Machine learning estimation of reaction energy barriers

Abstract

We describe a machine learning approach to approximate reaction energy barriers (E), requiring as input only estimates of geometry and energies of reactants and products. Using the dataset of Grambow, Pattanaik, and Green [Sci. Data 7 (137) (2020)] for reactions involving seven or fewer non-hydrogen atoms, 300 reaction features are computed, and an estimate of E is obtained by fitting a Kernel Ridge Regression (KRR) model with Laplacian kernel to a subset of Density Functional Theory reaction barriers. Our main interest is small energy barriers with the goal of modeling reactions in the interstellar medium and circumstellar envelope. We omitted reactions with E > 40 kcal mol−1 to obtain a subset of 5,276 reactions for 5-fold cross-validation. For this set, the KRR model predicts E with a mean absolute error of 4.13 kcal mol−1 and a root-mean square error of 6.02 kcal mol−1.