Using machine learning interatomic potentials to explore complex phenomena in energy materials

Abstract:

In this talk, I will showcase how machine learning interatomic potentials (MLIPs) based on modern neural network architectures, combined with large-scale molecular dynamics simulations, can be used to accurately capture and understand complex physical behavior in a variety of energy-related materials. I will discuss three case studies. The first is the fast Na-ion conductors Na3-xSb1-xWxS4,  where an MLIP is constructed able to describe the interplay between dopant concentration, phase transformation temperatures and ionic diffusion [1]. I will then touch on our work related to modelling the complex dynamics and phase transformations in halide perovskites [2]. Finally, I will showcase ongoing work relating to data efficient fine-tuning of so-called “foundational” MLIPs, which I will discuss in the context of barocaloric materials for solid-state cooling applications.

[1] Chemistry of Materials 2024 36 (19), 9406-9413
[2] J. Phys. Chem. C 127, 38, 19141–19151 (2023); Small 20, 230356 (2024); arXiv:2404.14598 (2024)