Barrier height prediction by machine learning correction of semiempirical calculations

Abstract

Different machine learning (ML) models are proposed in the present work to predict DFT-quality barrier heights (BHs) from semiempirical quantum-mechanical (SQM) calculations. The ML models include multi-task deep neural network, gradient boosted trees by means of the XGBoost interface, and Gaussian process regression. The obtained mean absolute errors (MAEs) are similar or slightly better than previous models considering the same number of data points. Unlike other ML models employed to predict BHs, entropic effects are included, which enables the prediction of rate constants at different temperatures. The ML corrections proposed in this paper could be useful for rapid screening of the large reaction networks that appear in Combustion Chemistry or in Astrochemistry. Finally, our results show that 70% of the bespoke predictors are amongst the features with the highest impact on model output. This custom-made set of predictors could be employed by future delta-ML models to improve the quantitative prediction of other reaction properties.