Predicting and harnessing unusual quantum effects in condensed-phase chemical processes via a synthesis of machine learning, path integration, and enhanced sampling

Abstract: Reliable theoretical prediction of complex chemical processes in condensed phases requires an accurate quantum mechanical description of interatomic interactions.  If these are to be used in a molecular dynamics calculation, they are often generated “on the fly” from approximate solutions of the electronic Schrödinger equation as the simulation proceeds, a technique known as ab initio molecular dynamics (AIMD).   However, due to the high computational cost of these quantum calculations, alternative approaches employing machine learning methods represent an attractive alternative and have become increasingly popular.  As the adoption of machine-learning potential becomes more widespread, it is important to consider how simulations employing them should be carried out.  Specifically, as they do not implicitly include nuclear quantum effects, these effects must be treated explicitly, for which the most efficient approach involves the use of Feynman path integral techniques.  This is especially important for processes involving light elements.  In this talk, I will discuss how state-of-the-art machine learning potential can be combined with path-integral molecular dynamics to address a variety of challenging chemical problems that have unexpected quantum behavior.  In particular, I will discuss a new class of battery electrolytes that, by harnessing their unusual quantum character, could lead to breakthrough performance.  Discovery of the mechanism of charge transport in these systems could only be achieved by the development of an equivariant transformer network interatomic potential model.  I will also discuss how enhanced sampling techniques can be applied to path integral molecular dynamics to describe the quantum diffusion of hydrogen in structure-II clathrates as a function of temperature, where unexpected inverse quantum effects control diffusion rates at higher temperatures.  Finally, I will discuss an open-chain path integral approach for the grand challenge problem of computing quantum time correlation functions.  The approach leads to an, in principle, exact positive-definite distribution function that can be sampled via Monte Carlo or molecular dynamics to yield exact quantum time correlation functions.  Various approximate and exact numerical schemes for sampling this distribution will be discussed and application to the problem of charge transfer reactions will be presented.