EDITS: Enhancing Digital Twins with Scalable and Interpretable AI

From physics-based modelling and simulation to adaptive learning, big-data analytics, and counterfactual reasoning.

Principal Investigator: Prof. Gianluca Bontempi (Machine Learning Group, ULB)

Funding: FNRS WEL-T Investigator Programme 2024 (Advanced)

Duration: 4 Years (2025–2029)

Key Keywords: Digital Twins, Causal Reasoning, Big Data Analytics, Physics-Informed ML, Counterfactuals.

Project Overview

A Digital Twin (DT) is a computational avatar of a real physical system (the Physical Twin, PT). Unlike conventional models that match a system in a static, offline manner, a DT relies on a seamless two-way data flow to track, monitor, and control the real-time evolution of the physical entity.

While DTs are often advertised as magical solutions for industry and smart cities, their deployment lacks a thorough foundation in scientific domains like adaptive learning and causal reasoning. EDITS aims to bridge the gap between traditional engineering (modelling and simulation) and modern Artificial Intelligence (AI).

Our goal is to transform DTs from simple monitoring tools into intelligent agents capable of explaining past behaviors and anticipating future outcomes through counterfactual reasoning (e.g., “What would have happened if we had performed maintenance yesterday?”).


EDITS Digital Twin Architecture
Figure 1: The EDITS Digital Twin Architecture integrating Big Data, Simulation, and Causal Reasoning.

Scientific Objectives

The EDITS project focuses on five main methodological contributions:

  • Big Data Infrastructure: Designing a scalable analytics infrastructure to support DT deployment in real streaming environments (handling high-dimensional data).
  • Adaptive Calibration: Developing assimilation techniques to calibrate physics-based models and simulators using streams of real-world data.
  • Foundational Models for DTs: Defining a strategy to transfer knowledge acquired from one DT to similar contexts (Transfer Learning).
  • Causal Inference: Moving beyond correlation by adding human interpretability and counterfactual reasoning capabilities to the models.
  • Validation: Assessing the methodology across three distinct physical case studies.

Methodology

The project is organized into tightly integrated work packages covering the full pipeline of a Digital Twin:

  • Sensing & Infrastructure: Implementation of Lambda architectures for real-time and batch processing.
  • Orchestration: Scalable, decentralized data handling and model storage.
  • Learning & Assimilation: Integration of physical constraints into Machine Learning models (Physics-Informed ML).
  • Simulation & Forecasting: Advanced “What-if” scenario analysis.
  • Causality: Discovery of causal graphs to enable interpretability and decision support.

Case Studies

To ensure real-world applicability, the EDITS methodology will be validated on three specific use cases:

1. IoT & Robotics

The “Sandbox”: Using accessible hardware (e.g., robotic arms, smart farm mockups) to validate the entire data-to-decision pipeline in a controlled environment. This allows for rigorous testing of control actions without safety risks.

2. TrafficTwin (Smart Cities)

Urban Mobility: Building on projects like TORRES and TULIPE, this case study focuses on traffic management. The DT will allow city planners to simulate roadworks or deviation plans and visualize their impact on congestion in 3D.

3. EnergyTwin (Wind Farms)

Sustainable Energy: In collaboration with the University of Sannio (Italy), we will model wind farm operations. The focus is on predicting maintenance needs and simulating critical scenarios (e.g., extreme weather events) to prevent blackouts.


Traffic Twin Visualisation
Figure 2: The TrafficTwin interface for urban decision making.

Collaboration & Resources

EDITS is coordinated by the Machine Learning Group (MLG) at ULB. It leverages the high-performance computing resources of the CECI infrastructure and collaborations within the TRAIL (Trusted AI Labs) and FARI ecosystems.

Internal Team: The project funds multiple Postdoctoral and PhD researchers focusing on Big Data, Forecasting, and Causal Inference.

External Partners:

  • University of Sannio (Italy) – Power Systems & Energy
  • SIRRIS – Manufacturing & IoT
  • Various academic partners in Switzerland, France, and Canada.

For more information on EDITS or to inquire about research opportunities, please contact Prof. Gianluca Bontempi.