, a new iteration of the deterministic strategy (Operatingc ) for each column (c ∈ C) in the flowsheet. The investment
begins. cost in Eq. (3B) considers the column diameter in meters (dc ), the total
column height in meters (hc .), and the number of catalytic stages (ncc ) if
4. Computational experiments any. The operating cost in Eq. (3c) represents the cost of the duties of the
reboiler (Qrc ) and the condenser (Qcc ) in kilowatt. This objective func­
This section presents two case studies that highlight the improve­ tion assumes 330 days of operation per year, an annual interest rate of
ments in terms of solution quality and computational times when using 5%, and payback period of 5 years. Note that search bounds are imposed
the proposed hybrid algorithm to optimize the total annual cost (TAC in over dc , hc and ncc whereas |Qrc |. and |Qcc |. are bounded below by 0, i.e.,
$/year) of intensified distillation flowsheets. In order to carry out the the objective function is bounded.
stochastic optimization process, we have selected the Differential Evo­ ∑
lution with Tabu List (DETL) as the SM. This algorithm combines the f = TAC = (Invesmentc + Operatingc ) (3A)
c∈C
advantages of Differential Evolution (DE), a powerful global search al­
gorithm, with the concept of tabu search, which is a local search method Invesmentc = 10000 + 292.67(3.28dc )1.066 (3.77hc )0.802 + 15.29(3.28dc )1.55 hc
that prevents the algorithm from revisiting previously explored solu­
tions (Srinivas and Rangaiah, 2016). Fig. 3 illustrates the method used to + 131.74(dc )2 ncc
implement the parallel algorithm proposed in this work. Two separate (3B)
computers with a processor Intel® Core™ i7-3770 CPU @ 3.4 GHz and
16 GB of RAM were used to run the tasks related to the stochastic Operatingc = 146.8|Qrc |+24.5|Qcc | (3C)
(computer 1) and deterministic (computer 2) algorithms. These com­ The distillation column models considered in this work involve
puters were allowed to interact by exchanging information through a multiple discrete DOF, continuous DOF, and constraints resulting from
mappable network folder with 10 GB of storage capacity using Microsoft multiple interconnected units in a process flowsheet. Rigorous stage-by-
Excel. The main codes of the hybrid algorithm (Fig. 2) were imple­ stage nonlinear models combining mass/energy balances and non-ideal
mented in Matlab. These codes were connected to an Excel spreadsheet equilibrium relationships for each unit are considered. The problems
that worked as an interface between Matlab and Aspen Plus, which was considered in this work are non-convex due to the nonlinear relation­
used for flowsheet simulation and NLP optimization with the SQP ships between the input variables (such as feed flow rate, reflux ratio,
method. In the case of the stochastic algorithm, Excel was also used to and tray temperatures) and the output variables (such as product purity

8
D.A. Liñán et al. Computers and Chemical Engineering 180 (2024) 108501

Table 1 and recovery). Additionally, the large search spaces considered for
Variables and their bounds for the first case study, and best solution found. discrete decisions and the nonlinear interactions between discrete and
Continuous DOF (x) continuous variables make the problems described above challenging
non-convex combinatorial optimization problems.
Variable Units Lower Upper Best solution
bounds bounds found (Fig. 5A)
(xL,0) (xU,0)
4.1. Case study 1: A thermally coupled system
Reflux ratio (C1) - 0.5 10 7.62
Distillate flow (C1) kmol/ 17.29 19.11 18.18
h The optimization of the thermally coupled system shown in Fig. 4 is
Distillate flow (C2) kmol/ 8.69 9.6 8.92 considered. In this work, stages are numbered from top to bottom, i.e.,
h the condenser corresponds to stage 1. The feed to the first column (C1)
Vapor flow (from kmol/ 11.34 34.02 18.84
has a total molar flow rate of F FEED total = 45 kmol/h, with mole fractions of
C1 to C2) h
Diameter (C1) m 0.1 6 1.68 χ FEED FEED FEED
isobutane = 0.4, χ n− butane = 0.2 and χ n− pentane = 0.4. The first column aims
Diameter (C2) m 0.1 6 0.37 to produce nearly pure isobutane at the top and nearly pure n-pentane at
Discrete DOF (y) the bottoms. The separation of n-butane from the mixture is performed
Variable Units Lower Upper Best solution in another column (C2) that is thermally coupled to C1 through inter­
bounds bounds found (Fig. 5A) connecting vapor and liquid flows that are connected to the last stage of
(yL,0) (yU,0)
C2. These interconnecting flows avoid the need of a reboiler for the
Number of stages - 5 70 55 second column, decreasing energy consumption with respect to a
(C1)
traditional distillation sequence (Contreras-Zarazúa et al., 2019, 2021).
Number of stages - 5 70 12
(C2) The packages available in Aspen Plus were used for thermodynamic
Feed stage (C1) - 2 69 24 calculations. The Chao-Seader property method with Lee-Kesler
Liquid feed stage - 2 69 46 enthalpy was considered, which uses the Redlich-Kwong equation of
(from C2 to C1) state for vapor phase properties, and the Scatchard-Hildebrand model
for liquid activity coefficients (Carlson, 1996).
The DOF of this problem (x and y), and their corresponding bounds

Fig. 5. Convergence plots for the optimal design of the thermally coupled system.

