Integrating MILP, Discrete-Event Simulation, and Data-Driven Models for Distributed Flow Shop Scheduling Using Benders Cuts

Abstract

Digitalization plays a crucial role in improving the performance of chemical companies. In this context, different modeling, simulation, and optimization techniques such as MILP, discrete-event simulation (DES), and data-driven (DD) models are being used. Due to their heterogeneity, these techniques must be executed individually, and holistic optimization is manual and time-consuming. We propose Benders decomposition to combine these techniques into one rigorous optimization procedure. The main idea is that heterogeneous models can simultaneously be optimized as Benders subproblems. We illustrate this concept with the distributed permutation flow shop scheduling problem (DPFSP) and assume that a MILP, DES, and DD model exist for three flow shops. Our approach can compute bounds and report gap information on the optimal makespan for five medium-sized literature instances. The approach is promising because it enables the optimization of heterogeneous models and makes it possible to build optimization capabilities on an existing model and tool landscape in chemical companies.

1. Introduction

In the chemical industry, advanced modeling and optimization techniques are crucial for operational decision-making. Their relevance is growing with the digital transformation of production, increasing sustainability requirements, and tougher global competition. Chemical companies often use a variety of different modeling and optimization techniques, each adapted for specific purposes and each with its own strengths and limitations. For example, mixed integer linear programming (MILP) can capture the discrete and continuous aspects of scheduling and planning problems, but it requires problem abstraction and simplification to have acceptable solution times. Discrete-event simulation (DES) can represent the dynamic behavior from the supply chain to the shop floor level in detail and can include stochastic effects, but it is costly to optimize DES models, and the solution quality can be inferior to MILP. Data-driven (DD) methods can exploit large amounts of data and bridge the gap if first principles are not known or are too complicated, but solutions may lack interpretability and causality, and the solution reliability can be inferior to MILP and DES.

As a result, modeling and optimization techniques may exist in many successful but siloed applications. A unifying, rigorous optimization framework is lacking, which would enable chemical companies to move from traditional, component-oriented optimization to system-wide optimization. Different research attempts on this topic have already been made. The need for integrated methods in optimization is studied in [1], and a unified treatment of mathematical programming, constraint programming, and global optimization is suggested. However, simulation techniques and data-driven approaches are not addressed. In [2], a bibliometric analysis of hybrid approaches involving optimization and machine learning algorithms is conducted. It shows that there is a growing interest in combining metaheuristic approaches, such as particle swarm optimization, with machine learning techniques. However, the integration of rigorous optimization techniques is not covered in the analysis. In [3], a comprehensive overview and taxonomy of simulation-optimization methods are given, and the synergies between the methods are discussed. However, the authors conclude that there are gaps in the simulation-optimization interaction matrix, and they specifically suggest investigating the AME-OSI combination. In this combination, the optimization proceeds in iterations that contain one or more simulation runs (OSI), and each run is used to refine and enhance the analytical model (AME). Furthermore, unifying optimization and decision-making frameworks have already been suggested for industrial settings, for example, to support collaborative supply chain decision-making in the chemical industry [4], to optimize supplier selection for green and sustainable supply chains [5], and to improve operational strategies and overall decision orchestration in pharmaceutical companies [6]. However, these frameworks operate on a conceptual level, and they lack a rigorous, algorithmic integration layer that interlinks the different methods, models, and tools.

Herein, we propose a novel framework that integrates MILP, DES, and DD models into one rigorous optimization framework for production scheduling in the chemical industry. The main idea is based on [7,8,9], which suggests using Benders decomposition for a MILP-DES integration. The advantage of Benders decomposition is that different model types can be considered in the different subproblems as long as Benders cuts can be generated from the subproblem solutions. In this work, we use this concept to integrate three different model types into one optimization framework. We show that the integration of MILP, DES, and DD models is possible and that the Benders master problem can guide the search process in these models efficiently while yielding dual bounds on the overall problem. We apply our approach to literature instances of the well-known distributed permutation flow shop scheduling problem (DPFSP) with a makespan minimization objective, as presented in [10]. The DPFSP is an important real-world problem in production and supply chain management in chemical companies and is one of the fastest-growing topics in the scheduling and operations research literature [11]. Production and supply chain management are very often related to heterogeneous methods, models, and tools that originate from the different regions, manufacturing sites, and production plants of chemical companies. To this end, we study DPFSP instances featuring three factories (plants) and assume that a MILP, a DES, and a DD model exist for these factories. To the best of our knowledge, this is the first approach that integrates different simulation and optimization methods rigorously and provides a comprehensive and flexible solution framework for the pre-existing and specialized models and tools in chemical companies.

