Improving accuracy of interatomic potentials: more physics or more data? A case study of silica

TL;DR

This study compares two strategies for enhancing machine-learning interatomic potentials for silica: increasing model complexity with more parameters versus incorporating a charge model, finding that more parameters improve accuracy while adding charge models does not.

Contribution

The paper provides a systematic comparison of adding more fitting parameters versus including a charge-equilibration model in machine-learning potentials for silica.

Findings

Adding more parameters improves accuracy and uncertainty.

Including a charge-equilibration model does not enhance performance.

Model complexity influences predictive reliability.

Abstract

In this paper we test two strategies to improving the accuracy of machine-learning potentials, namely adding more fitting parameters thus making use of large volumes of available quantum-mechanical data, and adding a charge-equilibration model to account for ionic nature of the SiO2 bonding. To that end, we compare Moment Tensor Potentials (MTPs) and MTPs combined with the charge-equilibration (QEq) model (MTP+QEq) fitted to a density functional theory dataset of alpha-quartz SiO2-based structures. In order to make a meaningful comparison, in addition to the accuracy, we assess the uncertainty of predictions of each potential. It is shown that adding the QEq model to MTP does not make any improvement over the MTP potential alone, while adding more parameters does improve the accuracy and uncertainty of its predictions.