(582c) A Learning Framework to Accelerate First Principles Simulations

First principles quantum mechanics based modeling schemes, such as density functional theory (DFT), are powerful tools to study the static and dynamical evolution of processes (e.g., chemical reactions, phase transformations, transport). Nevertheless, such methods have several practical drawbacks. For instance, owing to the computational expense, a typical DFT-based simulation can only handle system sizes of at most 100s of atoms or span timescales of the order of just picoseconds (with a maximum reachable timescale of about a nanosecond). Here, we show that DFT-based simulations can be significantly accelerated, by using machine-learning methods to make rapid high-fidelity predictions, based on past data or knowledge.

For many atomistic simulations, e.g. molecular dynamics, geometry optimization, identifying reaction barriers, materials properties, etc., determining the force on an atom is key. If we are able to rapidly estimate (with acceptable accuracy) the atomic forces of a new configuration, given past similar configurations (determined, say, using DFT), then we can significantly speed up the simulations. The ‘training’ underlying this force prediction capability requires a critical amount of prior information or results, adequate descriptors (or fingerprints) that uniquely represent our configurations, and a suitable measure of (dis)similarity between configurations. We show here that such training, prediction, and consequently, the acceleration of DFT-based simulations, are indeed possible. Illustrations of this new development are made for the case of the self-diffusion of vacancies in bulk Al, the diffusion of an Al adatom on an Al surface, finding optimal geometries of bulk Al structures with more than 100s of atoms, identifying reaction barriers for simple elementary processes, and lastly, determining thermodynamic properties of bulk Al. Such a framework lays the foundations, for understanding the behavior of chemical reactions or materials at increased time and length scales, all at first principles accuracy, a capability that is well beyond the current realm of first principles methods.