The rest of the article is structured as follows: In Section 2, we review current approaches to distributed flow shop (DFS) optimization and Benders decomposition. In Section 3, we present the optimization framework featuring the Benders master problem, Benders cuts, and different types of subproblems. In Section 4, we apply our approach to a literature case study of a DPFSP featuring three factories, represented by a MILP, DES, and DD model, and discuss the progression of the upper and lower bounds. In Section 5, we discuss the approach from an industrial perspective. In Section 6, we present ethical considerations and conclude with an outlook for future research in Section 7.

2. Literature

In the following, we briefly review current techniques for DFS optimization as well as Benders decomposition.

2.1. Distributed Flow Shop Optimization

A DFS consists of several flow shops arranged in parallel. DFS scheduling is the task of assigning each job to one of the flow shops and sequencing all jobs in each flow shop. The decision variables, therefore, include assignment and sequence variables. Each flow shop consists of multiple machines, and the sequence (permutation) of the jobs can either change between the machines (distributed non-permutation flow shop) or be fixed (distributed permutation flow shop). DFS scheduling, as initially mentioned in [12], can therefore be seen as a combination of parallel machine scheduling [13,14,15,16] and flow shop scheduling [17,18,19], which are both well-studied problems. Reviews and classifications of DFS scheduling problems have recently been proposed [11,20].

DFS scheduling problems are most commonly solved with the makespan minimization objective [21], which translates into the minimization of the maximum completion time among all jobs. A standard method for DFS scheduling is mathematical programming, such as MILP [10] or constraint programming (CP) [22]. A general advantage of mathematical programming lies in its ability to use the structure of closed-form models to optimize the solution process, for example, by generating cuts, solving the dual problem, and pruning the search space. However, a mathematical programming model must be set up, which is often a difficult and time-consuming task. Especially for complex systems, such as real-world DFS, building accurate and tractable mathematical models quickly becomes an impossible task from a practical point of view. In this context, heuristics and metaheuristics become important alternatives because they do not require mathematical programming models. Since they are based on the systematic evaluation of a black-box model, simulation and DD models can be used. Heuristic approaches have traditionally held a very strong position for DFS scheduling due to their simplicity and effectiveness. Based on the well-known Nawaz-Enscore-Ham (NEH) heuristic for simple flow shop problems [23], an improved heuristic that implements a novel insertion rule is proposed [24]. Closely related are iterated greedy methods, as initially proposed for permutation flow shops in [25]. The basic idea is to repeatedly construct an initial solution, destruct it by removing some jobs, and then reconstruct it by re-inserting the removed jobs in a new sequence to find a potentially better solution. This iterative process, combined with local search heuristics and optimization techniques, is one of the most successful approaches to DFS scheduling and has led to multiple research contributions. In [26], a modified iterated greedy procedure is adopted for the DPFSP. To this end, a bounded search iterated greedy algorithm is presented in [27], which exploits the specific structure of DPFSP. The authors claim that their heuristic improved the state-of-the-art of existing heuristics at that time. In [28], an enhanced two-stage iterated greedy method is suggested, which again led to a significant performance improvement. In addition to heuristic approaches, different metaheuristics, such as scatter search [29], taboo search [30], and genetic algorithms [31], have been adopted successfully. Although they are not always as efficient as heuristic approaches, metaheuristics generally require less customization as they are based on general and versatile search strategies. In addition, many standard packages are available as open-source codes for metaheuristics. Both heuristic and metaheuristic methods are compatible with a variety of different model types, as long as they can be evaluated as a black-box model. Therefore, they could also be used to optimize heterogeneous models.

However, heuristics and metaheuristics have two crucial drawbacks. Since they are primal search methods, they do not use dual bounds to reduce the search space and do not provide optimality gap information. Hence, finding good solutions relies on smart, but ultimately brute-force, primal search strategies. The solution quality cannot be assessed by optimality gap information, and the user must decide when to terminate the search process, leaving the potential solution improvement unknown. This work presents an alternative mathematical optimization framework that is capable of integrating heterogeneous models, uses dual bounds, and reports optimality gap information. The framework is based on Benders decomposition, which we review in the following.

