See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/365114260

The Role Of AI In Warehouse Digital Twins

Conference Paper · November 2022

DOI: 10.46354/i3m.2022.emss.024

CITATIONS READS
0 235

6 authors, including:

Adnane Drissi Elbouzidi Abdessamad Ait El Cadi

Ecole Nationale Supérieure d'Arts et Métiers Université Polytechnique Hauts-de-France
1 PUBLICATION 0 CITATIONS 57 PUBLICATIONS 244 CITATIONS

SEE PROFILE SEE PROFILE

Robert Pellerin Samir Lamouri

Polytechnique Montréal Ecole Nationale Supérieure d'Arts et Métiers
212 PUBLICATIONS 4,335 CITATIONS 153 PUBLICATIONS 3,179 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Study of Integration and coupling schemes of simulation and optimization. View project

Contextualization and Conceptualization of Collaborative Planning for Business Transformation Projects View project

All content following this page was uploaded by Robert Pellerin on 04 November 2022.

The user has requested enhancement of the downloaded file.

 Proceedings of the 34th European Modeling & Simulation Symposium (EMSS), 024
19th International Multidisciplinary Modeling & Simulation Multiconference

ISSN 2724-0029 © 2022 The Authors.

doi: 10.46354/i3m.2022.emss.024

The Role Of AI In Warehouse Digital Twins

Adnane Drissi Elbouzidi1,*, Marie-Jane Bélanger3, Abdessamad Ait el Cadi2,
Robert Pellerin3, Samir Lamouri1, Estefania Tobon Valencia4
1
LAMIH, Arts et Métiers ParisTech, 51 Bd de l'Hôpital, Paris, 75013, France
2
Univ. Polytechnique Hauts-de-France, LAMIH, CNRS, UMR 8201, F-59313 Valenciennes, France / INSA Hauts-
de-France, F-59313 Valenciennes, France
3
Polytechnique Montreal, 2500 Chemin de Polytechnique, Montréal, H3T 1J4, Canada
4
Groupe Square cabinet Flow&Co, Square Research Center, 173 Avenue Achille Peretti, Neuilly-sur-Seine, 92200,
France
*Corresponding author. Email address: [email protected]

Abstract
In the era of Industry 4.0, digital twins are at a pivotal phase. For a concept that is so inconsistently defined in the literature, it
has been used for many applications, especially in manufacturing, production, and operations. DT not only allows for
supervision and running simulations, but it also supports AI applications since it is mapped to all types of data and Intel on the
physical object. On the other hand, warehouses have been subject to little digitization over the years. Warehouse management
is at the very core of both manufacturing and retail operations, ensuring supply chain and production continuity. It is also a
conjunction of uncertain material handling activities. It could easily benefit from the Information visibility and the smart
features supplied by digital twins and machine learning. In this perspective, this paper examines the use cases of warehouse
digital twins (WDT). This study aims to assess the maturity of AI application within WDT, namely techniques, objectives, and
challenges. Consequently, inconsistencies are identified and research gaps are presented, making way for future development
and innovation.

Keywords: Digital twins; warehouse; material handling; artificial Intelligence, machine learning

for same-day delivery. They are also more demanding

1. Introduction of product personalization and ethical and
environmental beliefs. This requires a dynamic,
Intralogistics is at the center of both manufacturing straightforward, and completely transparent and
efficiency and customer satisfaction. Material visible process to one of the most unpredictable
handling could amount from 15% to 70% of activities, flawed by supply chain structure change,
production costs. In-store warehouse activities are demand seasonality, transportation cost change
also among the most hazardous industrial processes, trends, and many other variabilities (Gong & de
causing up to 50% of all kinds of industrial injuries Koster, 2011). Digital twin (DT) is one of the emerging
(Glatt et al., 2021). In a “post covid” world, companies technologies gaining popularity over the last decade
had to adapt their strategies to stay financially above and having the potential to better warehouse
water and satisfy impulsive and unheard customer management. The concept has stirred a lot of conflict
behavior. According to a survey done by LaserShip, in academia and industry as to how it should be
more than 60% of consumers are willing to pay more defined. DT should ensure an ever-evolving, fully