9
D.A. Liñán et al. Computers and Chemical Engineering 180 (2024) 108501

Table 2 Table 3
Best solution found by each run of the hybrid algorithm (Fig. 5B-C). First iteration of the hybrid algorithm.
Continuous DOF (x) Run Time [min] Iteration (k) fBk (DSDA-VB) fSk (DETL)
Variable Units Fig. 5B Fig. 5C Fig. 5D Fig. 5A 209 1 275351 645529
Fig. 5B 141 1 300204.4 537369.4
Reflux ratio (C1) - 7.40 7.64 7.61
Fig. 5C 150 1 276903.6 551489.4
Distillate flow (C1) kmol/h 18.18 18.19 18.18
Fig. 5D 252 1 297621.2 475507.7
Distillate flow (C2) kmol/h 8.92 8.92 8.92
Vapor flow (from C1 to C2) kmol/h 18.99 19.29 18.95
Diameter (C1) m 1.71 1.68 1.68
Diameter (C2 m 0.37 0.34 0.37 the last column of Table 1, and it is compared with the best (local) so­
lutions from the other runs (Fig. 5B–D) depicted in Table 2. The results
Discrete DOF (y)
Variable Units Fig. 5B Fig. 5C Fig. 5D from the different runs in Table 2 vary slightly, showing average per­
centage differences between the best-known solution (Fig. 5A) and the
Number of stages (C1) - 56 55 55
Number of stages (C2) - 12 11 12
remaining solutions (Fig. 5B–D) below 1% for all continuous DOF,
Feed stage (C1) - 25 25 25 except for the vapor flow from C1 to C2 and the diameter of C2 which
Liquid feed stage (from C2 to C1) - 47 46 46 show average percent differences of 1.3% and 2.7%, respectively. The
Objective function $/year 271519 271339 271281 discrete DOF also exhibits similar values between runs, differing only by
increments or decrements of 1 stage with respect to the best-known
solution. As a result, the resulting objective function values in Table 2
(xL,0 , xU,0 , yL,0 and yU,0 ) are summarized in Table 1. Note that the vapor
only differ by 0.05% on average compared to the best objective (271229
flow that goes from C1 to C2 is assumed to be located one stage below
$/year).
the liquid flow that goes from C2 to C1. Other variables of interest that
Based on Fig. 5, the proposed hybrid algorithm has a higher
are retrieved from the optimization (i.e., those in vector z) are the en­
convergence speed than the traditional DETL for this case study. The
ergy duties, which are required to calculate the objective function.
hybrid strategy closed the gap after 2435 minutes on average, while the
Vector z also contains those variables required to specify the problem’s
traditional DETL could not find the best solution reported by the hybrid
constraints. These variables include the molar fraction and flow of the
algorithm after 5000 minutes in any of the trials. For instance, the DETL
desired component for each output stream, i.e., molar flow (F ISOBUT isobutane ) objective found after 5000 minutes in Fig. 5A (275018 $/year) is still
and mole fraction (χ ISOBUTisobutane ) of isobutane at the top of C1; molar flow 1.4% larger than the solution obtained at 1825 minutes by the hybrid
PENTANE
(F n− pentane .) and mole fraction (χ PENTANE
n− pentane ) of n-pentane at the bottoms of algorithm (271229 $/year). Also, when the hybrid algorithm stops in
C1, and molar flow (F BUTANE BUTANE
n− butane ) and mole fraction (χ n− butane ) of n-butane
Fig. 5A–C, the objective function reported by the traditional DETL is on
at the top of C2. The constraints considered for this problem (g(x, y, z)) average 26% worse than the hybrid’s strategy objective. These im­
are shown in Eqs. (4A)–(4C), which establish that: 1) a minimum molar provements in convergence speed are mainly attributed to the “jump”
recovery of 94% must be attained for all components, 2) a minimum that the hybrid algorithm performs once the first execution of the DSDA-
molar product purity of 98%, 99% and 96% is enforced for isobutane, n- VB algorithm (k = 1) is performed and the first group of improved
pentane and n-butane respectively, and 3) the heigh to diameter ratio of variable bounds is returned to the SM. This is illustrated in Table 3,
the columns must be less or equal to 20 (Barker, 2018). where after 209, 141, 150, and 252 minutes of execution of the hybrid
algorithm, the DSDA-VB improves the DETL objective by 134%, 79%,
ISOBUT
99% and 59% for the trials shown in Fig. 5A, B, C, and D, respectively.
FEED PENTANE FEED BUTANE
F isobutane ≥ 0.94 F isobutane , F n− pentane ≥ 0.94 F n− pentane , F n− butane

≥ 0.94 F FEED
n− butane
Note that while the hybrid algorithm requires ∼ 102 minutes to obtain
(4A) the objective functions in column fBk shown in Table 3, the traditional
DETL needs ∼ 103 minutes to obtain solutions of this quality. Although
χ ISOBUT
isobutane ≥ 0.98, χ PENTANE
n− pentane ≥ 0.99, χ BUTANE
n− butane ≥ 0.96 (4B) the first iteration of the hybrid algorithm leads to a solution for which
DOF are locally optimal, the objective function is further improved in
hc /dc ≤ 20, ∀c ∈ {C1, C2} (4C) subsequent iterations of the hybrid algorithm due to the continuous
exploration of the DETL algorithm within the continuously refined
This case study was used to test the proposed hybrid algorithm and to
bounds obtained from the DSDA-VB strategy. Additional information
compare the performance of the proposed algorithm and the traditional
about the iterations of the hybrid algorithm for optimization runs in
DETL method. Given that both methods rely on random exploration,
Fig. 5A–D is available in section S5 of the supplementary material.
different solutions may be obtained from different runs of the same
The proposed hybrid algorithm is able to continuously identify new
optimization problem. Thus, this case study was solved four times under
locally optimal solutions because the DETL explores solution candidates
the same conditions using both the proposed method and DETL. A
within N∞ -neighborhoods of discrete decisions of size |N∞ | = 34 = 81,
relative optimality gap of ε=0.05 was established for the hybrid algo­
while DSDA-VB only guarantees local optimality within N2 -neighbor­
rithm; also, δ was set to 0.05, and a CPU time limit of 5000 min was
hoods of size |N2 | = 2(4) + 1 = 9. N2 -neighbors consider the effect of
considered for both strategies. More details about the implementation
varying one discrete decision at a time, while N∞ -neighbors consider
and execution of the hybrid algorithm are available in sections S1 and S3
interactions between variables, e.g., modifying both the number of
of the supplementary material.
stages and feed location of a column simultaneously. As a result, DETL
Despite their inherent random variability, the four trials showed a
can guide the local search towards solution candidates that are near the
similar qualitative performance, i.e., the parallel-hybrid method showed
current DSDA-VB solution but cannot be identified by this deterministic
faster convergence and better objective function values than the tradi­
algorithm. For instance, Fig. 6A illustrates the convergence plot of a
tional DETL. The evolution of the best objective function over time for
discrete DOF: the number of stages N of column C1. The variables and
these experiments is shown in Fig. 5A–D. Dotted lines represent the
bounds plotted in Fig. 6 correspond to the best objective found by each
traditional DETL method, while solid and dashed lines represent the
algorithm at each iteration. The bounds for the number of stages in
proposed DETL-based stochastic method and the proposed DSDA-VB
Fig. 6A start with the user-defined bounds (i.e., NL,0 = 5 and NU,0 = 70),
deterministic method, respectively. The best solution identified by the
hybrid algorithm is a local optimum and was obtained from the first trial but these are updated to NL,M = 53 and NU,M = 54 after the first local
(Fig. 5A). This solution corresponds to a design with the DOF reported in solution with N = 54 is identified by DSDA-VB at 209 minutes (see the
first row in Table 3). After 529 minutes of execution of the hybrid