2.2. Benders Decomposition

Benders decomposition [32] is a technique in mathematical programming that allows the splitting of the variables of an optimization problem into complicating and non-complicating variables to obtain a master problem with the complicating variables and independent subproblems with the non-complicating variables. The master and subproblems are solved individually in an iterative procedure in which constraints (Benders cuts) are generated from the dual information of subproblem solutions and added to the master problem. The advantage of Benders decomposition is that smaller, tractable subproblems are obtained, which can be treated, for example, using specialized algorithms. In [33], Benders decomposition is used to solve an improved, position-based MILP model of the DPSFP. The master problem contains the binary decision variables that encode the factory assignment and position of jobs. The subproblems contain the continuous decision variables, which are the completion times of jobs. A hybrid Benders decomposition algorithm is presented, which applies the local search algorithm (LS3) [28] to improve each master problem solution heuristically before sending it to the subproblems. With the improved and original master problem solution, two cuts are generated in each Benders iteration, which accelerates the convergence of the algorithm. In [34], the mixed no-idle permutation flow shop scheduling problem is solved using a similar concept. The referenced local search heuristic (RLS) [35] is used to generate several neighborhood master solutions, which lead to multiple Benders cuts. To the best of our knowledge, these are the only reported applications of Benders decomposition for (distributed) flow shop scheduling problems. Furthermore, in these applications, the subproblems are linear programs (LPs), which result from fixing the integer variables of the original MILP model. Therefore, we extend our literature analysis to Benders decomposition applied to scheduling problems in general, in which the subproblems do not directly come from the original MILP model but are of different types.

In [36], the Benders decomposition concept is generalized, and logic-based Benders cuts are proposed, which may be derived from any type of subproblem as long as the cuts represent valid inequalities. The combination of a MILP master problem and CP subproblem is well-known and shows superior performance over monolithic MILP or CP models for many types of scheduling problems, such as project scheduling with multi-skilled personnel [37], single batching machine scheduling [38], preemptive flexible job-shop scheduling [39], and two-stage flexible flow shop scheduling with unrelated parallel machines [40]. The idea of using a DES model as a subproblem in the Benders decomposition framework is relatively new. In one of the first examples [8], this concept is suggested to optimize the buffer allocation problem (BAP) in multi-stage serial-parallel manufacturing systems. Benders cuts are derived from a DES model to find the optimal design parameters of a joint workstation, workload, and buffer allocation system. In [9], this concept is further developed, and improved logic-based Benders cuts are proposed for the master problem to solve operational scheduling tasks. The approach is applied to a shift scheduling problem as well as a personnel allocation problem and can solve realistic-sized instances exactly in reasonable computing times.

This work presents a novel use case for Benders decomposition. In contrast to [7,8,9], we apply Benders decomposition to the DPFSP. In contrast to [33], we use three different model types in the subproblems and thereby leverage Benders decomposition as an integration framework, which can be used to combine pre-existing heterogeneous models in the chemical industry.

3. Methodology

We first introduce the general concept of using Benders decomposition with heterogeneous subproblems in Section 3.1 and then adopt the concept of distributed permutation flow shops in Section 3.2.

3.1. Benders Decomposition with Heterogeneous Subproblems

We consider the minimization problem shown in Equation (1) involving continuous variables $x\in {\mathbb{R}}^n$ and binary variables $y\in {\left\{0,1\right\}}^n$. Benders decomposition can be applied with respect to $x$ and $y$, such that (1) decomposes into a part (1d), which contains the combinatorial complexity and encodes the logical behavior, and a part (1c), which involves continuous decisions that are linked to the binary decisions.

$$\min z$$

(1a)

$$z=c^{T}x$$

(1b)

$$s.t. Ax+By\le b$$

(1c)

$$Cy\le d$$

(1d)

$$x\ge 0$$

(1e)

$$y\in {\left\{0,1\right\}}^n$$

(1f)

This yields the MILP Benders master problem in (2), which computes lower bounds $\mu$ on $z$. The master problem consists of the constraints block (1d) of the original problem, optimality cuts $\mu \ge$ $u^T\left(b-Bx\right)$ with $u \in O^{\prime }\subset O$, and feasibility cuts $0\ge u^T\left(b-Bx\right)$ with $u \in F^{\prime }\subset F$, that are generated based on solutions of the subproblems (1c). The set $O$ contains all dual solutions of (1) for which the dual problem is feasible and bounded, while the set $F$ contains all dual solutions for which the dual problem is unbounded. However, we only need to generate cuts from the subsets $O^{\prime }$ and $F^{\prime }$, since only a finite number of cuts are active constraints in the optimum of (2).

