An overview of reaction outcome prediction with physics-based and data-driven methods

Abstract

The prediction of reaction outcomes is a longstanding challenge in chemistry, with the ability to do so serving as a direct reflection of our understanding of chemical reactivity. Accurately predicting reaction products is crucial not only for synthetic planning but also for designing reaction pathways and experiments in silico. This review explores the diverse methodologies used to predict reaction outcomes, which can be broadly divided into two main categories. Some approaches predict reaction products and their likelihoods in a single step, while others break the task into two distinct parts: candidate enumeration and the subsequent prediction of product likelihoods. We examine both data-driven methods, such as graph-based and sequence-generation models, and physics-based methods, including potential energy surface exploration and reactive molecular dynamics. In addition, we discuss quantitative predictions of reaction selectivity, regioselectivity, stereoselectivity, and yield. This review summarizes trends and advances in reaction outcome prediction and briefly outlines future directions for the field.