Self-learning metabasin escape algorithm for supercooled liquids

Abstract

A generic history-penalized metabasin escape algorithm that contains no predetermined parameters is presented in this work. The spatial location and volume of imposed penalty functions in the configurational space are determined in self-learning processes as the 3N-dimensional potential energy surface is sampled. The computational efficiency is demonstrated using a binary Lennard-Jones liquid supercooled below the glass transition temperature, which shows an O(10(3)) reduction in the quadratic scaling coefficient of the overall computational cost as compared to the previous algorithm implementation. Furthermore, the metabasin sizes of supercooled liquids are obtained as a natural consequence of determining the self-learned penalty function width distributions. In the case of a bulk binary Lennard-Jones liquid at a fixed density of 1.2, typical metabasins are found to contain about 148 particles while having a correlation length of 3.09 when the system temperature drops below the glass transition temperature.