Download as pdf or txt
Download as pdf or txt
You are on page 1of 45

Annals of Operations Research

https://doi.org/10.1007/s10479-024-06088-0

ORIGINAL RESEARCH

Artificial intelligence powered predictions: enhancing supply


chain sustainability

Reza Farzipoor Saen1 · Farzaneh Yousefi2 · Majid Azadi3

Received: 23 March 2024 / Accepted: 29 May 2024


© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024

Abstract
Emerging advanced digital technologies, such as Blockchain and artificial intelligence (AI),
have had a substantial impact on performance improvement and operations optimization in
industrial organizations. This study presents a network model designed for sustainable supply
chains based on a real case study in the oil industry that deals with recursive outputs using
data envelopment analysis (DEA) approach. When designing this network, recurrent loops
are considered as factors that exit the stages and re-enter the previous stages as inputs. These
factors should be designed in a way to minimize their generation while maximizing their
utilization. The designed network model is then extended to a dynamic DEA model. Finally,
the performance of supply chains is predicted and evaluated with the least error for future
time periods using an explainable artificial neural network before they become inefficient.
The findings indicate that a rise in undesirable outputs notably impacts the efficiency of
decision-making units (DMUs) across different time periods. This paper’s approach not only
identifies these factors for forecasting trends in supply chain efficiency but also allows for
the observation of the effects of research and development budget allocations as a dual-
role factor influencing supply chain efficiency in future time frames. The model presented,
which takes into account the interaction between time periods, provides managers with a
framework to analyze the nature of each of these factors in the fluctuations seen in supply
chain efficiency. This paper emphasizes the role of explainable AI in forecasting supply chain
efficiency, enabling decision-makers to anticipate future trends beyond past performance. By
integrating growth trends, progress rates, and current efficiency levels, this approach refines
unit rankings. Analyzing projected efficiency trends, particularly in relation to investments

B Reza Farzipoor Saen


farzipour@yahoo.com
Farzaneh Yousefi
Farzaneh.Yousefi.or@gmail.com
Majid Azadi
azadim@aston.ac.uk
1 Department of Operations Management and Business Statistics, College of Economics and Political
Science, Sultan Qaboos University, Muscat, Oman
2 Faculty of Management, University of Tehran, Tehran, Iran
3 Operations and Information Management Department, Aston Business School, Aston University,
Birmingham, UK

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

in green research and development, highlights their significant impact on long-term supply
chain performance. Managers can use these insights to allocate resources effectively and
optimize strategies for sustained success.

Keywords Explainable artificial intelligence (XAI) · Sustainable supply chains · Dynamic


network data envelopment analysis (DEA) · Undesirable outputs · Oil industry

1 Introduction

The performance prediction of sustainable supply chains (SSCs) is very critical for orga-
nizations’ success. However, the lack of applying powerful approaches and techniques is
a key challenge for organizations to do so. The approaches and techniques applied for the
performance prediction of sustainable supply chains should first be able to consider each
component of the supply chain (SC), including human and physical resources. Not many
approaches and techniques can provide such capabilities to researchers and decision-makers.
Moreover, these approaches and techniques should identify inefficient resources in the per-
formance of sustainable supply chains. In addition, the accuracy of performance prediction
of sustainable supply chains should be reliable for managers and decision-makers to make
important decisions. However, the majority of the traditional forecasting approaches and
techniques might be unable to provide accurate results. Therefore, there is a critical need
to develop and apply powerful approaches and techniques to the performance prediction of
sustainable supply chains based on appropriate criteria.
One of the key factors in organizations’ success is to measure the performance of sup-
ply chains by considering economic, environmental and social aspects. However, in spite
of the increasing efforts in performance evaluation of SSCs, the disruption impact on it has
remained unexplored to date. In addition, in order to benefit from competitive advantages
in today’s markets considering sustainability and resilient factors in performance evalua-
tion supply chains is inevitable. Therefore, there is a need to develop efficiency evaluation
approaches and tools to improve the performance of resilient and sustainable SCs. In the
last two decades, there has been increasing growth in the number of proposed approaches
for performance evaluation of SC. However, the majority of the approaches deal with deter-
ministic environments and are unable to deal with an uncertain situation. Furthermore, most
of these approaches have been developed and applied in the manufacturing and industrial
sectors and little attention has been paid to service sectors such as public transport. Based
on the literature, the existing approaches are unable to evaluate the performance of transport
service providers’ SCs considering both sustainability and resilience criteria simultaneously.
Moreover, the proposed methods are unable to address both fuzzy data and stochastic data
simultaneously to evaluate the efficiency of a set of decision-making units (DMUs) in network
structures. Lastly, these approaches are unable to address both fuzzy data and stochastic data
simultaneously to evaluate the efficiency of a set of DMUs in network structures. In this study,
we deal with a two-stage supply chain in the public transport sector where the chain deals
with economic, environmental, social and resilient criteria simultaneously. To set the stages,
first, we considered each urban bus service providers as DMUs. Each of the chain stages in
this investigation has its own criteria. There are four inputs; the number of seats (economic),
the cost for training staff on safety and health issues (social), operating network (economic),
and staff cost (economic), three carryovers; preventive maintenance (economic), the aver-
age number of breakdowns (resilient), and environmental cost (environmental), and three

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

outputs; profit (economic), delay time average (resilient), the number of received warnings
(social) and CO2 emission (environmental).
The creation of computer systems that can carry out activities requiring human intellect
is known as artificial intelligence (AI), and it has been a major field for many years. AI
comprehends, processes, and analyzes large volumes of data. AI-based systems can learn and
intelligently mimic human behavior and decision-making (Goli et al., 2018, 2019, 2020, 2021;
Kim et al., 2023). AI has grown in significance for organizations, aiding in operational and
strategic decision-making across a range of sectors. Machine learning (ML), a popular branch
of AI, employs algorithms to enhance judgment. ML has been effectively used in various
fields requiring complex decision-making, such as predicting the sustainability of healthcare
supply chains (Azadi et al., 2023a), improving manufacturing systems (Demlehner et al.,
2021), predicting, optimizing, and controlling energy storage systems (Olabi et al., 2023),
forecasting and minimizing inventory distortions for resilient supply chains (Jauhar et al.,
2023), among others. Early attempts to mimic human decision-making processes using AI
were only partially effective due to the difficulty of properly defining the rules employed to
solve problems. However, the creation of algorithms to aid decision-making in various areas,
including energy, education, healthcare, economics, agriculture, banking, and medicine, has
made significant strides thanks to advances in AI (Zhdanov et al., 2022).
Improved prediction accuracy has been a major focus of recent AI developments, utiliz-
ing various advanced techniques such as deep learning, machine learning, text analysis, data
mining, and machine vision, among others (Dubey et al., 2022). Outputs generated by AI
are primarily comprehensible to AI experts who can decipher complex algorithms (Wang
et al., 2022). However, due to the frequent trade-off between explainability and predictive
performance, decision-makers often harbour mistrust and reject AI systems. In fact, conven-
tional AI algorithms present a significant dilemma for human decision-makers: how can they
have confidence in the results and advocate for the use of these algorithms? When humans
cannot obtain reasonable explanations for the internal computations that underlie AI deci-
sions, establishing trust and gaining validation becomes challenging (Kraus et al., 2020). For
instance, in a hypothetical scenario where AI is used to predict diseases (e.g., COVID-19), it
would be essential for the AI system to justify its predictions so that healthcare professionals
can place their trust in them. By offering explanations that are: (1) easily understandable
by both laypeople and professionals, (2) relevant to humans, (3) connected to contextual
information or previous knowledge, and (4) representative of human decision-making logic,
it is possible to enhance trust in AI decisions (Ding et al., 2022).
Regarding the provision of justifications for AI decisions, the research community employs
a variety of concepts, approaches, and methodologies. The definition of what constitutes an
"explanation" often depends on the perspectives of individual researchers, further contribut-
ing to this variation (Shin, 2021). Thus, explainability is a fundamental concept in AI. The aim
of explainable AI is to understand how AI systems make decisions, forecast outcomes, and
perform tasks (Kim et al., 2023). It underscores the merits and demerits of different decision-
making techniques and furnishes rationale for decision support systems. Many scholars and
industry experts have argued that explainable AI is indispensable for the development and
deployment of AI across various sectors, including marketing, retail, supply chains, elec-
tricity, manufacturing, healthcare, education, and agriculture. They have also recognized the
potential of explainable AI to enhance government productivity and decision-making capa-
bilities. In many cases, the accuracy of a model is not as critical as understanding why it
makes specific decisions and predictions. Managers can grasp the model parameters and use
them confidently with the assistance of model explainability. Consequently, managers are

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

better equipped to elucidate the analytical foundation for their decisions to stakeholders (Park
& Yang, 2022).
The objective of this paper is to develop a network model for sustainable supply chains
based on a real case study in the oil industry that deals with recursive outputs. To do so, we
apply data envelopment analysis (DEA) approach and consider recurrent loops in sustainable
supply chains. The factors used in designing such a network aim to minimize their generation
while maximizing their utilization. The designed network model is also extended to a dynamic
DEA model. In brief, the main contributions of this study are as follows:

• Designing a network model for sustainable supply chains that deals with recursive
outputs.
• Developing a new dynamic network DEA model that spans multiple time periods.
• Designing an artificial neural network model to predict the performance of sustainable
supply chains with the highest accuracy.
• Providing a real application in the oil industry to implement the proposed approach.

In the next section, we provide a background on existing research works. In Sects. 3 and
4, we develop our model along with a case study. In the last section, conclusions and future
research directions are provided.

2 Background

In this section, we discuss and review the existing literature on Explainable AI, its relationship
with decision-making, the application of AI techniques in supply chains, and network DEA
models. Then, we identify and discuss knowledge gaps that need to be addressed.

2.1 Explainable AI

The phrases "explainable AI," "interpretable AI," "transparent AI," "understandable AI,"
and "responsible AI" are frequently used interchangeably in academic writing (Rai, 2020).
The primary goal of explainable AI is to provide clear, reliable, and human-understandable
explanations (Lee et al., 2022). Black boxes, such as artificial neural networks, make it
challenging to understand the relationship between inputs and outputs. This has led to efforts
to make AI models more comprehensible and interpretable (Daglarli, 2020). Explainable AI
(XAI) is a branch of AI systems that prioritizes simple, reliable models and decisions for
end users, as defined by DARPA, a well-known research organization for XAI (Gunning &
Aha, 2019). The ease with which machine learning models can be accepted and understood
will determine their interpretability for the target audience (Chen et al., 2023). According
to Saeed and Omlin (2023), explainability in the context of explainable AI generally refers
to the AI’s ability to provide users with a comprehensive methodological basis for model
predictions. When explainable AI is presented to an audience, it clarifies and simplifies how it
operates (Pradhan et al., 2023), which is another way to describe it. The importance of making
machine model operations clear and unambiguous for both technical and non-technical users
is emphasized by the concept of interpretability (Joung & Kim, 2023).
Despite the frequent confusion between the terms interpretability and explainability, they
differ conceptually. Explainability refers to the capacity to convey the decisions made by
computer models in a manner that humans can understand, as opposed to interpretability,
which focuses on explaining how or why a model arrived at a specific prediction (Wang et al.,

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

2023). Another key component of explainable AI is transparency, which is the simplicity of


understanding. When an explainable AI model can explain each of its behaviors to human
users, it is said to be transparent (Grimes et al., 2021). However, understandability is the
ability of users to quickly recognize a model’s attributes and characteristics without needing
to understand the model’s internal organization. By collaborating, model developers and UI
designers could potentially create an AI architecture that is user-centric (Siering, 2022). The
governance system, which includes rules and regulations for AI technologies to ensure their
reliability, efficiency, and productivity, is the final component of responsible AI. This includes
responsible system design, adequate monitoring, and awareness, among other guidelines and
recommendations (Irarrázaval et al., 2021).

2.2 Explainable AI and decision-making

Cavdur and Sebatli (2019) proposed a decision support system (DSS) for allocating response
facilities of temporary disasters to decrease response time using AI. The DSS developed by
Cavdur and Sebatli (2019) consisted of three main components: a database, a decision engine,
and a user interface. Wang et al. (2021) utilized explainable AI to predict venue popularity on
location-based services. Additionally, they identified effective variables that drive venue pop-
ularity in location-based services. Aggarwal and Toshniwal (2021) employed deep learning
models to forecast the incidence of hazardous fine dust and related social expenses, providing
early warnings for enhanced decision-making and planning. Lee et al. (2022) proposed a con-
volutional neural network model for forecasting the concentration of Chlorophylla in major
rivers in Korea, using various water and weather variables. They also utilized Deep SHAP
to enhance policy decision-making and identify factors affecting Chlorophyll-a levels. Park
and Yang (2022) introduced two methods for improving economic forecasts and decision-
making. They presented a deep learning model for predicting economic growth and failure
rates using the network architecture of the long short-term memory. Zhdanov et al. (2022)
integrated principles such as transparency, accountability, and fairness to build and assess
AI systems aimed at enhancing decision-making frameworks. They developed a framework
based on these principles and a privacy-limited dataset, establishing a model to strike a bal-
ance among these principles. To enhance decision-making frameworks in high-risk tasks
based on AI, Leichtmann et al. (2023) evaluated the effects of explainable AI approaches
and educational interventions in the healthcare sector. They demonstrated the effectiveness
of intellectualized mushroom-picking tasks in explainable AI research for addressing various
research questions in high-risk decision-making contexts. Yagin et al. (2023) developed an
explainable AI model to identify COVID-19 gene biomarkers. The model assists physicians
in better understanding the COVID-19 genomic forecasting decision-making process and in
maximizing patient-specific early recognition and treatment decisions.