© 2022 The Authors. This article is an open access article distributed under the terms and conditions of the
Creative Commons Attribution (CC BY-NC-ND) license (https://creativecommons.org/licenses/by-nc-nd/4.0/).
 34th European Modeling & Simulation Symposium, EMSS 2022

connected, and reasonably identical entities linked

through IoT and allowing data analytics and
simulation. The digital twin has "smart" capabilities
allowing it to not only follow all the changes
happening in the warehouse but also be the one
suggesting modifications. Artificial Intelligence (AI)
proposes a set of methods that can aid decision
making, if not do it autonomously, based on collected
information. Since DTs have undergone the trouble of
collecting data and processing it, it only makes sense
to go the extra mile and use advanced data analytics to
grant it knowledge if not wisdom.
This paper aims to assess the maturity of AI
application within the DT paradigm in warehouse
management. Through a systematic review of
scientific literature, we attempt to answer the
following research questions: What are the AI
techniques most used for Warehouse management
under the DT paradigm? How is AI employed to ensure Figure 1. Research strategy
and elevate WDT functions? What are the challenges
and barriers to adopting both WDT and AI in 3.1. Digital twin
warehouses?
DTs are comprehensive digital representations of the
To do so the rest of the paper is organized as physical assets, comprising their design and
follows. Section 2 showcases the research configuration, state and behavior (Hribernik et al.,
methodology followed by section 3 that presents and 2021). Fuller et al. (2020) provided a more grounded
details the analysis framework. Section 4 presents and interpretation of the concept. They defined DT as the
discusses the findings of the review. Lastly, Section 5 effortless integration of data between a physical and
presents conclusions and open research, with overall virtual machine in either direction. Instead of focusing
challenges and perspectives to conclude the paper. on the definition, Fuller et al. made the distinction of
what is not a DT and why. This led to identification of
2. Research methodology the following:

This state of the art focuses on papers that have 1) Digital Model: There is no automatic data
covered the use of both AI and DT technologies to exchange between the physical model and digital
optimize in-store warehouse activities. We have model. This means once the simulation is created, a
adopted the machine learning categories described by change made to the physical object has no impact.
Usuga Cadavid et al. (2019) to construct our research 2) Digital Shadow: A digital shadow is a digital
query as follows: model with a one-way flow between the physical and
TITLE-ABS-KEY ( "digital twin" OR "digital twins" ) AND digital objects. A change in the state of the physical
TITLE-ABS-KEY ( "warehouse" OR "warehousing" OR object leads to a change in the digital representation
"material handling" OR "inventory" OR "packing" OR but not vice versa.
"store" OR "storage" ) AND TITLE-ABS-KEY ( "deep
learning" OR "artificial intelligence" OR "machine 3) Digital Twin: the data flow is bidirectional. A
learning" OR "AI" OR "ML" OR "neural networks" OR change made to the physical object automatically
"regression" OR "clustering" OR "sarsa" OR "nearest changes the digital object and vice versa.
neighbors" OR "Q-learning" OR "decision tree" )
In order to further differentiate the maturity level
We’ve studied the results through the phases of the studied papers we also identified the
detailed in figure 1. An additional reference was added characteristics of DTs, detailed in the literature. Zheng
upon analysis of the selected articles. The selected et al. (2021) synthetized and identified the
references either display a fully embedded AI and characteristics of cognitive digital twins as DT-based,
digital twin framework or explain the possible Cognition, Full lifecycle management, Autonomy
relationships between both technologies and how one capability, and Continuous evolving. On the other
could exploit the other. hand, Hribernik et al. (2021) described digital twins as
autonomous, context-aware, and adaptive. We have
3. Analysis framework identified DT characteristics relevant to our study,
which are a conjunction of the two visions presented
Our analysis differentiates what qualifies as a DT previously. Through our research, we’ve identified
and how it was used in collaboration with AI. We have how AI could be incorporated. Each characteristic is
also taken into consideration the data used in both AI defined as follows:
and DT to represent warehousing activities.
Context-awareness (CA) is the ability to distinguish
 Drissi Elbouzidi et al.

