Learning Reactive Islands of the Voter97 System

Abstract

In this paper, we assess the effectiveness of a widely used machine learning technique, support vector machines (SVM) for computing reactive islands in a benchmark system for testing molecular dynamics algorithms, the Voter97 model. Reactive islands are the phase space geometrical structure that mediate chemical reactions dynamics. The Voter97 model contains particular challenges for reaction dynamics methods as the reactant and product potential wells are separated by an intermediate well. We show that SVM can accurately compute the reactive islands in the Voter97 model and we assess the accuracy and the computational effort of the approach by comparing it with brute force methods for computing the reactive islands.