10
D.A. Liñán et al. Computers and Chemical Engineering 180 (2024) 108501

Fig. 6. Illustrative convergence plots for the number of stages of C1 (A) and the reflux ratio of C1 (B). These plots correspond to the optimization run in Fig. 5A.

algorithm, the bounds over N are recalculated again; this time around N
= 55, with NL,M = 55 and NU,M = 56. This change coincides with in­
crements of the number of stages of C2 and the location of the liquid
interconnection stage, i.e., multiple discrete decisions changed at 529
minutes, which is something that could not be achieved by the DSDA-VB
alone, which explores the search space using N2 -neighborhoods (|N2 | =
11). This indicates that this new solution identified at 529 minutes was
attained thanks to the exploration performed by DETL in the larger
N∞ -neighbrohood (|N∞ | = 243), which allowed varying multiple
discrete DOF simultaneously. Note that although DSDA-VB only per­
forms local optimality verifications within N2 -neighborhoods, it is
allowed to explore the user-defined search region of discrete decisions
(∼ 107 discrete solution candidates), while DETL seeks improved solu­
Fig. 7. Sequence of distillation processes designed for the production of MEK
tions within an N∞ -neighborhood of the current local DSDA-VB solution
from 2,3-BD, and the separation of by-products.
(81 discrete solution candidates).
In contrast to the discrete decisions, updating the bounds of the
continuous variables does not necessarily limit the search of continuous local optimality in the solutions. In contrast, the proposed parallel-
variables within a vicinity of the current best solution. Fig. 6B shows the hybrid approach addresses these issues offering both optimality gua­
behavior of the reflux ratio of unit C1. The upper bound of the reflux rantees and convergence improvements over the nested-hybrid meth­
ratio begins at RRU,0 , and then it its updated and kept at RRU,S , which is odology. More details about this comparison are provided in section S4
the maximum feasible reflux ratio stored in FS . Contrarily, the VB of the supplementary material.
strategy varied the lower bound of the reflux ratio (RRL,M ) during the
first 32 iterations of the algorithm (i.e., the first 615 minutes) and then 4.2. Case study 2: A sequence of reactive, extractive and conventional
RRL,M remained constant at RRL,S = 7.55. This avoided binding RR at its distillation columns
lower bound in the locally optimal solutions reported by DSDA, i.e., after
615 minutes, RRL,S = 7.55 guarantees that the best reflux ratio (7.62) is This case study aims to demonstrate the applicability of the proposed
locally optimal, given that the optimized reflux ratio is no longer algorithm to a more challenging process flowsheet optimization prob­
reaching its lower bound. Overall, updating bounds as illustrated in lem that involves a large number of DOF and constraints, i.e., 13 discrete
Fig. 6 allows to guide the hybrid algorithm towards locally optimal so­ DOF and 15 continuous DOF. This case study was adapted from the work
lutions faster than the traditional DETL, which requires more time to by Torres-Vinces et al. (2020). The sequence of distillation systems
converge for this case study. presented in Fig. 7 consists of a reactive distillation (RD) column that has
Additional computational experiments showcasing the performance an input of pure 2,3-butanediol (2,3-BD) equal to 1000 kg/h. The pure 2,
of the proposed parallel-hybrid algorithm against a nested-hybrid al­ 3-BD follows a series of dehydration reactions in the RD column, which
gorithm are presented in Fig. S5 of the supplementary material. The simultaneously generates Methyl Ethyl Ketone (MEK) as the main
nested-hybrid strategy optimizes discrete DOF using the DETL, which product in the top stream (T-RD) and separates the unreacted 2,3-BD as a
iteratively fixes those discrete DOF and optimizes continuous DOF bottom’s product. The dehydration reactions in the RD system also
through the NLP solver available in Aspen. As shown in Fig. S5 of the generate by-products in the T-RD stream such as 2-methylpropanal
supplementary material, the nested-hybrid approach was 464 minutes (2MPL), 3-buten-2-ol (3B2OL), water, and some 1,3-butadiene (1,
on average slower than the parallel-hybrid method at finding the best 3-BD). These by-products are separated from the MEK in subsequent
solution reported by each algorithm. Also, the best parallel-hybrid so­ separation stages. The first stage is an extractive distillation (ED) column
lution (271229 $/year) is 0.2% better than the best nested-hybrid so­ that uses glycerol as solvent to extract water from stream T-RD. The
lution (271669 $/year) in terms of its objective function value. solvent mixed with water is removed from the bottom of the ED column
However, the nested-hybrid exhibits two key limitations: this method (steam B-ED), while the remaining components in the mixture leave the
cannot refine search bounds for continuous DOF and cannot guarantee system from the top (steam T-ED). The latter stream (T-ED) is further

11
D.A. Liñán et al. Computers and Chemical Engineering 180 (2024) 108501

Table 4
Variables and their bounds for the second case study, and best solution found.
Continuous DOF (x)

