Transition state structure detection with machine learningś

Abstract

Transition structure calculations via quantum chemistry methods have become a staple in modern chemical reaction research. Yet, success rates in optimizing transition structures rely heavily on rational initial guesses and expert supervision. We develop a machine learning approach that utilizes a bitmap representation of chemical structures to generate high-quality initial guesses for modeling transition states of chemical reactions. The core of the approach comprises a convolutional neural network methodology with a genetic algorithm. An extensive dataset derived from quantum chemistry computations is built, providing sufficient data on which the model can be trained, validated and tested. By applying the method to typical bi-molecular hydrogen abstraction reactions involving hydrofluorocarbons, hydrofluoroethers, and hydroxyl radicals—reactions critical in atmospheric fluoride degradation and global warming potential evaluation, yet extremely challenging to model, we achieve transition state optimizations with an impressive, verified success rate of 81.8% for hydrofluorocarbons and 80.9% for hydrofluoroethers. The reported work demonstrates the effectiveness of employing visual representation in chemical space exploration tasks and opens new avenues for the transition structure modeling.