incoming stimuli clearly. It does not stop at the use of warehouses in the virtual space. Almost every feature
IoT and Information systems (IS), but also the way or application of AI and DTs can be traced back to data.
they are utilized. It is one of the key components to enable real time
connection, contextual information, and training the
Autonomy (Auto) is the DT’s ability to conduct
AI algorithms … It is also distinctively used depending
autonomous activities without human assistance or a
on the application, availability and the source.
minimum level of human intervention.
According to Leung et al. (2022), three different types
Continuous evolving (CE) is the ability to grow with of data, sourced from manufacturing information
the real system along the entire lifecycle. DT updates systems, IoT or manually entered are collected in a
itself according to the change of relevant data, warehouse:
information, and knowledge from the real system and
• Environmental data: include temperature,
all other interconnected software.
humidity, and light intensity ... These data could
Full lifecycle management (FLM) allows the model be useful in decision making, depending on the
to cover different phases across the entire lifecycle of types of goods stored in the warehouse.
the system, including the beginning of life (BOL, e.g.,
• Product data: include the inventory levels and
design, building, testing), middle of life (MOL, e.g.,
storage locations of goods. For instance, Radio
operating, usage, maintenance) and end-of-life (EOL,
Frequency Identification (RFID) technologies can
e.g., disassembly, recycling, remanufacturing).
be used to keep track of storage locations and
quantities linked to the warehouse management
3.2. Artificial Intelligence (AI) system (WMS) and then the DT for replenishment
AI was originated in 1956 at the Dartmouth Summer and stock-keeping purposes.
Research Project on Artificial Intelligence and had a • Handler data: include data related to both
tardy and rough evolution. To this day, there is still no equipment and workers, such as their real-time
consensus on what is properly considered as AI. locations. It can be data collected from workers'
According to Mehmood et al. (2019) the AI methods handheld devices given to workers to track their
include Analytic Hierarchy Process (AHP), fuzzy logic locations or measure some other physical
(FL), Genetic algorithms, Neural network (NN), and variable.
Simulated Annealing (SA). Modern subsections of AI
include machine learning (ML) and deep learning It is also important to take into consideration data
(DL). The three terms are often used interchangeably sources (WMS, IoT …) so as to judge the level of DT
in the literature. DL generally refers to deep artificial Interoperability and connectivity to other software.
neural networks and sometimes deep reinforcement
learning, both are primarily ML techniques; hence it is 3.4. Warehouse activities
considered a specific type of ML. Many developers of
In the scope of this study, we have distinguished two
AI systems now recognize that it can be far easier to
types of warehouses : Mechanical warehouses with
train a system by showing it examples of desired
automatic systems and mechanical tools that require
input-output behavior than to program it manually by
minimal human intervention (e.g., conveyors, stacker
anticipating the desired response for all possible
crane …) and non-mechanical warehouses in which
inputs (M. I. Jordan & T. M. Mitchell, 2015).
operators fulfil all processes through manual labor or
ML can be primarily categorized into three types. The while using machines and lifting tools (e.g. AGVs,
main discrepancy between these types lies in whether Forklifts …).
they are trained with labeled datasets or not (M. I.
Warehousing activities depend on the product,
Jordan & T. M. Mitchell, 2015):
industrial sector and the type of the warehouse itself.
a) Supervised learning (SL): The training set provides We have chosen to use Gong & de Koster's (2011)
clearly distinguished input features X and the adaptation of warehouse logistics. The general
corresponding output labels Y. material flow of In-house logistics goes through the
following activities: product/order arrival, put-away
b) Unsupervised learning (UL): the algorithm is
and preparation for storage, storage, order picking,
provided only input features X. It is up to the model to
and preparations for shipping (packaging,
classify all data in the sample space using techniques
accumulation, sortation) and shipping.
such as cluster analysis.
c) Reinforcement learning (RL): The learner performs 4. Results
a specific action in an interactive environment. Based
on experiments, the program can be rewarded or In the analyzed literature, AI applications in WDTs
punished. The goal is to obtain the maximum range from processes, strictly done in a warehouse, to
cumulative reward value through trial and error. warehouse-related activities in other fields such as
asset management and synthetic sensing (Bányai et
3.3. Data al., 2019; Corneli et al., 2019; Minerva et al., 2021;
Zacharaki et al., 2021). The papers studied here
This section details how data is used to describe present a mix of both manual and automatic stores
 34th European Modeling & Simulation Symposium, EMSS 2022