Variable Units Lower Upper bounds Best solution

bounds (xL,0) (xU,0) found

Reflux ratio (RD) - 1 3.2 3.016

Reflux ratio (ED) - 0.1 0.5 0.489
Reflux ratio (C1) - 1 14.7 1.306
Reflux ratio (C2) - 1 5.8 5.607
Reflux ratio (C3) - 1 1.6 1.000
Distillate flow (from kg/h 994.01 1004.00 999.258
RD to ED)
Distillate flow (from kg/h 792.57 800.53 797.241
ED to C1)
Bottoms flow (C1) kg/h 89.55 90.45 90.344
Bottoms flow (C2) kg/h 434.82 439.19 437.230
Bottoms flow (C3) kg/h 397.01 401.00 399.989
Diameter (RD) m 0.5 3.5 2.408
Diameter (ED) m 0.5 3.5 1.372
Diameter (C1) m 0.5 3.5 2.286
Diameter (C2) m 0.5 3.5 1.372
Diameter (C3) m 0.5 3.5 0.500

Discrete DOF (y) Fig. 8. Convergence plot for the MEK production process.
Variable Units Lower Upper bounds Best solution
bounds (yL,0) (yU,0) found
in the black-box constraints Ω that appear in problem (1). Note that this
Number of stages - 76 86 79 case study is more challenging than the problems usually considered to
(RD)
Number of stages (ED) - 45 55 45
test the hybrid algorithms for flowsheet optimization, which usually
Number of stages (C1) - 75 85 75 consider flowsheets involving one or two processing units as in case
Number of stages (C2) - 45 55 45 study 1, e.g., see (Gómez et al., 2006; Skiborowski et al., 2015; Ursel­
Number of stages (C3) - 13 23 13 mann et al., 2016).
Feed stage (RD) - 40 50 49
T-RD feed stage (ED) - 35 44 38 ωMEK
MEK ≥ 0.995 (5A)
Solvent feed stage - 2 10 2
(ED)
Feed stage (C1) - 45 55 53
ωsc ≥ 0.99, ∀(c, s)
Feed stage (C2) - 20 30 20 ∈ {(2, 3 − BD, 2, 3 − BD), (3B2OL, 3B2OL), (2MPL, 2MPL)} (5B)
Feed stage (C3) - 9 12 10
Location of first - 2 7 5
reactive stage (RD)
ωWATER GLYCEROL
Water ≥ 0.98, ωGlycerol ≥ 0.9999 (5C)
Location of last - 75 85 78
reactive stage (RD) M sc ≥ 0.98M T− RD
,
c
{ }
(MEK, MEK), (2MPL, 2MPL), (3B2OL, 3B2OL), (5D)
∀(c, s) ∈
purified in distillation columns C1 and C2, where column C1 generates a (Water, WATER), (Glycerol, GLYCEROL)
bottom product rich in 3B2OL, and C2 generates a top product with
2MPL and some 1,3-BD, and a bottom product with purified MEK. hc /dc ≤ 20, ∀c ∈ {RD, ED, C1, C2, C3} (5E)
Moreover, stream B-ED is separated in distillation column C3, which Due to the high number of variables and constraints involved, this
recovers the glycerol (bottom product) and produces water as the top problem requires much more time to converge than the previous case
product. The non-random two-liquid (NRTL) thermodynamic model was study. In the present case study, the proposed hybrid algorithm and the
used for the liquid phase, while the Redlich-Kwong equation of state was DETL strategy were left to run until a time limit of 5*104 minutes was
used for the vapor phase (Torres-Vinces et al., 2020). reached with δ=0.05. The convergence plot of both algorithms is shown
The DOF of this problem (x and y), and their corresponding bounds in Fig. 8. At 5*104 minutes, the relative gap between the traditional
(xL,0 , xU,0 , yL,0 and yU,0 ) are summarized in Table 4. Variables in vector z DETL and hybrid algorithms objectives is approximately 48%, while the
include the reboiler and condenser duties for all columns, as well as gap between the DETL and DSDA-VB algorithms within the proposed
those variables required to specify the problem’s constraints, i.e., the hybrid approach is around 1%. The best solution reported by the hybrid
mass fraction (ωsc ) and mass flow (M sc ) of the desired component c for algorithm has an objective function of 863423 $/year the DOF obtained
each output stream s. The user-defined constraints that are considered for this solution is depicted in the last column of Table 4. In contrast, the
for this problem (g(x,y,z)) include those proposed by Torres-Vinces et al. traditional DETL found a much more expensive design with an objective
(2020): a purity constraint of at least 99.5% (wt) is enforced over the function of 128280 $/year.
MEK output (Eq. (5A)); purity constraints of at least 99% (wt) are Due to the problem size and the search bounds selected, the perfor­
considered for the 2,3-BD, 3B2OL, and 2MPL streams (Eq. (5B)); purity mance of the traditional DETL deteriorated significantly for this case
constraints of at least 98% and 99.99% (wt) are imposed over the water study (see the dotted line in Fig. 8). Although DETL was able to identify
and glycerol output streams, respectively (Eq. (5C)), and a recovery of at feasible solutions during the early stages of its execution, it was unable
least 98% (wt) is enforced for all products (Eq. (5D)). Also, the heigh to to improve the objective function within the specified time limit. This
diameter ratio of the columns is constrained to be less or equal to 20, as performance may be attributed to the low ratio of individuals evalua­
stated in Eq. (5E) (Barker, 2018). In addition, hydrodynamic constraints tions that converged to feasible solutions, which in this case resulted in
that aim to maintain proper column hydrodynamic operation (e.g., stagnation of the traditional DETL algorithm, i.e., at 5 ∗ 104 minutes the
enough contact between vapor and liquid phases) such as entrainment DETL population still remains diverse and does not seem to be pro­
and downcomer flooding are imposed for all columns. These constraints gressing towards a particular solution. In contrast, the hybrid algorithm
are handled internally by Aspen Plus, i.e., these constraints are included managed to handle this convergence problem by automatically

12
D.A. Liñán et al. Computers and Chemical Engineering 180 (2024) 108501

Table 5
Convergence of objective functions fBk and fSk , number of stages (RD) and reflux ratio (RD) for iterations in which variable bounds are updated.
Number of stages (RD) Reflux ratio (RD)