2.3 The application of AI techniques in supply chains

AI has been applied in previous studies related to supply chains. Carbonneau et al. (2008)
explored the application of state-of-the-art AI techniques, such as support vector machines,
recurrent neural networks, and neural networks, for predicting distorted demand in supply
chains. Paul (2015) proposed a rule-based fuzzy inference model for selecting the most
appropriate supplier, considering a wide range of quantitative and qualitative indicators
to manage existing risks in supply chains. Kadadevaramath et al. (2012) applied particle
swarm intelligence algorithms in designing and optimizing supply chain networks. They

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

first developed a mathematical model for a supply chain and then used the particle swarm
optimization method to optimize the chain. Damoah et al. (2021) investigated the application
of AI-enhanced medical drones in healthcare supply chains to improve performance and
sustainability. They demonstrated how these drones can reduce CO2 emissions and improve
socio-economic indicators in healthcare supply chains. Kumar et al. (2023) identified key
success factors for applying AI in healthcare supply chains. They used the Step-Wise Assess-
ment Ratio Analysis (SWARA) method to weigh applied indicators based on their priority
in their study. Ghouri et al. (2023) used AI to formulate and test an advanced omnichan-
nel blood supply chain model. The model proposed by Ghouri et al. (2023) takes supply
and demand imbalances in unexpected conditions into account for the blood supply chain.
Liang et al. (2024) employed AI in selecting suppliers based on multi-stakeholders’ require-
ments. They identified stakeholders’ viewpoints using cluster analysis and determined group
coordination needs. Then, they evaluated and selected suppliers according to stakeholders’
group opinions. Mondal et al. (2023) proposed an integrated bio-energy supply chain and
sustainable biofuel model. In their model, they optimized green gas emissions credits and
health indicators. They also formulated a fuzzy-random robust flexible programming with
Me measure. Giri et al. (2023) considered responsiveness and resiliency indicators and devel-
oped a multi-objective mixed-integer programming model for designing a sustainable supply
chain network. Mondal et al. (2024) proposed a multi-objective model in the supply chain
to optimize sustainability and resilient goals simultaneously. They developed a prediction
model based on Dempster-Shafer evidence theory, Markov chain, and Shapely value. Giri
and Roy (2024) proposed a multi-objective mixed-integer programming model for construct-
ing a closed-loop renewable energy supply chain that is sustainable, considering some key
features. Mondal and Roy (2022) provided essential models for addressing the intricate chal-
lenges encountered in a supplier selection-order allocation dilemma, thereby mitigating risks
and disruptions. Mondal and Roy (2021) examined a multi-objective, multi-product, multi-
period, two-stage sustainable supply chain planning model, encompassing both open-loop
and closed-loop systems. The objective is to ensure a continuous supply of products between
production centers and multiple hospitals amid the COVID-19 pandemic.

2.4 Network data envelopment analysis

DEA as a nonprametric approach for measuring the performance of a number of DMUs


has been applied widely in the literature. In spite of this popularity in using DEA in perfor-
mance evalution programs its conventional models are unable to measure the performance
in network structures such as supply chains. To tackle this issue, Färe and Grosskopf (1996)
initially presented network DEA for measuring performance of DMUs when deal with net-
work strctures. By decomposing the performance of network DMUs Kao and Hwang (2008)
indiated that the overall perfroamnce of such DMUs obtained from the performance of each
layer. Another outstanding network DEA model proposed by Tone and Tsutsui (2009) by
integrating dynamic concept in the network structure. They proposed fixed and free forms
of dynamic network DEA and applied their model in energy sector. Fathi and Farzipoor
Saen (2018) presented a bidirectional network DEA model for evaluating sustainable supply
chains. The model deals with bounded connections in the network structure and applied in
the transport sector. Michali et al. (2023) incoportaed subsampling bootstrap technique in
two-stage series network DEA models. Azadi et al. (2023b) proposed a multiplier network

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

DEA model for assessing the performance of green supply chain. They also formulated com-
mon set of weight approach with respect to the cost production possibility set in their model.
They applied the model in the wire and cable industry.

2.5 Knowledge gaps

Our literature review indicates that sustainable supply chains and advanced digital tech-
nologies, such as AI, Blockchain, and deep learning, have received considerable attention.
Furthermore, it shows that network DEA, due to its numerous features, has been applied
in various sectors, including transportation, education, and energy. However, the literature
review reveals that the performance prediction of sustainable supply chains, using powerful
techniques like explainable artificial intelligence and a network DEA model, in strategic
industries, including the oil sector, has been explored to a much lesser extent. Hence, to
address this significant gap in the literature, we propose a framework for the performance
prediction of sustainable supply chains using explainable artificial intelligence and a network
DEA model and apply it in the oil industry.

3 Proposed methodology

In this section, we introduce a dynamic network DEA model with recursive outputs for
predicting the performance of sustainable supply chains. The model is designed to consider
all components of the supply chain, including inputs, intermediates, and outputs, and can
be generalized for various sustainable supply chains. We then extend the network model
to a dynamic DEA model. Subsequently, we evaluate the performance of supply chains for
future time periods using an explainable artificial neural network. This assessment allows
managers and decision-makers to take proactive measures before inefficiencies arise, thereby
enhancing the efficiency of supply chain operations.
As is shown in Fig. 1, the network model has inputs and outputs as follows:

• Inputs: Inputs that can enter any of the stages from outside the network.
• Desirable Outputs: Outputs that can exit from any stage of network.
• Undesirable Outputs: Outputs that exit from any of the stages but are either wastes or
cannot be considered as final outputs due to their low quality.

Fig. 1 Recurrent network topology

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

• Recoverable Undesirable Outputs: Undesirable outputs that can be recycled and re-entered
into one of the stages to be converted into final outputs through reprocessing or recycling.
Here, the supply chains are represented by j, where (j  1,2,…,J). Additionally, the DMU
under consideration is represented by o. The time periods are represented by p, where (p
 1,2,…,P). The stages in the supply chains are represented by s, where (s  1,2,…,S).

Table 1 shows the used notations.

Table 1 The notations

Notations Explanations

xisj p , (i  1,2,…,I) The i-th input related to the j-th supply chain in the p-th time period,
which enters the s-th stage
yt+s
j p , (t  1,2,…,T ) The t-th desirable output in supply chain j during time period p, which
has exited stage s and is considered as the final output of the network
yt−s
j p , (t  1,2,…,T ) The t-th undesirable output of supply chain j in time period p, which is
exited from stage s and considered as the final output of the network
wqss+1
j p , (q  1,2,…,Q) The q-th input of supply chain j in time period p that exits stage s and is
entered into stage s + 1
r zssj p , (z  1,2,…,Z) The q-th recyclable undesired output of the supply chain j in time period
p that exits from stage s and serves as an input to re-enter stage s
ytss+1
j p , (t  1,2,…,T) The t-th desirable output for the supply chain j in time period p, which
exits stage s and enters stage s + 1
λisp The coefficient of xisj p
βstp The coefficient of ytsj p
αqs p The coefficient of wqs j p and yt−s
jp

ηss
zp The coefficient of r zssj p
ϑ It represents the penalty assigned to the undesirable output that is
recyclable, and its value is between 0 and 1. If the value of 1 is selected
for this coefficient, it means that no penalty is applied, and if the value
of 0 is selected for this coefficient, it means that the highest penalty is
applied and this output is not considered in the efficiency calculation of
the stage in question
good
c f j p , (f  1, 2, …, F) The desired relationship of f with supply chain j, which is withdrawn
from time period p
chbad
j p ., (h  1, 2, …, H) The undesired relationship of h with supply chain j, which is withdrawn
from time period p
dual−r ole
cm jp , (m  1, 2, …, M) The dual-role relationship of m with supply chain j, which is withdrawn
from time period p
fp The f-th desired relationship coefficient associated with time period p
ξhp The h-th undesired relationship coefficient associated with time period p
θmp The dual-role relationship coefficient with the m-th desired entity
associated with time period p
The dual-role relationship coefficient with the m-th undesired entity
associated with me period p

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

The efficiency of the first stage in Fig. 1 can be calculated using Formula 1.
T T
t1 βt p yt j p + t1 βt p yt j p
s s ss+1 ss+1
ρP j  I
1
 Z T −s −s
(1)
i1 λi p x i j p + z1 ηzp r z j p + t1 αq p yt j p
s s s+2s s+2s

To calculate the efficiency of the first stage of Fig. 1, the weighted outputs that exit the
network as final output is added to the sum of weighted outputs that enter stage s + 1. In
the denominator of the fraction, there are two types of inputs. The first type is the sum of
weighted inputs that enter the supply chain from outside the network, and the second type is
the sum of weighted undesired and recoverable outputs that exit stage s + 2 and enter stage
s as inputs, which are considered in calculating the efficiency of the first stage. The second
stage of the network in the time period is calculated using Formula 2.
T Z
t1 βt p z1 ϑ(ηzp
s+1s+2 y s+1s+2 + s+1s+1 r s+1s+1 )
tjp zjp
ρP j  I
2
T Z (2)
λi p xi j p +
s+1 s+1
βt p yt j p +
ss+1 ss+1
ηs+1s+1 r s+1s+1
i1 t1 z1 zp zjp
Q T
−s+1
+ αqss+1
p wq j p +
ss+1
αq−s+1
p yt j p
q1 t1

To calculate the efficiency of the second stage in the network shown in Fig. 1, the numerator
is composed of two parts. The first part is the sum of weighted inputs entering this stage from
outside the network. The second part is the sum of weighted outputs leaving stage s + 1
and entering stage s + 2 as inputs. A penalty coefficient is applied to this value. Regarding
the recycled undesirable outputs, it should be noted that they are considered both inputs and
outputs in the network at the same time. In other words, the penalty for these recycled outputs
is that they are counted as inputs again. In the denominator of the fraction, the efficiency of the
second stage is composed of four parts. The first part is the sum of weighted inputs entering
this stage from outside the network. The second part is the sum of weighted outputs leaving
stage s and entering stage s + 1 as inputs. The third part is the sum of weighted undesirable
outputs leaving stage s and being recycled back as inputs to stage s + 1. The fourth part is
the sum of weighted undesirable outputs leaving stage s + 1 and being counted as the final
output, but due to their undesirable nature, they reduce the overall efficiency of this stage.
To calculate the third stage of supply chain, during the evaluation period, Formula 3 is
used.
T Z
t1 βt p yt j p + z1 ϑ(ηzp r z j p )
s+2 s+2 s+2s s+2s
ρ P j  T
3
 Q T −s+2 −s+2
(3)
t1 βt p yt j p q1 αq p
s+1s+2 w s+1s+2 +
t1 αq p yt j p
ss+2 s+1s+2 +
qjp

As it is evident, the fraction has two parts. The first part is the sum of weighted outputs that
exit as the final output from stage s+2, and the second part is the sum of weighted undesired
outputs that exit from stage s + 2 and enter stage s. In the denominator, the sum of weighted
inputs that exit from stage s + 1 and enter stage s + 2 is considered. Additionally, the provision
of undesired outputs that exit as the final output from the network is discussed. Based on this
basis, the efficiency of the stages is calculated according to Formula 4. If there is no agent in
a stage, the value zero is taken in this formula.
 P T I  
p1 t1 βts+1s+2
p yts+1s+2
j p + i1 ϑ ηzp
s+1s+1 r s+1s+1
z j p
ρ sP j  ⎛ I T T ⎞ (4)
P ⎜ λis+1 xis+1 + βtss+1 ytss+1 + αq−s+1 yt−s+1
i1 p j p t1 p j p t1 p j p ⎟
p1 ⎝  Z Q T ⎠
−s −s
+ ηzp
s+1s+1 s+1s+1
rz j p + αq p wq j p +
s+1 s+1
αq p yt j p
z1 q1 t1

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

Fig. 2 A supply chain and the connections between the loops

Figure 2 shows a typical supply chain and connections between loops.