which proves that we don’t need a highly automated, learning, namely deep neural networks, for it presents
fully mechanical system to make a DT or use AI. Table more computational power and is very compatible
1 summarizes the finding of this literature review. with the emergence of big data (Corneli et al., 2019;
Kegenbekov & Jackson, 2021; Melesse et al., 2022;
4.1. What are the AI techniques most used for Zhan et al., 2022). These algorithms were either
Warehouse management under the DT approved through tests and are envisioned to be used
paradigm? in a DT framework (Hayward, 2019; Kegenbekov &
Jackson, 2021; Melesse et al., 2022; Xiuyu & Tianyi,
2018; Bányai et al., 2019) or are already applied in a
4.1.1. Artificial Intelligence
case study presenting a united DT/AI embedded
ML has been used either for classification or system.
forecasting. For classification, neural networks (NN)
are abundantly exploited. Melesse et al. (2022) used 4.1.2. Data
convolutional NN to analyze thermal images of
bananas to monitor fruit freshness in stores. Zhan et Except for Kegenbekov & Jackson, (2021) and Bányai et
al. (2022) used a sparse autoencoder to differentiate al. (2019), all DTs and AI/ML algorithms were trained
abnormal stationary in cold storage warehouses. using real product data. They are also the only ones
Corneli et al. and Hayward (2019) both used YOLOv2 that discussed the disadvantages of running AI
for object detection, allowing inventory and asset algorithms in a simulate environment. Using
inspection applications in buildings. Wu et al. (2022) simulated data and environments might lead to
withdrew by long short-term memory (LSTM) unexpected results. RL is known for finding loop holes
network (deep learning) dependencies inside time in virtual models, in this case, it terminated ordering
series of the received signal to estimate locations additional inventory closer to the end of the
online. In the same manner, Zhao et al. (2021) used the simulation to minimize holding costs.
same resource with k-nearest neighbor to make Handler data is primarily used for online location
indoor location estimation for warehouse safety tracking. Environmental data, though heavily talked
management. Bányai et al. (2019) utilized Black Hole about in the literature, is not subject to much
Optimization-Based Clustering to group the available application. Temperature plays a heavy role in safety
supply demands based on time frame related objective monitoring (Zhan et al., 2022). Minerva et al. (2021)
function. states that we can exploit all data types for object
On the other hand, forecasting applications provide identification through AI. The DT develops auditory
a wider range of ML technics. Neural networks still and visual signatures based on all kinds of data
take the lead, having been used trice. Leung et al. collected to identify changes in an environment and
(2022) used a neuro-fuzzy model (a combination of act accordingly.
NN and fuzzy logic) to forecast the future arrivals of A common goal among all DT adaptations in
PI-SKUs. Xiuyu & Tianyi (2018) used backpropagation literature is to access information in real-time, this is
NN to make sales predictions. Other ML technics often not the case for real life implementations. Real-
comprise Gradient-boosting decision tree (GDBT) to time monitoring is necessary for safety applications
monitor anomaly detection and maintenance (Huang requiring immediate responses (Zhan et al., 2022).
et al., 2021), proximal policy optimization (PPO) to Leng et al. (2021) and Leung et al. (2022) opted for
make inventory predictions (Kegenbekov & Jackson, more realistic, periodic, and synchronized data
2021). Wang et al. (2020) used a combination of both updates, which are both efficient and effective for
time-weighted linear regression method (TWMLR) their applications. This proves that real-time data
and Non-dominated Sorting Genetic Algorithm acquisition is not always necessary depending on the
(NSGA-II) to predict the remaining time of processes level of abstraction and the objective.
and find an optimal allocation of trolleys for the
material handling tasks. DT could curate multiple 4.2. How to ensure DT characteristics through AI ?
algorithms in the same platform and subject problems
to this set of optimization tools. Gao et al. (2022) There is no standard mold for a DT. Before the
implemented multiple ML techniques, in the clarifications of Kritzinger et al. (2018) and Fuller et al.
“algorithm center”. The appropriate algorithm is (2020), the term DT was used interchangeably with
selected to match the problem from the Algorithm simulation or cyber-physical systems. Even now, this
Center. This architecture thus allows for multiple is still the case for some studies that reduce DT’s
adaptations and solutions, which ensures optimal potential to simply be a CAD model. AI Programs are
problem-solving. becoming fundamental to the proposed models and
frameworks of DTs. Almost all the studied papers
SL is the most used ML method having been used in developed models and architectures that use ML in the
twelve papers, either on its own or combined with DT platform or cloud. Otherwise, these algorithms are
other types of learning. UL was used trice, while RL closely linked to the virtual counterpart in order to
was only used once. There’s also a big interest in deep
 Drissi Elbouzidi et al.