$$\begin{gathered} \min \mu \\ \mathrm{s}.\mathrm{t}.\begin{array}{c} \\ \mu \ge u^T\left(b-By\right) \forall u\in O^{\prime }\\ 0\ge u^T\left(b-By\right) \forall u\in F^{\prime }\\ Cy\le d\\ \mu \in \mathbb{R}\\ y\in {\left\{0,1\right\}}^n\end{array} \end{gathered}$$

(2)

In each Benders iteration, we obtain a solution vector $\widehat{y}$ from (2). Based on $\widehat{y}$, we solve the individual subproblems of the form (3), where ${\widehat{y}}_s$ is the part of the solution vector $\widehat{y}$ that corresponds to subproblem $s\in S$, $x_s$ is the corresponding part of the decision vector $x$, and $A_s$, $B_s$, $b_s$, and $c_s^T$ describe subproblem $s$.

$$\begin{gathered} \min z_s \\ \mathrm{s}.\mathrm{t}.\begin{array}{c} \\ z_s= c_s^T{x}_s\\ A_s{x}_s+B_s{\widehat{y}}_s\le b_s\\ x_s\ge 0\end{array} \end{gathered}$$

(3)

We argue that this classical Benders decomposition framework can be used to integrate LP, DES, and DD models for the DPFSP, as illustrated in Figure 1. The basic idea is to generate Benders cuts from an LP model $f_{LP}$, a DES model $f_{DES}$, and a DD model $f_{DD}$ which act as the Benders subproblems, such that $S=\{LP, DES, DD\}$. For the LP subproblem, the dual multipliers $u_{LP}$ can be directly retrieved from the dual solution. For the DES subproblem, the dual multipliers $u_{DES}$ correspond to the sensitivity information of the DES solutions and can be derived, using a dual mapping. For the DD subproblem, no dual multipliers can be derived and only the objective value information $z_{DD}$ can be used to generate logic-based Benders cuts of the form $\mu \ge E_{DD}{y}_{DD}+e_{DD}$, where the cut parameters $E_{DD}$ and $e_{DD}$ are determined through the evaluation of the DD models. For all three models feasibility cuts can be formulated that exclude ${\widehat{y}}_s$ from the feasible set of the master problem. In the following, this concept is explained in more detail and applied to the DPFSP.

3.2. Application to the DPFSP

We consider the DPFSP as reviewed in Section 2.1, assume the makespan minimization objective, and apply Benders decomposition as discussed in Section 3.1, which leads to a mixed-integer programming (MIP) master problem and three subproblems of different model types. The task of the master problem is to assign each job to a factory and determine the sequence of the jobs for each factory. The task of the subproblems is to compute the makespan (primal information) and, if possible, the makespan sensitivity (dual information), from which the dual multipliers and Benders cuts are derived.

MIP master problem. The Benders master problem (4) is based on job sequencing variables $y_{ijf}$, which are equal to one if job $i$ precedes job $j$ on factory $f$ and are zero otherwise, as well as job assignment variables $a_{if}$, which are equal to one if job $i$ is assigned to factory $f$ and are zero otherwise. Constraints (4e) ensure that each job is assigned to exactly one factory, and constraints (4f) enforce the precedence relation between two jobs that are assigned to the same factory. Constraints (4g) and (4h) assign an artificial job with index zero to each factory and schedule it as the first job. As a result, it is possible to express the precedence relation of all real jobs $i,j$ by $y_{ijf}\ne 0$, which is needed to build the Benders optimality cuts. We do not need to consider feasibility cuts for the DPSFP because the flow shop subproblems always return primal feasible solutions when a set of jobs and permutations are assigned to them. Constraints (4b) are the optimality cuts from the LP subproblem; constraints (4c) are the optimality cuts from the DES subproblem; and constraints (4d) are the approximative optimality cuts from the DD subproblem. The parameter vector $t_s$ contains the processing times of all jobs assigned to subproblem $s$, and $T$ is a Big-M factor. The underlying assumption of the cuts (4d) is that each job assigned to the DD subproblem contributes to the makespan $z_{DD}$ equally, where $N_{DD}$ is the number of jobs assigned to the DD subproblem. Therefore, we compute the average makespan contribution per job ${\displaystyle \frac{z_{DD}}{N_{DD}}}$ and assume that a change in the job permutation causes a relaxation by ${\displaystyle \frac{z_{DD}}{N_{DD}}}$. While this approximation may exclude combinatorial effects on the makespan and therefore may cut off the global optimum, it strengthens the cuts. Furthermore, we select these approximative optimality cuts for the DD subproblem because we introduce some model inaccuracies compared to the LP and DES models, as we use a deep neural network to approximate the flow shop.

