Interatomic Potential in a Simple Dense Neural Network Representation

TL;DR

This paper demonstrates that classical interatomic potentials can be effectively represented using dense neural networks, offering a machine learning approach to improve atomic-scale simulations of materials.

Contribution

It introduces a novel method of representing interatomic potentials with dense neural networks, enhancing accuracy in atomic simulations.

Findings

Neural network potentials achieve high accuracy in modeling atomic interactions.

Machine learning methods can bypass traditional fitting difficulties.

Potential for improved simulations of metals and alloys.

Abstract

Simulations at the atomic scale provide a direct and effective way to understand the mechanical properties of materials. In the regime of classical mechanics, simulations for the thermodynamic properties of metals and alloys can be done by either solving the equations of motion or performing Monte Carlo sampling. The key component for an accurate simulation of such physical systems to produce faithful physical quantities is the use of an appropriate potential or a force field. In this paper, we explore the use of methods from the realm of machine learning to overcome and bypass difficulties encountered when fitting potentials for atomic systems. Particularly, we will show that classical potentials can be represented by a dense neural network with good accuracy.