Table 1. Overview of literature regarding digital twins and related concepts

Other AI
ML Level of DT DT characteristics Data types Data source Warehouse activities
technics
Manuel Put
SL UL RL FL GA DM DS DT CA FLM Auto CE Nature ED PD HD IS IoT Arrival Storage Picking Preparation Shipping
input away
Melesse et al.
X X X (X) (X) R X X X X X
(2022)
Pan et al. (2022) X (X) (X)
Leung et al. (2022) X X X X S X X X X X X X X
Huang et al. (2021) X X X R X X
Leng et al. (2021) X X (X) R X X X X X
Kegenbekov et al.
X (X) X S X X
(2021)
Zhan et al. (2022) X X X (X) X R X X X X X X X
Zacharaki et al.
X
(2021)
Minerva et al.
X X X X
(2021)
Sacks et al. (2020) X
Hayward & X
X R X X X X
Portugal (2019) X
Corneli et al. (2019) X X R X X X X X X
Xiuyu & Tianyi
X (X) R X X X X
(2018)
Wang et al. (2020) X X X X X R X X X X X
Gao et al. (2022) X X X X X X R X X X X X X
Wu et al. (2022) X X X X X R X X X X
Zhao et al. (2021) X X X X R X X X X X X
Bányai et al. (2019) X X X X S X X X X X X
DM, DS: digital model, shadow, FL: Fuzzy logic, GA: genetic algorithms, CA: context awareness, FLM: full lifecycle management, Auto: autonomy,
CE: continuous evolving, ED, PD, HD: environmental, product and handler data, R: real, S: simulated, IS: Information system,
information systems (Leng et al., 2021; Melesse et al.,
compare the results of the AI program with the
2022; Wang et al., 2020; Zhan et al., 2022). Through AI
expected status recorded in the DT’s database (Corneli
DT are an even more pertinent, perceptive, and to a
et al., 2019; Hayward, 2019).
degree autonomous decision-making systems, in the
sense that it could even exceed human perception in
4.2.1. Context awareness finding optimal solutions. But when it comes to acting
AI objectives, namely classification, and forecasting, on solutions, the literature is still lacking. Hribernik et
go beyond data collection by IoT and structuring. They al. (2021) defined autonomy from an architecture
were used for the elaboration of contextual point view, where the DT is made of multiple
information that DTs act on. Melesse et al. (2022) autonomous, interconnected, singular sub-systems.
used image classification to define banana quality. This usually influences the DT's architecture to make
Depending on the results, the DT can decide on the it able to access data singularly if need be.
actions to be taken moving forward: preserve, put on
sale or donate the goods before they are borderline 4.2.3. Continuous evolving
inconsumable. Leung et al. (2022) used machine
Warehouses are fast-evolving environments, prone to
learning to predict order arrivals within the hour,
daily if not hourly drastic changes. DTs need to adapt
allowing the system to organize the total inbound
to these changes and allow an accurate representation
synchronization strategy accordingly. RL is by nature
regardless of the degree of abstraction. The structure
to a degree context-aware. The set of rules and
of the DT should allow for a continuously evolving
interactions from the environment makes for a
adaptation of the system so as not to make false
predefined interactive, situational setting. Neural
predictions and flawed hypotheses. This is still not
networks were used to detect abnormal stationary,
very explored in the literature. When it comes to
regarded as dangerous when warehouse operations
adaptive DTs, collected data should be reintroduced
requiring urgent interventions, from false alarms.
into the cloud to reevaluate the virtual counterpart
Wang et al., (2020) used AI to predict the remaining
(Huang et al., 2021; Zhan et al., 2022). Huang et al.
processing times. The results are compared to the
(2021) created two data streams, one that goes to the
current status to optimize the material handling
edge layer which allows high performance data
method proactively.
processing and deployment of the ML algorithm (real
time application) and the second goes through the
4.2.2. Autonomy edge for preprocessing and feature extraction and
When it comes to autonomy, DT adaptations still then to the cloud to reevaluate and update the model
require (and depend on) human presence in the loop to when necessary. This also makes the DT aware of the
procure modifications in the physical system (by contextual changes in the warehouse. Zhan et al.
requesting a change or through notifications). (2022) stated having designed an online self-adapting
Autonomous control and data acquisition has been mechanism to ensure that the model is in line with the
environmental changes of the warehouse. Zhao et al.
facilitated through IoT and the interconnectivity of the (2021) used a closed-loop structure that permanently
DT with other software in the cloud, namely with
 34th European Modeling & Simulation Symposium, EMSS 2022