$$\min \mu$$

(4a)

$$\mathrm{s}.\mathrm{t}. \mu \ge u_{LP}^T\left(t_{LP}-T{(1-y}_{LP})\right)\forall u_{LP}\in O_{LP}^{\prime }$$

(4b)

$$\mu \ge u_{DES}^T\left(t_{DES}-T{(1-y}_{DES})\right)\forall u_{DES}\in O_{DES}^{\prime }$$

(4c)

$$\mu \ge y_{DD}^T\left({\displaystyle \frac{z_{DD}}{N_{DD}}},\dots ,{\displaystyle \frac{z_{DD}}{N_{DD}}}\right)\forall N_{DD},z_{DD}\in O_{DD}^{\prime }$$

(4d)

$$\sum _{f=1}^K{a}_{if} =1 \forall i\in \left\{1,\dots ,N\right\}$$

(4e)

$$y_{ijf}+y_{jif}=a_{if}{a}_{jf} \forall i,j\in \left\{1,\dots ,N\right\}\wedge i\ne j, f\in \left\{1,\dots ,F\right\}$$

(4f)

$$a_{0f}=0 \forall f\in \left\{1,\dots ,K\right\}$$

(4g)

$$y_{i0f}=0 \forall i\in \left\{1,\dots ,N\right\}, f\in \left\{1,\dots ,K\right\}$$

(4h)

In each Benders iteration, the master problem (4) is solved, which yields a lower bound on the makespan $c_{max}$ of the DPFSP as well as a job assignment and permutation for each factory subproblem. The subproblems compute the dual multipliers $\{u_{LP}$,$u_{DES}\}$ and individual makespans $\{z_{LP},z_{DES},z_{DD}\}$. Based on this primal and dual feedback, Benders cuts are formulated and added to (4), and an upper bound on $c_{max}$ is computed. The resulting algorithm is shown in Algorithm 1. Computational settings for the MIP master problem are shown in

Table 1. Computational settings for MIP master problem and LP subproblem.

Parameter	Value
Gurobi version	9.5.2
Gurobi hyperparameters	standard settings
Implementation environment	Gurobipy 11.0.1
Python version	3.7.6

, and general system settings are shown in Appendix B,

Table A2. General system settings.

Parameter	Value
CPU cores	64
CPU speed	Intel Xenon Gold 6338 @ 2.00 GHz
RAM	256 GB
Operating system	Windows 11 Enterprise

. In the following, we introduce the flow shop subproblems.

Algorithm 1: Benders Algorithm.
	Input: DPFSP featuring MILP, DES, and DD subproblem
	Output: Makespan-optimized schedule ($c_{max},\delta )$ containing job assignments $a_{if}$ and job sequences $y_{ijf}$  Initialize: ${\mathrm{c}}_{\mathrm{m}\mathrm{a}\mathrm{x}}=\mathrm{\infty }, \delta =\mathrm{\infty }$, $k=0$, t $=0$
	while ($\mathbf{k}$ < maximum iterations and $\mathbf{t}$< timeout and$\mathsf{\delta }$ > 0) do
		solve Benders master problem and obtain ${\mu }^k$
		solve subproblems with current master solution $a_{if}^k,y_{ijf}^k$
		obtain makespan ${\mathrm{c}}_{\mathrm{m}\mathrm{a}\mathrm{x}}^{\mathrm{k}}=\mathrm{m}\mathrm{a}\mathrm{x}\left({\mathrm{c}}_{\mathrm{m}\mathrm{a}\mathrm{x},1}^{\mathrm{k}},{\mathrm{c}}_{\mathrm{m}\mathrm{a}\mathrm{x},2}^{\mathrm{k}},{\mathrm{c}}_{\mathrm{m}\mathrm{a}\mathrm{x},3}^{\mathrm{k}}\right)$ derive dual multipliers $u_{LP}^k$,$u_{DES}^k$ and parameters $E_{DD},e_{DD}$
		build Benders cut for MILP, DES, and DD subproblem add Benders cuts to the master problem
		if (${\mathbf{c}}_{\mathbf{m}\mathbf{a}\mathbf{x}}^{\mathbf{k}}<{\mathbf{c}}_{\mathbf{m}\mathbf{a}\mathbf{x}}$) do
			update upper bound: ${\mathrm{c}}_{\mathrm{m}\mathrm{a}\mathrm{x}}$ ← ${\mathrm{c}}_{\mathrm{m}\mathrm{a}\mathrm{x}}^{\mathrm{k}}$
		update current gap: $\delta$ ← ${\displaystyle \frac{{\mathrm{c}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mu }^k}{{\mathrm{c}}_{\mathrm{m}\mathrm{a}\mathrm{x}}}}$
		increment $\mathrm{k}$ and $\mathrm{t}$