Supply chains are evaluated in different time periods, and each time period is connected
to another time period using three types of carry-overs:
Good relationship: A desirable relationship that exits from one time period and enters another
time period. Due to its desirability, the model seeks to maximize this relationship.
Bad relationship: An undesirable relationship that exits from one time period and enters
another time period. Due to its undesirability, the model seeks to minimize this relationship.
Dual-role relationship: A relationship that lacks a clear nature, so the nature of this relationship
will be determined by using pairwise comparison variables within the model.
Considering that the dual-role relationship coefficient is calculated in common with both
the desired and undesired entities, the following formula is also taken into account in order
to avoid repetitive calculation of the factor:
θmp + ϕmp  1
Given this, Formula (5) is used to calculate the efficiency of jth supply chain in time period
p.
F good  M
 f pc f jp dualr ole
θmp cm
f 1 m1 jp
P T Z
β s+1s+2 yts+1s+2
+ + s+1s+1 r s+1s+1
ϕ(ηzp
t1 t p jp p1 z1 zjp
ρ jp    
H bad M dualr ole P
ξ c ϕmp cm j p (
h1 hp h j p m1 p1
I T Z Q T
λs+1 x s+1 + β ss+1 ytss+1
jp +
s+1s+1 r s+1s+1 ) +
θ(ηzp α s+1 w s+1 + α −s y −s
i1 i p i j p t1 t p z1 zjp q1 qp q j p t1 qp t j p
(5)
The efficiency of the supply chain in the evaluated time period is presented according to
Model (6).
F M P  I 
good T
f 1  f pc f jp m1 θmp cm j p
dualr ole + p1 t1 βts+1s+2
p yts+1s+2
jp + i1 ϑ(ηzp
s+1s+1 r s+1s+1
zjp
Max  M P  T 
H I ss+1  Z ηs+1s+1 r s+1s+1 +  Q α s+1 w s+1 + T α −s y −s
h1 ξhp ch j p ϕmp cm λis+1 βtss+1
bad + dualr ole + ( s+1
m1 jp p1 i1 p xi j p + t1 p yt j p + z1 zp zjp q1 qp qjp t1 qp t j p

s.t.
P  I 
T
p1 βts+1s+2
t1p yts+1s+2
jp + i1 ϑ(ηzp
s+1s+1 r s+1s+1
zjp
P  T  ≤1
I
λis+1 s+1 βtss+1 ss+1  Z ηs+1s+1 r s+1s+1 +  Q α s+1 w s+1 + T α −s y −s
p1 i1 p xi j p + t1 p yt j p + z1 zp zjp q1 qp qjp t1 qp t j p
F good M
f 1  f pc f jp m1 θmp cm j p
dualr ole
H M ≥1 (6)
h1 ξhp ch j p ϕmp cm
bad dualr ole
m1 jp

θmp + ϕmp  1θmp ≥ 0, ϕmp ≥ 0, ξhp ≥ 0,  f p


≥ 0, θmp ≥ 0, αqs+1
p ≥ 0, ηzp ≥ 0 , βt p
s s+1 s+2
≥ 0 , λis+1
p ≥ 0 , βt p
s s+1
≥ 0.
Essentially, Model (6) is a nonlinear model that needs transformation into a linear one.
Since Model (6) is fractional, we apply Charnes and Cooper’s (1961) transformation to

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

convert it into linear Model (7). To do this, we maximize the objective function by setting its
denominator to one, serving as the primary constraint, while also maximizing the numerator.
Similarly, other constraints are converted into linear forms. Given that Model (6) is fractional,
it can be transformed into a linear model using the Charnes-Cooper transformation. The linear
model is denoted by Model (7).
 T 

F
good

M 
P  
I
Max  f pc f jp θmp cm
dualr ole
jp + βts+1s+2
p yts+1s+2
jp + ϑ(ηzp
s+1s+1 s+1s+1
rz j p
f 1 m1 p1 t1 i1

s.t.

H 
M
ξhp chbad
jp + ϕmp cm
dualr ole
jp
h1 m1
⎛ ⎛ ⎞⎞

P I 
T 
Z 
q 
T
+⎝ ⎝ λi p xi j p +
s+1 s+1
βt p yt j p +
ss+1 ss+1
ηzp
s+1s+1 s+1s+1
rz j p + αq p wq j p +
s+1 s+1 −s −s ⎠⎠
αq p y 1 tjp
p1 i1 t1 z1 q1 t1
 T 

P  
Z
βts+1s+2
p yts+1s+2
jp + ϑ(ηzp
s+1s+1 s+1s+1
rz j p
p1 t1 z1
⎛ ⎞

P 
I 
T 
Z 
Q 
T
≤ ⎝ λis+1 s+1
p xi j p + βtss+1 ss+1
p yt j p + ηzp
s+1s+1 s+1s+1
rz j p + αqs+1
p wq j p
s+1
αq−sp yt−s ⎠
jp
p1 i1 t1 z1 q1 t1


F
good

M 
H 
M
 f pc f jp θmp cm
dualr ole
jp ≤ ξhp chbad
jp ϕmp cm
dualr ole
jp (7)
f 1 m1 h1 m1

θmp + ϕmp  1θmp ≥ 0, ϕmp ≥ 0, ξhp ≥ 0,  f p ≥ 0, θmp ≥ 0, αqs+1


p
≥ 0, ηzp
s
≥ 0, βts+1s+2
p ≥ 0, λis+1
p ≥ 0, βt p
ss+1
≥ 0.

In the following, properties for the explanatory process of the conducted predictive arti-
ficial intelligence are introduced. It presents the results in the artificial neural network as
reliable predictions. These properties explain the approach from which the predictions are
derived and the reason these results are presented.

Definition 1 If, for all n ≥ 1, there exist positive real numbers cn > 0 and γn , then ε has a
stable distribution such that for independent random variables ε1 , ε2 , …, εn , we have (Nolan,
2016):
 n 

p εi ≤ u  P(cnε + yn ≤ u)∀u (8)
i1

A characteristic function of a stable variable (ε) is provided in the following formula for
predictive purposes (Nolan, 2016):
       
exp itβ − d|t|α 1 − iσ sgn(t) tan π2α , α   1
E e ψ(t) 
itε   
exp itβ − d|t| 1 − 2iσπ sgn(t) log|t| , α1
 
t
, t  0
sgn(t)  |t|
0, t  0
α ∈ [0, 2], β ∈ R, d ∈ [0, ∞], σ ∈ [−1, 1] (9)

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

As a result, ε ∼ S(α, σ , d, β) indicates that ε follows a stable distribution with parameters


α, σ , d, β, where α is the exponential parameter, σ is the symmetric (skewness) parameter,
d is the scale parameter, and β is the location parameter. The standard stable form is given
by ε−β (itβ−d|t|x ) , then ε follows a symmetric alpha
d ∼ S(α, σ , 1, 0). If σ  0 or ψ(t)  e
stable (SαS) distribution. The predicted efficiency has the following properties:

Property 1: If ε ∼ S(α, σ , d, β) and for every α < γ , the expected values of ε with a power
greater than or equal to γ do not exist (i.e., E|ε|τ  ∞, τ ≥ γ ).

Property 2: Let’s assume  ε ∼ S(α, 0, d, β) with α  2, which means ε follows a normal


distribution (i.e., ε ∼ N β, d 2 or ε ∼ S(2, 0, d, β)). The Pareto law is applied to approxi-
ξ2
mate the probabilities of sequences. For example, in N(0,1): P(ε > ξ ) → e√2
ξ 2π
as ξ → ∞.
In this case, when ξ → ∞ and α < 2, the sequences of stable distributions are non-normal
asymptotic distributions. If P(ε > δ) → d α C(1 + σ )δ −α : PPar eto (ε > δ) and δ → ∞,
then −1 < σ ≤ 1, which is sequence of non-normal stable distributions with a right-skewed
domain. If P(ε < −δ) → d α C(1 − σ )δ −α  PPar eto (ε < −δ) and δ → ∞, then −1α < 1,
which is sequence of non-normal stable distributions with a right-skewed domain. To measure
C, the following formula is used:
 
sin απ
2 (α)
C (10)
π

4 Case study

In the preceding section, we introduced our developed model. In this section, we aim to
demonstrate the capabilities of the model by applying it to a real case study within the
oil industry. SAP, as a knowledge-based organization in the field of Enterprise Resource
Planning (ERP) software production, has achieved success by designing and implementing
the opinions of managers and experts in the manufacturing and distribution industries. The
prominent features of SAP software, which distinguish it from other similar software, are the
utilization of years of experience in this field and the localization of state-of-the-art sciences.
In this case study, the efficiency of 28 refinary supply chains in Iran are examined, and
the evaluation is based on seasonal time periods from spring 2001 to winter 2021. Here,
the prediction of efficiency for four future time periods will be discussed. The dataset was
collected by looking at documents and archives of the refineries. The evaluated supply chain
network consists of two stages: extraction and refining. The process is schematically presented
in Fig. 3. Since this article focuses on the evaluation of supply chains based on sustainable
supply chain criteria, the criteria provided are based on economic, environmental, and social
responsibility aspects.
The inputs and outputs are presented in Table 2.
In this article, an evaluation of 28 supply chains has been conducted over 80 time periods
(seasons). Additionally, the time periods are connected to each other based on the relationships
shown in Fig. 4.
Table 3 presents the dataset for five time periods related to the 28 supply chains since
winter 2020 to winter 2021. Also, the information on the relationship between these time
periods is provided in Table 4.

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

Fig. 3 The supply chain network of extraction refineries

Using Model (7), the efficiency results for these 5 time periods are presented in Table 5.
Table 5 shows that the overall efficiency of supply chain #26 is the highest and can be
identified as a benchmark for other supply chains. Also, supply chain #1 is identified as the
weakest supply chain.
The efficiency of supply chains #1 and #26 during five time periods has been shown in
Fig. 5. In Fig. 5, the efficiency of supply chain #26 is presented in orange, which indicates
a low negative slope of the efficiency of this supply chain in the last 5 time periods. Also,
supply chain #1, which is identified as an inefficient supply chain, has a positive slope in
improving its performance and has improved its performance in recent periods. As is shown
in Fig. 5, although the efficiency of supply chain #26 has been introduced as a stable supply
chain in these time periods, it has a significant decline. Meanwhile, the efficiency of supply
chain #1 has grown. This example indicates that by predicting the efficiency in future time
periods, the performance of supply chains can be examined in future periods and prevent
their inefficiency. In the following section, we will discuss predicting efficiency in future
time periods.

4.1 Prediction using deep learning

Convolutional Neural Network (CNN) is a kind of artificial neural network (ANN) that its
inputs are two-dimension matrix. Due to the distinctive characteristics of CNN analysis, it has
a wide range of applications in various fields such as image recognition, video processing,
natural language processing, and expert systems. One of the recent areas of interest for
researchers is time series analysis. In the CNN architecture, different layers are used, including
convolutional, pooling, dropout, and fully connected (dense layers). Regarding the CNN’s
functionality in brief, it can be said that at the first level, there is a convolutional layer, whose
input is an array of numbers. In this layer, new features are extracted from the matrix using
various filters. It is worth noting that the filters themselves are arrays of weights that are

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Table 2 The criteria for evaluating the sustainability of supply chains

Supply chain stage Criteria Factors Type of factor Explanations References

123
Oil well Inputs Personnel cost Economic The cost of personnel Yousefi et al. (2017),
payment for services Yousefi et al. (2021),
rendered to the oil well Yousefi et al. (2019),
during the specified time Azadi et al., (2023a,
period, provided in 2023b)
thousands of dollars
Environmental design cost Environmental The cost of designing
measures to prevent
environmental damage
during the specified time
period, provided in
thousands of dollars
Employee safety cost Social responsibility The cost of designing
measures to prevent harm
to personnel during the
specified time period,
thousands of dollars
Desired output Number of barrels of extracted Economic This output represents the Chen et al. (2020), Ren
oil number of barrels of crude et al. (2020), Wang
oil produced during the et. (2022)
evaluated time period
Undesirable output Oil pollution Environmental and The amount of oil pollution Sawarkar et al. (2019),
Economic generated during the oil Ancheyta et al.
extraction process in the (2021)
environment, presented in
tons
Annals of Operations Research

Content courtesy of Springer Nature, terms of use apply. Rights reserved.


Table 2 (continued)

Supply chain stage Criteria Factors Type of factor Explanations References

Refinery Inputs Personnel cost Economic The cost of personnel Yousefi et al., (2017),
payment for services Yousefi et al. (2021),
rendered to the oil well Yousefi et al. (2019),
during the specified time Azadi et al., (2023a,
period, provided in 2023b)
Annals of Operations Research

thousands of dollars
Environmental design cost Environmental The cost of designing
measures to prevent
environmental damage
during the specified time
period, provided in
thousands of dollars
Employee safety cost Social responsibility The cost of designing
measures to prevent harm
to personnel during the
specified time period
Number of barrels of crude oil Economic The amount of barrels of Chen et al. (2020), Ren
crude oil extracted from the et al. (2020), Wang
oil well for the production et al. (2022)
of petrochemical products
Volume of products requiring Economic The petroleum products that Sawarkar et al. (2019),
reprocessing require reprocessing and Ancheyta et al.
are reintroduced into the (2021)
petrochemical operations,
injected into distillation
towers at specific levels.
This input is provided
based on tonnage

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Table 2 (continued)

Supply chain stage Criteria Factors Type of factor Explanations References

123
Desired output Value of produced products Economic The value of the products Chen et al., (2020);
produced during the Ren et al., (2020);
evaluated time period, Wang et al., (2022)
provided in thousands of
dollars
Tonnage of produced products Economic The tonnage of products
produced during the
evaluated time period
Undesired output Products requiring reprocessing Environmental and The tonnage of products that Chen et al. (2020), Ren
economic are outside the standard and et al. (2020), Wang
need to be reprocessed and et al. (2022)
reintroduced into
petrochemical operations
by injection into distillation
towers
Greenhouse gases Environmental The amount of greenhouse
gases emitted during the
specified time period, based
on cubic meters
Inter-temporal Relationship with Transferred liabilities Economic Transferred liabilities to De Graeve et al.
relationships desirable nature future periods include (2023), Esteve et al.
outstanding debts that are (2023), Biguri (2023)
deferred to a future time
period. It is presented in
thousands of dollars
Annals of Operations Research

Content courtesy of Springer Nature, terms of use apply. Rights reserved.


