Reduced-order surrogate models of physical systems: physics-constraints and temporal forecasting

MLG Seminar by Dr. Alberto Procacci

Abstract

Reduced-order modelling has become an important tool for the analysis and control of complex physical systems. Recent advances in machine learning have enabled the construction of surrogate models capable of approximating high-dimensional physical fields, such as pressure, velocity, or temperature, with high accuracy and at very low computational cost. These models are particularly attractive in applications requiring many model evaluations, such as uncertainty quantification, and they are increasingly used as building blocks for digital twins.

Despite their success, important challenges remain. In particular, many data-driven reduced-order models lack explicit physical constraints, which can lead to violations of basic physical laws such as mass conservation. Furthermore, the absence of physical structure often limits their ability to extrapolate beyond the training data, making reliable temporal forecasting of dynamical systems difficult.

In this seminar, I will present recent work aimed at addressing these limitations. We introduce modifications to standard reduced-order modelling techniques that incorporate physical consistency and enable temporal forecasting. The results show that embedding physical structure into surrogate models can significantly improve both their reliability and predictive capabilities.

Speaker Biography

Alberto Procacci is a professor in the Aero-Thermo-Mechanics Department at Université Libre de Bruxelles, specialising in the application of machine learning to combustion systems. He completed a PhD in Engineering at École Polytechnique of ULB (2024), focusing on reduced-order models for digital twins of reacting flows, integrating sparse sensing, modal analysis, physics-constrained machine learning, and dynamical system forecasting.

He holds a Master’s degree in Energy Engineering from the University of Bologna (2017). Alberto has previous professional experience as an R&D engineer in the refrigeration industry, combining strong engineering fundamentals with expertise in advanced data-driven modelling for energy and combustion applications. He was promoted to a professor in 2025.

Copyright © 2026 Machine Learning Group, All rights reserved.
Our mailing address is: Machine Learning Group, Bd de la Plaine, CP212, Ixelles 1050, Belgium