Time [min] k fkB NL,M NU,M Nk RRL,M RRU,M RRk NL,M

1833 1 878540.4 1134177 80 82 81 2.96 3.04 3.01

3692 49 878356.7 987397 80 82 81 2.93 3.03 2.97
5473 56 874157.8 959781.5 80 82 81 2.47 3.04 2.93
7826 161 868193.2 946637 79 81 80 2.89 3.04 2.95
10452 168 867731.8 946637 79 81 80 2.90 3.03 2.97
27378 272 863423.2 882641.1 78 80 79 2.89 3.04 3.02
50000 442 863423.2 875273.2 78 80 79 2.89 3.04 3.02

identifying variable bounds that, for this case study, allowed the DETL performance of this hybrid algorithm. The results show that the pro­
algorithm to improve its objective function and satisfy local optimality posed hybrid algorithm quickly identifies a good quality local solution in
requirements for discrete and continuous variables (see solid line in the early stages of its execution. After that, the hybrid method relies on
Fig. 8). the exploratory capabilities of the SM method to identify new feasible
Table 5 summarizes some of the iterations performed by the hybrid solutions that serve as starting points to find better local optima. Overall,
algorithm for this case study. These iterations correspond to those where the parallel hybridization of the DSDA-VB and an SM such as DETL re­
variable bounds are updated. Each iteration shows the locally optimal sults in an enhanced convergence performance of both algorithms to­
objective function obtained by DSDA-VB (fBk ), the best objective function ward improved solutions, with respect to the naïve execution of the SM.
identified by DETL (fSk ), and the bounds and optimal solution for the In contrast to traditional hybrid approaches (e.g., memetic algorithms),
number of stages (N) and reflux ratio (RR) of the RD. According to the discrete and continuous variables obtained from the present
Table 5, the frequency at which variable bounds are updated increases approach satisfy local optimality. Future work is needed to make the
as iterations progress, e.g., the first group of updated bounds is obtained proposed strategy accessible to a broader audience, through the devel­
after 1833 minutes at k=1, while the last group of updated bounds is opment of a toolbox that readily incorporates the different steps needed
obtained after 16926 minutes at k=272. Note that the hybrid algorithm to execute the proposed hybrid algorithm. In terms of applications,
spent 45% of the total execution time between iterations 272 and 442 (i. further research is also needed to improve the algorithm’s performance
e., 22622 minutes). During this time interval, DSDA-VB could not in handling multi-objective, and multimodal objective functions. The
identify better solutions, thus bounds were not updated, i.e., between algorithm must also be expanded to consider decisions that typically
iterations 272 and 442 DETL is aiming to improve the objective function appear in process synthesis problems, such as deciding between two
without recdback from DSDA-VB. This suggests that, although the DETL separation or reaction technologies, e.g., molecular sieve membrane
performance improves with respect to its naïve execution, DETL is still separation against distillation. Also, the proposed approach should be
the time-limiting factor when integrated with DSDA-VB. Concerning the tested in the future with integrated design and control applications
convergence of variables and their bounds, the behavior obtained is (Rafiei and Ricardez-Sandoval, 2020).
similar to that obtained for the first case study. For example, the lower
bounds are updated from the user-defined values (76 and 1) to 80 and Funding
2.96 at k=1, for N and RR, respectively. After the first iteration, variable
bounds and values do not change significantly, but they still have an This work was supported by The Natural Sciences and Engineering
impact on the objective function, e.g., fBk and fSk improve by 1.8% and Research Council of Canada (NSERC).
30% between iteration 1 and 442. Detailed data showing the evolution
CRediT authorship contribution statement
of the objective functions fBk and fSk is provided in section S5 of the
supplementary material.
David A. Liñán: Conceptualization, Methodology, Software, Vali­
dation, Visualization, Investigation, Writing – original draft. Gabriel
5. Conclusions Contreras-Zarazúa: Methodology, Software, Validation, Investigation,
Formal analysis, Writing – review & editing. Eduardo Sánchez-Ram­
A hybrid stochastic-deterministic algorithm for optimal flowsheet írez: Methodology, Software, Validation, Investigation, Formal analysis,
design involving ordered discrete decisions, e.g., the number of stages/ Writing – review & editing. Juan Gabriel Segovia-Hernández: Meth­
reactors in a superstructure, is presented in this work. This hybrid al­ odology, Investigation, Formal analysis, Writing – review & editing.
gorithm relies on the parallel execution of a deterministic algorithm that Luis A. Ricardez-Sandoval: Funding acquisition, Supervision, Project
incorporates a Discrete-Steepest Descent Algorithm with Variable administration, Formal analysis, Writing – review & editing.
Bounding (DSDA-VB) and a stochastic algorithm that implements a
Stochastic Method (SM), e.g., differential Evolution with Tabu List Declaration of Competing Interest
(DETL). The key feature of the proposed hybrid method is the parallel
interaction between DSDA-VB and the SM, through continuous ex­ The authors declare that they have no known competing financial
change of information. The SM is constantly generating new solution interests or personal relationships that could have appeared to influence
candidates that are optimized locally by the deterministic DSDA-VB the work reported in this paper.
method. In turn, the DSDA-VB generates new bounds for discrete and
continuous variables that guide the SM toward a local optimum, while Data availability
still allowing exploration by the SM within a neighborhood of the local
optimum. The above procedure is executed iteratively, which allows the Data will be made available on request.
identification of different local optima during the search. Contrary to
other hybrid methods, the proposed approach combines the advantages
of a pure SM with the streamlined efficiency of the DSDA-VB, resulting
in a practical and computationally efficient algorithm. Two case studies
involving intensified distillation units were considered to test the