LP subproblem. The LP subproblem represents factory one and consists of a MILP model (with fixed integer variables) of a permutation flow shop featuring $m$ machines as shown in (5). Given an input job permutation ${\widehat{y}}_{LP}$, the completion time $c_{im1}$ of each job $i$ assigned to this factory is computed for each machine $m$. The objective is to minimize the makespan ${z_{LP}=c}_{max,1}$. The intra-machine constraints (5b) enforce the job sequence induced by ${\widehat{y}}_{LP}$, where $t_{jm1}$ is the processing time of job $i$ on machine $m$ in factory 1 and $T$ is a Big-M factor. The inter-machine constraints (5c) enforce valid completion times as jobs move from one machine $m$ to the next $m+1$ in the flow shop. From each solution of this model, dual multipliers $u_{LP}$ are obtained, which correspond to the active constraints (5b) and (5c) given the job permutation ${\widehat{y}}_{LP}$. Based on $u_{LP},$ the Benders cuts (4b) are built. Computational settings for the LP subproblem are shown in Table 1.

$$\mathit{min}{c}_{max,1}$$

(5a)

$$s.t. c_{im1}+t_{jm1} \le c_{jm1}+T\left(1-{\widehat{y}}_{ij1}\right) \forall i,j\in \left\{1,\dots ,N\right\}\wedge i\ne j, m\in \left\{1,\dots ,M\right\}$$

(5b)

$$c_{im1}+t_{im1} \le c_{i(m+1)1}+T\left(1-{\widehat{a}}_{i1}\right) \forall i\in \left\{1,\dots ,N\right\}, m\in \left\{1,\dots ,M-1\right\}$$

(5c)

$$c_{max,1}\ge c_{iM1} \forall i\in \left\{1,\dots ,N\right\}$$

(5d)

DES subproblem. The DES subproblem represents factory two and consists of a DES model of a permutation flow shop featuring $m$ machines as shown in Figure 2. Given an input job permutation ${\widehat{y}}_{DES}$, the completion times $c_{im2}$ of each job $i$ assigned to this factory are computed for each machine $m$. The objective is to minimize the makespan $z_{DES}=c_{max,2}$ of this factory. From each solution of the DES model, dual multipliers $u_{DES}$ are computed using the concept of critical path mapping as described in [7]. Based on $u_{DES},$ the Benders cuts (4c) are built. The computational settings for the DES subproblem are shown in

Table 2. Computational settings for DES subproblem.

Parameter	Value
DES modeling package	SimPy 4.1.1
SimPy modules used	Resource, Container, Process, timeout, event
Python version	3.7.6

.

DD subproblem. The DD subproblem represents factory three and consists of a deep neural network model as specified in

Table 3. Computational settings for DD subproblem.

Parameter	Value
Input	job permutation ${\widehat{y}}_{DD}$
Output	makespan $c_{max,3}$
Data source	DES model evaluations
Python modeling package	sklearn.neural_network (1.5.0), MLPRegressor
Number of samples	400,000
Sampling mode	random uniform
Number of hidden layers	5
Number of nodes in hidden layers	600, 500, 400, 300, 100
max_iter	1000

. The model input is a job permutation ${\widehat{y}}_{DD}$ and the output is the makespan $z_{DD}=c_{max,3}$. The model was trained on 400,000 randomly generated job permutations, based on which makespans were computed using a DES model. The training progression of one exemplary literature instance is shown in Figure 3. Note that we had to build a DD model artificially by sampling the DES model because we wanted to illustrate the functionality of Benders decomposition for heterogeneous models. In real-world settings, DD models often pre-exist and are the only way to query complex production data sets for KPIs such as the makespan.

4. Computational Studies

We consider the five DPFSP instances (Instances and solutions taken from http://soa.iti.es/problem-instances, accessed on 2 February 2024), as presented in [10], featuring $K=3$ factories, $M=5$ machines, and $N=10$ jobs. The literature IDs of the instances we used are given in Appendix A,

Table A1. Literature ID of instances from http://soa.iti.es/problem-instances, accessed on 2 February 2024.

