International Journal of Production Research

ISSN: 0020-7543 (Print) 1366-588X (Online) Journal homepage: [Link]/journals/tprs20

Modelling and solving sustainable supply chain

network design based on graph autoencoder
clustering algorithm

Yuhan Guo, Runsheng Chen, Youssef Boulaksil & Hamid Allaoui

To cite this article: Yuhan Guo, Runsheng Chen, Youssef Boulaksil & Hamid Allaoui (05
Aug 2025): Modelling and solving sustainable supply chain network design based on graph
autoencoder clustering algorithm, International Journal of Production Research, DOI:
10.1080/00207543.2025.2542506

To link to this article: [Link]

Published online: 05 Aug 2025.

Submit your article to this journal

Article views: 89

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

[Link]
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH
[Link]

Modelling and solving sustainable supply chain network design based on graph
autoencoder clustering algorithm
Yuhan Guo a , Runsheng Chen b , Youssef Boulaksil c and Hamid Allaoui d

a School of Artificial Intelligence and Information Engineering, Zhejiang University of Science and Technology, Hangzhou, People’s Republic of

China; b School of Software, Liaoning Technical University, Huludao, People’s Republic of China; c College of Business and Economics, UAE
University, Al Ain, United Arab Emirates; d Laboratoire de Génie Informatique et d’Automatique de l’Artois (LGI2A), Univ. Artois, UR 3926,
Béthune, France

ABSTRACT ARTICLE HISTORY

The modelling of Sustainable Supply Chain Network Design (SSCND) is evolving significantly with Received 26 October 2024
increasing problem diversity and complexity. Existing algorithms face substantial challenges in Accepted 24 July 2025
accommodating various models and effectively handling large-scale instances. To address these KEYWORDS
challenges, we propose an intercity distances supply chain network model to reflect real-world sce- Sustainability; supply chain
narios, and develop a clustering mapping algorithm based on Graph Autoencoder (GAE) to solve design; graph autoencoder;
the model. The mathematical model incorporates both economic and environmental sustainability clustering; solution mapping
dimensions, while also including supply chain responsiveness metrics. The algorithm operates by mechanism
abstracting attribute information of supply chain potential participants, generating sparse graphs,
and applying GAE clustering to create abstract nodes. The classified abstract nodes are then pro-
cessed using the simplex method to generate preliminary solutions, which are subsequently mapped
back into real-world solutions through a mapping mechanism. Experimental results demonstrate the
high efficiency and stability of the proposed approach. For large-scale instances, it produces high-
quality solutions in substantially less time compared to commercial solvers like CPLEX, offering a
novel and practical approach for enterprises addressing sustainable supply chain design challenges.

1. Introduction While sustainable supply chain design increasingly

Due to ongoing global challenges such as increased pollu- embraces the triple bottom line framework, the oper-
tion, resource depletion, and global warming, many com- ationalisation of social sustainability continues to face
panies face significant difficulties in achieving their sus- measurement and integration challenges. Based on the
tainability goals (Hegab et al. 2023). In light of the United study of Yadav et al. (2021) who considered delivery
Nations’ Sustainable Development Goals (SDGs), the time as a social dimension, we adopt responsiveness to
need for sustainable supply chains has become increas- optimise operational efficiency and indirectly reflect the
ingly urgent (Sorooshian 2024). Traditionally, the evalu- impact on the social dimension. Integrating responsive-
ation of supply chain network design focussed primar- ness into sustainable supply chain design allows compa-
ily on economic objectives. However, there has been nies to achieve multiple benefits that contribute to a more
a paradigm shift towards a triple bottom line (TBL) resilient and sustainable business model (Setiawan, Tari-
approach, which includes economic, environmental, and gan, and Siagian 2023). Enhanced responsiveness leads
social factors (Arda et al. 2023). Sustainable supply chains to reduced waste and improved product quality, as well
are essential in addressing these global challenges by as lower operational costs, which in turn enhances cor-
ensuring responsible sourcing, reducing waste, and pro- porate competitiveness (Ekren et al. 2021). Moreover,
moting ethical business practices. As a result, many optimising delivery time can significantly reduce car-
companies are now revising their sustainability strate- bon emissions, improve product safety, and enhance cus-
gies to mitigate their negative environmental and soci- tomer satisfaction (Granillo-Macías 2021). Importantly,
etal impacts, thereby contributing to a more sustainable high responsiveness increases the feasibility of sustain-
future (Ali et al. 2023). able supply chain network designs. In contrast, a network

CONTACT Yuhan Guo [Link]@[Link] School of Artificial Intelligence and Information Engineering, Zhejiang University of Science and
Technology, Hangzhou 310023, People’s Republic of China
This article has been corrected with minor changes. These changes do not impact the academic content of the article.

© 2025 Informa UK Limited, trading as Taylor & Francis Group

2 Y. GUO ET AL.

design with low responsiveness may be impractical and issues (Chattopadhyay, Sengupta, and Sahay 2016). How-
difficult to implement in real-world scenarios. ever, traditional clustering algorithms lack a learning
Supply chain responsiveness is conventionally mea- mechanism, which limits their ability to maintain solu-
sured through production time and transportation time, tion quality and stability, especially when solving large-
two operational metrics that capture temporal efficiency scale instances (Ran et al. 2024). These limitations arise
across manufacturing and distribution processes (Moh’d because traditional algorithms do not adapt to changes
Anwer 2022). This temporal dimension of responsive- over time or adjust based on new data inputs, leading to
ness has been widely recognised as a critical precur- decreased effectiveness in dynamic environments.
sor to organisational resilience (Bronzo et al. 2024). Graph autoencoder (GAE) offers a powerful solution
While comprehensive resilience frameworks incorpo- to the challenges faced in sustainable supply chain net-
rate multi-phase capabilities from disruption antici- work design through the use of graph neural networks.
pation to system recovery (Emami, Lorenzoni, and Unlike traditional clustering algorithms, GAE provides
Turchetti 2024), optimising temporal responsiveness several key advantages when addressing large-scale sus-
establishes essential infrastructure for achieving these tainable supply chain challenges. First, GAE captures the
capabilities. Empirical evidence further confirms that intricate interconnections among nodes in the supply
accelerated production-transportation cycles enhance chain network, enabling a more accurate representation
supply chain agility and structural robustness under of the system’s topology Dong et al. (2023). Second, GAE
uncertainty (Ha and Lee 2024). The established correla- can handle both structured and unstructured data, mak-
tion between responsiveness and resilience thus justifies ing them versatile tools for dealing with the diverse types
using these time-based metrics as practical operational of information encountered in supply chain manage-
proxies in resilience optimisation. ment. Third, GAE is capable of learning latent represen-
Traditional supply chain design is a well-known tations that preserve important features while reducing
NP-hard problem. Incorporating sustainability criteria dimensionality, thus enhancing the interpretability and
requires the inclusion of numerous additional variables, efficiency of the clustering process (Ren et al. 2024). Most
constraints, and objective functions, which significantly importantly, GAE leverage the high-level information
increase the model’s complexity. Madani, Saihi, and embedded in the data of sustainable supply chains and
Abdelfatah (2024) reviewed 102 papers on sustainable fully exploit the non-Euclidean structure, which cannot
supply chain network design and found that most stud- be achieved with traditional clustering methods such as
ies rely on conventional solvers to address the generated k-means. Moreover, some studies have shown that neural
models. However, the size and scope of variables and con- networks exhibit significant potential in solving mixed-
straints in real-world sustainable supply chain models integer programming problems (Nair et al. 2020) and
make it infeasible to obtain solutions within a reasonable that graph clustering algorithms outperform other meth-
timeframe. While heuristic and metaheuristic methods ods in terms of stability when applied to general data
can significantly reduce computational complexity, they objects (Li, Zhang, and Zhang 2021), indicating their
often result in solutions of lower quality and stability. suitability for optimisation tasks in complex supply chain
Therefore, there is an urgent need for efficient and robust scenarios. Despite these advantages, however, there has
solution methods to address large-scale, complex sustain- been limited substantial research applying GAE specifi-
able supply chain problems, which is the focus of this cally to the problem of sustainable supply chain network
study. design. By integrating GAE into sustainable supply chain
Given that the nodes within a sustainable supply chain design, we address a critical gap in the literature, pro-
network are interconnected, clustering algorithms can viding a novel perspective that enhances both academic
leverage these relationships to reduce the problem’s size research and practical applications.
by breaking down large supply chain challenges into Based on the aforementioned motivations, we propose
smaller, more manageable subproblems. As a result, they a novel clustering mapping algorithm based on graph
have been adopted to address complex issues across var- autoencoder in this study. We integrate GAE with cluster-
ious levels of sustainable supply chain problems. For ing algorithms to capture key features and relationships
instance, researchers have used clustering algorithms within supply chain networks, thereby constructing a
to evaluate the sustainable performance of suppliers more flexible and efficient supply chain network model
and regions (Jabbarzadeh, Fahimnia, and Sabouhi 2018; capable of addressing complex sustainability issues. The
Sabouhi, Jabalameli, and Jabbarzadeh 2021), identify proposed algorithm tackles the challenges of high com-
spatial locations for national carbon-to-fuel infrastruc- plexity and low solution efficiency in large-scale models.
ture (Kim, Do, and Kim 2024), form production route Experimental results demonstrate that the algorithm not
sets (Geiger 2024), and tackle other complex supply chain only provides high-quality solutions close to optimal in
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 3

large-scale instances but also exhibits fast runtime and Wang et al. 2023). The application domains of these mod-
robust stability. Notably, this study represents the first els show a shift in focus over time, reflecting emerging
attempt to model and solve sustainable supply chain net- global challenges and trends, with particular attention
work design problems using graph clustering algorithms. to topics such as mining, food security and pandemics.
The key contributions of this study are summarised as However, although these models can help decision-
follows: makers optimise their sustainable supply chains in vari-
ous fields, few models provide a universal framework that
• We model the supply chain network structure based can be applied across different sectors.
on intercity distances, extending the traditional point- Most studies have focussed on economic objectives
to-point problem and making the model more suitable within sustainability. For example, Sabogal-De La Pava
for clustering representation. et al. (2021) used market value maximisation as the mea-
• We incorporate the responsiveness indicator into the sure of economic performance. Some studies have also
sustainable supply chain model, thereby filling the gap integrated environmental sustainability into economic
in relevant research. goals, balancing profitability and pollution reduction. Cai
• We propose a novel method that abstracts potential and Jiang (2023) introduced decision models of pricing
participants’ attributes to generate a feature matrix, and carbon emission reduction for low-carbon supply
uses a generative model to create sparse graphs, and chain under cap-and-trade regulation. Elhedhli, Gzara,
employs GAE to learn and classify the graph informa- and Waltho (2021) developed a profit-maximisation
tion, leading to efficient problem-solving. model that examined how carbon emissions at differ-
• We introduce an accurate mapping mechanism to ent stages of the supply chain affect consumer demand.
translate abstract solutions into real-world decisions, Similarly, Valderrama et al. (2020) proposed a multi-
providing timely decision support for enterprises. level, multi-period, multi-product mining supply chain
network design model aimed at reducing greenhouse
The remainder of this paper is organised as fol- gas emissions while minimising costs. It is evident
lows. In Section 2, we review relevant models and algo- that the models for both economic and environmen-
rithms from the literature. In Section 3, we present the tal dimensions have innovated in response to chang-
mathematical model for sustainable supply chains. The ing global mechanisms, such as the proposal of market
GAE-based clustering method and its three-step solu- value maximisation in economics and the integration
tion process are described in Section 4. In Section 5, of cap-and-trade in environmental regulation. However,
we discuss experimental results and managerial insights; the key challenge remains in developing a measure-
Section 6 provides conclusions and suggestions for future ment framework for the social dimension and achiev-
research. ing a balanced approach across all three sustainability
dimensions.
A few studies have sought to balance all three sus-
2. Literature review
tainability dimensions, offering more comprehensive
This section reviews the mathematical models, solving approaches. For instance, Varsei and Polyakovskiy (2017)
and clustering algorithms in sustainable supply chain introduced a general model for sustainable wine supply
network design. chain network design, incorporating economic, environ-
mental, and social objectives. Their findings showed that
the model helped decision-makers balance competing
2.1. Mathematical models
sustainability goals more effectively. Naeeni and Sab-
Numerous studies have proposed mathematical mod- baghi (2022) proposed a multi-objective mixed-integer
els that incorporate sustainability dimensions across dif- programming model for sustainable supply chain net-
ferent contexts, such as wine (Varsei and Polyakovskiy work design, addressing economic, environmental, and
2017), mining (Valderrama et al. 2020), perishable prod- social objectives. Their analysis of experimental results
ucts (Biuki, Kazemi, and Alinezhad 2020), biofuels suggested a more sustainable and balanced network
(Zarrinpoor and Khani 2021), food (Reusken, Cruijssen, design that outperformed existing configurations in all
and Fleuren 2023; Yadav et al. 2021), natural gas (Baz- three sustainability dimensions. The ability to success-
yar, Zarrinpoor, and Safavian 2021), municipal solid fully balance these three dimensions requires not only
waste (Eghbali, Arkat, and Tavakkoli-Moghaddam 2022), advanced modelling techniques but also an understand-
epidemic (Hosseini and Ivanov 2022; Katsoras and ing of the interactions and trade-offs between them.
Georgiadis 2022), maritime transport (Kong, Liu, and For example, while reducing environmental impact may
Chen 2023), and personal protective equipment (Bill lead to short-term economic costs, it may simultaneously
4 Y. GUO ET AL.

