HetSys News

In his talk, Professor Kermode outlined recent rapid advances in machine-learning interatomic potentials (MLIPs) for tackling so-called chemomechanical problems — phenomena in which local chemical bonding and long-range mechanical stress are tightly coupled, such as fracture and plastic deformation in structural materials. These problems have traditionally been extremely challenging for computational modelling, as they require both high chemical accuracy and the ability to represent large-scale stress fields.

The presentation highlighted several recent developments, including hybrid approaches that couple quantum-mechanical calculations with MLIPs, as well as the use of MLIPs alone to model complex plasticity processes in materials such as tungsten. Professor Kermode also discussed new methods that allow multiple interatomic potentials of differing accuracy and computational cost to be combined within a single large-scale simulation, enabling efficient yet reliable modelling of complex systems.

A major focus of the talk was the emergence of foundation MLIP models, trained on large datasets using deep learning, which are capable of describing a wide range of elements across the periodic table with promising accuracy. Professor Kermode critically evaluated the performance of the MACE MP0 and MPA models for chemomechanical applications and presented new results demonstrating how fine-tuning these models can significantly improve their predictive capability for materials failure processes.

Finally, the lecture emphasised the importance of robust uncertainty quantification when using machine-learning-based surrogate models in scientific simulations, outlining recent progress in assessing and managing uncertainty in atomistic modelling.

Professor Kermode is a leading figure in materials modelling and predictive simulation. He is Professor of Materials Modelling in the School of Engineering at the University of Warwick, Research Cluster Leader for the Predictive Modelling cluster, Director of the EPSRC Centre for Doctoral Training in Modelling of Heterogeneous Systems (HetSys), and Director of the Warwick Centre for Predictive Modelling (WCPM). His research focuses on multiscale modelling, machine learning, and uncertainty quantification, with a particular emphasis on predicting chemomechanically driven materials failure, such as crack propagation where chemical reactions and mechanical stress are intimately linked.

This colloquium offered a compelling insight into how machine learning is transforming the way materials scientists understand and predict failure at the most fundamental scales.