13
D.A. Liñán et al. Computers and Chemical Engineering 180 (2024) 108501

Acknowledgments Herrera Velázquez, J.J., Zavala Durán, F.M., Chávez Díaz, L.A., Cabrera Ruiz, J.,
Alcántara Avila, J.R., 2022. Hybrid two-step optimization of internally heat-
integrated distillation columns. Journal of the Taiwan Institute of Chemical
The financial support provided by the Natural Sciences and Engi­ Engineers 130, 103967. https://doi.org/10.1016/j.jtice.2021.06.061. J. Taiwan
neering Research Council of Canada (NSERC) is gratefully Inst. Chem. Eng.
acknowledged. Holtbruegge, J., Kuhlmann, H., Lutze, P., 2015. Process analysis and economic
optimization of intensified process alternatives for simultaneous industrial scale
production of dimethyl carbonate and propylene glycol. Chem. Eng. Res. Des. 93,
Supplementary materials 411–431. https://doi.org/10.1016/j.cherd.2014.05.002.
Hong, J., Van Duc Long, N., Harvianto, G.R., Haider, J., Lee, M., 2019. Design and
optimization of multi-effect-evaporation-assisted distillation configuration for
Supplementary material associated with this article can be found, in recovery of 2,3-butanediol from fermentation broth. Chem. Eng. Process. Process
the online version, at doi:10.1016/j.compchemeng.2023.108501. Intensif. 136, 107–115. https://doi.org/10.1016/j.cep.2019.01.002.
Javaloyes-Antón, J., Kronqvist, J., Caballero, J.A., 2022. Simulation-based optimization
of distillation processes using an extended cutting plane algorithm. Comput. Chem.
References Eng. 159, 107655. https://doi.org/10.1016/j.compchemeng.2021.107655.
Jia, S., Cao, X., Qian, X., Liu, X., Luo, Y., Yuan, X., 2023. Boundary function method for
stage number optimization for multi-stage distillation process design. Chem. Eng.
Aliaga-Vicente, A., Caballero, J.A., Fernández-Torres, M.J., 2017. Synthesis and
Sci. 276, 118773 https://doi.org/10.1016/j.ces.2023.118773.
optimization of membrane cascade for gas separation via mixed-integer nonlinear
Kronqvist, J., Bernal, D.E., Lundell, A., Grossmann, I.E., 2019. A review and comparison
programming. AlChE J. 63, 1989–2006. https://doi.org/10.1002/aic.15631.
of solvers for convex MINLP. Optim. Eng. 20, 397–455. https://doi.org/10.1007/
Barker, G., 2018. The engineer’s guide to plant layout and piping design for the oil and
s11081-018-9411-8.
gas industries. Chapter 11 - Towers 285–308. https://doi.org/10.1016/B978-0-12-
Kruber, K.F., Grueters, T., Skiborowski, M., 2021. Advanced hybrid optimization
814653-8.00011-4. Barker, G. (Ed.).
methods for the design of complex separation processes. Comput. Chem. Eng. 147,
Bernal, D.E., Carrillo-Diaz, C., Gómez, J.M., Ricardez-Sandoval, L.A., 2018. Simultaneous
107257 https://doi.org/10.1016/j.compchemeng.2021.107257.
design and control of catalytic distillation columns using comprehensive rigorous
Liberti, L., Kucherenko, S., 2005. Comparison of deterministic and stochastic approaches
dynamic models. Ind. Eng. Chem. Res. 57, 2587–2608. https://doi.org/10.1021/acs.
to global optimization. Int. Trans. Oper. Res. 12, 263–285. https://doi.org/10.1111/
iecr.7b04205.
j.1475-3995.2005.00503.x.
Carlson, E.C., 1996. Don’t gamble with physical properties for simulations. Chem. Eng.
Liñán, D.A., Bernal, D.E., Gómez, J.M., Ricardez-Sandoval, L.A., 2021. Optimal synthesis
Prog. 92, 35–46.
and design of catalytic distillation columns: a rate-based modeling approach. Chem.
Chen, X., Li, M., 2021. M♮-convexity and its applications in operations. Oper. Res. 69,
Eng. Sci. 231, 116294 https://doi.org/10.1016/j.ces.2020.116294.
1396–1408. https://doi.org/10.1287/opre.2020.2070.
Liñán, D.A., Bernal, D.E., Ricardez-Sandoval, L.A., Gómez, J.M., 2020. Optimal design of
Chia, D.N., Duanmu, F., Sorensen, E., 2021. Optimal Design of Distillation Columns
superstructures for placing units and streams with multiple and ordered available
Using a Combined Optimisation Approach. Elsevier, pp. 153–158. https://doi.org/
locations. Part II: rigorous design of catalytic distillation columns. Comput. Chem.
10.1016/B978-0-323-88506-5.50025-5.
Eng. 139, 106845 https://doi.org/10.1016/j.compchemeng.2020.106845.
Ciric, A.R., Gu, D., 1994. Synthesis of nonequilibrium reactive distillation processes by
Liñán, D.A., Ricardez-Sandoval, L.A., 2023. A Benders decomposition framework for the
MINLP optimization. AlChE J. 40, 1479–1487. https://doi.org/10.1002/
optimization of disjunctive superstructures with ordered discrete decisions. AlChE J.
aic.690400907.
69, e18008. https://doi.org/10.1002/aic.18008.
Contreras-Zarazúa, G., Jasso-Villegas, M.E., Ramírez-Márquez, C., Sánchez-Ramírez, E.,
Ma, Y., Li, J., 2022. Homotopy continuation enhanced branch and bound algorithms for
Vázquez-Castillo, J.A., Segovia-Hernandez, J.G., 2021. Design and intensification of
strongly nonconvex mixed-integer nonlinear optimization. AlChE J. 68, e17629.
distillation processes for furfural and co-products purification considering economic,
https://doi.org/10.1002/aic.17629.
environmental, safety and control issues. Chem. Eng. Process. Process Intensif. 159,
Ma, Y., Yang, Z., El-Khoruy, A., Zhang, N., Li, J., Zhang, B., Sun, L., 2021. Simultaneous
108218 https://doi.org/10.1016/j.cep.2020.108218.
synthesis and design of reaction–separation–recycle processes using rigorous models.
Contreras-Zarazúa, G., Martin-Martin, M., Sánchez-Ramirez, E., Segovia-Hernández, J.
Ind. Eng. Chem. Res. 60, 7275–7290. https://doi.org/10.1021/acs.iecr.1c00250.
G., 2022. Furfural production from agricultural residues using different intensified
Munawar, S.A., Gudi, R.D., 2005. A nonlinear transformation based hybrid evolutionary
