The widespread use of artificial intelligence (AI) to extract value from historical data is transforming the supply chain landscape, adding complexity to operational management and risk mitigation. This paper provides a discussion of the SmarTwin research project and its methodology, with the aim of implementing an intelligent Digital Supply Chain Twin (iDSCT) to address the challenges of modern supply chains. The SmarTwin architecture has been designed to achieve several objectives, including establishing a traceable automated system that can improve the overall reliability of the supply chain. Furthermore, the system is intended to develop a comprehensive representation of a specific supply chain instance that serves as a controlled virtual environment to implement, simulate and optimize the decision support system (DSS). SmarTwin aggregates data and information from multiple sources and integrates them into a semantic framework. This unified vision is continuously monitored by a predictive analytics layer capable of issuing early warning alerts, which trigger coordinated responses based on combined simulation and optimization strategies to support data-driven, resilient, and sustainable supply chain operations.
The SmarTwin project, started in 2023, is structured around a series of milestones (OR), with
responsibilities distributed among research partners, and now reached its 24th month. The project has produced
publications outlining its goals to investigate an innovative service model for cost optimization, risk
reduction, micro-traceability of product processing steps, certification of the ecological footprint and
ifnancial support for complex supply chains [
This work focuses on the Smartwin objective of developing an intelligent Digital Supply Chain Twin
(iDSCT) [
To achieve these goals, the system was designed around pivotal non-functional requirements. It adopts a decentralized, Web3-based, event-driven architecture, and comprises loosely coupled, specialized subsystems for scalable and flexible configurations. The subsystems are responsible for distinct functional domains, and can be added or removed as needed. The macro-functionalities of SmarTwin emerge from the coherent integration of the capabilities of the active subsystems within the current configuration.
The AI Subsystem represents a crucial component of the system, ofering a pluggable set of AI tools delivered as a service that support core functionalities to address various supply chain challenges.
Section 2 presents the architectural design and its core elements. Section 3 presents the AI as a service approach and its role in enabling system functionalities.
This section outlines the architectural design principles adopted to support the system operations and meet key non-functional requirements. As shown in Figure 1, the architecture comprises infrastructural components (UML components) and specialized subsystems (UML packages), integrated in a decentralized manner through the Blockchain Subsystem, which supports the event-driven supply chain workflow. The following high-level overview describes the main elements of the architecture, with reference numbers corresponding to those in the figure (noted in parentheses).
The Generic Stakeholder component (1) is intended to abstract any participant in the supply chain who contributes to the realization of the final product and shares data related to specific transformation phases. It provides a multitude of data types, including silos data (ERP, financial), operational data, formal declarations, IoT, and video streams.
The Data Lake component (2) stores data shared by stakeholders and any information produced by
the SmarTwin subsystems. The employment of AI functionalities facilitates the classification of data [
The IoT-queue component (3) is a middleware that collects the data provided by IoT devices installed in the work areas of stakeholders. These measurements are then integrated into the digital twin of the corresponding supply chain instance. It supports various communication protocols for noninvasive interoperability with stakeholder infrastructures, thereby ensuring seamless integration. It is implemented using Eclipse Hono2.
The IoT Manager Subsystem (4) is responsible for processing IoT-Queue sensor data to build a semantic representation that enables the creation of a digital twin for a supply chain instance. It is also responsible to manage real-time alerts and prepare time-series data for machine learning tasks. It uses Eclipse Ditto3 and InfluxDB 4.
The Industrial Knowledge Management component (5) ingests data from the Data Lake component to
construct a financial knowledge graph, assessing supply chain and stakeholders financial health and
risk. The use of AI techniques facilitates the extraction and inference of knowledge from data [
The Visual Data Manager Subsystem (6) is responsible for retrieving the video streams shared by the stakeholders. It is also tasked with the preparation and analysis of them using computer vision models that are provided by the AI Subsystem. The results are transmitted to the IoT queue to enrich the digital twin managed by the IoT Subsystem.
The Blockchain Subsystem (7) is designed to handle data certification and the decentralized logging
of system and workflow events, supporting the system’s Web3 and event-driven foundation [
The Fintech Subsystem (8) is a portal that ofers financial technology services based on credits derived
from supply chain transactions [
The API Subsystem (9) ofers standardized, secure REST interfaces to access blockchain and fintech services. A technological solution is currently under development, based on Hyperledger FireFly7 and aligned with Open Banking guidelines.
The Supply Chain instance manager (10) is responsible for managing workflow execution, data verification, and product quality assessment, managing multiple SC instances concurrently via blockchain event monitoring.
The Authoring Subsystem (11) is responsible for building the system configuration. Provides tools for domain experts to manage the complexity of the task minimizing the need for developer support. A domain-specific language (DSL) is under development to configure supply chain instances and define related smart contracts in Solidity for the Blockchain Subsystem. The use of a fine-tuned LLM is in consideration to translate contract requirements into the Solidity DSL through a conversational interface.
