Prediction of Hydrogen Abstraction Rate Constants at the Allylic Site between Alkenes and OH with Multiple Machine Learning Models
Lei Zhang,Lili Ye,Fan Wang,Wei Gao,Jinhui Yu,Lidong Zhang
DOI: https://doi.org/10.1021/acs.jpca.3c06917
2024-01-18
The Journal of Physical Chemistry A
Abstract:Hydrogen abstraction reactions between hydrocarbons and hydroxyl radicals are important propagation steps in radical chain reactions, playing a crucial role in atmospheric and combustion chemistry. This study focuses on predicting the rate constants of the prototype of the reaction class of hydrogen abstractions, i.e., the primary allylic hydrogen abstraction from alkenes by the OH radical, via utilizing machine learning (ML) methods. Specifically, three distinct models, namely, feedforward neural network (FNN), support vector regression (SVR), and Gaussian process regression (GPR), have been employed to construct robust ML models for prediction. We proposed a novel strategy that seamlessly integrates descriptor preprocessing, a pairwise linear correlation analysis, and a model-specific Wrapper method to enhance the effectiveness of the feature selection procedure. The selected feature subset was then evaluated using two cross-validation techniques, i.e., leave-one-group-out (LOGO) and K-fold cross-validations, for each of the three ML models (FNN, SVR, and GPR) to assess their predictive and stability performance. The results demonstrate that the FNN model, trained with seven representative descriptors, achieves superior performance compared to the other two methods. For the FNN model, the average percentage deviation is 39.06% on the test set by performing LOGO cross-validation, while the repeated 10-fold cross-validation achieves a percentage prediction deviation of 19.1%. Two larger alkenes with 10 carbons were selected to test the prediction performance of the trained FNN model on primary allylic hydrogen abstraction. Results show that the kinetic predictions follow well the modified three-parameter Arrhenius equation, indicating the reliable performance of FNN in predicting hydrogen abstraction rate constants, especially for the primary allylic site. Hopefully, this work can shed useful light on the application of ML in generating chemical kinetic parameters of hydrocarbon combustion chemistry.
chemistry, physical,physics, atomic, molecular & chemical