Database optimization for empirical interatomic potential models

TL;DR

This paper introduces an objective function for optimizing empirical interatomic potential fitting databases based on testing set errors, improving transferability and model accuracy for complex crystal structures.

Contribution

It proposes a novel method to optimize fitting database weights and structure inclusion using testing set errors, enhancing potential transferability.

Findings

Different optimal databases are found depending on the objective function used.

The method improves the transferability of Lennard-Jones potentials for Ti.

It enables comparison and assessment of fitting databases.

Abstract

Weighted least squares fitting to a database of quantum mechanical calculations can determine the optimal parameters of empirical potential models. While algorithms exist to provide optimal potential parameters for a given fitting database of structures and their structure property functions, and to estimate prediction errors using Bayesian sampling, defining an optimal fitting database based on potential predictions remains elusive. A testing set of structures and their structure property functions provides an empirical measure of potential transferability. Here, we propose an objective function for fitting databases based on testing set errors. The objective function allows the optimization of the weights in a fitting database, the assessment of the inclusion or removal of structures in the fitting database, or the comparison of two different fitting databases. To showcase this technique, we consider an example Lennard-Jones potential for Ti, where modeling multiple complicated crystal structures is difficult for a radial pair potential. The algorithm finds different optimal fitting databases, depending on the objective function of potential prediction error for a testing set.