Adaptive Learning Reaction Pathways Search

Computational simulations have become essential for understanding reaction mechanisms, identifying optimal catalysts, studying materials, and discovering chemical pathways. Even with the current advanced computational resources, determining competing reaction pathways and the associated transition states is extremely challenging. The problem is further exasperated as the system size increases due to the increasing number of comparable energy isomers and intermediates. In this work, we present a novel computational protocol to efficiently explore the potential energy surface and identify reaction pathways, by combining an adaptive learning global optimization procedure and neural networks for predicting transition states. For a given reaction, the global optimizer is first used to obtain putative low-lying minima in the basins of reactants and products. In the second step, generative adversarial networks (GAN) are used to estimate the transition states that connect any two initial and final states of a given reaction. This computational interface (GlobOptRx) can quickly identify the transition state structures along a reaction pathway and aid in the discovery of new reactions, determine generalized systems descriptors and facilitate the construction of kinetic models.