Table 2 (continued)

Supply chain stage Criteria Factors Type of factor Explanations References

Relationship with dual The green research and Environmental The green research and Yousefi et al. (2017),
role development budget development budget Yousefi et al. (2021)
includes a budget allocated
for environmental
developments in future time
Annals of Operations Research

periods. From a cost


perspective, this budget is
identified as an undesirable
factor, while from a
development perspective, it
is recognized as a desirable
factor. It is presented in
thousands of dollars
Relationship with Debts that are transferred from Economic Debts that are transferred De Graeve et al.
undesirable nature one time period from one time period to a (2023), Esteve et al.
future time period. It is (2023), Biguri (2023)
presented in thousands of
dollars

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

Fig. 4 Inter-temporal relationships between two periods

updated during the training stages and undergo pooling operations, which are responsible for
reducing the dimensions and the number of network parameters (Sezer et al., 2020). Deep
neural networks are often susceptible to overfitting due to the addition of extra layers, which
can capture rare dependencies in the training data. By using the dropout method, it is possible
to mitigate rare dependencies in the training data and prevent overfitting. The output of this
layer, after being transformed into a one-dimensional vector, is sent to a fully connected
layer in the network, where a common neural network algorithm is used. The operation
of the convolutional layer is represented by the following formula (Poorzaker Arabani &
Ebrahimpour Komleh, 2019):


s(t)  (xw)(t)  x(i)w(t − i)
i−∞

where t represents time. s(t) is the feature map. w is the filter. x represents the input. i is
a counter that iterates over values from negative infinity to positive infinity. This equation
represents the mathematical expression for the convolution operation between the input signal
x and the filter w to produce the feature map s at a specific time t. It is a fundamental operation
in CNNs used for tasks like image processing and feature extraction. The operation involves
sliding the filter w over the input signal x at various time positions t, multiplying corresponding
elements, and summing up the results to generate the feature map.
The performance of these supply chains over 84 periods has been utilized. The results
for predicting the next 4 periods are presented in Table 6. Considering that the data related
to the periods ending in winter 2021 were collected in this study, the four upcoming time
periods, namely spring 2022 to winter 2022, have been forecasted. The predicted performance
indicates that supply chain 26 will experience a significant decline in the upcoming periods,
and supply chain 21 will be introduced as the top-performing supply chain. Additionally,
supply chain 1 will improve its ranking from 28th place to 24th place. These changes might
be realized during the four forecasted time periods.

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Table 3 The dataset related to the supply chain in five time periods

Supply Time Oil well Refinery


chain period
(DMUs) Inputs Desired Undesirable Inputs Desired outputs Undesirable outputs
output output

Environmental Employee Personnel Number Oil Environmental Employee Personnel Value of Products Greenhouse Tonnage of
design cost safety cost of barrels pollution design cost safety cost produced requiring gases (cubic produced
(1000 $) cost (1000 (1000 $) of (tons) (1000 $) cost (1000 (1000 $) products reprocessing meters) products
Annals of Operations Research

$) extracted $) (1000 $) (tons) (tons)


oil

1 Winter 73 27 19 4260 14 98 24 14 5154.6 298.2 25 118,929


2 2020 62 39 10 3204 12 82 14 16 3876.84 224.28 36 134,365
3 91 30 10 3977 16 44 28 13 4812.17 278.39 21 145,445
4 92 45 15 4984 11 59 13 17 6030.64 348.88 19 137,168
5 26 41 6 5698 13 36 26 16 6894.58 398.86 22 122,628
6 33 41 16 3723 10 69 39 11 4504.83 260.61 24 111,985
7 38 36 8 6821 12 40 42 15 8253.41 477.47 28 100,567
8 44 22 17 2682 14 66 37 7 3245.22 187.74 24 145,691
9 90 20 7 3044 12 91 14 12 3683.24 213.08 20 95,684
10 40 43 5 5885 11 76 12 9 7120.85 411.95 20 134,286
11 60 20 7 3941 10 30 38 19 4768.61 275.87 29 118,923
12 92 18 8 4460 15 71 14 10 5396.6 312.2 25 138,136
13 20 21 15 5751 13 60 21 17 6958.71 402.57 31 105,551
14 40 44 17 2785 14 33 16 6 3369.85 194.95 32 119,291
15 66 44 6 5031 16 84 45 19 6087.51 352.1 25 138,815
16 30 38 9 4636 14 56 29 7 5609.56 324.52 24 113,801
17 38 19 16 2945 12 69 36 5 3563.45 206.15 21 147,630
18 96 42 9 2508 14 57 12 15 3034.68 175.56 25 134,556
19 36 21 16 3546 11 74 22 13 4290.66 248.22 22 133,273

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Table 3 (continued)

Supply Time Oil well Refinery

123
chain period
(DMUs) Inputs Desired Undesirable Inputs Desired outputs Undesirable outputs
output output

Environmental Employee Personnel Number Oil Environmental Employee Personnel Value of Products Greenhouse Tonnage of
design cost safety cost of barrels pollution design cost safety cost produced requiring gases (cubic produced
(1000 $) cost (1000 (1000 $) of (tons) (1000 $) cost (1000 (1000 $) products reprocessing meters) products
$) extracted $) (1000 $) (tons) (tons)
oil

20 87 28 17 5004 10 28 33 15 6054.84 350.28 20 149,907


21 40 30 13 2739 9 17 23 6 3314.19 191.73 21 121,958
22 79 9 10 5285 13 83 18 7 6394.85 369.95 23 148,593
23 59 45 17 6227 14 82 16 14 7534.67 435.89 24 129,983
24 20 11 6 3345 12 84 25 15 4047.45 234.15 24 112,021
25 60 40 17 5098 10 69 14 9 6168.58 356.86 25 140,496
26 21 18 10 5078 11 45 14 12 6144.38 355.46 26 136,926
27 91 16 19 3171 10 84 45 15 3836.91 221.97 24 116,356
28 82 33 8 5441 12 72 42 5 6583.61 380.87 28 143,337
1 Spring 72.27 26.73 18.81 4217.4 16 97.02 23.76 13.86 5103.054 295.218 24.75 117,739.71
2 2021 61.38 38.61 9.9 3171.96 14 81.18 13.86 15.84 3838.072 222.037 35.64 133,021.35
3 90.09 29.7 9.9 3937.23 12 43.56 27.72 12.87 4764.048 275.606 20.79 143,990.55
4 91.08 44.55 14.85 4934.16 11 58.41 12.87 16.83 5970.334 345.391 18.81 135,796.32
5 25.74 40.59 5.94 5641.02 10 35.64 25.74 15.84 6825.634 394.871 21.78 121,401.72
6 32.67 40.59 15.84 3685.77 12 68.31 38.61 10.89 4459.782 258.004 23.76 110,865.15
7 37.62 35.64 7.92 6752.79 14 39.60 41.58 14.85 8170.876 472.695 27.72 99,561.33
8 43.56 21.78 16.83 2655.18 14 65.34 36.63 6.93 3212.768 185.863 23.76 144,234.09
Annals of Operations Research

Content courtesy of Springer Nature, terms of use apply. Rights reserved.


Table 3 (continued)

Supply Time Oil well Refinery


chain period
(DMUs) Inputs Desired Undesirable Inputs Desired outputs Undesirable outputs
output output

Environmental Employee Personnel Number Oil Environmental Employee Personnel Value of Products Greenhouse Tonnage of
design cost safety cost of barrels pollution design cost safety cost produced requiring gases (cubic produced
(1000 $) cost (1000 (1000 $) of (tons) (1000 $) cost (1000 (1000 $) products reprocessing meters) products
Annals of Operations Research

$) extracted $) (1000 $) (tons) (tons)


oil

9 89.1 19.8 6.93 3013.56 12 90.09 13.86 11.88 3646.408 210.949 19.8 94,727.16
10 39.6 42.57 4.95 5826.15 13 75.24 11.88 8.91 7049.642 407.831 19.8 132,943.14
11 59.4 19.8 6.93 3901.59 14 29.7 37.62 18.81 4720.924 273.111 28.71 117,733.77
12 91.08 17.82 7.92 4415.4 12 70.29 13.86 9.9 5342.634 309.078 24.75 136,754.64
13 19.8 20.79 14.85 5693.49 11 59.4 20.790 16.83 6889.123 398.544 30.69 104,495.49
14 39.6 43.56 16.83 2757.15 10 32.67 15.840 5.94 3336.152 193.001 31.68 118,098.09
15 65.34 43.56 5.94 4980.69 12 83.16 44.550 18.81 6026.635 348.648 24.75 137,426.85
16 29.7 37.62 8.91 4589.64 14 55.44 28.710 6.93 5553.464 321.275 23.76 112,662.99
17 37.62 18.81 15.84 2915.55 12 68.31 35.640 4.95 3527.816 204.089 20.79 146,153.7
18 95.04 41.58 8.91 2482.92 13 56.43 11.88 14.85 3004.333 173.804 24.75 133,210.44
19 35.64 20.79 15.84 3510.54 14 73.26 21.78 12.87 4247.753 245.738 21.78 131,940.27
20 86.13 27.72 16.83 4953.96 16 27.72 32.67 14.85 5994.292 346.777 19.8 148,407.93
21 39.6 29.7 12.87 2711.61 17 16.83 22.77 5.94 3281.048 189.813 20.79 120,738.42
22 78.21 8.91 9.9 5232.15 18 82.17 17.82 6.93 6330.902 366.251 22.77 147,107.07
23 58.41 44.55 16.83 6164.73 15 81.18 15.84 13.86 7459.323 431.531 23.76 128,683.17
24 19.8 10.89 5.94 3311.55 17 83.16 24.75 14.85 4006.976 231.809 23.76 110,900.79
25 59.4 39.6 16.83 5047.02 14 68.31 13.86 8.91 6106.894 353.291 24.75 139,091.04

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Table 3 (continued)

Supply Time Oil well Refinery

123
chain period
(DMUs) Inputs Desired Undesirable Inputs Desired outputs Undesirable outputs
output output

Environmental Employee Personnel Number Oil Environmental Employee Personnel Value of Products Greenhouse Tonnage of
design cost safety cost of barrels pollution design cost safety cost produced requiring gases (cubic produced
(1000 $) cost (1000 (1000 $) of (tons) (1000 $) cost (1000 (1000 $) products reprocessing meters) products
$) extracted $) (1000 $) (tons) (tons)
oil

26 20.79 17.82 9.9 5027.220 18 44.55 13.86 11.88 6082.936 351.905 25.74 135,556.74
27 90.09 15.84 18.81 3139.290 19 83.16 44.55 14.85 3798.541 219.75 23.76 115,192.44
28 81.18 32.67 7.92 5386.590 20 71.28 41.58 4.95 6517.774 377.061 27.72 141,903.63
1 Summer 71.547 26.463 18.622 4175.226 12 96.05 23.522 13.721 5052.023 292.266 24.503 116,562.313
2 2021 60.766 38.224 9.801 3140.240 11 80.368 13.721 15.682 3799.691 219.817 35.284 131,691.137
3 89.189 29.403 9.801 3897.858 10 43.124 27.443 12.741 4716.408 272.850 20.582 142,550.645
4 90.169 44.105 14.702 4884.818 10 57.826 12.741 16.662 5910.63 341.937 18.622 134,438.357
5 25.483 40.184 5.881 5584.610 12 35.284 25.483 15.682 6757.378 390.923 21.562 120,187.703
6 32.343 40.184 15.682 3648.912 14 67.627 38.224 10.781 4415.184 255.424 23.522 109,756.499
7 37.244 35.284 7.841 6685.262 15 39.204 41.164 14.702 8089.167 467.968 27.443 98,565.717
8 43.124 21.562 16.662 2628.628 16 64.687 36.264 6.861 3180.64 184.004 23.522 142,791.749
9 88.209 19.602 6.861 2983.424 12 89.189 13.721 11.761 3609.944 208.840 19.602 93,779.888
10 39.204 42.144 4.901 5767.889 14 74.488 11.761 8.821 6979.145 403.752 19.602 131,613.709
11 58.806 19.602 6.861 3862.574 14 29.403 37.244 18.622 4673.715 270.38 28.423 116,556.432
12 90.169 17.642 7.841 4371.246 12 69.587 13.721 9.801 5289.208 305.987 24.503 135,387.094
13 19.602 20.582 14.702 5636.555 13 58.806 20.582 16.662 6820.232 394.559 30.383 103,450.535
14 39.204 43.124 16.662 2729.579 11 32.343 15.682 5.881 3302.790 191.070 31.363 116,917.109
Annals of Operations Research

Content courtesy of Springer Nature, terms of use apply. Rights reserved.