offer long-term social benefits, such as improving public et al. (2021) introduced a fresh food distribution network
health or enhancing community well-being. This balance model, one of the few that considered delivery time,
is essential for creating truly sustainable supply chains using it as a social sustainability metric. Moreover, trans-
that can meet the needs of both present and future gen- portation plays a critical role in supply chain network
erations. design and is one of the major contributors to green-
Regarding uncertainty, most research has focussed house gas emissions (Joshi 2022). While some stud-
on uncertainties in raw material supply, facility capac- ies have explored how selecting appropriate transporta-
ity, and product demand. For example, Das, Shaw, and tion modes can reduce economic and environmental
Irfan (2020) introduced a multi-product, multi-echelon impacts (Brahami et al. 2022; Sherafati et al. 2020), few
supply chain network design model addressing sustain- have considered transportation time as a key factor in
ability goals under uncertain demand and capacity con- their models.
ditions. Lahri, Shaw, and Ishizaka (2021) proposed a Integrating responsiveness into sustainable supply
two-stage, multi-objective possibilistic integer linear pro- chain design can produce multiple benefits, promoting a
gramming model aimed at maximising social sustain- more resilient and sustainable operational model (Seti-
ability while minimising economic and environmental awan, Tarigan, and Siagian 2023). Specifically, our
objectives. This model tackled uncertainties arising from emphasis on responsiveness distinguishes our model
fluctuations in demand, capacity, and emissions. Jiyao from others.
Gao and You (2017) developed a stochastic mixed-integer
linear fractional programming model to handle various
2.2. Solving and clustering algorithms
uncertainties related to raw material supply and product
demand. The Sustainable Supply Chain Network Design (SSCND)
Additionally, many researchers have addressed trans- problem is typically a combinatorial optimisation prob-
portation issues related to sustainable supply chain. For lem, where the objective is to find the optimal solu-
example, Sherafati et al. (2020) identified transportation tion from a finite set of feasible options. For small-
as a major source of greenhouse gas emissions and pro- scale instances, exact algorithms have been proposed
posed sustainable supply chain network designs to miti- to solve these problems, such as Benders Decom-
gate these issues, along with appropriate carbon emission position Shaw et al. (2016) and Lagrangian Relax-
regulatory mechanisms. Vafaei et al. (2020) introduced a ation (Moheb-Alizadeh, Handfield, and Warsing 2021).
mixed-integer programming model for a sustainable dis- For example, Jiyao Gao and You (2017) developed a
tribution network, considering multiple products, multi- parametric decomposition multi-cut L-shaped method,
ple echelons, and various transportation modes, particu- and Guo et al. (2019) proposed a decomposition-based
larly third-party logistics. Brahami et al. (2022) proposed distributed approximation method. While these exact
a multi-objective model aimed at minimising both trans- algorithms are effective for small-scale problems, they
portation costs and environmental impact by consider- face significant challenges in large-scale instances due
ing the environmental implications of different trans- to the exponential growth in computational complexity.
portation modes. Chadha, Ülkü, and Venkatadri (2022) Their reliance on precise formulations and exhaustive
proposed a physical internet supply chain model with searches limits their practicality in real-world sustain-
peddling and shipment consolidation to optimise freight able supply chain networks, which involve a vast num-
delivery. Considering the growing importance of trans- ber of variables and constraints. Therefore, more and
portation in sustainable supply chains, we fully con- more studies have focussed on improving heuristic or
sider the transportation aspect, separately analyse the metaheuristic methods.
key transportation-related indicators and provide valu- Heuristic and metaheuristic methods have been
able managerial insights into optimising transportation widely applied to solve large-scale supply chain net-
in sustainable supply chain networks. work problems. For instance, Zahiri, Zhuang, and
An overview of the model from the literature is Mohammadi (2017) proposed a Pareto-based meta-
presented in Table 1. According to the latest review heuristic algorithm, and Sabogal-De La Pava et al. (2021)
by Madani, Saihi, and Abdelfatah (2024), only a few designed a primal heuristic approach based on optimisa-
studies have incorporated responsiveness into sustain- tion. Habibi, Abbasi, and Chakrabortty (2023) developed
able supply chain models. For example, Moheb-Alizadeh, a population-based simulated annealing metaheuristic
Handfield, and Warsing (2021) was one of the few studies algorithm and improved the Grey Wolf Optimiser. Guo
focussed on balancing efficiency and sustainability, but it et al. (2023) proposed a heuristic algorithm that inte-
only focussed on improving facility resource efficiency in grates a Markov decision-making process and informa-
closed-loop supply chain networks. Furthermore, Yadav tion accumulation mechanism with the exploration of
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 5

Table 1. An overview of the model in the literature.

Sustainability Dimension Responsiveness
Literature Application area Economic Environmental Social Production Transportation
Hosseini and Ivanov (2022) COVID-19 pandemic C, P
Katsoras and Georgiadis (2022) COVID-19 pandemic C, P
Reusken, Cruijssen, and Fleuren (2023) food I
Brahami et al. (2022) – C CE
Cai and Jiang (2023) – C CE
Elhedhli, Gzara, and Waltho (2021) – P CE
Jiyao Gao and You (2017) biofuel C CE
Kong, Liu, and Chen (2023) maritime C CE
Valderrama et al. (2020) mining I, C GHGE
Chadha, Ülkü, and Venkatadri (2022) automotive C CE •
Bazyar, Zarrinpoor, and Safavian (2021) natural gas C GHGE, EU, W WC, CI
Biuki, Kazemi, and Alinezhad (2020) perishable product C GHGE WC
Das, Shaw, and Irfan (2020) – C CE SC
Eghbali, Arkat, and municipal solid waste C GHGE RD
Tavakkoli-Moghaddam (2022)
Lahri, Shaw, and Ishizaka (2021) – C CE SC
Naeeni and Sabbaghi (2022) glass C GHGE, EU, W RD
Sabogal-De La Pava et al. (2021) toilet paper market value CE RD
Sherafati et al. (2020) – C GHGE RD
Vafaei et al. (2020) electronic commerce C CE WC
Varsei and Polyakovskiy (2017) wine C CE SC
Gang Wang (2023) personal protective equipment C CE WC
Yadav et al. (2021) food C CE delivery time
Zarrinpoor and Khani (2021) biofuel P CE RD, WC
Moheb-Alizadeh, Handfield, and electronic industry C CE WC •
Warsing (2021)
This paper general C CE – • •
Notation: C – cost, P – profit, I – investment, CE – carbon emission, GHGE - greenhouse gas emission, EU – energy use, W – waste, RD – regional development, WC
– work condition, SC – societal commitment, CI – customer issues.

the solution space. Benmamoun et al. (2024) introduced leading to potential premature convergence or subop-
a novel bio-inspired metaheuristic algorithm called the timal solutions. Additionally, GA-based methods can
Wombat Optimiser, specifically designed for supply be computationally intensive for large-scale problems,
chain optimisation problems. Kumar and Kumar (2024) and their performance heavily depends on parameter
proposed a greedy heuristic method tailored for solv- tuning, which can be time-consuming and problem-
ing large-scale decision problems. While these meth- specific. Furthermore, these approaches may lack adapt-
ods provide computational efficiency and flexibility, they ability to dynamic and uncertain environments, which
often suffer from several limitations, such as premature are common in real-world sustainable supply chain
convergence, lack of global optimality guarantees, and scenarios.
sensitivity to parameter settings. Moreover, many studies have integrated multiple algo-
Among various heuristic and metaheuristic appro- rithms to improve performance. Quddus et al. (2017)
aches, genetic algorithms (GA) are the most widely combined the Sample Average Approximation method
applied and refined. Shen (2020) proposed a hybrid with an enhanced progressive hedging hybrid decom-
GA based on variable-length chromosome encoding, position algorithm, employing rolling horizon heuris-
while Brahami et al. (2022) developed an improved tics and variable fixing techniques to enhance the PH
algorithm based on Non-dominated Sorting Genetic algorithm. Rabbani et al. (2018) used a combination of
Algorithm II. Yao Gao et al. (2024) proposed a two- NSGA and multi-objective particle swarm optimisation
stage adaptive GA to address the challenges of scale and to solve multi-objective problems. Biuki, Kazemi, and
complexity in supply chain networks. Sun et al. (2024) Alinezhad (2020) solved supply chain problems through
designed a model combining a minimum spanning tree hybrid metaheuristic algorithms, showing that parallel
with a two-layer adaptive GA, and Zhen Wang, Ye, and hybrid algorithms outperform serial ones. Mohammed,
Guo (2024) tackled multi-objective optimisation prob- Al-Shaibani, and Duffuaa (2023) proposed a hybrid
lems using an enhanced multi-objective GA. The exten- metaheuristic method combining simulated annealing,
sive focus on GA-based methods demonstrates their tabu search, and variable neighbourhood search. Gang
adaptability and effectiveness. However, such improve- Wang (2023) developed a hybrid algorithm combining
ments based on GA still have certain limitations. For branch-and-price with a primal heuristic approach. Mar-
instance, these enhancements often struggle with balanc- tins and Pato (2024) introduced three decomposition
ing exploration and exploitation in the solution space, and heuristic methods to successfully address large-scale
6 Y. GUO ET AL.

real-world problem instances. Although these hybrid Sengupta, and Sahay (2016) proposed an algorithm based
algorithms have demonstrated notable improvements in on hierarchical self-organising maps that used visual
solving efficiency and solution quality, they primarily clustering diagrams to decompose large-scale problems
focus on integrating and enhancing existing methods into independent sub-problems. An overview of the algo-
rather than achieving significant breakthroughs in algo- rithms from the literature is presented in Table 2.
rithmic innovation. Although these clustering algorithms generate promi-
Research has shown that the complexity of variables sing results in various applications of sustainable sup-
and constraints in sustainable supply chain network ply chain management, they exhibit instability when
models makes finding feasible solutions within a reason- solving large-scale instances due to their lack of learn-
able timeframe a significant challenge (Madani, Saihi, ing capabilities (Ran et al. 2024). Graph autoencoder
and Abdelfatah 2024). As mentioned before, clustering (GAE) can learn data features through the training of
algorithms can leverage the relationships between nodes graph neural networks and demonstrate good stability
in a supply chain network to classify similar connec- on general data, which motivate us to adopt them in this
tions, reducing the problem size and decomposing large- study.
scale problems into smaller, solvable sub-problems. Many
studies have investigated the application of clustering
algorithms in various areas of sustainable supply chain
problems, such as assessing sustainability performance, 3. Problem description and model
facility location, and production route construction. For establishment
instance, Jabbarzadeh, Fahimnia, and Sabouhi (2018) 3.1. Problem description
developed a stochastic bi-objective optimisation model
that used the fuzzy c-means clustering method to quan- This paper examines a four-tier sustainable supply chain
tify and assess the sustainability performance of suppli- network comprising suppliers, producers, distributors,
ers. Sabouhi, Jabalameli, and Jabbarzadeh (2021) pro- and retailers. Raw materials procured from suppliers are
posed a multi-period multi-objective optimisation model transported to the respective producers. After being pro-
that applied k-means clustering to evaluate regional sus- cessed by producers, the products are then transported
tainability performance and inform procurement and to the corresponding distributors for temporary storage.
network design decisions. Kim, Do, and Kim (2024) Subsequently, distributors fulfil orders by transporting
applied the Expectation-Maximisation algorithm based goods to retailers. Each transportation selects suitable
on Gaussian Mixture Models to cluster carbon emission transporters to complete various stages. In this paper, the
sources and determine supply regions. For production model is refined based on three sustainable indicators:
routing problems, Geiger (2024) utilised several cluster- cost, carbon emissions and lead time. These indicators
ing methods to construct route sets, and Chattopadhyay, are further subdivided into the following specific metrics:

• Cost: The cost of procuring raw materials from suppli-

Table 2. An overview of the methods used in the literature. ers, processing raw materials into products by produc-
Type of method Literatures ers, storing products by distributors, transporting raw
Exact method Jiyao Gao and You (2017), Guo et al. (2019),
materials from suppliers to producers, transporting
Moheb-Alizadeh, Handfield, and products from producers to distributors and trans-
Warsing (2021), Shaw et al. (2016) porting products from distributors to retailers.
Metaheuristic Benmamoun et al. (2024), Brahami et al. (2022), Yao
Gao et al. (2024), Guo et al. (2023), • Carbon emissions: The carbon emissions from pro-
Habibi, Abbasi, and Chakrabortty (2023), Kumar and cessing raw materials into products by producers,
Kumar (2024),
Sabogal-De La Pava et al. (2021), Shen (2020), Sun storing products by distributors and transporting raw
et al. (2024), materials or products by transporters.
Zhen Wang, Ye, and Guo (2024), Zahiri, Zhuang, and
Mohammadi (2017) • Lead time: The time of processing products by pro-
Mixed method Biuki, Kazemi, and Alinezhad (2020), ducers and transporting raw materials or products by
Martins and Pato (2024), Mohammed, Al-Shaibani,
and Duffuaa (2023),
transporters.
Quddus et al. (2017), Rabbani et al. (2018), Gang
Wang (2023)
Clustering Chattopadhyay, Sengupta, and
Sahay (2016), Geiger (2024), Kim, Do, and Kim (2024),
Jabbarzadeh, Fahimnia, and Sabouhi (2018), Sabouhi, 3.2. Model assumptions
Jabalameli, and Jabbarzadeh (2021)
Neural network Nair et al. (2020) The development of our model is based on the following
GAE This paper
assumptions:
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 7

