Accelerating ab initio atomic relaxations using machine-learning interatomic potentials

Details

Supervisors

Rignanese, Gian-Marco ; Giantomassi, Matteo

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en chimie et science des matériaux, à finalité spécialisée

Abstract

enAb initio materials simulations, used for a wide class of calculations of technological interest, rely on computationally costly DFT methods, making them unaffordable for systems of larger sizes. Machine-learning interatomic potentials (MLIPs) have been developed as an alternative solution, attempting to approach the accuracy of DFT, while keeping the computational cost similar to empirical potentials. The objective of this work is to implement universal machine-learning interatomic potentials within the computational framework of ABINIT. To this purpose, we evaluate three related MLIPs based on the Point Edge Transformer architecture in terms of accuracy of their predictions and improvements in computational speed. In the later stages, we discuss implementation strategies for a module allowing to use these potentials for calculations directly in ABINIT.