Table 3 (continued)

Supply Time Oil well Refinery


chain period
(DMUs) Inputs Desired Undesirable Inputs Desired outputs Undesirable outputs
output output

Environmental Employee Personnel Number Oil Environmental Employee Personnel Value of Products Greenhouse Tonnage of
design cost safety cost of barrels pollution design cost safety cost produced requiring gases (cubic produced
(1000 $) cost (1000 (1000 $) of (tons) (1000 $) cost (1000 (1000 $) products reprocessing meters) products
Annals of Operations Research

$) extracted $) (1000 $) (tons) (tons)


oil

15 64.687 43.124 5.881 4930.883 10 82.328 44.105 18.622 5966.369 345.162 24.503 136,052.582
16 29.403 37.244 8.821 4543.744 12 54.886 28.423 6.861 5497.93 318.062 23.522 111,536.36
17 37.244 18.622 15.682 2886.395 13 67.627 35.284 4.901 3492.537 202.048 20.582 144,692.163
18 94.09 41.164 8.821 2458.091 14 55.866 11.761 14.702 2974.29 172.066 24.503 131,878.336
19 35.284 20.582 15.682 3475.435 15 72.527 21.562 12.741 4205.276 243.28 21.562 130,620.867
20 85.269 27.443 16.662 4904.420 15 27.443 32.343 14.702 5934.349 343.309 19.602 146,923.851
21 39.204 29.403 12.741 2684.494 12 16.662 22.542 5.881 3248.238 187.915 20.582 119,531.036
22 77.428 8.821 9.801 5179.829 10 81.348 17.642 6.861 6267.592 362.588 22.542 145,635.999
23 57.826 44.105 16.662 6103.083 11 80.368 15.682 13.721 7384.73 427.216 23.522 127,396.338
24 19.602 10.781 5.881 3278.435 10 82.328 24.503 14.702 3966.906 229.490 23.522 109,791.782
25 58.806 39.204 16.662 4996.550 12 67.627 13.721 8.821 6045.825 349.758 24.503 137,700.130
26 20.582 17.642 9.801 4976.948 12 44.105 13.721 11.761 6022.107 348.386 25.483 134,201.173
27 89.189 15.682 18.622 3107.897 12 82.328 44.105 14.702 3760.555 217.553 23.522 114,040.516
28 80.368 32.343 7.841 5332.724 14 70.567 41.164 4.901 6452.596 373.291 27.443 140,484.594
1 Fall 70.116 25.933 18.249 4091.721 14 94.129 23.052 13.447 4950.983 286.421 24.012 114,231.067
2 2021 59.551 37.459 9.605 3077.436 15 78.761 13.447 15.368 3723.697 215.420 34.578 129,057.314
3 87.405 28.815 9.605 3819.901 15 42.262 26.894 12.486 4622.080 267.393 20.170 139,699.632

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Table 3 (continued)

Supply Time Oil well Refinery

123
chain period
(DMUs) Inputs Desired Undesirable Inputs Desired outputs Undesirable outputs
output output

Environmental Employee Personnel Number Oil Environmental Employee Personnel Value of Products Greenhouse Tonnage of
design cost safety cost of barrels pollution design cost safety cost produced requiring gases (cubic produced
(1000 $) cost (1000 (1000 $) of (tons) (1000 $) cost (1000 (1000 $) products reprocessing meters) products
$) extracted $) (1000 $) (tons) (tons)
oil

4 88.366 43.222 14.407 4787.122 14 56.669 12.486 16.328 5792.418 335.099 18.249 131,749.590
5 24.973 39.38 5.763 5472.918 12 34.578 24.973 15.368 6622.23 383.104 21.131 117,783.949
6 31.696 39.38 15.368 3575.934 12 66.274 37.459 10.565 4326.880 250.315 23.052 107,561.369
7 36.499 34.578 7.684 6551.557 10 38.42 40.341 14.407 7927.384 458.609 26.894 96,594.402
8 42.262 21.131 16.328 2576.056 11 63.393 35.538 6.723 3117.027 180.324 23.052 139,935.914
9 86.445 19.21 6.723 2923.756 10 87.405 13.447 11.526 3537.745 204.663 19.21 91,904.291
10 38.42 41.301 4.802 5652.531 10 72.998 11.526 8.644 6839.562 395.677 19.21 128,981.434
11 57.63 19.21 6.723 3785.323 12 28.815 36.499 18.249 4580.24 264.973 27.854 114,225.304
12 88.366 17.289 7.684 4283.821 14 68.195 13.447 9.605 5183.424 299.867 24.012 132,679.352
13 19.21 20.17 14.407 5523.824 14 57.630 20.17 16.328 6683.827 386.668 29.775 101,381.524
14 38.42 42.262 16.328 2674.987 14 31.696 15.368 5.763 3236.734 187.249 30.736 114,578.767
15 63.393 42.262 5.763 4832.265 12 80.682 43.222 18.249 5847.041 338.259 24.012 133,331.530
16 28.815 36.499 8.644 4452.869 13 53.788 27.854 6.723 5387.971 311.701 23.052 109,305.633
17 36.499 18.249 15.368 2828.667 11 66.274 34.578 4.802 3422.687 198.007 20.170 141,798.32
18 92.208 40.341 8.644 2408.929 10 54.748 11.526 14.407 2914.804 168.625 24.012 129,240.769
19 34.578 20.17 15.368 3405.926 12 71.077 21.131 12.486 4121.17 238.415 21.131 128,008.45
20 83.563 26.894 16.328 4806.332 14 26.894 31.696 14.407 5815.662 336.443 19.210 143,985.374
Annals of Operations Research

Content courtesy of Springer Nature, terms of use apply. Rights reserved.


Table 3 (continued)

Supply Time Oil well Refinery


chain period
(DMUs) Inputs Desired Undesirable Inputs Desired outputs Undesirable outputs
output output

Environmental Employee Personnel Number Oil Environmental Employee Personnel Value of Products Greenhouse Tonnage of
design cost safety cost of barrels pollution design cost safety cost produced requiring gases (cubic produced
(1000 $) cost (1000 (1000 $) of (tons) (1000 $) cost (1000 (1000 $) products reprocessing meters) products
Annals of Operations Research

$) extracted $) (1000 $) (tons) (tons)


oil

21 38.42 28.815 12.486 2630.804 12 16.328 22.091 5.763 3183.273 184.156 20.170 117,140.415
22 75.879 8.644 9.605 5076.232 12 79.721 17.289 6.723 6142.241 355.336 22.091 142,723.279
23 56.669 43.222 16.328 5981.021 14 78.761 15.368 13.447 7237.035 418.671 23.052 124,848.412
24 19.21 10.565 5.763 3212.866 15 80.682 24.012 14.407 3887.568 224.901 23.052 107,595.946
25 57.63 38.42 16.328 4896.619 16 66.274 13.447 8.644 5924.909 342.763 24.012 134,946.127
26 20.17 17.289 9.605 4877.409 14 43.222 13.447 11.526 5901.665 341.419 24.973 131,517.149
27 87.405 15.368 18.249 3045.739 18 80.682 43.222 14.407 3685.344 213.202 23.052 111,759.705
28 78.761 31.696 7.684 5226.07 12 69.156 40.341 4.802 6323.544 365.825 26.894 137,674.902
1 Winter 71.54 26.46 18.62 4174.8 12 96.04 23.52 13.720 5051.508 292.236 24.5 116,550.42
2 2021 60.76 38.22 9.8 3139.92 14 80.36 13.72 15.680 3799.303 219.794 35.28 131,677.7
3 89.18 29.4 9.8 3897.46 14 43.12 27.44 12.740 4715.927 272.822 20.58 142,536.1
4 90.16 44.1 14.7 4884.32 20 57.82 12.74 16.660 5910.027 341.902 18.62 134,424.64
5 25.48 40.18 5.88 5584.04 15 35.28 25.48 15.680 6756.688 390.883 21.560 120,175.44
6 32.34 40.18 15.68 3648.54 16 67.62 38.22 10.78 4414.733 255.398 23.52 109,745.3
7 37.24 35.28 7.84 6684.58 14 39.2 41.16 14.7 8088.342 467.921 27.44 98,555.66
8 43.12 21.56 16.66 2628.36 12 64.68 36.26 6.86 3180.316 183.985 23.52 142,777.18
9 88.2 19.6 6.86 2983.12 11 89.18 13.72 11.76 3609.575 208.818 19.6 93,770.32

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Table 3 (continued)

Supply Time Oil well Refinery

123
chain period
(DMUs) Inputs Desired Undesirable Inputs Desired outputs Undesirable outputs
output output

Environmental Employee Personnel Number Oil Environmental Employee Personnel Value of Products Greenhouse Tonnage of
design cost safety cost of barrels pollution design cost safety cost produced requiring gases (cubic produced
(1000 $) cost (1000 (1000 $) of (tons) (1000 $) cost (1000 (1000 $) products reprocessing meters) products
$) extracted $) (1000 $) (tons) (tons)
oil

10 39.2 42.14 4.9 5767.3 10 74.48 11.76 8.82 6978.433 403.711 19.6 131,600.28
11 58.8 19.6 6.86 3862.18 11 29.4 37.24 18.62 4673.238 270.353 28.42 116,544.54
12 90.16 17.64 7.84 4370.8 11 69.58 13.72 9.8 5288.668 305.956 24.5 135,373.28
13 19.6 20.58 14.7 5635.98 12 58.8 20.58 16.66 6819.536 394.519 30.38 103,439.98
14 39.2 43.12 16.66 2729.3 14 32.34 15.68 5.88 3302.453 191.051 31.36 116,905.18
15 64.68 43.12 5.88 4930.38 15 82.32 44.1 18.62 5965.760 345.127 24.5 136,038.7
16 29.4 37.24 8.82 4543.28 16 54.88 28.42 6.86 5497.369 318.030 23.52 111,524.98
17 37.24 18.62 15.68 2886.1 14 67.62 35.28 4.9 3492.181 202.027 20.58 144,677.4
18 94.08 41.16 8.82 2457.84 15 55.86 11.76 14.7 2973.986 172.049 24.5 131,864.88
19 35.28 20.58 15.68 3475.08 14 72.52 21.56 12.74 4204.847 243.256 21.56 130,607.54
20 85.26 27.44 16.66 4903.92 11 27.44 32.34 14.7 5933.743 343.274 19.6 146,908.86
21 39.2 29.4 12.74 2684.22 12 16.66 22.54 5.88 3247.906 187.895 20.58 119,518.84
22 77.42 8.82 9.8 5179.3 12 81.34 17.64 6.86 6266.953 362.551 22.54 145,621.14
23 57.82 44.1 16.66 6102.46 14 80.36 15.68 13.72 7383.977 427.172 23.52 127,383.34
24 19.6 10.78 5.88 3278.1 15 82.32 24.5 14.7 3966.501 229.467 23.52 109,780.58
25 58.8 39.2 16.66 4996.04 17 67.62 13.72 8.82 6045.208 349.723 24.5 137,686.08
26 20.58 17.64 9.8 4976.44 15 44.1 13.72 11.76 6021.492 348.351 25.48 134,187.48
27 89.18 15.68 18.62 3107.58 11 82.32 44.1 14.7 3760.172 217.531 23.52 114,028.88
28 80.36 32.34 7.84 5332.18 10 70.56 41.16 4.9 6451.938 373.253 27.44 140,470.26
Annals of Operations Research

Content courtesy of Springer Nature, terms of use apply. Rights reserved.


Table 4 The relationship between time periods

DMUs Winter 2020 to Spring 2021 Spring 2021 to Summer 2021 Summer 2021 to Fall 2021 Fall 2021 to Winter 2021

Debts that Transferred The green Debts that Transferred The green Debts that Transferred The green Debts that Transferred The green
are liabilities research and are liabilities research and are liabilities research and are liabilities research and
transferred development transferred development transferred development transferred development
from one budget from one budget from one budget from one budget
time time time time
period period period period
Annals of Operations Research

1 116 300 32 149 269 26 128 389 16 141 512 35


2 147 389 11 107 468 36 153 341 14 99 371 22
3 108 266 35 155 534 3 135 451 27 149 297 12
4 155 283 3 148 467 25 126 457 2 154 439 16
5 127 435 16 110 461 34 105 391 35 121 499 14
6 105 352 26 130 398 19 113 429 15 111 323 32
7 126 480 9 148 425 29 133 527 9 111 248 20
8 116 451 36 117 364 29 133 252 30 145 468 28
9 126 499 3 107 468 35 103 254 19 141 397 16
10 127 265 16 105 247 18 156 359 32 100 243 19
11 145 490 27 133 437 23 118 346 10 136 532 19
12 126 462 23 150 502 12 118 268 7 135 444 10
13 130 527 16 108 292 2 130 451 8 153 305 18
14 122 358 9 131 342 8 154 261 22 141 396 30
15 111 313 34 112 382 17 133 536 15 148 485 31
16 148 368 34 140 522 30 140 224 24 113 448 34
17 141 388 30 132 365 2 128 291 36 149 526 31
18 137 366 32 129 223 3 141 460 6 135 417 25
19 114 276 35 141 359 27 136 232 23 111 317 16

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Table 4 (continued)

DMUs Winter 2020 to Spring 2021 Spring 2021 to Summer 2021 Summer 2021 to Fall 2021 Fall 2021 to Winter 2021

123
Debts that Transferred The green Debts that Transferred The green Debts that Transferred The green Debts that Transferred The green
are liabilities research and are liabilities research and are liabilities research and are liabilities research and
transferred development transferred development transferred development transferred development
from one budget from one budget from one budget from one budget
time time time time
period period period period

20 147 363 18 109 418 20 116 285 6 127 390 35


21 116 431 30 102 402 32 130 287 11 107 299 24
22 132 484 18 141 281 10 139 449 24 155 251 21
23 101 336 10 123 260 34 116 379 28 154 480 17
24 153 408 13 151 513 6 105 501 6 129 396 31
25 142 222 30 105 425 28 115 478 18 126 436 21
26 156 360 19 100 440 26 137 487 29 132 423 12
27 117 491 8 132 462 9 106 246 8 153 398 28
28 100 362 12 136 469 18 150 484 6 150 376 29
Annals of Operations Research

Content courtesy of Springer Nature, terms of use apply. Rights reserved.