The DSCT Subsystem (12) provides a holistic view of the entire supply chain by aggregating data from all components and subsystems. It establishes a semantic integration layer that is used to generate multiple views highlighting key aspects such as financial performance, product quality, carbon footprint, and ethical considerations, among others. Its core function is to enable proactive monitoring through an AI-powered prediction layer that supports real-time analysis, identifies critical data paths, and forecasts potential scenarios, such as financial disruptions. Upon detecting risks, the subsystem issues early warnings to the monitoring system, triggering contingency management through simulation and optimization. These processes yield prescriptive suggestions to support decision-making. The subsystem also includes a suite of Grafana-based8 dashboards, providing a comprehensive visualization of interconnected components and the current state of supply chain.
The AI Subsystem (13) manages and delivers as a service AI models (machine learning, computer vision) to other subsystems and components of SmarTwin, exposing them as REST/gRPC microservices using the Seldon Core framework9. provides more information on its AI as a service approach.
The Explainable AI Subsystem (14) provides tools to interpret black-box AI decisions and translate them into human-understandable explanations, improving transparency and stakeholder trust.
This section discusses the application of artificial intelligence and data analysis techniques in the development of SmarTwin functionalities. Table 1 summarizes the methodologies that will be implemented, organized by subsystem and infrastructural component. These techniques support key system operations, as detailed below. ibKnanfseoerwdelnelecdaegrenvGiinargae)pmhbReedadsionngisngan(rdelagtriaopnhal- IMndanuastgreiamleKnntowledge
ence and algorithms) Management vTiesxetdCLleaassrinfiicnagt)ion and Labeling (Super- Data Lake TNeaxttuGraelneLraantgiouna)ge Processing (LLMs, Authoring Subsystem
ties and condition assessment of assets.
communities) to extract insights.
ment using predefined taxonomies.
ply chain instances configuration.
A main project goal is the construction of a prediction layer for real-time monitoring and risk
8https://grafana.com
9https://www.seldon.io/solutions/core
identification within the DSCT subsystem to provide proactive suggestions about anomalies and
potential disruptions [
The concept of AI as a Service (AIaaS) refers to the delivery of artificial intelligence capabilities as modular, API-accessible services, hosted on cloud or hybrid infrastructure, and consumable on demand by external applications. This paradigm enables organizations to integrate intelligent capabilities, such as classification, anomaly detection, prediction, or language processing, without having to deal directly with managing the entire machine learning pipeline, from development to deployment.
For SmarTwin, which operates in complex scenarios like supply chain control in the transport of perishable fruits and vegetables, adopting a Seldon Core-based deployment architecture ofers key advantages. These include timely, scalable, and reliable model orchestration—especially for use cases involving computer vision, time series analysis, and predictive maintenance.
At this late stage of the project, it has been possible to appreciate one of the main advantages of using
Seldon Core: its ability to orchestrate ML models in Kubernetes environments, providing a flexible,
cloud-native infrastructure that facilitates the deployment and management of heterogeneous models
in complex manufacturing environments. In contexts such as agribusiness, where data can come from
multiple sources, such as images acquired by cameras for automated visual inspection of product
surfaces and IoT sensors for supply chain monitoring [
Seldon Core also ofers robust model versioning, testing, release and rollback mechanisms for services, which are key aspects in scenarios where models need to be frequently updated or new ones added without compromising service continuity. This is particularly relevant when considering models that are adaptive or subject to periodic retraining based on recent data, as in the case of visual quality classification of vegetables and fruits, which is subject to seasonal and environmental variations.
However, despite the benefits, its adoption also presents challenges. Integration demands advanced Kubernetes and DevOps knowledge, increasing initial project complexity and slow down prototyping phases, especially when rapid deployment is required. Additionally, managing distributed components and maintaining extensive YAML configurations introduces technical overhead that is unnecessary in simpler or low-scale scenarios. Another limitation involves latency, introduced by containerization and the orchestration of complex pipelines. In applications requiring real-time inferences with subsecond response times, as quality checks on high-speed packaging lines, the architectural overhead may necessitate a trade-of between model accuracy and speed of response.
Currently, we believe that Seldon Core is a powerful and scalable solution for deploying artificial intelligence models in highly complex industrial settings such as the transportation of perishable fruit and vegetables. However, its success depends on a proper assessment of the trade-of between architectural power and operational complexity, as well as a harmonious integration with the digital ecosystem of the agrifood supply chain.
SmarTwin is a research project aimed at addressing complex challenges in the supply chain domain through the development of a configurable and scalable digital solution. At its core lies an intelligent digital twin of the supply chain, designed to ensure certification, traceability, and transparency across all operational stages. The system integrates a holistic view of supply chain dynamics with a predictive analytics layer capable of monitoring anomalies and forecasting potential disruptions.
Following 24 months of research and development, the project has now entered its final phase, during which validation, integration, prediction and optimization activities will be completed.
The project was funded by the Italian Ministry of Economic Development (MISE): it started in 2023 and it will end in 2026. Project reference no. F/310218/05/X56.
During the preparation of this work, the author(s) used GPT-4o and DeepL in order to: grammar and spelling check, paraphrase, reword and improve writing style. The author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.