Machine learning for accelerated materials design
Title: Machine learning for accelerated materials design
SNIC Project: Berzelius-2021-83
Project Type: LiU Berzelius
Principal Investigator: Igor Abrikosov <igor.abrikosov@liu.se>
Affiliation: Linköpings universitet
Duration: 2022-01-14 – 2022-08-01
Classification: 10304
Homepage: https://liu.se/organisation/liu/ifm/teofy
Keywords:

Abstract

Artificial intelligence and machine learning methods are expected to revolutionize the field of materials modeling, increasing time and length scale of the simulations and greatly shortening the time it takes to design new materials. The aim of this project is to get access to novel computational resources provided by Berzelius system and to investigate and develop AI/ML based techniques for accelerated materials design through the use of the so-called machine learning potentials. We focus mainly on molecular dynamics (MD) simulations of thermodynamic and mechanical properties of refractory nitride materials with potential industrial and societal interest. In our approach we actively train and use machine learned interatomic potentials to run simulations of exceedingly large scales (thousands of atoms, or nanosecond time scales) with ab-initio accuracy. We also utilize advanced neural networks to explore the properties of alloys and find new potential candidates for thorough investigation with our active learning scheme. We aim with this project to test the viability of implementing our methods in a purely GPU-accelerated or hybrid setup. The scientific part of the project is supported with Wallenberg Scholar award from Knut and Alice Wallenbergs Foundation (KAW) to Prof. Abrikosov. The simulation results are essential for our industrial collaborations with, e.g. leading Swedish companies Sandvik Coromant and Seco Tools in the framework of VINNOVA supported program FUNMATII. The large scale MD simulations will be carried out using LAMMPS, which benefits from GPU-acceleration. For the active learning of interatomic potentials, our data-sets will be improved by running small scale ab-initio calculations, carried out using the GPU-accelerated version of VASP.