Machine Learning Transition State Geometries and Applications in Reaction Property Prediction

Abstract

The calculation of transition state (TS) geometries is essential for understanding reaction mechanisms and rational synthetic methodology design. However, traditional methods like density functional theory are often too computationally expensive for large-scale TS identification and are significantly slower than high-throughput experimental screening methods. Recent advancements in machine learning (ML) offer promising alternatives, enabling the direct prediction of TS geometries, reducing the reliance on expensive quantum mechanical (QM) calculations, and affording predictions ahead of experiments. The works explored here include the broader application of ML in reaction property prediction, emphasizing how accurate TS geometries can serve as vital input data to improve model accuracy. A comprehensive review of ML methods developed to explicitly predict TS geometries is then presented, with attention to their application in downstream tasks, such as energy barrier calculations, and their use as initial structures for further optimization via QM methods. Finally, a critical evaluation of the accuracy and limitations of existing TS prediction methods is discussed, highlighting challenges that impede wider adoption and areas where further research is needed.