Machine-Learning Interatomic Potentials on the Way to High-Throughput Calculations

Alexander Shapeev*

Skolkovo Institute of Science and Technology, Moscow, Russia

ABSTRACT:  Machine-learning interatomic potentials are a promising tool combining the efficiency of empirical interatomic potentials and accuracy of quantum-mechanical models such as the density functional theory (DFT). The most resource-consuming part of using machine-learning potentials is the manual labor put in tediously constructing the training set. Active learning algorithms come to rescue – combined with such algorithms, machine-learning potentials can be constructed automatically for a given problem thus seamlessly accelerating DFT – see Figure 1. This paves the way for machine-learning potentials to become an important part of high-through materials design. Machine-learning potentials commit a small error compared to DFT, but this error is often smaller than the error of DFT as compared to the experimental data. However, in some applications machine-learning potentials can be used as a screening tool thus completely eliminating their approximation error in the final answer.

In my presentation I will introduce these algorithms and demonstrate how these algorithms can yield a speedup of several orders of magnitude in a number of applications, including construction of convex hulls of stable alloy structures, computing vibrational and configurational free energy of alloys, and diffusion of point defects.

Figure 1.  Active learning of interatomic potential automatically accelerating quantum-mechanical calculations.

Keywords：machine-learning interatomic potentials, active learning, moment tensor potentials