updates the datasets and regenerates the programs AI. For the most part, ML algorithms are black box
following gene structure. From this perspective, the models that we do not fully comprehend the workings
algorithms are self-conscious and self-modifying. of. This can make the technic untrustworthy.
However, this might actually help us discover patterns
4.2.4. Full lifecycle management that we did not consider in the first place. Pan et al.
(2022) discussed the field of data-centric engineering,
None of the applications take into consideration the leveraging the best of both physics, simulation, and
full lifecycle of the warehouses. The scope of the data science. This helps ground AI and makes a little
studies analyzed is mainly focused on the middle of more predictable. AI has also proven Itself
life. Zacharaki et al. (2021) present through the consequential to identifying contextual information
RECLAIM project a framework that covers the whole and continuous evolving. Another potential research
lifecycle of the equipment. Their objective is to gap is the use of continual learning (CL) in a DT
elongate through DTs the MOF. During that time and framework (Hashash et al., 2022). CL is in fact a
through data collection and analysis, we can prescribe branch of machine learning representing the
refurbishment and remanufacturing actions on the capability of a model to continue evolving and
machines, allowing to restore the functionality of the regenerating from a data stream. this is also a degree
product or a part of it to an “as-new” state and of autonomy since it no longer requires human
optimize EOL. intervention. Nevertheless, autonomy is still poorly
covered In scientific literature. Having autonomous
5. Discussion and conclusion systems presents more safety hazards in the workshop
since we lose control over what and when actions
ML technics have been highly applied in this
happen.
literature review. Few articles that have not used AI
explicitly still consider it as an important part of their DTs have been associated with product life cycle
future studies (Pan et al., 2022; Zacharaki et al., 2021). management since the very beginning. The first time
AI can help DTs reach maturity and wisdom the notion was ever explicitly used is by Grieves,
throughout the entire product lifecycle by playing two during a PLM lecture (CoBuilder, 2018). None of the
roles : papers covered the entire lifecycle generally or focused
on the end of life. None of the applications use
• Reconstruction: AI can be an important tool
artificial intelligence or simulation to model the
for the reconstruction process; the process of
beginning of life of a warehouse, revamping and
creating and revisiting the virtual
redesigning the building, or discuss what would
representation based on the raw data from the
become of the warehouse by the end of its life.
sensors;
IoT aside, warehouse operations can easily reach a
• Application: Once the Digital Twin is
huge amount of records and data used for
reconstructed, another AI algorithm can be
management or assurance purposes. Tufano et al.
applied to the semantically rich
(2022) used machine learning models to predict
representation of the Digital Twin to support
warehouse components design based on data and
the business goals (Slama, 2021).
metrics collected through the life on a another storage
Intralogistics are lightly covered in research system. The algorithms could be used to assess the
relating to WDTs. None of the papers selected worked effectiveness of the current, up and running
on optimizing package preparation (packaging, warehouse to either duplicate or avoid making similar
accumulation, sorting …) and shipping. These mistakes. Which brings us back to the ultimate
activities present a research gap related to co-packing, research question of all time : What came first, the
prospective package preparations, testing the quality digital twin or the physical system?
of packages based on clients appreciation, and finding
If we are to consider digital twins as a simulation
links between the packaging and the shipping method.
based concept, then making of the digital twin doesn’t
DT has often been described as the perfect replica, necessarily need the physical counterpart. It is
being able to copy every change in the physical twin irrelevant whether the real counterpart already exists
and anticipate it. This futuristic ambition is not very in the physical world or is about to exist.(
doable because we can neither have real-time “Gesellschaft für Informatik (GI): Digitaler Zwilling”).
connectivity yet (Fuller et al., 2020), nor can we Prior simulation will test if the soon to be warehouse
capture all the states and minimal changes of a can actually handle the stock, if resources are enough
system. Such declarations presume that we have for ramp up, to test if we can get away with a
sensors everywhere to capture every shift in the air traditional manual warehouse or if we ought to invest
and omits connectivity and simulation times. This is at in a mechanical one … All of these previous attributes
once impossible, costly, and unnecessary. With a are to be consolidated and evaluated, If not changed as
certain level of abstraction and operational both twins grow and evolve. Figure 2 showcases the
synchronization, it should be possible to effectively warehouse twins evolution through time. This model
design and manage a stochastic WDT. It is important was inspired by Sacks et al. (2020) representation of
to keep realistic expectations when discussing DTs and the lifecycle of the twins for building construction.
 Drissi Elbouzidi et al.