• Due to differences in geographical location, scale, pur- • Capacity: Production capacity is measured based on
chasing capacity, and cooperative relationships among the workload of each producer’s workers. Storage
various suppliers, the procurement costs of suppliers capacity is determined by the number of pallets (or
are different. shelves). Transportation capacity is calculated based
• Some raw materials do not require processing; there- on the number of lorries in the fleet and the number
fore, it is assumed that suppliers do not bear raw of containers each lorry can accommodate. Addition-
material processing costs. Additionally, for certain raw ally, each producer, distributor, and transporter has a
materials, protective measures during transportation carbon emission quota that must not be exceeded.
are unnecessary, leading to the additional assumption • Location: To enhance the model’s compatibility with
that no containers are required when transporting raw the graph autoencoder clustering algorithm, we mod-
materials. ify the distance attribute used in traditional supply
• Producers have different production capacities, types chains and substitute it with the location attribute.
of energy, and processing technologies, resulting in This modification is explained in detail in the subse-
differences in costs, carbon emissions and processing quent section.
times for producing products. • Economic: Production costs encompass all expenses
• Distributors possess different storage capacities and related to product manufacturing, including labour
storage methods, resulting in differences in costs and costs, machinery costs, and energy consumption costs.
carbon emissions of storing products. Storage costs cover all expenses associated with stor-
• Transporters have different transportation capacities, ing products, such as handling costs, packaging costs,
methods, efficiencies, and used fuels. These differ- and energy consumption costs. Transportation costs
ences lead to variations in transportation costs, carbon include all expenses related to the transportation, such
emissions, and times in transporters. as rental costs for transport vehicles, container costs,
and energy consumption costs.
To maintain the tractability and solvability of the pro- • Environmental: For clustering purposes, we focus on
posed model, we adopt several simplifications commonly the carbon emissions from each stage in the model,
used in the sustainable supply chain network design lit- which include emissions from processing, storing, and
erature. Specifically, the model assumes deterministic transporting.
demand, fixed facility capacities, and linear cost struc- • Lead time: We select two components most closely
tures. This controlled modelling design follows estab- associated with lead time–production time and trans-
lished practices in computational logistics research where portation time–to ensure the responsiveness of the
model simplification serves to evaluate novel algorithmic sustainable supply chain network. Production time
architectures (Shoaib, Mustafee, and Ramamohan 2025). is influenced by factors such as facility size, worker
Specifically, maintaining linear program formula- speed, and machinery production efficiency. Trans-
tions and static parameters isolates the graph clus- portation time is affected by factors including work-
tering method’s topological decomposition capability ers’ packaging speed, transportation methods, and
from confounding optimisation effects – a critical step weather conditions.
when benchmarking computational efficiency in large-
scale SSCND problems. While acknowledging that real-
world implementations require stochastic programming
3.4. Objective functions
extensions, the current abstraction enables: polynomial-
time complexity analysis of cluster generation algo- Based on the problem description, assumptions, and
rithms, unambiguous attribution of solution quality symbol definitions provided, the objective functions of
improvements to network partitioning mechanisms, the sustainable supply chain network model established
and establishment of baseline performance metrics for are as follows:
future comparative studies with robust optimisation
variants. • Economic objective function comprises two compo-
nents, functional cost and transportation cost. Func-
tional cost encompasses the expenses of procuring
3.3. Symbol description
raw materials from suppliers, processing raw materi-
This section provides a detailed explanation of relevant als into products by producers and storing products
symbols. The sets, constants, and decision variables asso- by distributors. Transportation cost contains logistical
ciated with the model are summarised in Table 3. expenses between different levels in the supply chain
8 Y. GUO ET AL.

Table 3. Symbol description.

Symbol Description Type
S set of suppliers, each individual supplier i ∈ S set
P set of producers, each individual producer j ∈ P set
D set of distributors, each individual distributors k ∈ D set
R set of retailers, each individual retailer l ∈ R set
T set of transporters, each individual transporter m ∈ T set
DPRl demand of products by the retailer l constant
LRl location of the retailer l constant
CapMSi maximum capacity of providing raw materials from supplier i constant
LSi location of the supplier i constant
CSi cost of purchasing one unit of raw materials from supplier i constant
CapPj maximum capacity of processing products from producer j constant
CQPj carbon quota of the producer j constant
LPj location of the producer j constant
CPj cost of processing one unit of product from producer j constant
CEPj carbon emissions of processing one unit of product from producer j constant
TPj time of processing one unit of product from producer j constant
CapDk maximum capacity of storing products from distributor k constant
CQDk carbon quota of the distributor k constant
LDk location of the distributor k constant
SCDk cost of storing one unit of product from distributor k constant
SCEDk carbon emissions of storing one unit of product from distributor k constant
CapTm maximum capacity of transporting raw materials or products from transporter m constant
CQTm carbon quota of the transporter m constant
CTm cost of transporting one unit of raw material from transporter m constant
CETm carbon emissions of transporting one unit of raw material from transporter m constant
NCTm number of containers of transporting one unit of product from transporter m constant
CCTm cost of transporting one container one unit distance from transporter m constant
CECTm carbon emissions of transporting one container one unit distance from transporter m constant
TTm time of transporting one unit of raw material or product one unit distance from transporter m constant
di,j distance from supplier i to producer j’s location constant
dj,k distance from producer j to distributor k’s location constant
dk,l distance from distributor k to retailer l’s location constant
qsi pj quantity of raw material needed from supplier i to producer j variable
qpj dk quantity of product needed from producer j to distributor k variable
qdk rl quantity of product needed from distributor k to retailer l variable
qsi tm pj quantity of raw material transported from supplier i to producer j by transporter m variable
qpj tm dk quantity of product transported from producer j to distributor k by transporter m variable
qdk tm rl quantity of product transported from distributor k to retailer l by transporter m variable


network. + qdk rl SCDk
k∈K l∈L

⎛  
 + qdk tm rl NCTm CCTm dk,l
Objeco = min ⎝ qsi pj CSi m∈M k∈K l∈L
i∈I j∈J (1)
 
+ qsi tm pj CTm di,j • Environmental objective function encompasses two
m∈M i∈I j∈J
 components, functional carbon emissions and trans-
+ qpj dk CPj portation carbon emissions. Functional carbon emis-
j∈J k∈K sions entail the emissions of processing raw materials
  into products by producers and storing products by
+ qpj tm dk NCTm CCTm dj,k distributors. Transportation carbon emissions denote
m∈M j∈J k∈K
the emissions of conveying raw materials or products
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 9


by transporters. qsi tm pj + qpj tm dk + qdk tm rl
⎛ i∈I j∈J k∈K
 
Objenv = min ⎝ qsi tm pj CETm di,j ≤ CapTm (m ∈ M) (9)
m∈M i∈I j∈J
 Equation (5) represents the constraint on the maxi-
+ qpj dk CEPj mum supply capacity of suppliers; Equation (6) rep-
j∈J k∈K resents the constraint on the maximum production
  capacity of producers; Equation (7) represents the
+ qpj tm dk NCTm CECTm dj,k
constraint on the maximum storage capacity of dis-
m∈M j∈J k∈K
 tributors; Equation (8) represents the demand con-
+ qdk rl SCEDk straint of retailers; Equation (9) represents the con-
k∈K l∈L straint on transportation capacity of transporters.
 • Carbon emission constraints
 
+ qdk tm rl NCTm CECTm dk,l 
qpj dk CEPj ≤ CQPj (j ∈ J) (10)
m∈M k∈K l∈L
k∈K
(2) 
qdk rl SCEDk ≤ CQDk (k ∈ K) (11)
• Responsiveness objective function comprises two ele- k∈K
ments, the production time and transportation time. 
qsi tm pj CETm di,j
⎛
 i∈I j∈J k∈K
Objeff = min ⎝ qpj dk TPj + qpj tm dk NCTm CECTm dj,k
j∈J k∈K
  + qdk tm rl NCTm CECTm dk,l ≤ CQTm (m ∈ M)
+ qsi tm pj TTm di,j (12)
m∈M i∈I j∈J
  Equation (10) represents the carbon emission con-
+ qpj tm dk TTm dj,k straint when producers produce products; Equa-
m∈M j∈J k∈K tion (11) represents the carbon emission constraint

  when distributors store products; Equation (12) rep-
+ qdk tm rl TTm dk,l (3) resents the carbon emission constraint when trans-
m∈M k∈K l∈L porters transport raw materials or products.
• Consistency constraints
Utilising weighted aggregation, the multi-objective  
optimisation model is converted into a single-objective qsi pj = qpj dk (13)
optimisation model, as illustrated in Equation (4). Here, i∈I k∈K
α, β and γ denote the weights assigned to each objective.  
qpj dk = qdk rl (14)
Obj = αObjeco + βObjenv + γ Objeff (4) j∈J l∈L
 
qsi tm pj = qsi pj (i ∈ I) (15)
3.5. Constraints m∈M j∈J j∈J
 
The constraints of the model in this paper are as follows: qpj tm dk = qpj dk (j ∈ J) (16)
m∈M k∈K k∈K
 
• Capacity constraints qdk tm rl = qdk rl (k ∈ K) (17)
 m∈M l∈L l∈L
qsi pj ≤ CapSi (i ∈ I) (5)
j∈J Equations (13) and (14) represent constraints on the
 quantity of products produced by producers; Equa-
qpj dk ≤ CapPj (j ∈ J) (6) tions (15)–(17) represent the constraints on quantity
k∈K of raw materials or products transported by trans-

qdk rl ≤ CapDk (k ∈ K) (7) porters.
l∈L
• Non-negativity constraints

qdk rl = DRl (l ∈ L) (8) qsi pj , qpj dk , qdk rl ≥ 0 (18)
k∈K
10 Y. GUO ET AL.

qsi tm pj , qpj tm dk , qdk tm rl ≥ 0 (19) at layer l, W (l) are trainable weight matrices, and σ is an
di,j , dj,k , dk,l ≥ 0 (20) activation function. The decoder in the GAE reconstructs
the adjacency matrix or predicts other graph properties
Equations (18)–(20) represent integer constraints. from the latent embeddings:

A = σ (ZZT ), (22)
4. Solution method
where  A is the reconstructed adjacency matrix. In the
In traditional supply chain architectures, intricate net-
context of SCND, the problem involves multiple objec-
works of internodal links are prevalent, which poses a
tives such as minimising cost, carbon emissions, and lead
significant complexity. Empirical observation indicates
time. These objectives are formulated as functions of the
that these connections can be systematically categorised
latent embeddings Z:
and merged, thereby significantly improving efficiency.
This paper adopts the AdaGAE clustering method of Li, min L = Lrecon + λLregularisation , (23)
Zhang, and Zhang (2021) and applies it to supply chain
network optimisation. The approach reconstructs con- where Lrecon represents the reconstruction loss that eval-
nections between participants by probabilistically clus- uates the fidelity of the graph structure, Lregularisation
tering similar data points into classes, thereby effectively ensures smoothness and prevents overfitting in the
addressing the scalability problem. embeddings, and λ controls the balance between these
To prepare for clustering, we modify the traditional two terms. Based on these theoretical foundations, we
model attributes. More specifically, the distance attribute propose a modified AdaGAE algorithm.
is replaced with the location attribute to reduce complex- The adapted AdaGAE clustering algorithm (Li, Zhang,
ity and redundancy. This is because distance constants are and Zhang 2021) integrates the GAE into a graph neu-
readily obtained from the locations of upper and lower- ral network for clustering tasks, expanding its application
level roles. Additionally, a category attribute is added beyond graph-structured data to more diverse types. This
to each role to facilitate the recording of clustering and approach adaptively constructs graphs from a genera-
operation results. tive perspective, leveraging non-Euclidean structures to
uncover high-level information for clustering.
The algorithm applies the following general steps:
4.1. AdaGAE clustering algorithm
The Sustainable Supply Chain Network Design (SCND) • Feature extraction: Initially, feature matrices are
problem involves various participants (e.g. suppliers, pro- extracted from node attributes. A sparse graph is then
ducers, distributors, retailers) and their relationships (e.g. formed by employing a generative model under con-
similarity of participants’ attributes). This structure can trolled sparsity levels.
be represented as a graph G = (V, E), where V denotes • GAE encoding-decoding: The graph undergoes enco-
the set of nodes (participants) and E denotes the set of ding and decoding phases within the GAE frame-
edges (relationships). Each node i ∈ V is associated with work, converting the weighted undirected graph into
features such as cost, carbon emissions, and lead time, embedded node representations.
which are captured in a feature matrix X ∈ RN×F . Here, • Graph updating: Following GAE training, the algor-
N represents the number of nodes, and F denotes the ithm adjusts the graph at a higher sparsity level based
number of features per node. The graph structure is fur- on node embeddings.
ther characterised by an adjacency matrix A ∈ RN×N , • Iterative training: GAE training persists iteratively
encoding the connections between nodes. until the clustering results converge.
The traditional Graph Autoencoder (GAE) employs
a Graph Convolutional Network (GCN) to encode node The execution flowchart of the AdaGAE algorithm is
features and graph structure into latent embeddings Z ∈ depicted in Figure 1.
RN×D , where D represents the dimensionality of the Prior to clustering process, we filter out pertinent
latent space. The encoding process is mathematically role attributes from identical locations, including capac-
expressed as: ity, carbon quota, cost, carbon emissions, and lead
 1  time. These attributes are stored in node vector xi . The
D− 2 
1
H (l+1) = σ  A D− 2 H (l) W (l) , (21) resulting attribute dataset constitutes the feature matrix
X = [x1 , x2 , . . . , xn ]T ∈ Rn×d , where n denotes the node
where A = A + I is the adjacency matrix with self-loops, count, and d represents the attribute dimensionality. Usu-
D is the degree matrix of 
 A, H (l) is the node embeddings ally,a smaller distance xi − xj 22 should be assigned a
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 11