separation and pretreatment alternatives. Economic and environmental assessment.
method for MINLP solution. Chem. Eng. Res. Des. 83, 1218–1236. https://doi.org/
Chem. Eng. Process. Process Intensif. 171, 108569 https://doi.org/10.1016/j.
10.1205/cherd.04286.
cep.2021.108569.
Murota, K., 2003. Discrete convex analysis, discrete mathematics and applications. Soc.
Contreras-Zarazúa, G., Sánchez-Ramírez, E., Vázquez-Castillo, J.A., Ponce-Ortega, J.M.,
Ind. Appl. Math. https://doi.org/10.1137/1.9780898718508.
Errico, M., Kiss, A.A., Segovia-Hernández, J.G., 2019. Inherently safer design and
Murota, K., Tamura, A., 2023. Recent progress on integrally convex functions. Jpn. J.
optimization of intensified separation processes for furfural production. Ind. Eng.
Indust. Appl. Math. https://doi.org/10.1007/s13160-023-00589-4.
Chem. Res. 58, 6105–6120. https://doi.org/10.1021/acs.iecr.8b03646.
Peccini, A., Jesus, L.F.S., Secchi, A.R., Bagajewicz, M.J., Costa, A.L.H., 2023. Globally
Costa, A.L.H., Bagajewicz, M.J., 2019. 110th anniversary: on the departure from
optimal distillation column design using set trimming and enumeration techniques.
heuristics and simplified models toward globally optimal design of process
Comput. Chem. Eng. 174, 108254 https://doi.org/10.1016/j.
equipment. Ind. Eng. Chem. Res. 58, 18684–18702. https://doi.org/10.1021/acs.
compchemeng.2023.108254.
iecr.9b02611.
Rafiei, M., Ricardez-Sandoval, L.A., 2020. New frontiers, challenges, and opportunities in
Cremaschi, S., 2015. A perspective on process synthesis: challenges and prospects.
integration of design and control for enterprise-wide sustainability. Comput. Chem.
Comput. Chem. Eng. 81, 130–137. https://doi.org/10.1016/j.
Eng. 132, 106610 https://doi.org/10.1016/j.compchemeng.2019.106610.
compchemeng.2015.05.007. Special Issue: Selected papers from the 8th
Ramapriya, G.M., Selvarajah, A., Jimenez Cucaita, L.E., Huff, J., Tawarmalani, M.,
International Symposium on the Foundations of Computer-Aided Process Design
Agrawal, R., 2018. Short-cut methods versus rigorous methods for performance-
(FOCAPD 2014), July 13-17, 2014, Cle Elum, Washington, USA.
evaluation of distillation configurations. Ind. Eng. Chem. Res. 57, 7726–7731.
Deb, K., 2009. Multi-Objective Optimization Using Evolutionary Algorithms. Wiley,
https://doi.org/10.1021/acs.iecr.7b05214.
Chichester Weinheim, 1st edition. ed.
Rangaiah, G.P., 2008. Multi-Objective Optimization: Techniques and Applications in
Douglas, J., 1988. Conceptual Design of Chemical Processes. McGraw-Hill Science/
Chemical Engineering, Har/Cdr edition. World Scientific Publishing Company,
Engineering/Math, New York, 1st edition. ed.
Hackensack, N.J ed.
Franke, M.B., 2017. Design of dividing-wall columns by mixed-integer nonlinear
Rangaiah G.P., Feng Z., Hoadley A.F., 2020. Multi-Objective Optimization Applications
programming optimization. Chem. Ing. Tech. 89, 582–597. https://doi.org/
in Chemical Process Engineering: Tutorial and Review. Processes 8, 508. doi:1
10.1002/cite.201700005.
0.3390/pr8050508.
Gómez, J.M., Reneaume, J.M., Roques, M., Meyer, M., Meyer, X., 2006. A mixed integer
Romero-García, A.G., Mora-Morales, C., Chargoy-Amador, J.P., Ramírez-Corona, N.,
nonlinear programming formulation for optimal design of a catalytic distillation
Sánchez-Ramírez, E., Segovia-Hernández, J.G., 2022. Implementing CO2 capture
column based on a generic nonequilibrium model. Ind. Eng. Chem. Res. 45,
process in power plants: Optimization procedure and environmental impact. Chem.
1373–1388. https://doi.org/10.1021/ie0504506.
Eng. Res. Des. 180, 232–242. https://doi.org/10.1016/j.cherd.2022.02.023.
González-Navarrete, C., Sánchez-Ramírez, E., Ramírez-Márquez, C., Hernández, S.,
Ryu, J., Kong, L., Pastore de Lima, A.E., Maravelias, C.T., 2020. A generalized
Cossío-Vargas, E., Segovia-Hernández, J.G., 2022. Innovative reactive distillation
superstructure-based framework for process synthesis. Comput. Chem. Eng. 133,
process for the sustainable purification of lactic acid. Ind. Eng. Chem. Res. 61,
106653 https://doi.org/10.1016/j.compchemeng.2019.106653.
621–637. https://doi.org/10.1021/acs.iecr.1c04050.
Sánchez-Ramírez, E., Hernández, S., Romero-García, A.G., Alcocer-García, H., Segovia –
Gooty, R.T., Agrawal, R., Tawarmalani, M., 2022. Advances in MINLP to identify energy-
Hernández, J.G., 2022. Synthesis and optimization of sustainable processes based on
efficient distillation configurations. Oper. Res. https://doi.org/10.1287/
liquid-liquid extraction to purify methyl ethyl ketone. Chem. Eng. Process. Process
opre.2022.2340.
Intensif. 171, 108522 https://doi.org/10.1016/j.cep.2021.108522.
Göttl, Q., Grimm, D.G., Burger, J., 2022. Automated synthesis of steady-state continuous
Segovia-Hernández, J.G., Hernández, S., Bonilla Petriciolet, A., 2015. Reactive
processes using reinforcement learning. Front. Chem. Sci. Eng. 16, 288–302. https://
distillation: a review of optimal design using deterministic and stochastic techniques.
doi.org/10.1007/s11705-021-2055-9.
Chem. Eng. Process. 97, 134–143. https://doi.org/10.1016/j.cep.2015.09.004.
Hernández-Pérez, L.G., Ramírez-Márquez, C., Segovia-Hernández, J.G., Ponce-Ortega, J.
Sharma S., Rangaiah G.P., Cheah K.S., 2012. Multi-objective optimization using MS Excel
M., 2020. Simultaneous structural and operating optimization of process flowsheets
with an application to design of a falling-film evaporator system. Food Bioprod.
combining process simulators and metaheuristic techniques: the case of solar-grade
Process. 90, 123–134. doi:10.1016/j.fbp.2011.02.005.
silicon process. Comput. Chem. Eng. 140, 106946. https://doi.org/10.1016/j.
compchemeng.2020.106946.