The model ensures information saving and visibility in evolution detailed in figure 2; From a digital model to
a structured and evolutive configuration. Our vision a digital shadow and lastly, a twin. The concept is to be
for WDT will chronologically follow the natural constructed based on simulation software allowing us

Figure 2. Full lifecycle management of WDT based on (Sacks et al., 2020)

https://doi.org/10.22260/ISARC2019/0045
to control the environment and run experiments in the
first place, import datasets, and use machine learning Fuller, A., Fan, Z., Day, C., & Barlow, C. (2020). Digital
based optimization and forecasts depending on the Twin: Enabling Technologies, Challenges and Open
problematic at hand secondly. Also depending on the Research. IEEE Access, 8, 108952–108971.
problem, we need to define the degree of https://doi.org/10.1109/ACCESS.2020.2998358
synchronization required for the optimization to make
Gao, Y., Chang, D., Chen, C.-H., & Xu, Z. (2022). Design
sense. This way the “digital shadow” will be equipped
of digital twin applications in automated storage
with automatic data exchange and become a twin.
yard scheduling. Advanced Engineering Informatics,
This literature review showcase the ever growing 51, 101477.
potential of both AI and DT to optimize warehousing. https://doi.org/https://doi.org/10.1016/j.aei.2021.10
The algorithms are utilized in different ways to ensure 1477
WDT characteristics and potential. And on the other
Glatt, M., Sinnwell, C., Yi, L., Donohoe, S., Ravani, B., &
hand DT presents a mature and sophisticated and
Aurich, J. C. (2021). Modeling and implementation
interactive modeling environment for AI.
of a digital twin of material flows based on physics
simulation. Journal of Manufacturing Systems, 58,
References 231–245.
Bányai, Á., Illés, B., Glistau, E., Coello Machado, N. I., https://doi.org/10.1016/j.jmsy.2020.04.015
Tamás, P., Manzoor, F., & Bányai, T. (2019). Smart Gong, Y., & de Koster, R. B. M. (2011). A review on
cyber-physical manufacturing: Extended and real- stochastic models and analysis of warehouse
time optimization of logistics resources in matrix operations. Logistics Research, 3(4), 191–205.
production. Applied Sciences (Switzerland), 9(7). https://doi.org/10.1007/s12159-011-0057-6
https://doi.org/10.3390/app9071287
Hashash, O., Chaccour, C., & Saad, W. (2022). Edge
CoBuilder, 2018. The ‘digital Twin’ – A Bridge Continual Learning for Dynamic Digital Twins over
Between the Physical and the Dig-ital World. Wireless Networks.
https://cobuilder.com/en/the-digital-twin-a- https://doi.org/10.48550/ARXIV.2204.04795
bridge-between-the-physical-and-the-digital-
world/. Hayward, N. (2019). Machine Learning Image Analysis
for Asset Inspection. SPE Offshore Europe
Corneli, A., Naticchia, B., Cabonari, A., & Bosché, F. Conference and Exhibition, Aberdeen, UK.
(2019, May 24). Augmented Reality and Deep
Learning towards the Management of Secondary https://doi.org/10.2118/195773-MS
Building Assets. Hribernik, K., Cabri, G., Mandreoli, F., & Mentzas, G.
 8 | 34th European Modeling & Simulation Symposium, EMSS 2022

(2021). Autonomous, context-aware, adaptive 117271. https://doi.org/10.1016/j.ces.2021.117271