Figure 1. The execution process of the AdaGAE algorithm.

larger probability pij ; thus, it provides a natural method of graph convolutional neural network layers; its goal
to determine the probabilities by solving the following is to transform the original features into latent features
problem: using the constructed graph structure. We define the
latent features generated by m layers as:

n
min xi − xj 22 pij + γi p2ij , (24)
Z = ϕm (Âϕm−1 (· · · ϕ1 (ÂXW1 ) · · · )Wm ). (27)
i,j=1

where γi controls the sparsity of each node. It can be • Decoder: Our goal is to recover the connectivity distri-
proven that the solution has precisely k nonzero entries bution based on Euclidean distance. First, the distance
(i.e. it is k-sparse) if of the latent features, Z, is calculated by d̂ij = z i −
⎛ ⎞ z j 22 . Then, the connectivity distribution is recon-
1⎝ x 
k structed by a normalisation step:
γi = kdi,k+1 − dijx ⎠ , (25)
2
j=1 exp(−d̂ij )
qji = n
. (28)
where dijx= xi − xj 22 . We denote dix ∈ Rn×1 as a vec- j=1 exp(−d̂ij )

tor with the jth element as dijx . Furthermore, it results in a

closed-form solution: This process is viewed as inputting −d̂ij to the Soft-
Max layer. Hence, we change the way similarity is mea-
(di,k+1
x − dijx )+ sured from the common inner product used in GAE
pij = , (26)
x
kdi,k+1 − k x to the Euclidean distance. To measure the difference
j=1 dij
between the two distributions, Kullback-Leibler diver-
where (·)+ = max(·, 0). After obtaining the sparse con- gence is used, and its objective function is defined as
nectivity distribution by solving Equation (24), we con-
vert the directed graph to an undirected graph via Wij = 
n
1
min KL(pq) ⇔ min pji log . (29)
(pij + pji )/2 and use the connectivity distribution as the qji
i,j=1
reconstruction target for the GAE. Next, we elaborate on
the specific implementation details of the GAE. This is equivalent to the widely used minimisation of
cross-entropy in clustering tasks. Finally, to avoid the col-
• Encoder: Generally, we simply set 
A = W and  = lapse caused by simple updates, sparsity k is increased
− 21 − 21
D AD . The encoder comprises of multiple layers when updating the weight graph. Generally, we define its
12 Y. GUO ET AL.

step size t and sparsity upper bound km as follows: where A ∈ Rm×n , and n ≥ m. Generally, we have
 
x
km − k0 x= B , (34)
t= , (30) xN
T
n
km = , (31) where B ∈ Rm×m , N ∈ Rm×(n−m) , xB ∈ Rm represents
c basic variables and xN ∈ R(n−m) represents non-basic
variables. Rewrite Equations (32) and (33) in the follow-
where T is the number of iterations to update the weight ing form
adjacency, n represents the original data size and c is the
number of clusters. min cBT xB + cN
T
xN , (35)
x∈Rn
The clustering outcomes derived from the GAE are
stored in the category attribute. We extract role attributes s.t. BxB + NxN = b, xB ≥ 0, xN ≥ 0. (36)
belonging to the same category at the same location, com-
Hence, from the Equation (36), we obtain
pute their averages, and generate new data for subsequent
solving tasks. xB = B−1 b − B−1 NxN . (37)

We select a component xq from the non-basic variables

4.2. The simplex method to enter the basic variables
Simplex serves as a general approach for solving linear
programming problems. It begins with a basic feasible xB = B−1 b − xq B−1 Aq , (38)
solution, scrutinises its optimality and iteratively pro- where Aq is the column of A corresponding to xq .
gresses towards an improved solution until the optimum Equation (38) reflects the relationship between basic and
is achieved. Otherwise, the problem is deemed infeasible non-basic variables. Let the current solution be x, and the
or unbounded. next solution we want to obtain is denoted as x(λ)
The algorithm unfolds as follows:
x(λ) := x + λdq , (39)
• Canonical formulation: This expresses the linear pro-  −1 
−B Aq
gramming problem constraints in canonical form and dq = , (40)
identifies an initial basic feasible solution. eq
• Feasibility check: If no basic feasible solution exists, where dq is the search direction, λ ≥ 0 is the step size,
which indicates contradictory constraints, then the dq ∈ Rn , eq is a vector with a value of 1 at the posi-
problem lacks a solution. tion corresponding to xq and 0 elsewhere. Substituting
• Iteration for improvement: Starting from the initial, Equation (39) into the objective function yields
basic feasible solution, this step introduces non-basic
variables to enhance the objective function, aiming for cT x(λ) = cT (x + λdq )
a better iterative solution.
• Optimality condition: This process iterates until the = cT x + λ(cq − CBT B−1 Aq )
optimality conditions are met, signifying that the = cT x + λrq . (41)
objective function value can not be further improved,
thereby reaching the optimal solution. If rq < 0, then this is a good movement direction. Here,
• Unboundedness check: If the objective function value
appears unbounded during iteration, the process is rq = cq − CBT B−1 Aq , (42)
halted.
where rq represents the amount by which adding the non-
basic variable q can decrease the objective function. We
After clustering and averaging, the newly acquired
refer to this as reduced cost. Then, we use the minimum
data serves as input for the Simplex. The ensuing calcu-
index rule to determine the entering basic variable
lation commences by addressing the following standard
form linear programming problem: c T dj
min . (43)
j∈N dj 
min cTx , (32)
x∈Rn In the above objective function, cT dj represents minimis-
s.t. Ax = b, x ≥ 0, (33) ing the objective function for the next step search, and
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 13

Algorithm 1 Revised Simplex Method. in six bins, labelled sequentially as Qsi pj , Qpj dk , Qdk rl ,
Require: B, N, cB , cN , xB , xN , b. Qsi tm pj , Qpj tm dk , and Qdk tm rl . Although all sets share a
Find a basic feasible solution(bfs) x with A = [B | N]. similar structure, we focus on describing Qsi pj for illus-
while True do trative purposes.
Compute rq by Equation (42). To facilitate the restoration process, adjustments are
if All rq ≥ 0 then made to the id attribute of the abstracted objects. Rather
return The current bfs x is optimal solution. than sequentially numbering them in the conventional
break; manner, their id values comprise two parts represented
else by a two-digit number. The first digit, idl , denotes the
Choose rq by Equation (43). location of the abstracted object, and the second digit,
end if idk , signifies the category to which the abstracted object
Compute dq by Equation (40). belongs. Before commencing the restoration process, the
if All dq ≥ 0 then number of abstracted objects, m, is calculated as follows:
return The LP problem is unbounded. n
break; m= , (45)
r
else
Compute λ by Equation (44), where n denotes the original data size and r is the scaling
Update x ← x + λdq , B and N for the next itera- factor. The size of Qsi pj is m2 . To traverse the set of suppli-
tion. ers, the range of the traversal number i is set to [0, m − 1].
end if Consequently, we have:
end while  
i
Ensure: Optimal solution. idl = , (46)
c
idk =i mod c, (47)
dividing by dj  represents normalisation. We determine
the step length using the minimum ratio rule id =10idl + idk , (48)
  where c represents the number of clusters. While travers-
xi q
λ = min q | − di < 0 , (44) ing the Qsi pj set, converting from the traversal number, i,
i∈B −di to the id of the abstract object is completed. Based on the
structure of the id attribute, we then have
where xi is the ith component of the current search point,
q  
and di is the ith component of dq . id
After updating of x ← x + λdq , we adjust the basis l= , (49)
10
matrix and iterate through the calculation process until
all rq ≥ 0. Once achieved, we output the optimal solu- k =id mod 10, (50)
tion. The detailed algorithmic steps are outlined in
where l signifies the location of the real solution, and k
Algorithm 1.
denotes the category to which the real solution belongs.
At this juncture, we attain the optimal solution for the
An example of the restoration process is depicted in
abstracted entities after clustering, which might not cor-
Figure 2.
respond to the optimal solution for the original problem.
At this stage, we map the abstract solution to the spe-
Therefore, we proceed by filtering the optimal solution
cific solution. Subsequently, by utilising the location and
derived by the solver, which lays the groundwork for the
category of the real solution as filtering criteria, we iden-
subsequent restoration of the true solution.
tify the set of real solutions corresponding to the abstract
solution. From this set, we select the solution with the
4.3. Mapping from abstract solutions to real-world best weighted attributes as the true solution.
solutions
Mapping abstract solutions back to real-world solutions 5. Experiments and insights
involves transforming abstract representations into their
5.1. Experimental environment, test cases and
corresponding real-world counterparts. Initially, we fil-
parameters
ter nonzero decision variable values from the clustering
phase based on their type. As this process involves six We implement our algorithm using Python 3.9 on a com-
quantity decision variables, the filtered results are stored puter equipped with an 11th Gen Intel(R) Core(TM)
14 Y. GUO ET AL.

Figure 2. An example of the restoration, where n = 90, r = 3, m = 30, c = 3.

Table 4. The vale of experimental parameters. Table 5. The constant unit and rage of test instances.
Symbol Description Value Symbol Unit Range
α the weight of economic side 0.39 DPRl ten thousand [1000, 3500]
β the weight of environment side 0.2806 LRl , LSi , LPj , LDk – [1, 10]
γ the weight of lead time side 0.3294 CapMSi ten thousand [8000, 10,000]
k0 initial sparsity 1 CSi dollar [3000, 7000]
ti number of iterations to update GAE 50 CapPj ten thousand [4000, 8000]
n CQPj ton [50,000, 60,000]
km upper bound of sparsity
c CPj dollar [5000, 10,000]
T number of iterations to update weight adajacency 10 CEPj ton [0.15, 0.35]
r scaling factor 3 TPj hour [1, 24]
CapDk ten thousand [2000, 4000]
CQDk ton [10,000, 20,000]
SCDk dollar [300, 500]
SCEDk ton [0.01, 0.05]
i7-11800H processor running at 2.30 GHz, with 16 GB CapTm ten thousand [10,000, 12,000]
CQTm ton [500, 1000]
RAM and a Windows 11 operating system. The algorithm CTm dollar [0.1, 0.5]
parameters are summarised in Table 4. CETm ton [0.01, 0.02]
NCTm unit [1, 10]
To illustrate the effectiveness of the proposed algor- CCTm dollar [5, 10]
ithm, three large-scale instances are generated based on CECTm ton [0.01, 0.05]
a food processing industry scenario. These instances are TTm hour [0.01, 0.1]
configured with supplier/producer/distributor/retailer/
transporter quantities set at 90/90/90/90/90 (90), 120/120
Figure 3 visualises the producers after AdaGAE clus-
/120/120/120 (120), and 150/150/150/150/150 (150),
tering. The clustering results of each producer are anal-
respectively. The instance data are obtained from the
ysed based on three key dimensions: cost, carbon emis-
collaborative research platform developed by the Pub-
sions, and production time. A three-dimensional surface
lic Environmental Research Center and the Natu-
plot is used to visualise the clustering results without clus-
ral Resources Conservation Association for sustainable
ter labels. To facilitate a clearer observation of the cluster-
green supply chains. Further details are provided in
ing results, a three-dimensional bar chart is employed to
Table 5.
visualise the clustering results with cluster labels.
Table 8 provides the chosen virtual producers by sim-
plex method to potential real producers through map-
5.2. Case study
ping algorithm. Table 9 displays the real-world solution
In this section, we present an illustrative example to results, which includes all facility locations and material
explain the studied instances. Specifically, we consider an flows.
instance with a size of 90, where S1–S90 represent sup-
pliers, P1–P90 denote producers, D1–D90 correspond to
5.3. Experimental results and analysis
distributors, and R1–R90 are retailers. Due to the large
scale of the instance, we focus on providing detailed 5.3.1. Solution quality analysis
information about the producers, while the details of We execute the k-means (Sabouhi, Jabalameli, and
other participants are omitted for brevity. Jabbarzadeh 2021), k-medians, k-medoids, Fuzzy C-
Table 6 presents the attribute information of producers means (Jabbarzadeh, Fahimnia, and Sabouhi 2018),
along with their cluster labels after AdaGAE clustering. Gaussian mixture (Kim, Do, and Kim 2024), and Ada-
Table 7 provides the attribute information and cluster GAE algorithms five times each and calculate the mean of
labels of the virtual producers generated by the AdaGAE the outputs, with the final results presented in Table 10.
clustering algorithm. The comparisons of solution quality and differences for
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 15

Table 6. The attribute information with cluster labels of producers.