Annals of Operations Research

Table 5 The results

DMUs Winter Spring Summer Fall 2021 Winter Overall


2020 2021 2021 2021 efficiency

1 0.438 0.574 0.600 0.616 0.626 0.5708


2 0.555 0.634 0.584 0.599 0.600 0.5944
3 0.643 0.694 0.701 0.725 0.841 0.7208
4 0.595 0.606 0.616 0.598 0.564 0.5958
5 0.886 0.877 0.870 0.901 0.839 0.8746
6 0.541 0.639 0.707 0.634 0.630 0.6302
7 0.585 0.600 0.622 0.634 0.687 0.6256
8 0.671 0.659 0.615 0.660 0.714 0.6638
9 0.499 0.596 0.612 0.634 0.639 0.5960
10 0.846 0.693 0.848 0.839 0.822 0.8096
11 0.602 0.663 0.747 0.760 0.739 0.7022
12 0.581 0.609 0.660 0.634 0.711 0.6390
13 0.624 0.606 0.598 0.714 0.801 0.6686
14 0.752 0.706 0.806 0.784 0.834 0.7764
15 0.529 0.619 0.606 0.661 0.609 0.6048
16 0.692 0.700 0.708 0.724 0.771 0.7190
17 0.719 0.725 0.801 0.834 0.902 0.7962
18 0.565 0.874 0.639 0.816 0.903 0.7594
19 0.666 0.701 0.634 0.736 0.771 0.7016
20 0.695 0.644 0.747 0.722 0.739 0.7094
21 0.893 0.936 0.964 1 1 0.9586
22 0.629 0.644 0.701 0.706 0.708 0.6776
23 0.586 0.586 0.636 0.655 0.693 0.6312
24 0.593 0.789 0.714 0.612 0.801 0.7018
25 0.676 0.707 0.780 0.636 0.700 0.6998
26 1.000 1.000 1.000 0.980 0.975 0.9910
27 0.444 0.601 0.596 0.609 0.636 0.5772
28 0.609 0.636 0.648 0.569 0.609 0.6142

Table 7 reports the prediction accuracy considering different network topologies. In


Table 7, the prediction accuracy is broken down by different network topologies and train-
ing algorithms. The number of scenarios examined to observe the prediction accuracy is
much greater than the reported cases. The table highlights the top-performing topologies,
which have the lowest errors. The "Valid percentage" column shows the percentage of data
used for validation, the "Test percentage" indicates the percentage of data used for testing,
and the "Train percentage" represents the percentage of data used for training. "Number of
hidden neurons" refers to the number of neurons in the hidden layer of the designed net-
work, "Number of delays" indicates the number of delays in the designed network, "Training
algorithm" specifies the algorithm used for training the neural network, and finally, in the
mean squared error (MSE) column, the error value is presented. As is seen, the training

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

Fig. 5 The supply chains’ efficiency during time periods

algorithm Levenberg–Marquardt has the lowest MSE, which means the training algorithm
Levenberg–Marquardt has a high accuracy.
As is shown in Table 7, the topology presented in the first row exhibits the lowest error and
is used as the reference network for prediction. Based on this network, Fig. 6 is presented, in
which data convergence is depicted considering the test, validation, and ultimately the entire
dataset. In Fig. 6, validation data is represented by the green line, indicating that 15 data
points have had the highest accuracy in prediction, meaning that the network has been able
to predict the data trend with an acceptable level of accuracy.
Finally, as is shown in Fig. 6, the shape "all," considering all the data, exhibits the high-
est convergence across all test, validation, and training data. The predictive convergence is
highest across all data, which is also evident in the overall data. Therefore, the predictions
made with the provided network have the highest accuracy.
As is shown in Table 8, the accuracy of the prediction is provided for each supply chain.
Increasing the r variable from 1 to 3 for all DMUs simulates the goodness of fit of three
classical distributions. As shown in Table 8, the accuracy of the prediction for the variable r
 2 has the highest accuracy of prediction.
Figure 7 shows the histograms of three classical distribution functions with mean and
standard deviation. Figure 7 represents the distribution function of predicted efficiency val-
ues. Three classical distribution functions with mean and standard deviation are presented.
Figure 7 shows that assuming a sequence distribution with skewness and kurtosis in ε24 k

cannot be a mistake; therefore, we can compare sequence distributions with high kurtosis to
other classical distributions that can be assumed for ε24k . Non-parametric tests like the Kol-

mogorov–Smirnov (K-S) test can also be used to identify the type of probability distribution
(Esteve et al., 2023). For instance, the K-S test result for the Lévy distribution is 0.00918.
This indicates that assuming an asymmetric Lévy distribution (α  0.5 and σ  1) with β 
32.18 and d  33.574 cannot be rejected as a hypothesis with 95% confidence. Comparing

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Table 6 Predicting the efficiency of supply chains

DMUs Achieved Predicted

Winter 2020 Spring 2021 Summer 2021 Fall 2021 Winter 2021 Spring 2022 Summer 2022 Fall 2022 Winter 2022 Overall efficiency

1 0.438 0.574 0.600 0.616 0.626 0.6962 0.738 0.7798 0.8216 0.645
2 0.555 0.634 0.584 0.599 0.600 0.6109 0.6164 0.6219 0.6274 0.597
Annals of Operations Research

3 0.643 0.694 0.701 0.725 0.841 0.8489 0.8916 0.9343 0.977 0.795
4 0.595 0.606 0.616 0.598 0.564 0.5748 0.5678 0.5608 0.5538 0.574
5 0.886 0.877 0.870 0.901 0.839 0.8536 0.8466 0.8396 0.8326 0.849
6 0.541 0.639 0.707 0.634 0.630 0.6821 0.6994 0.7167 0.734 0.655
7 0.585 0.600 0.622 0.634 0.687 0.697 0.7208 0.7446 0.7684 0.664
8 0.671 0.659 0.615 0.660 0.714 0.6899 0.6986 0.7073 0.716 0.672
9 0.499 0.596 0.612 0.634 0.639 0.6914 0.7232 0.755 0.7868 0.650
10 0.846 0.693 0.848 0.839 0.822 0.839 0.8488 0.8586 0.8684 0.818
11 0.602 0.663 0.747 0.760 0.739 0.8135 0.8506 0.8877 0.9248 0.766
12 0.581 0.609 0.660 0.634 0.711 0.7245 0.753 0.7815 0.81 0.686
13 0.624 0.606 0.598 0.714 0.801 0.8072 0.8534 0.8996 0.9458 0.750
14 0.752 0.706 0.806 0.784 0.834 0.849 0.8732 0.8974 0.9216 0.813
15 0.529 0.619 0.606 0.661 0.609 0.6654 0.6856 0.7058 0.726 0.636
16 0.692 0.700 0.708 0.724 0.771 0.7736 0.7918 0.81 0.8282 0.745
17 0.719 0.725 0.801 0.834 0.902 0.9387 0.9862 1.0337 1.0812 0.879
18 0.565 0.874 0.639 0.816 0.903 0.9448 1.0066 1.0684 1.1302 0.871
19 0.666 0.701 0.634 0.736 0.771 0.7751 0.7996 0.8241 0.8486 0.740
20 0.695 0.644 0.747 0.722 0.739 0.7592 0.7758 0.7924 0.809 0.732

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Table 6 (continued)

DMUs Achieved Predicted

123
Winter 2020 Spring 2021 Summer 2021 Fall 2021 Winter 2021 Spring 2022 Summer 2022 Fall 2022 Winter 2022 Overall efficiency

21 0.893 0.936 0.964 1 1 1.042 1.0698 1.0976 1.1254 1.000


22 0.629 0.644 0.701 0.706 0.708 0.7436 0.7656 0.7876 0.8096 0.711
23 0.586 0.586 0.636 0.655 0.693 0.7161 0.7444 0.7727 0.801 0.678
24 0.593 0.789 0.714 0.612 0.801 0.7735 0.7974 0.8213 0.8452 0.739
25 0.676 0.707 0.780 0.636 0.700 0.6929 0.6906 0.6883 0.686 0.685
26 1.000 1.000 1.000 0.980 0.975 0.97 0.963 0.956 0.949 0.963
27 0.444 0.601 0.596 0.609 0.636 0.6948 0.734 0.7732 0.8124 0.646
28 0.609 0.636 0.648 0.569 0.609 0.5941 0.5874 0.5807 0.574 0.592
Annals of Operations Research

Content courtesy of Springer Nature, terms of use apply. Rights reserved.


Annals of Operations Research

Table 7 The results of different tests

MSE Training algorithm Number Number Train Test Valid


of of percentage percentage percentage
delays hidden
neurons

0.04785 Levenberg–Marquardt 2 10 15 70 15
0.08673 Bayesian 2 10 15 70 15
regularization
0.08870 Scaled conjugate 2 10 15 70 15
gradient
0.09930 Levenberg–Marquardt 4 8 10 80 10
0.09674 Bayesian 4 8 10 80 10
regularization
0.07398 Scaled conjugate 4 8 10 80 10
gradient
0.07698 Levenberg–Marquardt 2 10 15 80 5
0.08450 Bayesian 8 15 80 5
regularization
0.07000 Scaled conjugate 22 12 15 80 5
gradient
0.07829 Levenberg–Marquardt 2 12 20 60 20
0.07748 Bayesian 4 10 20 60 20
regularization
0.08630 Scaled conjugate 2 8 20 60 20
gradient

the values of the empirical distribution function with the quantiles of hypothetical distribu-
tion functions in a scatter plot is a method for goodness-of-fit testing (Wang et al., 2023).
In this plot, if two-dimensional points are closer to the first quadrant, one cannot reject the
predictive distribution’s performance.
Based on the dispersion and range of attraction of each of the functions provided, the
distribution functions that have the best fit with the predicted values are presented in Table 9.
To prove the accuracy of the prediction, Table 9 is presented using the nonparametric Kol-
mogorov–Smirnov fit test. In this regard, the goodness of fit for different distribution functions
related to supply chain #24 is presented. Each distribution function has parameters. The accu-
racy of the prediction for the nonparametric Kolmogorov–Smirnov fit test is also presented,
which shows that the exponential distribution function has highest accuracy.

Property 3: This property is related to the attraction domain. Consider {εi }n1 a sequence of
independent
∞  with stable distribution and scale parameter 0 < α < 2. If
 αrandom variables
θ j  < ∞, ω  ∞ θ j ε j is defined. The distribution of ω should be stable
j−∞ j−∞
  α  α1
and ω
d θ j  ε, where d means equality in distribution.
α
Property 4: If ε ∼ S(α, 0, C1 , 0), then ψε (t)  eC1 |t| , where C1 is the dispersion of ε,
  α  α1    α 
(disp(ε)  C1 ). Also, if ω
d θ j  ε1 , then disp(ω)  C1 θ j  .

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

Fig. 6 Convergence in predicting the efficiency of supply chains

In Fig. 8, probability density function models are presented with respect to the predicted
empirical probability function of supply chain 24, in which classical functions and accuracy
rates are provided. In Fig. 8, the horizontal axis shows the empirical P value, and the vertical
axis shows the model P value. For each P value, the closeness of the predicted distribution
efficiency to the provided distribution is shown. As it is obvious, the Levy function has the
most repetitions in the simulation of the efficiency of supply chain #26.
In Table 10, the fluctuations of efficiency with respect to different distribution functions
for each supply chain are presented. This table indicates the amount of changes in efficiency
with respect to the changes in the prediction accuracy and which supply chain is ranked as the
best in these changes. The best supply chain with the least changes in supply chain efficiency
is number 9. Similarly, the largest fluctuation in efficiency prediction is related to supply
chain number 2, which has appropriate prediction accuracy despite the largest fluctuation.

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

Table 8 The accuracy of predicted


k r
efficiency εrk4 in terms of r and k
1 2 3

1 25 115 35
2 24 113 34
3 29 110 36
4 26 111 30
5 28 111 33
6 20 95 33
7 21 96 38
8 25 115 34
9 31 134 36
10 34 94 32
11 26 92 30
12 28 88 29
13 29 86 29
14 31 87 21
15 30 85 25
16 24 80 30
17 25 89 33
18 26 91 26
19 33 92 29
20 34 111 28
21 36 116 34
22 38 95 36
23 31 92 28
24 30 93 34
25 20 79 40
26 29 77 41
27 36 75 42
28 34 116 44

Based on Table 9 and Fig. 8, the best-fitting probability distributions could be Cauchy or
Levy. However, to choose between them, the Hill estimator will be used (Hill, 1975). This
estimator estimates the exponential parameter (α) of the stable distribution (we know that
when α  1, the distribution is Cauchy, and when α  0.5, the distribution is Levy). Let’s
assume that α̂ kH ill represents the estimator for α. Then,
⎛ ⎛ ⎞⎞−1
1 k
ln ε28
j
α̃ kH ill ⎝ ⎝ ⎠⎠ , ε (1) ≥ ε (2) ≥ . . . ≥ ε (k) (11)
(k+1) 28 28 28
k ln ε
j1 28

The values of α̂ kH ill against different values of k show us that as k approaches 9, α̂ kH ill
approaches 0.5. Table 11 demonstrates that as the value of k increases, the value of α

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

k with goodness of fit for three classical distributions


Fig. 7 Histogram of predicted efficiency values ε24

Table 9 Nonparametric Kolmogorov–Smirnov fit test for the accuracy of predicted efficiency of supply chain
#26

Distribution Parameters Kolmogorov–Smirnov test

Cauchy Scale  18.761, Location  114.09 0.3749


Exponential λ  0.00664 0.4489
Levy Scale  33.574, Location  32.18 0.4196
Lognormal σ  0.0313, μ  0.849 0.4321
Normal σ  0.0211, μ  0.794 0.348

approaches 0.5. Therefore, it can be assumed that the predictive performance follows an
asymmetric Levy distribution. However, by using Property 3, we can consider ε24
k in stan-

ε24
k −32.506
dard form 30.595 . Replacing the standard form with the original form will not lead to
incorrect results. In other words, based on the relationship α̃ kH ill , it can be assumed that for
each k  1,…,10, we have ε24
k ∼ S(0.5, 1, 1, 0). These results can also be found for ε k in
24
Fig. 8.
k ∼ S(0.5, 1, 1, 0) follows a standard stable distribution for
Now, let’s assume that ε24
k
all values of k. Therefore, the assumption of a Levy distribution will be meaningful (ε14, S,
ε20,
k
S , and ε24, S have distribution of S(0.5, 1, 1, 0)), and according to the former discussions,
k

sin π4 (0.5) ∗  1. If we continue


C π ≈ √1 .