14
D.A. Liñán et al. Computers and Chemical Engineering 180 (2024) 108501

Skiborowski, M., Rautenberg, M., Marquardt, W., 2015. A Hybrid Urselmann, M., Barkmann, S., Sand, G., Engell, S., 2011. A memetic algorithm for global
Evolutionary–Deterministic Optimization Approach for Conceptual Design. Ind. Eng. optimization in chemical process synthesis problems. IEEE Trans. Evol. Comput. 15,
Chem. Res. 54, 10054–10072. https://doi.org/10.1021/acs.iecr.5b01995. 659–683. https://doi.org/10.1109/TEVC.2011.2150753.
Spall, J.C., 2003. Introduction to Stochastic Search and Optimization: Estimation, Urselmann, M., Barkmann, S., Sand, G., Engell, S., 2011. Optimization-based design of
Simulation, and Control, 1st edition. Wiley-Interscience, Hoboken, N.J. ed. reactive distillation columns using a memetic algorithm. Comput. Chem. Eng. 35,
Srinivas, M., Rangaiah, G.P., 2016. Differential evolution with tabu list for global 787–805. https://doi.org/10.1016/j.compchemeng.2011.01.038. Selected Papers
optimization: evaluation of two versions on benchmark and phase stability from ESCAPE-20 (European Symposium of Computer Aided Process Engineering -
problems, in: differential evolution in chemical engineering. Advances in Process 20), 6-9 June 2010, Ischia, Italy.
Systems Engineering. World Scientific, pp. 91–127. https://doi.org/10.1142/ Urselmann, M., Janus, T., Foussette, C., Tlatlik, S., Gottschalk, A., Emmerich, M.T.M.,
9789813207523_0004. Bäck, T., Engell, S., 2016. Derivative-free chemical process synthesis by memetic
Srinivas, M., Rangaiah, G.P., 2007. Differential evolution with tabu list for solving algorithms coupled to aspen plus process models. Kravanja, Z., Bogataj, M. (Eds.).
nonlinear and mixed-integer nonlinear programming problems. Ind. Eng. Chem. Res. Comput. Aided Chem. Eng. 187–192. https://doi.org/10.1016/B978-0-444-63428-
46, 7126–7135. https://doi.org/10.1021/ie070007q. 3.50036-9, 26 European Symposium on Computer Aided Process Engineering.
Srinivas, M., Rangaiah, G.P., 2006. An integrated stochastic method for global Wang, Z., Parhi, S.S., Rangaiah, G.P., Jana, A.K., 2020. Analysis of weighting and
optimization of continuous functions. Marquardt, W., Pantelides, C. (Eds.). selection methods for pareto-optimal solutions of multiobjective optimization in
Computer Aided Chemical Engineering, 16th European Symposium on Computer chemical engineering applications. Ind. Eng. Chem. Res. 59, 14850–14867. https://
Aided Process Engineering and 9th International Symposium on Process Systems doi.org/10.1021/acs.iecr.0c00969.
Engineering. Elsevier, pp. 439–444. https://doi.org/10.1016/S1570-7946(06) Wang, Z., Rangaiah, G.P., 2017. Application and analysis of methods for selecting an
80085-4. optimal solution from the pareto-optimal front obtained by multiobjective
Staudt, P.B., Soares, R., 2009. Reliability vs. efficiency when solving multiphase optimization. Ind. Eng. Chem. Res. 56, 560–574. https://doi.org/10.1021/acs.
equilibrium problems with hybrid optimization codes. Computer Aided Chemical iecr.6b03453.
Engineering. In: Proceedings of the 10th International Symposium on Process Zhang, H., Zheng, Z., Lavaei, J., 2022. Gradient-based algorithms for convex discrete
Systems Engineering: Part A. Elsevier, pp. 585–590. https://doi.org/10.1016/ optimization via simulation. Oper. Res. https://doi.org/10.1287/opre.2022.2295.
S1570-7946(09)70318-9. Zhang, X., Song, Z., Zhou, T., 2018. Rigorous design of reaction-separation processes
Storn, R., 1996. On the usage of differential evolution for function optimization. In: using disjunctive programming models. Comput. Chem. Eng. 111, 16–26. https://
Proceedings of the North American Fuzzy Information Processing. Presented at the doi.org/10.1016/j.compchemeng.2017.12.013.
Proceedings of North American Fuzzy Information Processing, pp. 519–523. https:// Zhou, T., Zhou, Y., Sundmacher, K., 2017. A hybrid stochastic–deterministic
doi.org/10.1109/NAFIPS.1996.534789. optimization approach for integrated solvent and process design. Chem. Eng. Sci.
Torres-Vinces, L., Contreras-Zarazua, G., Huerta-Rosas, B., Sánchez-Ramírez, E., Segovia- 159, 207–216. https://doi.org/10.1016/j.ces.2016.03.011 iCAMD – Integrating
Hernández, J.G., 2020. Methyl ethyl ketone production through an intensified Computer-Aided Molecular Design into Product and Process Design.
process. Chem. Eng. Technol. 43, 1433–1441. https://doi.org/10.1002/ Zielinski, K., Laur, R., 2008. Stopping criteria for differential evolution in constrained
ceat.201900664. single-objective optimization. Chakraborty, U.K. (Ed.). Advances in Differential
Tula, A.K., Babi, D.K., Bottlaender, J., Eden, M.R., Gani, R., 2017. A computer-aided Evolution, Studies in Computational Intelligence. Springer, Berlin, Heidelberg,
software-tool for sustainable process synthesis-intensification. Comput. Chem. Eng. pp. 111–138. https://doi.org/10.1007/978-3-540-68830-3_4.
105, 74–95. https://doi.org/10.1016/j.compchemeng.2017.01.001. Process
Intensification.

15