Producer Location Capacity Cost Carbon quota Carbon emission Production time Cluster label
P1 1 5506 9409.62 52,152 0.29 15 2
P2 1 6249 7159.83 51,880 0.29 18 0
P3 1 7004 7956.78 54,951 0.31 9 1
P4 1 4309 8586.77 52,448 0.21 21 2
P5 1 6274 7561.52 58,288 0.24 7 1
P6 1 5539 6795.53 55,966 0.29 16 1
P7 1 4700 5719.27 50,776 0.21 5 0
P8 1 4815 6013.08 53,035 0.22 18 0
P9 1 5984 8006.94 57,426 0.30 3 1
··· ··· ··· ··· ··· ··· ··· ···
P82 10 7474 9253.48 53,668 0.27 9 2
P83 10 5691 8144.14 50,431 0.22 14 1
P84 10 7669 9167.23 52,787 0.33 10 2
P85 10 4914 5735.96 50,666 0.16 13 1
P86 10 4274 7859.96 54,823 0.16 15 1
P87 10 5204 5215.72 52,531 0.31 15 1
P88 10 7592 5548.07 52,345 0.26 18 1
P89 10 4599 6780.56 59,536 0.23 22 0
P90 10 6961 5110.15 58,300 0.24 6 0

Table 7. The AdaGAE clustering results of producers.

Virtual producer Location Capacity Cost Carbon quota Carbon emission Production time Cluster label
P1 1 5254.67 6297.39 51, 897 0.24 13.66 0
P2 1 6200.25 7580.19 56, 657 0.28 8.75 1
P3 1 4907.5 8998.2 52, 300 0.25 18 2
P4 2 6108.25 6114.27 57, 036.5 0.24 16.75 0
P5 2 6179.33 6928.55 53, 846 0.31 8.33 1
P6 2 7599.5 9147.39 55, 395 0.20 16 2
··· ··· ··· ··· ··· ··· ··· ···
P25 9 7192.5 6220.4 59, 402 0.25 8 0
P26 9 5311.8 7990.04 55, 605.6 0.24 15.4 1
P27 9 7427 8967.61 55, 349 0.20 11 2
P28 10 5780 5945.36 58, 918 0.23 14 0
P29 10 4599 6500.77 52, 159 0.22 15 1
P30 10 7571.5 9210.36 53, 227 0.30 9.5 2

Table 8. The chosen virtual producers to potential real producers instances (90), the maximum deviation using the Ada-
through mapping algorithm. GAE algorithm is 13.02%, with a minimum of 8.33% and
Chosen virtual producer Potential real producer an average of 10.76%. Notably, in some smaller instances,
P1 P2, P7, P8 the K-means algorithm yields superior solution quality.
P4 P10, P12, P14 In medium scale instances (120), the maximum deviation
P5 P13, P15, P17
P7 P21, P22, P25, P26 with AdaGAE is 12.89%, the minimum is 4.95%, and the
P8 P19, P20, P27 average is 8.84%. The experimental results suggest that
P10 P29, P32, P33
P13 P37, P39, P40 the solution quality achieved by the AdaGAE algorithm
P16 P46, P47, P48, P54 outperforms all other methods. For larger scale instances
P19 P56, P60, P61, P62
P20 P55, P57, P59 (150), the maximum deviation using AdaGAE is 12.10%,
P22 P64, P66, P68, P70 the minimum is 8.60%, and the average is 10.42%, indi-
P25 P73, P75, P79
P28 P89, P90
cating that AdaGAE maintains optimal solution quality
P29 P83, P85, P86, P87, P88 in most cases.
Figure 4 demonstrates that, except for a few relatively
small and large scale instances where the solution qual-
ity of the AdaGAE algorithm is inferior to that of K-
each instance across various algorithms are illustrated in means, the solution quality advantage of AdaGAE is sig-
Figures 4 and 5. nificantly more pronounced in medium scale and most
As shown in Table 10, for the same instance scale, other instances compared to other clustering algorithms.
the solution quality of most datasets using the AdaGAE Figure 5 illustrates that for medium and larger scale
algorithm is closer to the optimal solution. The max- instances, the solution gap of the AdaGAE algorithm
imum deviation from the optimal solution is 13.02%, gradually decreases with different data training, a char-
while the minimum is 4.95%. For relatively smaller scale acteristic not exhibited by other clustering algorithms.
16 Y. GUO ET AL.

Figure 3. The visualised AdaGAE clustering results of producers.

5.3.2. Multi-objective analysis approximation compared to the other clustering

The average values of the objective functions–including methods.
costs (excluding transportation costs), transportation
costs, carbon emissions (excluding emissions related to
transportation), transportation carbon emissions, pro- 5.3.3. Time analysis
duction time, and transportation time–are summarised With all other parameters kept constant, we run
in Table 11. The magnitudes presented in the table the Distributed Approximation Approach (DAA) (Guo
clearly indicate that transportation is indeed the most et al. 2019), the Multi-Neighbourhood Descent Traversal
critical aspect to consider when designing a supply Algorithm (MNDTA) (Guo et al. 2021), Reinforcement
chain network, as it significantly contributes to various Learning (RL) (Guo et al. 2023), Graph Neural Networks
greenhouse gas emissions (Joshi 2022). Consequently, (GNN) (Chen et al. 2024), CPLEX, and the AdaGAE clus-
we conduct a separate analysis of the performance tering algorithm five times for each instance. The average
of each objective function related to transportation solution times are reported in Table 12, with the trend
under different clustering algorithms, as illustrated in illustrated in Figure 7.
Figure 6. As shown in Table 12, AdaGAE consistently achieves
From Table 11, it is evident that, for instances of the the shortest solution time across all solving algorithms
same size, the costs calculated using the AdaGAE cluster- and instance sizes. For smaller instances (90), AdaGAE
ing algorithm are the closest to the optimal solution, with demonstrates a remarkable improvement, solving prob-
a maximum deviation of 1.18%, a minimum deviation lems 20.97 times faster than CPLEX, 7.40 times faster
of 0.03%, and an average deviation of 0.53%. Regarding than DAA, 7.20 times faster than MNDTA, 5.85 times
carbon emissions, the AdaGAE algorithm achieves the faster than RL, and 13.97 times faster than GNN. Simi-
lowest emissions, with a maximum deviation of 0.86%, larly, for medium-sized instances (120), AdaGAE’s solu-
a minimum deviation of 0.04%, and an average devia- tion time is 37.62 times faster than CPLEX, 7.05 times
tion of 0.51%. In terms of production time, the AdaGAE faster than DAA, 4.63 times faster than MNDTA, 7.87
algorithm yields shorter production times compared to times faster than RL, and 25.08 times faster than GNN.
all other clustering algorithms, with a maximum devi- For larger instances (150), AdaGAE maintains its domi-
ation of 2.57%, a minimum deviation of 0.01%, and an nance, solving problems 61.84 times faster than CPLEX,
average deviation of 1.44%. 6.39 times faster than DAA, 3.19 times faster than
Figure 6 illustrates that while the AdaGAE algorithm MNDTA, 10.35 times faster than RL, and 41.22 times
provides optimal results for transportation costs and faster than GNN.
transportation carbon emissions, all clustering algo- Figure 7 provides a clear visual comparison of the
rithms display suboptimal performance in this area, solution times across different algorithms and instance
exhibiting noticeable deviations from the optimal solu- sizes. It highlights AdaGAE’s superior performance, with
tion. However, for transportation time, it is clear its advantage over other algorithms becoming increas-
that the AdaGAE clustering algorithm offers a closer ingly pronounced as the problem size grows. This trend
Table 9. The solution results of the test instance with size 90.
Supplier Producer Material Producer Distributor Product Distributor Retailer Product
S7 P8, P27, P87 1322, 3627, 4995 P7 D6, D33, D71 938, 1823, 1939 D1 R7, R21, R23, R34, R49 277, 706, 780, 1974
S8 P10, P46, P56 5575, 1853, 574 P8 D9, D12, D71 2923, 305, 1587 D2 R17, R77 1509, 1542
S16 P8, P73 2493, 6865 P10 D12, D41, D72 1443, 2115, 2047 D3 R6, R53, R58 1443, 1530, 445
S17 P15, P37, P66 675, 3338, 5861 P11 D10, D22 1556, 724 D4 R53, R55 657, 1506
S19 P12, P90 1823, 6961 P12 D1, D54, D85 214, 3575, 2560 D6 R37, R86 1231, 2098
S23 P12, P54 4125, 5493 P15 D31, D32, D43 3169, 2081, 1838 D9 R5, R32, R85 414, 3103, 407
S24 P20, P25, P64 5809, 3176, 245 P19 D20, D60 2501, 1519 D10 R29, R36, R48 1699, 250, 1057
S25 P12, P40, P48 401, 6858, 2410 P20 D11, D38, D56 994, 2170, 2645 D11 R11, R87 1855, 1485
S26 P15, P39 6413, 1994 P21 D1, D3, D11, D30 889, 843, 2346, 2762 D12 R1, R9 1462, 286
S27 P10 30 P25 D9, D10, D13, D27, D58, D73 1001, 1450, 2207, 784, 211, 1760 D13 R85 2207
S35 P21, P46 4858, 4444 P27 D3, D55, D63 1631, 3788, 1896 D14 R13 2621
S37 P29, P37 6981, 1098 P29 D14, D25, D74 2621, 3902, 840 D16 R49, R60 2573, 806
S47 P21, P59 1982, 7145 P32 D6, D24 1422, 3455 D19 R7, R31, R43, R47 30, 986, 635, 811
S48 P33, P55 5042, 3773 P33 D23, D63 3860, 1182 D20 R4, R9, R58, R63 1689, 1066, 2099, 191
S50 P48, P88 4514, 3761 P37 D27, D39 500, 3936 D21 R23, R80 803, 1932
S62 P19, P25 4020, 4237 P39 D27, D48 2459, 2789 D22 R25, R38 1060, 1158
S64 P11, P32, P55 1606, 3912, 3802 P40 D19, D52, D83, D86 1063, 3900, 1426, 469 D23 R75, R88 2810, 1050
S66 P79, P87, P88 5244, 214, 3831 P46 D4, D19, D21 2163, 1399, 2735 ··· ··· ···
S70 P60, P64 4367, 4388 P48 D6, D35, D42 969, 2664, 3291 D68 R30, R59, R76 1984, 1252, 539
S71 P32, P57 965, 7971 P54 D3, D43, D76, D86 879, 761, 2297, 1566 D69 R26, R28, R46 2129, 1104, 638
S74 P70, P85 3957, 4914 P55 D65, D72, D80 2735, 1605, 3235 D70 R19, R63 1107, 2772

INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH

S75 P7, P11, P29, P56, P60 4700, 674, 382, 3669, 18 P56 D20, D75 744, 3499 D71 R18, R89 1451, 2075
S77 P27, P39, P70 3688, 3254, 1986 P57 D59, D73, D77 3237, 1416, 3318 D72 R20, R56 1727, 1925
S78 P8, P75 1000, 7520 P59 D16, D62, D83 3379, 3054, 712 D73 R36, R78 1834, 1342
P60 D45, D60 2173, 2212 D74 R21, R52 1122, 2189
P64 D30, D69 762, 3871 D75 R36, R88 1314, 2419
P66 D33, D36, D66 1743, 3656, 462 D76 R31 2287
P70 D2, D90 3051, 2892 D77 R6, R42, R50, R72 206, 389, 1371, 1352
P73 D3, D53, D66, D90 65, 2913, 2984, 903 D79 R8, R15 2988, 312
P75 D43, D61, D70 937, 2704, 3879 D80 R42, R68 980, 2255
P79 D79, D81 3300, 1944 D81 R23, R33, R64, R73 1394, 165, 1187, 259
P85 D48, D58, D75 2304, 2376, 234 D83 R67 2138
P87 D1, D67 2858, 2346 D85 R45, R84 1167, 1393
P88 D28, D33, D68, D81 2628, 128, 3775, 1061 D86 R12, R78 31, 2004
P90 D22, D57, D74 1494, 2996, 2471 D90 R16, R19, R40 1400, 600, 1795

17
18 Y. GUO ET AL.

