(747g) QM/ML: A Hybrid Quantum-Mechanics/Machine-Learning Scheme

Hybrid quantum-mechanics / molecular-mechanics (QM/MM) simulations are popular tools for the simulation of extended atomistic systems, in which the atoms in a core region of interest are treated with a QM calculator and the surrounding atoms are treated with an empirical potential. Recently, a number of atomistic machine-learning (ML) tools have emerged that provide functional forms capable of reproducing the output of more expensive electronic-structure calculations. Such ML tools are intriguing candidates for the MM calculator in QM/MM schemes, and have some obvious advantages such as the avoidance of mis-matches (e.g., between lattice constants, mechanical responses, or liquid structures) between the QM and MM calculators, and the removal of the need for the researcher to gain expertise in the MM literature. However, accurate ML models generally require a substantial amount of training data that anticipates atomic configurations that will be encountered during the course of a simulation, provided on the basis of QM methods. Here, we present a simple quantum-mechanics / machine-learning algorithm created with the goal of minimizing the amount of QM training data that needs to be anticipated and generated. Importantly, by re-training the ML calculator to the new QM information at each step, we can identically cancel the overlapping QM and MM energy and forces in a subtractive scheme, producing a simple algorithm for treating boundary interactions between the regions. The approach is analogous to the subtractive framework of the QM/MM methodology, but is not dependent on empirical parameterization, and allows the simulation of arbitrary atomic configurations while keeping QM precision, within certain physical limitations. In this talk, we outline the basic theory and algorithm and also demonstrate the application of QM/ML on the basis of simple examples.