Digital Twins—State of the art and roadmap.
Sacks, R., Brilakis, I., Pikas, E., Xie, H. S., & Girolami,
Computers in Industry, 133, 103508.
M. (2020). Construction with digital twin
https://doi.org/10.1016/j.compind.2021.103508
information systems. Data-Centric Engineering, 1.
Huang, H., Yang, L., Wang, Y., Xu, X., & Lu, Y. (2021). https://doi.org/10.1017/dce.2020.16
Digital Twin-driven online anomaly detection for
Tufano, A., Accorsi, R., & Manzini, R. (2022). A
an automation system based on edge intelligence.
machine learning approach for predictive
Journal of Manufacturing Systems, 59, 138–150.
warehouse design. International Journal of Advanced
https://doi.org/10.1016/j.jmsy.2021.02.010
Manufacturing Technology, 119(3–4), 2369–2392.
Kegenbekov, Z., & Jackson, I. (2021). Adaptive supply https://doi.org/10.1007/s00170-021-08035-w
chain: Demand-supply synchronization using deep
Usuga Cadavid, J. P., Lamouri, * Samir, Grabot, B., &
reinforcement learning. Algorithms, 14(8).
Fortin, A. (2019). Machine Learning in Production
https://doi.org/10.3390/a14080240
Planning and Control: A Review of Empirical
Kritzinger, W., Karner, M., Traar, G., Henjes, J., & Sihn, Literature.
W. (2018). Digital Twin in manufacturing: A
Wang, W., Zhang, Y., & Zhong, R. Y. (2020). A proactive
categorical literature review and classification. 51(11),
material handling method for CPS enabled shop-
1016–1022.
floor. Robotics and Computer-Integrated
https://doi.org/10.1016/j.ifacol.2018.08.474
Manufacturing, 61.
Leng, J., Yan, D., Liu, Q., Zhang, H., Zhao, G., Wei, L., https://doi.org/10.1016/j.rcim.2019.101849
Zhang, D., Yu, A., & Chen, X. (2021). Digital twin-
Wu, W., Zhao, Z., Shen, L., Kong, X. T. R., Guo, D.,
driven joint optimisation of packing and storage
Zhong, R. Y., & Huang, G. Q. (2022). Just Trolley:
assignment in large-scale automated high-rise
Implementation of industrial IoT and digital twin-
warehouse product-service system. International
enabled spatial-temporal traceability and visibility
Journal of Computer Integrated Manufacturing, 34(7–
for finished goods logistics. Advanced Engineering
8), 783–800.
Informatics, 52.
https://doi.org/10.1080/0951192X.2019.1667032
https://doi.org/10.1016/j.aei.2022.101571
Leung, E. K. H., Lee, C. K. H., & Ouyang, Z. (2022). From
Xiuyu, C., & Tianyi, G. (2018). Research on the
traditional warehouses to Physical Internet hubs: A
Predicting Model of Convenience Store Model
digital twin-based inbound synchronization
Based on Digital Twins. Proceedings - 2018
framework for PI-order management. International
International Conference on Smart Grid and Electrical
Journal of Production Economics, 244.
Automation, ICSGEA 2018, 224–226.
https://doi.org/10.1016/j.ijpe.2021.108353
https://doi.org/10.1109/ICSGEA.2018.00062
M. I. Jordan, & T. M. Mitchell. (2015). Machine
Zacharaki, A., Vafeiadis, T., Kolokas, N., Vaxevani, A.,
learning: Trends, perspectives, and prospects.
Xu, Y., Peschl, M., Ioannidis, D., & Tzovaras, D.
Science, 349(6245), 255–260.
(2021). RECLAIM: Toward a New Era of
https://doi.org/10.1126/science.aaa8415
Refurbishment and Remanufacturing of Industrial
Mehmood, M. U., Chun, D., Zeeshan, Han, H., Jeon, G., Equipment. Frontiers in Artificial Intelligence, 3.
& Chen, K. (2019). A review of the applications of https://doi.org/10.3389/frai.2020.570562
artificial intelligence and big data to buildings for
Zhan, X., Wu, W., Shen, L., Liao, W., Zhao, Z., & Xia, J.
energy-efficiency and a comfortable indoor living
(2022c). Industrial internet of things and
environment. Energy and Buildings, 202, 109383.
unsupervised deep learning enabled real-time
https://doi.org/10.1016/J.ENBUILD.2019.109383
occupational safety monitoring in cold storage
Melesse, T. Y., Bollo, M., Pasquale, V. di, Centro, F., & warehouse. Safety Science, 152, 105766.
Riemma, S. (2022). Machine Learning-Based https://doi.org/10.1016/j.ssci.2022.105766
Digital Twin for Monitoring Fruit Quality
Zhao, Z., Shen, L., Yang, C., Wu, W., Zhang, M., &
Evolution. Procedia Computer Science, 200, 13–20.
Huang, G. Q. (2021). IoT and digital twin enabled
https://doi.org/https://doi.org/10.1016/j.procs.202
smart tracking for safety management. Computers
2.01.200
and Operations Research, 128.
Minerva, R., Awan, F. M., & Crespi, N. (2021). https://doi.org/10.1016/J.COR.2020.105183
Exploiting Digital Twin as enablers for Synthetic
Zheng, X., Lu, J., & Kiritsis, D. (2021). The emergence
Sensing. IEEE Internet Computing.
of cognitive digital twin: vision, challenges and
https://doi.org/10.1109/MIC.2021.3051674
opportunities. International Journal of Production
Pan, I., Mason, L. R., & Matar, O. K. (2022). Data- Research.
centric Engineering : integrating simulation, https://doi.org/10.1080/00207543.2021.201459
machine learning and statistics. Challenges and
opportunities. Chemical Engineering Science, 249,

View publication stats