Figure 4. The optimal solution is compared with the results obtained by K-means, K-medians, K-medoids, Fuzzy C-means, Gaussian
mixture, and AdaGAE algorithms when the number of examples is 90, 120, 150.

Figure 5. The solution gap is compared with the results obtained by K-means, K-medians, K-medoids, Fuzzy C-means, Gaussian mixture,
and AdaGAE algorithms when the number of examples is 90, 120, 150.
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 19

demonstrates AdaGAE’s robustness and scalability in

−12.89%
−8.68%
−8.33%
−13.02%

−12.10%
−4.95%

−8.60%
−10.94%

−10.57%
Gap
handling large-scale sustainable supply chain network
problems effectively.

AdaGAE

5.3253
5.1503
3.7072
4.2332

4.9328

6.4102

7.1865
4.4220

6.6004
5.3.4. Sensitivity analysis
Sensitivity analysis is a crucial step in evaluating the

−10.21%
−19.71%
−8.24%
−24.40%
−14.96%

−12.88%
−9.01%
−9.44%
−9.77%
robustness of clustering algorithms, as it provides insights
Gap

into the consistency and reliability of the solutions across

multiple executions. Under identical conditions, each
clustering algorithm is executed five times, and the vari-
Gaussian Mixture

ation in solutions across the number of runs is presented

4.3145

5.4881
5.5924

6.5069
7.2425
3.7564
4.4833

5.4357

6.4553
in Figure 8.
As illustrated in Figure 8, the Gaussian Mixture Clus-
tering algorithm stands out as the most stable one among
the evaluated methods. This stability is evident in the
−14.77%
−5.25%
−16.83%
−12.73%

−8.48%
−9.13%
−8.65%
−17.02%

−13.80%

minimal variance of its solutions across multiple runs.

Gap

However, it is important to note that despite its stabil-

ity, the solution values generated by this algorithm are
consistently higher for all instance sizes, indicating sub-
Fuzzy C-means

optimal performance in terms of solution quality. This

5.8234
4.1950
5.1045
5.3426

7.2216
3.7178
4.3827

6.5075
6.4756

observation aligns with earlier findings from the solution

quality analysis, reinforcing that stability alone does not
guarantee effectiveness in achieving optimal results.
The K-means and AdaGAE clustering algorithms
−9.49%
−6.62%
−17.85%
−11.82%
−12.61%
−13.36%

−13.29%
−10.30%
−9.64%
Gap

exhibit the next highest levels of stability. However, a

direct comparison reveals that AdaGAE outperforms
K-means in both solution quality and convergence
K-medoids

4.2497
5.1490
5.2992
5.5556

6.5842
7.2556
3.8537
4.2457

6.4785

behaviour. As the number of runs increases, AdaGAE

consistently produces lower and more refined solution
values, demonstrating a strong convergence trend. This
characteristic highlights AdaGAE’s ability to adapt and
−8.76%

−12.71%
−12.44%

−9.91%
−9.66%
−9.63%
−16.05%

−17.74%

−13.31%
Gap

improve its solutions over multiple iterations, making it

a more robust and reliable choice for sustainable supply
chain network design.
K-medians

In contrast, K-medians, K-medoids, and Fuzzy C-

4.3351

5.3440
5.7053

6.5610
7.2570
3.7518
4.3465

5.1445

6.4797

means clustering algorithms are the least stable. Their

solution values show significant variability across runs,
reflecting their inability to produce consistent outcomes.
−5.16%

−8.00%
−9.21%
−8.33%

−9.85%
−10.10%
−17.48%

−15.50%

−14.27%

While these algorithms occasionally achieve optimal

Gap

results, their solutions can deviate drastically, sometimes

Table 10. Analysis of the solution quality.

exceeding the optimum by several magnitudes. More-

K-means

over, the instability of these algorithms persists even as

4.1917

6.4468
5.0462
5.1758
5.4967

7.2695
3.7679
4.4000

6.5348

the number of runs increases, suggesting a lack of con-

vergence and reliability. This unpredictable performance
poses a significant challenge for practical applications,
Optimal solution

where consistency is critical for decision-making.

4.73916
5.07416

5.96930
6.61764
3.42221
3.74531
3.98587
4.36919

5.71852

Overall, the sensitivity analysis highlights the trade-

offs between stability and solution quality among the
evaluated clustering algorithms. While Gaussian Mixture
Clustering demonstrates excellent stability, it falls short
Instance

in achieving competitive solution values. On the other

150-1
120-1
120-2
120-3

150-2
150-3
90-1
90-2
90-3

hand, AdaGAE strikes a balance between stability and

20 Y. GUO ET AL.

Figure 6. The solution results of transportation cost, carbon emission and time under different algorithms are compared when the
number of examples is 120.

Table 11. Analysis of multi-objective solution quality.

Transportation Carbon Transportation Carbon Production Transportation
Instance Clustering algorithm Cost(109 ) cost(109 ) emission(104 ) emission(106 ) time(106 ) time(106 )
120-1 Optimal solution 2.50942 1.84441 7.58620 7.74217 3.04991 4.48304
K-means 2.5554 2.4813 7.6214 10.516 3.3711 4.5372
K-medians 2.7162 2.4366 7.7691 10.355 3.2655 4.5332
K-medoids 2.6670 2.7929 7.6014 11.442 3.2562 4.8120
Fuzzy C-means 2.5962 2.5122 7.6627 11.165 3.2865 4.8345
Gaussian Mixture 2.6718 2.7438 8.0817 11.875 3.4977 4.6881
AdaGAE 2.5392 2.3631 7.5896 10.239 3.1298 4.5210
120-2 Optimal soulution 2.64489 2.07798 6.98767 8.72277 3.13164 4.35580
K-means 2.6682 2.4854 7.1841 10.228 3.1497 4.4523
K-medians 2.8776 2.5776 7.4778 10.431 3.4206 4.4241
K-medoids 2.7084 2.4699 7.4547 10.182 3.4632 4.5285
Fuzzy C-means 2.7303 2.6058 7.2420 10.599 3.5955 4.4058
Gaussian Mixture 2.8316 2.5999 7.1751 10.432 3.5727 4.4613
AdaGAE 2.6460 2.4327 6.9940 9.8619 3.1320 4.3589
120-3 Optimal solution 2.75081 2.30604 7.78013 8.89584 3.6124 4.72021
K-medians 2.8488 2.8002 7.9092 10.253 3.7902 4.8552
K-means 2.7750 2.7366 7.9025 10.511 3.7087 4.9392
K-medoids 2.8074 2.7282 8.0457 11.319 3.7529 4.8466
Fuzzy C-means 2.8572 2.9502 8.0163 11.551 3.7790 4.8971
Gaussian Mixture 2.8395 2.7333 8.0427 11.244 3.8780 4.8943
AdaGAE 2.7615 2.5389 7.7888 10.019 3.6755 4.7214

quality, delivering superior results with consistent per- 5.4. Discussion

formance. These findings underscore the importance of
The experimental findings presented in this study
considering both robustness and accuracy when select-
underscore the effectiveness of the proposed AdaGAE
ing clustering algorithms for sustainable supply chain
algorithm across multiple performance dimensions in
network design.
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 21

Practically, AdaGAE also outperforms competing

methods in terms of multi-objective optimisation, espe-
cially regarding transportation costs, emissions, and
time-related objectives. For instance, transportation cost
deviations average only 0.18%; transportation emission
deviations average 0.32%; and production time devia-
tions average 0.84% (Table 11). These results highlight
AdaGAE’s potential for real-world applications, where
achieving a balance between environmental sustainabil-
ity and operational efficiency is critical.
Furthermore, the time analysis confirms the algor-
ithm’s scalability and computational efficiency. Across
all instance sizes, AdaGAE achieves the shortest solution
times, with speedups of 20.97× over CPLEX, 7.40× over
Figure 7. The trend of average solving time among different DAA, 7.20× over MNDTA, 5.85× over RL, and 13.97×
algorithms. over GNN on 90-node instances; and 61.84×, 6.39×,
3.19×, 10.35×, and 41.22× on 150-node instances,
respectively (Table 12, Figure 7). Such efficiency is vital
Table 12. Analysis of solution time. for large-scale supply chain problems that require timely
Instance CPLEX DAA MNDTA RL GNN AdaGAE decision-making.
90-1 4.5549 1.5858 1.5462 1.0539 3.0366 0.2107 The sensitivity analysis (Figure 8) further reveals the
90-2 4.0869 1.5227 1.4809 1.4400 2.7246 0.2009 algorithm’s robustness. Although the Gaussian Mixture
90-3 4.4901 1.5281 1.4805 1.1683 2.9934 0.2147
120-1 16.3968 3.1598 2.0314 3.0605 10.9312 0.4445 model exhibits the smallest variance across five runs, its
120-2 16.5420 3.2567 2.1257 4.1815 11.028 0.4562 mean solution value remains up to 16.52% higher than
120-3 17.9032 3.1024 2.0988 3.3925 11.9354 0.4505
150-1 45.0293 4.9290 2.5692 6.7021 30.0195 0.7448 AdaGAE’s. Conversely, AdaGAE’s run-to-run variance is
150-2 49.4145 4.5978 2.2434 9.1570 32.943 0.7769 only 0.5%–1.2%, while consistently producing the low-
150-3 44.7284 4.8613 2.3732 7.4291 29.8189 0.7287
est average objective values. K-means shows a variance
of 1.5%–2.0% with occasional optimal peaks of 5.16%,
underlining AdaGAE’s superior balance between stabil-
sustainable supply chain network design. This section ity and optimality.
offers a critical interpretation of the results, emphasising In summary, the results validate AdaGAE as a robust,
their theoretical and practical implications. scalable, and high-performing algorithm for clustering
From a theoretical standpoint, the superior perfor- in sustainable supply chain network design. Its abil-
mance of AdaGAE can be attributed to its integration ity to deliver high-quality solutions (deviations below
of graph autoencoders with adaptive learning mecha- 11%), efficiently solve large-scale instances(up to 62×
nisms. This design allows the algorithm to better capture speedups), and maintain low variance (≤1.2%) positions
the underlying structure and heterogeneity of the supply it as a valuable tool for both researchers and practitioners.
chain data, enabling more accurate clustering and opti-
misation. Compared to traditional clustering methods,
5.5. Implications
AdaGAE demonstrates a consistently lower deviation
from the optimal solution–particularly in medium (120- This paper presents three key implications for compa-
node) and large-scale (150-node) instances–where aver- nies seeking sustainable development through modelling
age deviations are 8.84% and 10.42%, respectively, versus and solving sustainable supply chains using the Graph
up to 12.89% and 14.10% for other methods. In smaller Autoencoder (GAE) clustering algorithm.
instances (90-node), AdaGAE’s average deviation of
10.76% remains competitive, even though K-means occa- 5.5.1. Theoretical implications
sionally achieves a lower deviation of 5.16% (Table 10). This study explores the application of artificial intelli-
As observed in Figure 5, the solution gap associated with gence, particularly graph neural networks (GNNs), in
AdaGAE narrows with repeated runs–dropping from a the domain of sustainable supply chain management,
maximum of 12.10% on the first run to below 8.6% by providing critical theoretical insights into the transfor-
the fifth–suggesting the model’s capacity for continuous mative potential of emerging technologies. The findings
learning and refinement, a trait not observed in baseline underscore that GNNs demonstrate remarkable robust-
algorithms. ness and stability in processing and classifying complex,
22 Y. GUO ET AL.

