Machine learning for analysing ab initio molecular dynamics simulations

Abstract

Post-calculation analyses are often required to extract physical insights from ab initio molecular dynamics simulations. In the present work, we use different machine learning classifiers to take a new perspective on the decomposition reaction of dioxetane. Upon thermally activated decomposition, dioxetane can form products in an electronically excited state and can thus chemiluminesce. Simulated dynamics trajectories exhibit both successful and frustrated dissociations. As an exhaustive and systematic study of the decomposition mechanism “by hand” is beyond feasibility, machine learning models have been employed to study the relevant nuclear distortions governing molecular dissociation. According to all classifiers used in the study, the two sets of geometries differ by the in-phase planarisation of the two formaldehyde moieties. New insights are obtained from this analysis: if both moieties are not planar enough when the dissociation is attempted, it is frustrated and the molecule remains trapped. The postponing of the decomposition reaction by the so-called entropic trap enhances the chemiexcitation efficiency.