Assessing the efficiency of first-principles basin-hopping sampling

TL;DR

This paper systematically analyzes the performance of first-principles basin-hopping algorithms in identifying low-energy isomers of small Si and Cu clusters, providing insights and guidelines for optimizing sampling efficiency.

Contribution

It offers a detailed performance analysis of basin-hopping methods with simple move classes, revealing universal rules for move settings independent of cluster chemistry.

Findings

Performance limited by revisits to dominant isomers

Simple move classes yield very good sampling efficiency

Rules-of-thumb for move settings are broadly applicable

Abstract

We present a systematic performance analysis of first-principles basin-hopping (BH) runs, with the target to identify all low-energy isomers of small Si and Cu clusters described within density-functional theory. As representative and widely employed move classes we focus on single-particle and collective moves, in which one or all atoms in the cluster at once are displaced in a random direction by some prescribed move distance, respectively. The analysis provides detailed insights into the bottlenecks and governing factors for the sampling efficiency, as well as simple rules-of-thumb for near-optimum move settings, that are intriguingly independent of the distinctly different chemistry of Si and Cu. At corresponding settings, the observed performance of the BH algorithm employing two simple, general-purpose move classes is already very good, and for the small systems studied essentially limited by frequent revisits to a few dominant isomers.