CATALYST: the Reliable Hybrid Model Forge

Overview

A Digital Twin (DT) is a coordinated collection of digital models that mirrors a physical system throughout its lifecycle. Continuously refreshed with real-time data, it relies on reasoning techniques such as simulation and machine learning to support tasks ranging from prediction to real-time decision-making.

Because no single model can capture every aspect of a complex system, a DT brings together diverse disciplines — architecture description, modeling & simulation, and data science — each contributing models with different structures, fidelities, and purposes.

These models can be classified along several dimensions, such as inference method (inductive vs. deductive), modeling approach (empirical vs. mechanistic), concern (behavioral, structural, safety, etc.), purpose (descriptive vs. prescriptive), fidelity, and temporal scope.

Most DTs rely on hybridization, combining complementary models — data-driven, physics-based, or mixed — to improve accuracy, robustness, and adaptability. Techniques such as physically informed neural networks and uncertainty quantification illustrate how heterogeneous models can be fused to exploit their respective strengths. Nowadays, this hybridization is realized in an ad-hoc manner, and consequently requires a more systematic and principled approach.

To address this issue, the Catalyst project focuses on three related research axes of model hybridization:

• the definition of consistent interfaces for querying and manipulating models;
• the specification of composition operators to hybridize and federate models;
• the realization of proper verification and validation tools to populate these interfaces.

Through these research axes, the objective of Catalyst is to develop a set of methods and tools to manipulate deductive and inductive models homogeneously, keeping their main characteristics accessible in order to reliably hybridize and federate them.

Associated Use Cases

Use cases in Catalyst will include diverse Digital Twins representing various physical entities. Use cases are currently under elicitations.

Investigator & Project Partners

The project involves nowadays the following partners: Université Côte d’Azur, I3S, Inria, CETIM, INRAE, Telecom Paris, Aniti, Université de Pau et des Pays de l’Adour, IMT-Atlantique, CEA et IRIT.

Principal Investigator:

Julien Deantoni

Full Professor at Université Côte d'Azur and Head of the Kairos Team

Julien Deantoni is full professor at Université Côte d’Azur and head of the Kairos team. He applies rigorous modelings and abstractions for heterogeneous models since his PhD thesis in 2007. He published on the importance of handling the globalization of modeling languages more than 10 years ago. Since then, he has published various papers and developed tools to facilitate this globalization. He has also published papers highlighting the need to consider the semantics of the different models for more precise and efficient collaborative simulations. More recently, he worked on the notion of multi-fidelity in Digital Twins and already hybridized DT’s models in order to mitigate uncertainty and consequently provide better control in the PT. Conjointly, he participated to numerous collaborative projects

Participating Partners:

External Partners:

Project Implementation

The project is structured into three cohesive work packages, with multi-fidelity and uncertainty addressed throughout.