Machine learned interatomic potentials using random features

This paper introduces a computationally efficient method for modeling interatomic interactions, such as energy and forces, using a linear combination of random features. This approach allows for fast parameter estimation by solving a linear least-squares problem. The authors explore the use of random features based on stationary and non-stationary kernels for energy approximation, demonstrating results for 2D materials, metals, and semiconductors. Force and energy predictions show close agreement with density functional theory (DFT) calculations, with a 96% reduction in training time compared to standard kernel models. Molecular dynamics simulations using the proposed potentials agree well with experimental and DFT values, and phonon frequencies are within 0.1% of DFT results. The method also addresses scalability issues in machine learning for interatomic potentials.

Gurjot Dhaliwal, Prasanth B. Nair, Chandra Veer Singh