Instance	Literature Name
81	l_3_10_5_1.txt
82	l_3_10_5_2.txt
85	l_3_10_5_3.txt
86	l_3_10_5_4.txt
87	l_3_10_5_5.txt

. According to [10], the number of possible solutions for instances of this size is $66\mathrm{*}10!=239500800$.

We analyze the upper- and lower-bound progression reported by the Benders algorithm with respect to the Benders iterations. Figure 4 shows that for all instances, meaningful lower bounds are found. Both lower bounds (LB) and upper bounds (UB) improve as the algorithm progresses. For four of the five instances, the Benders algorithm approximates the best-known literature solution through LBs and UBs. For instance, in 86, a better UB than the best-known literature solution is found.

For all instances, the remaining optimality gap $\delta =(UB-LB)/UB$ after 3000 Benders iterations and approximately ten hours of solution time is still more than 29 percent, as shown in

Table 4. Remaining optimality gap of Benders algorithm after 3000 iterations.

Instance	$\mathbf{Remaining} \mathbf{Gap} \mathit{\delta }$
81	0.29
82	0.35
85	0.36
86	0.44
87	0.32

, which shows some limitations of the approach. The relatively large remaining gap is due to the large number of possible solutions compared to the relatively small number of iterations and the fact that the Benders cuts cannot cover the large, combinatorial decision space sufficiently. We build the Benders cuts based on precedence variables $y_{ijf}$ and Big-M factors $T$ for the LP and DES subproblems to obtain a strong relaxation mechanism in the cuts. This mechanism is necessary, because a master solution that differs in one element $y_{ijf}$ from a previously evaluated master solution might lead to a completely different makespan. Therefore, a relaxation by $T$ is required to account for combinatorial effects for the exact LP and DES models. Furthermore, as the processing times of jobs are not factory-specific and we only use simple Benders cuts, symmetric solutions arise, which may add to the slow convergence. They could be addressed using symmetry-breaking constraints, which could lead to an improvement in the effectiveness of the cuts and tighter LBs.

Despite these areas of improvement, we conclude that the Benders algorithm yields acceptable UBs after moderate solution time, which proves our concept of rigorous optimization on a set $S$ of heterogeneous models $\{LP, DES, DD\}$. In addition to the UBs, remaining optimality gaps $\delta$ are reported, which determine when to terminate the search process. The gap information is not available with primal search methods and therefore qualitatively distinguishes the Benders algorithm from heuristic and meta-heuristic approaches. On the other hand, it is known that these approaches are very effective when adapted to scheduling problems such as the DPFSP. They remain the best alternative if no gap information is required, since in this case no dual information has to be extracted from the subproblems, Benders cuts have to be formed, or a Benders master problem has to be solved, which is a computational disadvantage of the Benders algorithm compared to primal search methods.

5. Discussion and Practical Examples

The computational results show that Benders decomposition can be used to integrate heterogeneous flow shop problems and solve DFS scheduling problems with a makespan minimization objective. In the following, we discuss this approach from an industrial perspective and highlight two practical examples in supply chain and operations management in the chemical industry to which the optimization framework can be applied.

In the decision-making hierarchy of chemical companies, the link between supply chain planning and site manufacturing is critical for optimal operations. A common approach is to optimize a global supply chain model, resulting in setpoints and production targets for sites and plants. The shop floor level is not considered in detail in the global model and usually must be flexible enough to fulfill the propagated targets. On the other hand, detailed process models may exist on the site or plant level, which include all relevant shop floor constraints, such as process utilities, personnel, logistics, purchasing, maintenance, and downtime constraints. The Benders algorithm can be used to integrate these models with the global supply-chain model, ensuring optimized and feasible operations. Assessing the impact of changes at the supply chain level on production can then help reduce the cost of flexibility in production.

In chemical companies, often a range of different, heterogeneous models and tools pre-exist. The integration of these models into a single optimization framework is one advantage of the proposed approach. In addition, the ability to combine different model types gives modeling experts the freedom to select the best-fitting modeling technique for a given problem. While DES models may be suitable to describe large-scale systems with many constraints in continuous time, MILP can be used when systems are easy to describe but difficult to optimize. Similarly, DD or hybrid models might be the best-fitting solution for specific problems. As long as these models can be queried as black-box models and Benders cuts can be formulated, potentially based on dual information, they can be used in the proposed framework.

