Apprentissage Actif Multimodal Explicable pour la découverte des mécanismes de plasticité en science des matériaux

Keywords

Active Learning, Explainable Artificial Intelligence (XAI), Data-Efficient Learning, Multimodal Learning, Physics-Aware Machine Learning, Sobriety and Sustainability

Subject details

The objective of studying material plasticity is to understand, predict, and control the irreversible deformation of solids under mechanical loading. In metallic materials, plastic deformation results from complex microstructural mechanisms involving crystal defects and their evolution under stress, which are still only partially captured by existing constitutive models and numerical simulations especially in local, multiscale, or non-standard deformation regimes. In parallel, physical simulators of deformation [3] and damage rely on simplified laws, while advanced experimental techniques generate rich but costly datasets that remain largely under-exploited. Positioned at the interface between materials science and ML, and in close collaboration with the LEM3, this thesis aims to exploit such data to identify and characterize latent parameters or mechanisms governing metallic plasticity, with the goal of improving physical understanding and predictive modeling capabilities. Research in ML has long focused on improving model performance. However, with recent breakthroughs in artificial intelligence and its integration into an ever-increasing number of sometimes safety-critical systems, new challenges are emerging. Growing attention is now being paid to the ecological, economic, and societal costs of AI [2], as well as to the need for more frugal and trustworthy models. For systems already achieving near-optimal performance, it becomes crucial to quantify and communicate uncertainty, particularly for rare but critical predictions, where a single failure may have severe consequences. In parallel, the widespread use of black-box models raises concerns about trust and interpretability, especially when models exhibit overconfidence. The first scientific objective is to develop explainable, data-efficient, multimodal active learning [1] methods to discover latent physical parameters that are not explicitly modelled in current simulators, by comparing simulated predictions with experimental observations. Active learning [4] strategies will guide the acquisition of the most informative experimental samples by combining model uncertainty, simulator–reality mismatch, and expert knowledge from LEM3, thereby minimizing experimental cost. The second objective is to explore an alternative, simulation-free modeling paradigm, where ML models directly predict material deformation at the image level. Given an initial microstructural image and loading conditions, the model will learn to predict the deformed microstructure. This approach aims to capture complex, non-modelled phenomena while remaining computationally efficient, enabling fast predictions where classical finite-element simulations are too costly or inaccurate. Multimodal architectures will combine images, metadata, and possibly simulator outputs, while embedded XAI mechanisms [5,6] will allow experts to validate predictions against physical intuition. Throughout the project, sobriety and sustainability are central concerns. Lightweight architectures, active data selection, and model compression will be employed to ensure a reduced computational and energy footprint. The expected outcomes include improved physical simulators enriched with newly identified parameters, a significant reduction in experimental campaigns through targeted data acquisition, and the development of a robust and explainable AI framework. Beyond materials science, the proposed methodology is generic and transferable to many application domains where data acquisition is costly and uncertainty-aware decision-making is critical. If successful, future work will aim to apply this framework to reducing societal and ecological costs associated with large-scale AI systems [7], as well as to disease detection in expensive multimodal settings, such as EEG and MRI acquisition or microbiological cultures, where active, explainable, and data-efficient learning strategies are particularly relevant [1].