Gruppo AIMETA proponente

GIMC

Titolo

Physics, Data, and Hybrids: Rethinking Modeling Strategies in Computational Mechanics

Organizzatori

• Mauro Corrado, Politecnico di Torino, Questo indirizzo email è protetto dagli spambots. È necessario abilitare JavaScript per vederlo.
• Stefano Mariani, Politecnico di Milano, Questo indirizzo email è protetto dagli spambots. È necessario abilitare JavaScript per vederlo.
• Francesco Regazzoni, Politecnico di Milano, Questo indirizzo email è protetto dagli spambots. È necessario abilitare JavaScript per vederlo.
• Salvatore Sessa, Università degli Studi di Napoli Federico II, Questo indirizzo email è protetto dagli spambots. È necessario abilitare JavaScript per vederlo.

Descrizione

The rapid diffusion of machine learning and model-free approaches, driven by their growing popularity across applied sciences, but also by EU strategic initiatives promoting data-centric innovation, trustworthy AI, and digital sovereignty, has brought renewed attention to their promised advantages in terms of flexibility, automation, and computational efficiency. Such an accelerating shift toward datadriven modeling naturally has raised fundamental and sometimes controversial questions within the community:

• Are data-driven models merely sophisticated interpolators, effective only for relatively simple or well-posed problems in mechanics, or are they able to generalize to significantly different scenarios?
• Is the often observed loss of physical interpretability in data-driven approaches an intrinsic limitation of these methods, or does it simply reflect our still immature understanding of their potential physical meaning?
• To what extent can physical principles be relaxed, embedded, or even replaced by data?
• Do hybrid physics–data approaches offer a robust and reliable compromise, or do they risk diluting the strengths of both paradigms, ultimately inheriting their limitations rather than their advantages?
• Do the often-claimed gains in computational efficiency withstand careful quantitative assessment, sufficiently justifying the additional modeling, data collection, and training effort?

This minisymposium aims to stimulate an open and critical debate on the role of data-driven strategies – ranging from model-free approaches to machine learning techniques – versus classical physics-based models in theoretical and applied mechanics.

Moreover, a distinctive feature of mechanics, compared to many other fields where machine learning has thrived, is the absence of true big data. Available datasets are often sparse, noisy, heterogeneous, and multi-fidelity, originating from simulations, experiments, or operational measurements with varying levels of uncertainty. This raises a further key question that the minisymposium seeks to address: should data scarcity and inconsistency be regarded as a fundamental bottleneck for data-driven approaches, or rather as an opportunity to develop novel methodologies that explicitly leverage physics, structure, and prior knowledge to learn effectively from limited information?

The minisymposium welcomes contributions presenting recent advances in pure data-driven methods, physics-informed and hybrid strategies, and systematic comparisons between machine learning, modelfree, and traditional physics-based approaches. Emphasis will be placed on methodological soundness, physical interpretability, robustness to data limitations, uncertainty quantification, and rigorously quantified computational performance. The overarching goal is to clarify strengths, limitations, and open challenges, and to foster a constructive dialogue on the future role of data-driven modeling in mechanics.

Aree di interesse

• Mechanics of solids and materials
• Fluid mechanics and fluid–structure interaction
• Dynamics of mechanical systems and structural mechanics
• Computational methods and numerical modeling in mechanics
• Model identification, uncertainty, and validation