Driving Data Generation in Molecular Discovery Through Development of Benchmark Reinforcement Learning Environments

Abstract

Molecular and materials discovery is an area of great technological significance that continues to greatly benefit from the data revolution. Here we present a series of benchmark reinforcement learning environments that allow mixing and matching different design goals, representation with different reinforcement learning agents. These provides both a set of standards to evaluate different reinforcement learning algorithms applied to molecular design but also a standarised way of generating molecular datasets capturing the molecular design process.