Regensburg 2019 – wissenschaftliches Programm
MM 8.5: Topical Talk
Montag, 1. April 2019, 17:30–18:00, H43
Machine-learning interatomic potentials for multicomponent alloys — •Alexander Shapeev — Skolkovo Institute of Science and Technology, Skolkovo Innovation Center, Moscow 143026, Russia
Multicomponent alloys are a challenge to materials design. It is time- and resource-consuming to exhaust the space of possible compositions experimentally, and equally time- and resource-consuming to do it via ab initio modeling. Moreover, many empirical and data-driven approximants to ab initio models also fail because the configurational space is huge and it is hard to avoid extrapolation when using such approximants for modeling multicomponent alloys.
In my talk I will present a machine-learning framework for the discovery of stable phases of multicomponent alloys and computation of their free energy and derivative properties including thermodynamic stability of phases. The framework is based on (1) machine-learning interatomic potentials capable of very accurately approximating ab initio models, and (2) an active-learning algorithm capable of detecting extrapolation in configurational space attempted when predicting interatomic interaction, and through additional fitting ensure reliability of the predictions.