Training Equivariant Neural Network Interatomic Potentials to Model Phase Transformations in Complex Materials
Title: Training Equivariant Neural Network Interatomic Potentials to Model Phase Transformations in Complex Materials
DNr: Berzelius-2024-177
Project Type: LiU Berzelius
Principal Investigator: Johan Klarbring <>
Affiliation: Linköpings universitet
Duration: 2024-05-03 – 2024-12-01
Classification: 10304


This project aims to train machine learning interatomic potentials (MLIPS) based on recently developed equivariant graph neural network architectures [1], [2]. These will be used to model phase transformations in complex materials. We will focus on two classes of materials: (1) So-called plastic crystals, which are materials that contain molecular units which are randomly oriented, but whose centers-of-mass retain the long-range ordering characteristic of crystalline solids. These systems often have phase transformations from an ordered low-temperature state to a high-temperature phase with disordered molecular units. Since these transformations are highly sensitive to external pressure, they can be used to produce efficient solid-state cooling devices through the barocaloric effect [3]. (2) Halide perovskites, which are highly promising materials for photovoltaic applications [4]. These systems are known to have complex atomistic dynamics, and an intricate sequence of phase transformations. Although they have been studied very intensively over the last decade, the connection between their unusual atomistic dynamics and their outstanding photovoltaic properties, remains elusive. Ideally, atomistic modelling of these material classes would employ large-scale DFT-based ab-initio molecular dynamics (AIMD) simulations based on density functional theory (DFT). The time- and length scales that need to be simulated are, however, prohibitively expensive with this methodology. The development of accurate MLIPs over the last ~decade, such as the ones we propose to use, show promise in resolving this issue. In this project, we will use DFT-based data-sets for a set of materials to train MLIPs based on two state-of-the-art GNN architectures: Allegro [1] and MACE [2]. After training the MLIPs on Berzelius, production MD runs will be performed on other computational resources. A successful project is expected to result in a deepening of our understanding of the atomistic dynamics of these two classes of important energy materials. Allegro and MACE have both been installed, tested and used on Berzelius in our previous proposal. We apply for the default allocation. [1]; [2] ; [3] [4]