The predictive accuracy of machine learning force fields under active learning conditions

Speaker: Prof. Damien Vandembroucq (University of Twente, and the MESA+ Institute for Nanotechnology)

Abstract:

Upcoming automated laboratories for accelerated materials discovery will have an important simulation component. In many laboratory concepts, Machine Learning Force Fields (MLFF) with an active learning mode, are proposed as the preferred simulations workhorse. In such labs, seed crystal structures, for example for novel battery materials, are generated by combinatorial synthesis and then passed to simulations to rapidly explore the broader thermodynamic phase space from the known stable point. It is crucial to assess predictive accuracy of these methods in order to minimize the number of (expensive time/equipment) experiments and prevent the simulations steering the materials exploration into the wrong direction. In this talk, I will introduce our implementation of on-the-fly learning in a commonly used Density Functional Theory software package. I will showcase some successful application examples from the (in)organic halide perovskites, where we have simulated, using (non)-equilibrium molecular dynamics, crystal phase transitions, thermal conductivity, and other physical observables that would have been impossible using fully ab-initio methods. These soft materials are illustrative of a much larger class of ‘Dynamic Solids’, and are characterized by strong anharmonic lattice dynamics and wide variety of rare-events. I will conclude with some very recent results on the predictive accuracy of superionic phase transitions in AX2 fluorites. In this study we emulate the situation where the MLFF is part of an automated materials discovery pipeline.

This seminar will take place in room D1.111