While we discussed two examples in the chemical industry, the proposed approach has broader implications that could extend to other industries such as pharmaceuticals and energy. The integration of diverse modeling techniques can enhance manufacturing science and technology (MSAT) in pharmaceutical companies as well as sector coupling and energy transition modeling in energy companies. Additionally, the framework has the potential to advance future technological developments, such as Industry 4.0 and smart manufacturing, by enabling more efficient and integrated decision-making across functional areas of companies.

6. Ethical Considerations

The integration of MILP, DES, and DD models into a unified optimization framework for production scheduling in the chemical industry brings with it several ethical considerations. These considerations revolve around data usage, model assumptions, and the potential impacts on stakeholders and are reviewed in the following.

6.1. Data Usage

The framework relies heavily on data to drive decisions, especially in the DD models. It is essential to ensure that the data used is obtained, stored, and processed in compliance with relevant legal and ethical guidelines, such as data protection regulations such as GDPR. The use of large datasets, potentially involving sensitive information, such as processing times and personnel constraints, must be handled with strict confidentiality to avoid unauthorized access or misuse. Another important consideration is the potential for biases in the data that feeds into the models. For instance, DD models for some chemical plants might be built using real process data, capturing the true operational conditions and challenges. In contrast, other plants might rely on DES and MILP models that are based on idealized process data, which may not fully reflect the complexities and variability of actual operations. This discrepancy can introduce biases, leading to solutions that are more tailored to some plants than others, potentially resulting in uneven performance outcomes and inefficiencies. Addressing these biases is crucial to ensuring that the optimization framework provides fair and accurate solutions across all plants and does not favor certain job-plant assignments $a_{if}$ due to the nature of the underlying data and model type.

6.2. Model Assumptions

The assumptions and modeling depth of MILP, DES, and DD models can have significant implications. For instance, simplifying complex real-world problems for tractability might overlook critical factors that could impact the validity and reliability of the solutions. The ethical challenge here is to find a balance between model simplification for computational efficiency and the accuracy of real-world representation. Transparency in documenting and communicating these assumptions is essential to ensuring that users are aware of the limitations and potential implications of the models. Furthermore, the integration of models using Benders decomposition introduces its own set of assumptions and limitations in the form of Benders cuts, which need to be critically validated to ensure reliable results.

6.3. Potential Impact on Stakeholders

The deployment of the framework in chemical companies can have wide-ranging impacts on various stakeholders, including employees, customers, and the environment. For employees, the introduction of the framework could lead to changes in job roles, especially if production planning and scheduling become automated. Ethically, it is important to consider the social responsibility of companies in managing such transitions, including providing retraining and other forms of support to affected groups. From an environmental perspective, while the framework aims to optimize the make-up of batch processes, there is a risk that the primary focus on efficiency and throughput could overshadow aspects of sustainability. Ensuring that environmental impact is factored into the optimization process is crucial to prevent negative consequences, such as increased changeover costs, waste or energy consumption due to job sequencing decisions $y_{ijf}$, and increased logistics costs due to job-plant assignments $a_{if}$. Lastly, the use of this framework in decision-making must be carefully monitored to prevent any form of discrimination or inequity, particularly if the models favor certain outcomes over others due to biased data or assumptions. The framework must include mechanisms for ongoing verification and validation to ensure fairness and ethical integrity in its application.

7. Conclusions and Outlook

In this study, we showed how Benders decomposition can be used as an integration concept for heterogeneous models, that is, to combine models of different types into one rigorous optimization framework. We illustrated this concept with the class of the DPFSP and assumed that a MILP, DES, and DD model exist for three different flow shops. To this end, we solved five literature instances of the DPFSP from [10] and observed that the proposed algorithm can report upper and lower bounds as well as remaining gap information on the optimal makespan.

This work is early conceptual research, along with the first computational studies. There are areas for improvement and potential for future work. The remaining gap after 3000 Benders iterations and approximately ten hours of solution time is still more than 29%, which highlights the issues of convergence that still need to be addressed. Therefore, future work should focus on cut generation, for example, by combining the approach with a branch and Benders cut algorithm [41] or by deriving more no-good cuts in each Benders iteration using similar procedures as the CB separation heuristic in [42]. In addition, with the topic of DFS scheduling gaining more attention over the last few years, it would be interesting to apply the approach to a larger, more diverse, real-world case study. The ability to perform rigorous optimization on systems with heterogeneous models also has interesting practical advantages, such as optimality gap information and system integration, which could be explored further on this basis.