As a result, according to Eq. (11), we have φ21
this procedure for other DMUs, the predicted efficiency results in Table 6 and the fluctuations

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

Fig. 8 The PP plot of probability density function models with respect to the predicted empirical probability
k
function of efficiency of ε24

presented in Table 10 will not be significantly different. Therefore, the predicted efficiency
is explainable.

4.2 Managerial implications

In this paper, explainable AI is used to explain the predictability of supply chain efficiency in
future periods. This approach helps decision-makers avoid being misled by the past perfor-
mance of units and make decisions based on the future. It also takes into account the growth
trends and progress rates of units in addition to their achieved efficiency when ranking them.
The obtained efficiency, predicted using data functions for 84 cycles, is presented in his-
togram form in Fig. 7. Based on the goodness-of-fit analysis for three classic distributions, it
appears to be most closely related to one of them. Therefore, the predicted data also has the
highest convergence with the previous classic distribution in forecasting efficiency.
Figure 9 displays the convergence of predicted efficiency for time periods 84 to 100.
It is explained in this graph that the predicted efficiency has the highest error in the 48th
time period and then has the highest convergence afterward. Therefore, based on Fig. 9, the
achieved efficiency appears to have good accuracy.
The same method is applied to 28 supply chains, as shown in Fig. 5, depicting this trend.
This trend remains consistent in the predictions, and these predictions are based on an explain-
able approach. This approach highlights that the impact of investment values associated with
green research and development as well as employee safety is more influential than the
efficiency of supply chains. A decrease in these investment values directly correlates with

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

Table 10 The efficiency


fluctuations DMUs Fluctuation Rank

1 0.04 11
2 0.075 28
3 0.066 24
4 0.016 2
5 0.039 10
6 0.049 15
7 0.058 19
8 0.054 17
9 0.014 1
10 0.052 16
11 0.048 14
12 0.02 4
13 0.018 3
14 0.044 13
15 0.069 26
16 0.029 7
17 0.042 12
18 0.055 18
19 0.035 9
20 0.025 6
21 0.064 23
22 0.059 20
23 0.061 21
24 0.063 22
25 0.022 5
26 0.033 8
27 0.071 27
28 0.067 25

Table 11 Hill estimator in identifying predictive distribution values

K 1 2 3 4 5 6 7 8 9

k
α̃Hill 1 0.988 0.924 0.81 0.757 0.699 0.634 0.523 0.514

reduced efficiency over time. The influence of research and development on outputs that
require rework is evident. Furthermore, not only is there an increase in these outputs, but
there is also a stronger emphasis on maximizing their utilization. Hence, this factor is depicted
as a positive element across the time periods in the supply chains.
It is essential to note that an increase in undesirable outputs significantly affects the
efficiency of DMUs in each time period. The functions for these criteria are referenced, and

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

Fig. 9 Convergence trend in predicting efficiency

the trends in the supply chain, such as increasing rework, point to a weakness in efficiency.
The same methodology is also applied to supply chain #26, with the subsequent time period
indicating that supply chain #21 will surpass supply chain #26. This is due to not only an
increase in indebtedness in future time periods but also a negative impact stemming from
changes in green research and development budgets. Therefore, supply chain #21 is expected
to outperform supply chain #26.
The approach presented in this paper not only identifies these factors for predicting trends
in supply chain efficiency but also enables the observation of the effects of investments made
in allocating research and development budgets as a dual-role factor in supply chain efficiency
for future time periods. The model presented, which considers the interplay between time
periods, offers managers the perspective to examine the nature of each of these factors in the
changes occurring in supply chain efficiency.

5 Conclusions and future research

In this section, we draw conclusions from our paper and provide guidance for future
researchers. In recent years, there has been a remarkable surge in the development and
widespread applications of cutting-edge digital technologies, such as AI, Blockchain, and
the Internet of Things (IoT). These technologies have garnered significant attention from
governments, industries, and the academic community due to their potential to revolution-
ize various aspects of modern life. One area where these technologies have demonstrated
great promise is in the realm of supply chain management, particularly in enhancing the
performance of industrial companies while adhering to sustainability criteria (Qi et al.,
2023). Supply chain management is a critical component of any industrial organization,
and optimizing its operations can lead to significant improvements in overall performance.
The integration of advanced digital technologies into supply chains has enabled companies
to operate more efficiently, reduce costs, and minimize their environmental impact. As a

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

result, many managers and decision-makers in industrial organizations have recognized the
potential benefits of incorporating these technologies into their supply chain strategies. In
this study, we proposed a novel network model specifically designed for sustainable supply
chains within the oil industry. Our approach leverages DEA, a widely recognized methodol-
ogy for evaluating the relative efficiency of DMUs, in this case, the components of a supply
chain. However, our model goes beyond traditional DEA by considering recurrent loops in
its design, a feature often present in real-world supply chains but not adequately addressed in
existing research. By accounting for these recurrent loops, our model aims to provide a more
accurate representation of the complex dynamics at play within supply chains. Moreover, our
proposed model is not limited to a static analysis; it is extended to a dynamic network DEA
model. This extension allows for the consideration of changes and fluctuations in supply
chain performance over time, an essential feature for adapting to evolving market conditions
and customer demands. By employing a dynamic approach, our model can offer valuable
insights into the long-term sustainability and efficiency of supply chains, which is crucial
in today’s rapidly changing business landscape. In addition to our network DEA model, we
have incorporated the use of an Explainable AI for the prediction of sustainable supply chain
performance. Explainable AI is a rapidly developing field that focuses on providing trans-
parent and interpretable AI models. By using an Explainable AI, we aim to bridge the gap
between the complex, data-driven nature of AI and the need for human decision-makers to
understand and trust the predictions made by these systems. This approach can significantly
enhance the acceptance and adoption of AI-based decision support tools in the context of
supply chain management.
This paper underscores the value of explainable AI in forecasting supply chain efficiency,
empowering decision-makers to base their decisions on future trends rather than solely on
past performance. By considering growth trends, progress rates, and achieved efficiency, this
approach enhances the accuracy of ranking units. The analysis of predicted efficiency trends,
particularly in relation to investment values in green research and development, highlights
their significant impact on supply chain performance over time. Managers can leverage these
insights to allocate resources effectively and anticipate future efficiencies, thereby optimizing
supply chain management strategies for sustained success.
We acknowledge the importance of comparative analysis in evaluating the effectiveness
of our methodology. Firstly, our approach stands out due to its incorporation of advanced
digital technologies, such as Blockchain and AI, which have been demonstrated to signifi-
cantly enhance performance improvement and operations optimization in industrial settings.
By leveraging these technologies, our study offers a more sophisticated framework for sus-
tainable supply chain management. Moreover, our model addresses the challenge of recursive
outputs in supply chain networks using a DEA approach. Unlike traditional methods, which
may overlook the dynamic nature of supply chain processes, our model accounts for recur-
rent loops by minimizing their generation and maximizing their utilization. This innovative
feature leads to more efficient resource allocation and decision-making within the supply
chain. Furthermore, we extend our network model to a dynamic DEA model, allowing for
the prediction and evaluation of supply chain performance across future time periods. This
predictive capability is augmented by an explainable AI, which enables managers to antici-
pate inefficiencies before they arise and take proactive measures to mitigate them. Lastly, our
approach facilitates a dual-role analysis of research and development investments, wherein
these expenditures are not only evaluated for their immediate impact on efficiency but also for
their long-term implications on supply chain dynamics. This holistic perspective empowers
managers to make informed decisions regarding resource allocation and strategic investments.

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

In summary, our proposed approach offers a comprehensive and forward-looking methodol-


ogy for sustainable supply chain management, leveraging advanced technologies, dynamic
modeling techniques, and dual-role analysis to enhance decision-making and performance
optimization.
Looking ahead, there are several promising avenues for further research in this domain.
One potential direction is the exploration of various network DEA models, including slack-
based measures and range-adjusted measures. Comparing the performance and suitability
of these models in the context of sustainable supply chains could offer valuable insights for
decision-makers. Furthermore, the incorporation of deep learning technology is another area
worth investigating. Deep learning, a subfield of AI, has demonstrated remarkable capabilities
in handling complex, non-linear relationships within data. Applying deep learning techniques
to the design and optimization of sustainable supply chains may open up new possibilities for
enhancing efficiency and sustainability, albeit at the potential cost of some interpretability.
Finally, in terms of methodology, our proposed DEA model operates as an envelopment
version, which means it does not produce coefficients (weights) for inputs and outputs. To
derive weights for inputs and outputs, it is necessary to construct a dual model of the proposed
model and then apply it to the dataset.
Acknowledgements Authors would like to appreciate two anonymous Reviewers for their insightful and
constructive comments and suggestions.

Declarations

Conflict of interest Authors declare that they have no conflict of interest.

Ethical approval This article does not contain any studies with human participants or animals performed by
any of the authors.

References
Aggarwal, A., & Toshniwal, D. (2021). A hybrid deep learning framework for urban air quality forecasting.
Journal of Cleaner Production, 329, 129660.
Ahmad, K., Younas, Z. I., Manzoor, W., & Safdar, N. (2023). Greenhouse gas emissions and corporate social
responsibility in the USA: A comprehensive study using dynamic panel model. Heliyon, 9(3), 13979.
Ancheyta, J. (2021). API-Barrel Yield: A new index for evaluating heavy oil upgrading technologies. Fuel,
294, 120476.
Azadi, E., Moghaddas, Z., Farzipoor Saen, R., Mardani, A., & Azadi, M. (2023a). Green supply chains and
performance evaluation: A multiplier network analytics model with common set of weights. Journal of
Cleaner Production, 411, 137377.
Azadi, M., Yousefi, S., Farzipoor Saen, R., Shabanpour, H., & Jabeen, F. (2023b). Forecasting sustainability of
healthcare supply chains using deep learning and network data envelopment analysis. Journal of Business
Research, 154, 113357.
Biguri, K. (2023). How does access to the unsecured debt market affect investment? Journal of Banking &
Finance, 152, 106856.
Bošković, I., & Radivojević, A. (2023). Life cycle greenhouse gas emissions of hemp-lime concrete wall
constructions in Serbia: The impact of carbon sequestration, transport, waste production and end of life
biogenic carbon emission. Journal of Building Engineering, 66, 105908.
Carbonneau, R., Laframboise, K., & Vahidov, R. (2008). Application of machine learning techniques for
supply chain demand forecasting. European Journal of Operational Research, 184(3), 1140–1154.
Cavdur, F., & Sebatli, A. (2019). A decision support tool for allocating temporary-disaster-response facilities.
Decision Support Systems, 127, 113145.
Chen, S., Ngai, E. W., Ku, Y., Xu, Z., Gou, X., & Zhang, C. (2023). Prediction of hotel booking cancellations:
Integration of machine learning and probability model based on interpretable feature interaction. Decision
Support Systems, 170, 113959.

