A local Bayesian optimizer for atomic structures

TL;DR

This paper introduces a local Bayesian optimization method using Gaussian Processes for atomic structures, demonstrating improved efficiency over traditional algorithms across various systems.

Contribution

It presents a novel Bayesian Gaussian Process-based local optimizer for atomic structures, enhancing optimization speed and accuracy compared to standard methods.

Findings

Outperforms conjugate gradient and BFGS in various atomic systems

Uses surrogate potential energy surfaces for faster optimization

Hyperparameter tuning can further improve speed

Abstract

A local optimization method based on Bayesian Gaussian Processes is developed and applied to atomic structures. The method is applied to a variety of systems including molecules, clusters, bulk materials, and molecules at surfaces. The approach is seen to compare favorably to standard optimization algorithms like conjugate gradient or BFGS in all cases. The method relies on prediction of surrogate potential energy surfaces, which are fast to optimize, and which are gradually improved as the calculation proceeds. The method includes a few hyperparameters, the optimization of which may lead to further improvements of the computational speed.