Figure 8. Comparison the sensitivity of solutions under different algorithms across various problem sizes.

non-linear data, which is pivotal for modelling the intri- 5.5.2. Policy implications
cate structures and relationships inherent in sustainable This study underscores the critical role of innovative
supply chains. This establishes GNNs as a promising the- technologies in enhancing the speed and effectiveness
oretical framework for addressing the challenges posed of decision-making in sustainable supply chain manage-
by sustainability objectives, such as balancing economic, ment. By adopting our approach, corporate managers
environmental, and social dimensions. can rapidly generate actionable solutions that balance
The study also highlights the need for theoretical economic, environmental, and operational goals. The
advancements in integrating AI-driven methodologies approach’s ability to provide near-optimal results within
into sustainable supply chain design and optimisation. significantly reduced timeframes highlights its suitability
By leveraging the capabilities of graph neural networks, for addressing the dynamic and urgent demands of sus-
researchers can better capture the relationships among tainability initiatives. This is particularly relevant in the
supply chain entities, enabling the development of more current context, where corporations are under increasing
adaptive and efficient models. Furthermore, this explo- pressure to respond swiftly to environmental regulations,
ration encourages a shift in academic discourse toward market demands, and stakeholder expectations.
a deeper examination of how AI techniques, like GNNs, Policymakers can draw from these insights to advo-
can be systematically incorporated into sustainability cate for and support the integration of advanced AI-
paradigms. based tools in corporate practices. Encouraging the adop-
The theoretical contributions of this research extend tion of such technologies through incentives, training
beyond immediate practical applications, offering a foun- programmes, or partnerships can enhance the overall
dation for future studies aimed at integrating AI tech- responsiveness and adaptability of industries to sustain-
nologies into broader sustainability frameworks, ulti- ability challenges. Furthermore, our approach’s demon-
mately contributing to societal and national advance- strated superiority over traditional clustering algorithms
ments. This work paves the way for further exploration of in terms of solution speed and fitting effectiveness sug-
how AI can drive sustainable innovation and foster theo- gests that investing in research and development for sim-
retical growth at the intersection of artificial intelligence ilar AI-driven methods could yield significant societal
and supply chain management. benefits.
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 23

The research also calls for a paradigm shift in corpo- with increasing problem scales but also demonstrate
rate and policy frameworks, emphasising the need for superior predictive accuracy in solution quality com-
infrastructure and regulatory environments that facili- pared to conventional clustering methods, particularly
tate the widespread use of AI technologies. By priori- on large-scale datasets. In the comparative analysis of dif-
tising policies that enable faster adoption of advanced ferent objective functions, the function values derived
computational techniques, governments can foster an from the AdaGAE clustering algorithm show signifi-
ecosystem where businesses are better equipped to align cant advantages across various dimensions, although its
with sustainability objectives. This alignment not only performance in optimising transportation costs and car-
supports corporate competitiveness but also accelerates bon emissions remains suboptimal. Overall, with the
progress toward broader national and global sustainabil- advancement of artificial intelligence, clustering meth-
ity targets. ods present a novel approach to addressing supply chain
challenges, paving the way for enhanced efficiency and
5.5.3. Managerial implications quality in solutions.
The study emphasises the critical role of transporta- The operationalisation of social sustainability in this
tion network design in achieving sustainable supply work focuses on supply chain responsiveness. While this
chains. Corporate managers must prioritise the search selection captures time-sensitive social impacts, compre-
for advanced algorithms capable of simultaneously hensive social sustainability assessment requires integra-
addressing multiple transportation objectives, includ- tion of complementary indicators. Subsequent research
ing cost minimisation and carbon footprint reduc- will focus on utilising clustering algorithms to tackle
tion. A robust transportation network not only sup- more social parameters, enabling simultaneous optimi-
ports sustainability objectives but also enhances opera- sation of workforce, community, and consumer welfare
tional resilience. Managers should integrate algorithm- dimensions. Moreover, the current deterministic mod-
driven insights into their strategic planning to drive elling paradigm employs linear programming formula-
more efficient and environmentally friendly logistics tions with static parameters to establish computational
operations. baselines. This foundational work paves the way for crit-
Moreover, the findings of this study highlight the ical extensions including stochastic demand modelling,
superior solution quality and stability of our approach time-dependent capacity constraints, and non-convex
when applied to large-scale sustainable supply chain chal- cost behaviour integration. Future research will syn-
lenges. For managers, this underscores the importance of thesise these elements and transform the current static
leveraging artificial intelligence to address the inherent framework into an adaptive decision support system
complexity of sustainable supply chain design. Effective capable of handling real-world volatility.
decision-making relies on tools that provide consistent,
high-quality results across diverse scenarios. Managers
should not only adopt GAE but also foster an organi-
Acknowledgments
sational culture that values continuous experimentation The authors thank the anonymous reviewers for their valuable
and iteration to unlock the full potential of such advanced contribution to improve the quality and clarity of this work.
Chat GPT 4.o was used for language improvement.
technologies.

Disclosure statement
6. Conclusion
No potential conflict of interest was reported by the author(s).
In this study, we employ an adaptive Graph Autoen-
coder (GAE) clustering algorithm for the modelling and
solution of sustainable supply chain network problem. Funding
We introduce a sustainable supply chain model based This work is funded by the Natural Science Foundation
on inter-city clustering and conduct a comprehensive of Zhejiang Province (LMS25G010003) and the Scientific
modelling of transporters. To enhance clustering perfor- Research Project of Zhejiang Provincial Department of Edu-
cation (Y202454572). This work is also partially supported by
mance, we modify the attributes traditionally used in
the Research and Innovation on Ecological Responsibility and
supply chain models. Furthermore, we provide methods Autonomous Transport and Mobility (RITMEA) project. The
for reverting to true solutions. The experimental results RITMEA project is co-financed by the European Union with
indicate that this clustering approach not only funda- the European Regional Development Fund, the French state
mentally addresses the high time complexity associated and the Hauts de France Region Council.
24 Y. GUO ET AL.

Notes on contributors Data availability statement

Yuhan Guo is Associate Professor at the The data that support the findings of this study are avail-
School of Artificial Intelligence and Infor- able from the corresponding author, Y. Guo, upon reasonable
mation Engineering of the Zhejiang Uni- request.
versity of Science and Technology. He
holds a Ph.D. in Computer and Automa-
tion from the University of Artois, France. ORCID
His main research interests include intelli-
Yuhan Guo [Link]
gent transportation optimisation, sustain-
able supply chain management, metaheuristics, and machine Runsheng Chen [Link]
learning. He has authored two books and published research Youssef Boulaksil [Link]
papers in leading international academic journals such as TRE, Hamid Allaoui [Link]
EJOR, IJPR, C&IE, and C&OR. He teaches various courses in
undergraduate and post graduate programs.
References
Runsheng Chen was born in Henan, China
in 1998. He received his B.S. degree in Ali, Syed Mithun, Andrea Appolloni, Fausto Cavallaro, Idi-
Software Engineering from Zhengzhou ano D’Adamo, Assunta Di Vaio, Francesco Ferella, Massimo
University, China in 2020. He received his Gastaldi, et al. 2023. “Development Goals towards Sustain-
M.S. degree in Software Engineering from ability.” Sustainability 15 (12): 9443. [Link]
Liaoning Technical University, China in su15129443.
2025. His main research interests include Arda, Ozlem Ayaz, Frank Montabon, Ekrem Tatoglu, Ismail
intelligent transportation and graph neu- Golgeci, and Selim Zaim. 2023. “Toward A Holistic Under-
ral network. standing of Sustainability in Corporations: Resource-Based
View of Sustainable Supply Chain Management.” Sup-
Youssef Boulaksil is Associate Professor ply Chain Management: An International Journal 28 (2):
of Operations Management and Logis- 193–208. [Link]
tics at the College of Business and Eco- Bazyar, Arash, Naeme Zarrinpoor, and Amir Safavian. 2021.
nomics of the United Arab Emirates Uni- “Optimal Design of A Sustainable Natural Gas Supply Chain
versity. He holds a Ph.D. in Industrial Network under Uncertainty.” Chemical Engineering Research
Engineering from the Eindhoven Univer- and Design 176:60–88. [Link]
sity of Technology, the Netherlands. His 09.006.
research focuses mainly on sustainability Benmamoun, Zoubida, Khaoula Khlie, Mohammad Dehghani,
in supply chain management, retail operations and logistics and Youness Gherabi. 2024. “WOA: Wombat Optimization
in emerging markets, and dual sourcing inventory manage- Algorithm for Solving Supply Chain Optimization Prob-
ment. He published several research papers in leading inter- lems.” Mathematics 12 (7): 1059. [Link]
national academic journals, such as EJOR, IISE Transactions, math12071059.
IJPE, IJPR, OR Spectrum, TRA, and IJOPM. He teaches various Biuki, Mehdi, Abolfazl Kazemi, and Alireza Alinezhad. 2020.
courses in the undergraduate BA program, as well as in MBA, “An Integrated Location-Routing-inventory Model for Sus-
MEM (Master of Engineering Management), and DBA (Doc- tainable Design of A Perishable Products Supply Chain Net-
torate of Business Administration) programs. He is a member work.” Journal of Cleaner Production 260:120842. https://
of the editorial board of the International Journal of Logistics [Link]/10.1016/[Link].2020.120842.
Research and Applications. For more details, visit his home- Brahami, Mustapha Anwar, Mohammed Dahane, Mehdi
page: [Link] Souier, and M’hammed Sahnoun. 2022. “Sustainable Capac-
itated Facility Location/Network Design Problem: A Non-
Hamid Allaoui is a Full Professor at the dominated Sorting Genetic Algorithm Based Multiobjective
University of Artois in France and Direc- Approach.” Annals of Operations Research 311 (2): 821–852.
tor of the Computer and Automation [Link]
Engineering Laboratory (LGI2A) since Bronzo, Marcelo, Marcelo Werneck Barbosa, Paulo Renato
2017. With an engineering degree in de Sousa, Noel Torres Junior, and Marcos Paulo Val-
maintenance engineering and a PhD in adares de Oliveira. 2024. “Leveraging Supply Chain Reac-
Computer Science, his journey spans both tion Time: The Effects of Big Data Analytics Capabil-
industry and academia. He began his ities on Organizational Resilience Enhancement in the
career at STMicroelectronics as a manufacturing engineer Auto-parts Industry.” Administrative Sciences 14 (8): 181.
before transitioning to research and teaching. His exper- [Link]
tise lies in the design, management, and optimisation of Cai, Jianhu, and Feiying Jiang. 2023. “Decision Models
sustainable supply chains, with a focus on scheduling and of Pricing and Carbon Emission Reduction for Low-
operations planning. Professor Allaoui has contributed to Carbon Supply Chain under Cap-and-trade Regulation.”
numerous international journals and led various collabo- International Journal of Production Economics 264:108964.
rative research and industrial projects, making significant [Link]
impact in the fields of industrial engineering and supply chain Chadha, Simranjeet S., M. Ali Ülkü, and Uday Venkatadri.
sustainability. 2022. “Freight Delivery in A Physical Internet Supply Chain:
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 25

An Applied Optimisation Model with Peddling and Ship- the Production Routing Problem.” Computers & Industrial
ment Consolidation.” International Journal of Production Engineering 196:110455. [Link]
Research 60 (16): 4995–5011. [Link] 110455.
543.2021.1946613. Granillo-Macías, Rafael. 2021. “Logistics Optimization through
Chattopadhyay, Manojit, Sourav Sengupta, and B. S. Sahay. A Social Approach for Food Distribution.” Socio-Economic
2016. “Visual Hierarchical Clustering of Supply Chain Using Planning Sciences 76:100972. [Link]
Growing Hierarchical Self-organising Map Algorithm.” 2020.100972.
International Journal of Production Research 54 (9): Guo, Yuhan, Fangxia Hu, Hamid Allaoui, and Youssef Boulak-
2552–2571. [Link] sil. 2019. “A Distributed Approximation Approach for Solv-
Chen, Jingfang, Ling Wang, Yile Liang, Yang Yu, Jie Feng, Jiuxia ing the Sustainable Supply Chain Network Design Prob-
Zhao, and Xuetao Ding. 2024. “Order Dispatching via GNN- lem.” International Journal of Production Research 57 (11):
Based Optimization Algorithm for on-Demand Food Deliv- 3695–3718. [Link]
ery.” IEEE Transactions on Intelligent Transportation Systems 6412.
25 (10): 13147–13162. [Link] Guo, Yuhan, Junyu Yu, Youssef Boulaksil, Hamid Allaoui, and
3389090. Fangxia Hu. 2021. “Solving the Sustainable Supply Chain
Das, Rakhi, Krishnendu Shaw, and Mohd Irfan. 2020. “Sup- Network Design Problem by the Multi-neighborhoods
ply Chain Network Design considering Carbon Foot- Descent Traversal Algorithm.” Computers & Industrial Engi-
print, Water Footprint, Supplier’s Social Risk, Solid Waste, neering 154:107098. [Link]
and Service Level under the Uncertain Condition.” Clean 7098.
Technologies and Environmental Policy 22(2): 337–370. Guo, Yuhan, Tao Chen, Youssef Boulaksil, Linfan Xiao, and
[Link] Hamid Allaoui. 2023. “Collaborative Planning of Multi-
Dong, Guimin, Mingyue Tang, Zhiyuan Wang, Jiechao Gao, tier Sustainable Supply Chains: A Reinforcement Learn-
Sikun Guo, Lihua Cai, Robert Gutierrez, Bradford Camp- ing Enhanced Heuristic Approach.” Computers & Industrial
bel, Laura E. Barnes, and Mehdi Boukhechba. 2023. “Graph Engineering 185:109669. [Link]
Neural Networks in LoT: A Survey.” ACM Transactions 109669.
on Sensor Networks 19 (2): 1–50. [Link] Ha, Young-Kyou, and Changjoon Lee. 2024. “Supply Chain
3565973. Agility: Achieving Robustness and Logistics Performance.”
Eghbali, Hamed, Jamal Arkat, and Reza Tavakkoli-Mogh- Journal of Distribution Science 22 (9): 65–72. [Link]
addam. 2022. “Sustainable Supply Chain Network Design for 10.15722/JDS.22.09.202409.65.
Municipal Solid Waste Management: A Case Study.” Journal Habibi, Farhad, Alireza Abbasi, and Ripon Kumar Chakra-
of Cleaner Production381:135211. [Link] bortty. 2023. “Designing an Efficient Vaccine Supply Chain
jclepro.2022.135211. Network Using A Two-Phase Optimization Approach: A
Ekren, Banu Yetkin, Sachin Kumar Mangla, Ecem Eroglu Case Study of COVID-19 Vaccine.” International Journal
Turhanlar, Yigit Kazancoglu, and Guo Li. 2021. “Lat- of Systems Science: Operations & Logistics 10 (1): 2121623.
eral Inventory Share-Based Models for LoT-enabled E- [Link]
commerce Sustainable Food Supply Networks.” Computers Hegab, Hussien, Ibrahim Shaban, Muhammad Jamil, and
& Operations Research 130:105237. [Link] Navneet Khanna. 2023. “Toward Sustainable Future: Strate-
[Link].2021.105237. gies, Indicators, and Challenges for Implementing Sustain-
Elhedhli, Samir, Fatma Gzara, and Cynthia Waltho. 2021. able Production Systems.” Sustainable Materials and Tech-
“Green Supply Chain Design with Emission Sensitive nologies 36:e00617. [Link]
Demand: Second Order Cone Programming Formulation e00617.
and Case Study.” Optimization Letters 15(1): 231–247. Hosseini, Seyedmohsen, and Dmitry Ivanov. 2022. “A Multi-
[Link] layer Bayesian Network Method for Supply Chain Dis-
Emami, Seyedeh Gelareh, Valentina Lorenzoni, and Giuseppe ruption Modelling in the Wake of the COVID-19 Pan-
Turchetti. 2024. “Towards Resilient Healthcare Systems: A demic.” International Journal of Production Research 60 (17):
Framework for Crisis Management.” International Journal 5258–5276. [Link]
of Environmental Research and Public Health 21 (3): 286. 3180.
[Link] Jabbarzadeh, Armin, Behnam Fahimnia, and Fatemeh Sabouhi.
Gao, Yao, Shaojun Lu, Hao Cheng, and Xinbao Liu. 2024. 2018. “Resilient and Sustainable Supply Chain Design:
“Data-Driven Robust Optimization of Dual-Channel Sustainability Analysis under Disruption Risks.” Interna-
Closed-Loop Supply Chain Network Design considering tional Journal of Production Research 56 (17): 5945–5968.
Uncertain Demand and Carbon Cap-and-trade Policy.” [Link]
Computers & Industrial Engineering 187:109811. [Link] Joshi, Sidharath. 2022. “A Review on Sustainable Supply Chain
org/10.1016/[Link].2023.109811. Network Design: Dimensions, Paradigms, Concepts, Frame-
Gao, Jiyao, and Fengqi You. 2017. “Modeling Framework work and Future Directions.” Sustainable Operations and
and Computational Algorithm for Hedging against Uncer- Computers 3:136–148. [Link]
tainty in Sustainable Supply Chain Design Using Functional- 01.001.
Unit-based Life Cycle Optimization.” Computers & Chemi- Katsoras, Efthymios, and Patroklos Georgiadis. 2022. “An Inte-
cal Engineering 107:221–236. [Link] grated System Dynamics Model for Closed Loop Supply
chemeng.2017.05.021. Chains under Disaster Effects: The Case of COVID-19.”
Geiger, Andreas. 2024. “Enhancing Supply Chain Coordina- International Journal of Production Economics 253:108593.
tion: A Comparative Analysis of Clustering Techniques for [Link]
26 Y. GUO ET AL.

