Abstract

Over the past 15 years, machine-learning interatomic potentials have evolved from a promising idea to a wide field of materials modeling. The idea is that a true interatomic interaction energy (or an accurate quantum-mechanical model of it) can be approximated as a function of positions of neighbors of each atom with a flexible (systematically improvable) functional form. Within this field, researchers are working on different directions: studying ways to improve accuracy, efficiency, and the range of applicability of potentials, constructing potentials for different atomistic systems, or combining these potentials with other algorithms to enable materials properties calculation that was out-of-reach for more traditional methods. I will present my work in this field under my favorite angle: machine-learning potentials as a computational technology to seamlessly accelerate quantum-mechanical calculations.

Namely, I will present my version of machine-learning potentials, moment tensor potentials, and an active learning algorithm automating the procedure of assembling the dataset. I will show how the two algorithms combined allow for an automatic acceleration by orders of magnitude in such applications as constructing convex hulls of stable alloy structures or computing vibrational and configurational free energy of alloys. Moreover, machine-learning potentials can be used as a screening tool before the final quantum-mechanical calculation, offering a speedup of several orders of magnitude without committing any numerical error.