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

Chen, Q., Wang, T., Tang, L., Zeng, Z., & Zhu, B. (2020). Study on the structure-activity relationship between
oil dewetting self-cleaning and surface morphology for crude oil pollution treatment and crude oil/water
separation. Journal of Environmental Chemical Engineering, 11(1), 109092.
Damoah, I. S., Ayakwah, A., & Tingbani, I. (2021). Artificial intelligence (AI)-enhanced medical drones in
the healthcare supply chain (HSC) for sustainability development: A case study. Journal of Cleaner
Production, 328, 129598.
De Graeve, F., & Mazzolini, G. (2023). The maturity composition of government debt: A comprehensive
database. European Economic Review, 154, 104438.
Demlehner, Q., Schoemer, D., & Laumer, S. (2021). How can artificial intelligence enhance car manufactur-
ing? A Delphi study-based identification and assessment of general use cases. International Journal of
Information Management, 58, 102317.
Dikmen, M., & Burns, C. (2022). The effects of domain knowledge on trust in explainable AI and task
performance: A case of peer-to-peer lending. International Journal of Human-Computer Studies, 162,
102792.
Ding, W., Abdel-Basset, M., Hawash, H., & Ali, A. M. (2022). Explainability of artificial intelligence methods,
applications and challenges: A comprehensive survey. Information Sciences, 615, 238–292.
Dubey, R., Bryde, D. J., Dwivedi, Y. K., Graham, G., & Foropon, C. (2022). Impact of artificial intelligence-
driven big data analytics culture on agility and resilience in humanitarian supply chain: A practice-based
view. International Journal of Production Economics, 250, 108618.
Esteve, V., & Prats, M. A. (2023). Testing explosive bubbles with time-varying volatility: The case of Spanish
public debt. Finance Research Letters, 51, 103330.
Färe, R., & Grosskopf, S. (1996). Productivity and intermediate products: A frontier approach. Economics
Letters, 50(1), 65–70.
Fathi, A., & Farzipoor Saen, R. F. (2018). A novel bidirectional network data envelopment analysis model
for evaluating sustainability of distributive supply chains of transport companies. Journal of Cleaner
Production, 184, 696–708.
Ghouri, A. M., Khan, H. R., Mani, V., ul Haq, M. A., & de Sousa Jabbour, A. B. L. (2023). An Artificial-
Intelligence-Based omnichannel blood supply chain: A pathway for sustainable development. Journal
of Business Research, 164, 113980.
Giri, B. K., & Roy, S. K. (2024). Fuzzy-random robust flexible programming on sustainable closed-loop
renewable energy supply chain. Applied Energy, 363, 123044.
Giri, B. K., Roy, S. K., & Deveci, M. (2023). Fuzzy robust flexible programming with Me measure for electric
sustainable supply chain. Applied Soft Computing, 145, 110614.
Goli, A., Aazami, A., & Jabbarzadeh, A. (2018). Accelerated cuckoo optimization algorithm for capacitated
vehicle routing problem in competitive conditions. International Journal of Artificial Intelligence, 16(1),
88–112.
Goli, A., Khademi, Z. H., Tavakkoli-Moghaddam, R., & Sadeghieh, A. (2019). Hybrid artificial intelligence
and robust optimization for a multi-objective product portfolio problem Case study: The dairy products
industry. Computers and Industrial Engineering, 137, 106090.
Goli, A., Khademi-Zare, H., Tavakkoli-Moghaddam, R., Sadeghieh, A., Sasanian, M., & Malekalipour,
Kordestanizadeh R. (2021). An integrated approach based on artificial intelligence and novel meta-
heuristic algorithms to predict demand for dairy products: A case study. Network Computation in Neural
Systems, 32(1), 1–35.
Goli, A., Moeini, E., Shafiee, A. M., Zamani, M., & Touti, E. (2020). Application of improved artificial intelli-
gence with runner-root meta-heuristic algorithm for dairy products industry: A case study. International
Journal on Artificial Intelligence Tools, 29(5), 2050008.
Grimes, G. M., Schuetzler, R. M., & Giboney, J. S. (2021). Mental models and expectation violations in
conversational AI interactions. Decision Support Systems, 144, 113515.
Gunning, D., & Aha, D. W. (2019). DARPA’s explainable artificial intelligence program. AI Magazine, 40(2),
44–58.
Hill, B. (1975). A simple general approach to inference about the tail of a distribution. The Annals of Statistics,
3(5), 1163–1174.
Hu, G., & Jiang, H. (2023). Time-varying jumps in China crude oil futures market impacted by COVID-19
pandemic. Resources Policy, 82, 103510.
Irarrázaval, M. E., Maldonado, S., Pérez, J., & Vairetti, C. (2021). Telecom traffic pumping analytics via
explainable data science. Decision Support Systems, 150, 113559.
Jauhar, S. K., Jani, S. M., Kamble, S. S., Pratap, S., Belhadi, A., & Gupta, S. (2023). How to use no-code artificial
intelligence to predict and minimize the inventory distortions for resilient supply chains. International
Journal of Production Research. https://doi.org/10.1080/00207543.2023.2166139

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

Joung, J., & Kim, H. (2023). Interpretable machine learning-based approach for customer segmentation for new
product development from online product reviews. International Journal of Information Management,
70, 102641.
Kadadevaramath, R. S., Chen, J. C., Shankar, B. L., & Rameshkumar, K. (2012). Application of particle
swarm intelligence algorithms in supply chain network architecture optimization. Expert Systems with
Applications, 39(11), 10160–10176.
Kao, C., & Hwang, S. N. (2008). Efficiency decomposition in two-stage data envelopment analysis: An
application to non-life insurance companies in Taiwan. European Journal of Operational Research,
185(1), 418–429.
Kim, D., Song, Y., Kim, S., Lee, S., Wu, Y., Shin, J., & Lee, D. (2023). How should the results of artificial intel-
ligence be explained to users?-Research on consumer preferences in user-centered explainable artificial
intelligence. Technological Forecasting and Social Change, 188, 122343.
Kraus, M., Feuerriegel, S., & Oztekin, A. (2020). Deep learning in business analytics and operations research:
Models, applications and managerial implications. European Journal of Operational Research, 281(3),
628–641.
Kumar, A., Mani, V., Jain, V., Gupta, H., & Venkatesh, V. G. (2023). Managing healthcare supply chain through
artificial intelligence (AI): A study of critical success factors. Computers & Industrial Engineering, 175,
108815.
Lee, D., Kim, M., Lee, B., Chae, S., Kwon, S., & Kang, S. (2022). Integrated explainable deep learning
prediction of harmful algal blooms. Technological Forecasting and Social Change, 185, 122046.
Liang, D., Cao, W., Zhang, Y., & Xu, Z. (2024). A two-stage classification approach for AI technical service
supplier selection based on multi-stakeholder concern. Information Sciences, 652, 119762.
Leichtmann, B., Humer, C., Hinterreiter, A., Streit, M., & Mara, M. (2023). Effects of Explainable Artificial
Intelligence on trust and human behavior in a high-risk decision task. Computers in Human Behavior,
139, 107539.
Michali, M., Emrouznejad, A., Dehnokhalaji, A., & Clegg, B. (2023). Subsampling bootstrap in network DEA.
European Journal of Operational Research, 305(2), 766–780.
Mondal, A., & Roy, S. K. (2021). Multi-objective sustainable opened-and closed-loop supply chain under
mixed uncertainty during COVID-19 pandemic situation. Computers & Industrial Engineering., 159,
107453.
Mondal, A., & Roy, S. K. (2022). Application of Choquet integral in interval type-2 Pythagorean fuzzy
sustainable supply chain management under risk. International Journal of Intelligent Systems, 37(1),
217–263.
Mondal, A., Giri, B. K., & Roy, S. K. (2023). An integrated sustainable bio-fuel and bio-energy supply chain: A
novel approach based on DEMATEL and fuzzy-random robust flexible programming with Me measure.
Applied Energy, 343, 121225.
Mondal, A., Giri, B. K., Roy, S. K., Deveci, M., & Pamucar, D. (2024). Sustainable-resilient-responsive
supply chain with demand prediction: An interval type-2 robust programming approach. Engineering
Applications of Artificial Intelligence, 133, 108133.
Naseri H. (2004). Linear prediction for electricity consumption with levy distribution. In: IWMS2004. In 13th
International Workshop on Matrices and Statistics, Poznan, Poland (pp. 18–21).
Nolan, J. P. (2016). Stable distributions, models for heavy-tailed data. American University.
Olabi, A. G., Abdelghafar, A. A., Maghrabie, H. M., Sayed, E. T., Rezk, H., Al Radi, M., Obaideen, K.,
& Abdelkareem, M. A. (2023). Application of artificial intelligence for prediction, optimization, and
control of thermal energy storage systems. Thermal Science and Engineering Progress, 39, 101730.
Pradhan, B., Dikshit, A., Lee, S., & Kim, H. (2023). An explainable AI (XAI) model for landslide susceptibility
modelling. Applied Soft Computing, 142, 110324.
Park, S., & Yang, J. S. (2022). Interpretable deep learning LSTM model for intelligent economic decision-
making. Knowledge-Based Systems, 248, 108907.
Paul, S. K. (2015). Supplier selection for managing supply risks in supply chain: A fuzzy approach. The
International Journal of Advanced Manufacturing Technology, 79, 657–664.
Poorzaker Arabani, S., & Ebrahimpour Komleh, H. (2019). The optimization of forecasting ATMs cash demand
of Iran banking network using LSTM deep recursive neural network. Journal of Operational Research
and Its Applications, 16(3), 69–88.
Qi, B., Shen, Y., & Xu, T. (2023). An artificial-intelligence-enabled sustainable supply chain model for B2C
E-commerce business in the international trade. Technological Forecasting and Social Change, 191,
122491.
Rai, A. (2020). Explainable AI: From black box to glass box. Journal of the Academy of Marketing Science,
48, 137–141.

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Annals of Operations Research

Ren, H., Yang, F., Li, C., & Deng, C. (2020). Controllable dewetting transition on graphene-based nanotextured
surfaces. Applied Surface Science, 520, 146374.
Saeed, W., & Omlin, C. (2023). Explainable AI (XAI): A systematic meta-survey of current challenges and
future opportunities. Knowledge-Based Systems, 263, 110273.
Sawarkar, A. N. (2019). Cavitation induced upgrading of heavy oil and bottom-of-the-barrel: A review. Ultra-
sonics Sonochemistry, 58, 10469.
Sezer, O. B., Gudelek, M. U., & Ozbayoglu, A. M. (2020). Financial time series forecasting with deep learning:
A systematic literature review: 2005–2019. Applied Soft Computing, 90, 106181.
Sezer, O. B., & Ozbayoglu, A. M. (2018). Algorithmic financial trading with deep convolutional neural
networks: Time series to image conversion approach. Applied Soft Computing, 70, 525–538.
Shin, D. (2021). The effects of explainability and causability on perception, trust, and acceptance: Implications
for explainable AI. International Journal of Human-Computer Studies, 146, 102551.
Siering, M. (2022). Explainability and fairness of RegTech for regulatory enforcement: Automated monitoring
of consumer complaints. Decision Support Systems, 158, 113782.
Wang, L., Gopal, R., Shankar, R., & Pancras, J. (2022). Forecasting venue popularity on location-based services
using interpretable machine learning. Production and Operations Management, 31(7), 2773–2788.
Wang, S., Jia, H., Lu, J., & Yang, D. (2023). Crude oil transportation route choices: A connectivity reliability-
based approach. Reliability Engineering & System Safety, 235, 109254.
Wang, B., Li, W., Bradlow, A., Bazuaye, E., & Chan, A. T. (2023). Improving triaging from primary care into
secondary care using heterogeneous data-driven hybrid machine learning. Decision Support Systems,
166, 113899.
Yagin, F. H., Cicek, İB., Alkhateeb, A., Yagin, B., Colak, C., Azzeh, M., & Akbulut, S. (2023). Explainable
artificial intelligence model for identifying COVID-19 gene biomarkers. Computers in Biology and
Medicine, 154, 106619.
Yousefi, S., Shabanpour, H., & Farzipoor Saen, R. (2021). Sustainable clustering of customers using capacitive
artificial neural networks: A case study in Pegah Distribution Company. RAIRO-Operations Research,
55(1), 51–60.
Yousefi, S., Soltani, R., Bonyadi Naeini, A., & Farzipoor Saen, R. (2019). A robust hybrid artificial neural
network double frontier data envelopment analysis approach for assessing sustainability of power plants
under uncertainty. Expert Systems, 36(5), 12435.
Yousefi, S., Soltani, R., Farzipoor Saen, R., & Pishvaee, M. S. (2017). A robust fuzzy possibilistic programming
for a new network GP-DEA model to evaluate sustainable supply chains. Journal of Cleaner Production,
166, 537–549.
Zhdanov, D., Bhattacharjee, S., & Bragin, M. A. (2022). Incorporating FAT and privacy aware AI modeling
approaches into business decision making frameworks. Decision Support Systems, 155, 113715.

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under
a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of such publishing agreement and applicable
law.

123
Content courtesy of Springer Nature, terms of use apply. Rights reserved.
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center
GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers
and authorised users (“Users”), for small-scale personal, non-commercial use provided that all
copyright, trade and service marks and other proprietary notices are maintained. By accessing,
sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of
use (“Terms”). For these purposes, Springer Nature considers academic use (by researchers and
students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and
conditions, a relevant site licence or a personal subscription. These Terms will prevail over any
conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription (to
the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of
the Creative Commons license used will apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may
also use these personal data internally within ResearchGate and Springer Nature and as agreed share
it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not otherwise
disclose your personal data outside the ResearchGate or the Springer Nature group of companies
unless we have your permission as detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial
use, it is important to note that Users may not:

1. use such content for the purpose of providing other users with access on a regular or large scale
basis or as a means to circumvent access control;
2. use such content where to do so would be considered a criminal or statutory offence in any
jurisdiction, or gives rise to civil liability, or is otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association
unless explicitly agreed to by Springer Nature in writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a
systematic database of Springer Nature journal content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a
product or service that creates revenue, royalties, rent or income from our content or its inclusion as
part of a paid for service or for other commercial gain. Springer Nature journal content cannot be
used for inter-library loans and librarians may not upload Springer Nature journal content on a large
scale into their, or any other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not
obligated to publish any information or content on this website and may remove it or features or
functionality at our sole discretion, at any time with or without notice. Springer Nature may revoke
this licence to you at any time and remove access to any copies of the Springer Nature journal content
which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or
guarantees to Users, either express or implied with respect to the Springer nature journal content and
all parties disclaim and waive any implied warranties or warranties imposed by law, including
merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published
by Springer Nature that may be licensed from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a
regular basis or in any other manner not expressly permitted by these Terms, please contact Springer
Nature at

onlineservice@springernature.com

You might also like