Kim, Changsu, Thai Ngan Do, and Jiyong Kim. 2024. “Spa- Quddus, Md Abdul, Niamat Ullah Ibne Hossain, Marufuz-
tially Explicit Supply Chain for Nationwide CO2-to-fuel zaman Mohammad, Raed M. Jaradat, and Mohammad S.
Infrastructure: Data-Driven Optimization with Gaussian Roni. 2017. “Sustainable Network Design for Multi-purpose
Mixture Model-Based Region Screening and Clustering.” Pellet Processing Depots under Biomass Supply Uncer-
Journal of Cleaner Production 471:143390. [Link] tainty.” Computers & Industrial Engineering 110:462–483.
10.1016/[Link].2024.143390. [Link]
Kong, Yudan, Jiaguo Liu, and Jihong Chen. 2023. “Exploring Rabbani, Masoud, Fatemeh Navazia, Hamed Farrokhi-Asl, and
the Carbon Abatement Measures in Maritime Supply Chain: Mohammad Hasan Balali. 2018. “A Sustainable Transpor-
A Scenario-Based System Dynamics Approach.” Interna- tation-Location-routing Problem with Soft Time Windows
tional Journal of Production Research 61 (18): 6131–6152. for Distribution Systems.” Uncertain Supply Chain Manage-
[Link] ment 6 (3): 229–254. [Link]
Kumar, Amit, and Kaushal Kumar. 2024. “An Uncertain 12.002.
Sustainable Supply Chain Network Design for Regulating Ran, Xiaojuan, Naret Suyaroj, Worawit Tepsan, Jianghong Ma,
Greenhouse Gas Emission and Supply Chain Cost.” Cleaner Xiangbing Zhou, and Wu Deng. 2024. “A Hybrid Genetic-
Logistics and Supply Chain10:100142. [Link] Fuzzy Ant Colony Optimization Algorithm for Automatic
16/[Link].2024.100142. K-means Clustering in Urban Global Positioning System.”
Lahri, Vijay, Krishnendu Shaw, and Alessio Ishizaka. 2021. Engineering Applications of Artificial Intelligence 137:109237.
“Sustainable Supply Chain Network Design Problem: Using [Link]
the Integrated BWM, TOPSIS, Possibilistic Programming, Ren, Yazhou, Jingyu Pu, Zhimeng Yang, Jie Xu, Guofeng
and -constrained Methods.” Expert Systems with Appli- Li, Xiaorong Pu, Philip S. Yu, and Lifang He. 2024.
cations 168:114373. [Link] “Deep Clustering: A Comprehensive Survey.” IEEE Trans-
4373. actions on Neural Networks and Learning Systems 36 (4):
Li, Xuelong, Hongyuan Zhang, and Rui Zhang. 2021. “Adap- 5858–5878. [Link]
tive Graph Auto-encoder for General Data Clustering.” IEEE Reusken, Meike, Frans Cruijssen, and Hein Fleuren. 2023. “A
Transactions on Pattern Analysis and Machine Intelligence 44 Food Bank Supply Chain Model: Optimizing Investments to
(12): 9725–9732. [Link] Maximize Food Assistance.” International Journal of Produc-
5687. tion Economics 261:108886. [Link]
Madani, Batool, Afef Saihi, and Akmal Abdelfatah. 2024. “A 23.108886.
Systematic Review of Sustainable Supply Chain Network Sabogal-De La Pava, Martha Lucía, Carlos Julio Vidal-Holguín,
Design: Optimization Approaches and Research Trends.” Diego Fernando Manotas-Duque, and Juan José Bravo-
Sustainability 16 (8): 3226. [Link] Bastidas. 2021. “Sustainable Supply Chain Design consider-
3226. ing Indicators of Value Creation.” Computers & Industrial
Martins, C. L., and M. V. Pato. 2024. “Decomposition Heuris- Engineering 157:107294. [Link]
tics for Multiobjective Problems. The Food Bank Network 107294.
Redesign Case.” International Journal of Production Eco- Sabouhi, Fatemeh, Mohammad Saeed Jabalameli, and Armin
nomics 268:109121. [Link] Jabbarzadeh. 2021. “An Optimization Approach for Sus-
9121. tainable and Resilient Supply Chain Design with Regional
Mohammed, Awsan, Maged S. Al-Shaibani, and Salih O. Duf- Considerations.” Computers & Industrial Engineering 159:
fuaa. 2023. “A Meta-heuristic-Based Algorithm for Design- 107510. [Link]
ing Multi-objective Multi-echelon Supply Chain Network.” Setiawan, Hendry Sugianto, Zeplin Jiwa Husada Tarigan, and
Applied Soft Computing 147:110774. [Link] Hotlan Siagian. 2023. “Digitalization and Green Supply
16/[Link].2023.110774. Chain Integration to Build Supply Chain Resilience toward
Moh’d Anwer, A. L.-Shboul. 2022. “An Investigation of Trans- Better Firm Competitive Advantage.” Uncertain Supply
portation Logistics Strategy on Manufacturing Supply Chain Chain Management 11 (2): 683–696. [Link]
Responsiveness in Developing Countries: The Mediating 67/[Link].2023.1.012.
Role of Delivery Reliability and Delivery Speed.” Heliyon 8 Shaw, Krishnendu, Mohd Irfan, Ravi Shankar, and Surendra
(11): e11283. [Link] S. Yadav. 2016. “Low Carbon Chance Constrained Supply
Moheb-Alizadeh, Hadi, Robert Handfield, and Donald Wars- Chain Network Design Problem: A Benders Decomposition
ing. 2021. “Efficient and Sustainable Closed-Loop Supply Based Approach.” Computers & Industrial
Chain Network Design: A Two-Stage Stochastic Formu- Engineering 98:483–497. [Link]
lation with A Hybrid Solution Methodology.” Journal of 06.011.
Cleaner Production 308:127323. [Link] Shen, Jiayu. 2020. “An Uncertain Sustainable Supply Chain
epro.2021.127323. Network.” Applied Mathematics and Computation 378:
Naeeni, Hannan Sadjady, and Navid Sabbaghi. 2022. “Sustain- 125213. [Link]
able Supply Chain Network Design: A Case of the Glass Sherafati, Mahtab, Mahdi Bashiri, Reza Tavakkoli-Moghaddam,
Manufacturer in Asia.” International Journal of Production and Mir Saman Pishvaee. 2020. “Achieving Sustainable
Economics 248:108483. [Link] Development of Supply Chain by Incorporating Various
108483. Carbon Regulatory Mechanisms.” Transportation Research
Nair, Vinod, Sergey Bartunov, Felix Gimeno, Ingrid von Glehn, Part D: Transport and Environment 81:102253. [Link]
Pawel Lichocki, Ivan Lobov, Brendan O’Donoghue, et al. org/10.1016/[Link].2020.102253.
2020. “Solving Mixed Integer Programs Using Neural Net- Shoaib, Mohd, Navonil Mustafee, and Varun Ramamohan.
works.” arXiv Preprint ArXiv :2012.13349. 2025. “A Framework for Predicting Runtime Savings from
 INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH 27

Discrete-Event Simulation Model Simplification Opera- ogy to Ensure Compliance with Sustainability Standards in
tions.” arXiv Preprint ArXiv: 2503.19264. the PPE Multi-tier Supply Chain.” International Journal of
Sorooshian, Shahryar. 2024. “The Sustainable Development Production Research 61 (14): 4934–4950. [Link]
Goals of the United Nations: A Comparative Midterm 10.1080/00207543.2022.2025944.
Research Review.” Journal of Cleaner Production 453:142272. Wang, Gang. 2023. “Designing A Stochastic Supply Chain
[Link] Network: An Error-Bound-based Heuristic.” Computers &
Sun, Junyan, Zefei Chen, Zhirui Chen, and Xiaopeng Li. Industrial Engineering 184:109570. [Link]
2024. “Robust Optimization of A Closed-Loop Supply Chain [Link].2023.109570.
Network Based on an Improved Genetic Algorithm in Wang, Zhen, Chunming Ye, and Jianquan Guo. 2024. “Robust
an Uncertain Environment.” Computers & Industrial Engi- Optimization of Multi-objective Multi-cycle Remanufactur-
neering 189:109997. [Link] ing Supply Chain Network considering Incentive Compat-
997. ibility Theory under Improved Algorithm.” Computers &
Vafaei, Arezoo, Saeed Yaghoubi, Javad Tajik, and Farnaz Barz- Industrial Engineering 187:109777. [Link]
inpour. 2020. “Designing A Sustainable Multi-channel Sup- [Link].2023.109777.
ply Chain Distribution Network: A Case Study.” Journal Yadav, Vinay Surendra, A. R. Singh, Rakesh D. Raut, and
of Cleaner Production251:119628. [Link] Naoufel Cheikhrouhou. 2021. “Design of Multi-objective
jclepro.2019.119628. Sustainable Food Distribution Network in the Indian Con-
Valderrama, Claudio Vergara, Ernesto Santibanez-González, text with Multiple Delivery Channels.” Computers & Indus-
Bruno Pimentel, Alfredo Candia-Véjar, and Linda Canales- trial Engineering 160:107549. [Link]
Bustos. 2020. “Designing an Environmental Supply Chain 2021.107549.
Network in the Mining Industry to Reduce Carbon Zahiri, Behzad, Jun Zhuang, and Mehrdad Mohammadi. 2017.
Emissions.” Journal of Cleaner Production 254:119688. “Toward an Integrated Sustainable-Resilient Supply Chain:
[Link] A Pharmaceutical Case Study.” Transportation Research
Varsei, Mohsen, and Sergey Polyakovskiy. 2017. “Sustainable Part E: Logistics and Transportation Review 103:109–142.
Supply Chain Network Design: A Case of the Wine Industry [Link]
in Australia.” Omega 66:236–247. [Link] Zarrinpoor, Naeme, and Aida Khani. 2021. “Designing A Sus-
omega.2015.11.009. tainable Biofuel Supply Chain by considering Carbon Poli-
Wang, Bill, Zhiyu Lin, Michael Wang, Fangyi Wang, Peng cies: A Case Study in Iran.” Energy, Sustainability and Society
Xiangli, and Zhi Li. 2023. “Applying Blockchain Technol- 11(1): 1–22. [Link]