Predicting Rate Constants of Hydrogen Abstraction Reactions between OH/HO₂ and Alkanes by Machine Learning Models

Area of Science:

• Combustion Chemistry
• Computational Chemistry
• Chemical Kinetics

Background:

• Hydrogen abstraction reactions are crucial in fuel oxidation.
• Experimental and theoretical determination of rate constants is challenging due to high radical reactivity.
• Machine learning presents a viable alternative for predicting these rate constants.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting rate constants of hydrogen abstraction reactions.
• To investigate the performance of XGB, FNN, and a hybrid XGB-FNN approach.
• To assess the models' ability to predict reactions involving larger alkanes and differentiate reactivity at various sites.

Main Methods:

• Utilized three machine learning algorithms: XGBoost (XGB), Feedforward Neural Network (FNN), and a hybrid XGB-FNN.
• Selected six key descriptors based on correlation coefficients, descriptor importance, and clustering heatmaps.
• Trained and tested models on rate constants for hydrogen abstraction reactions between alkanes and hydroxyl (OH) or hydroperoxyl (HO2) radicals.

Main Results:

• The XGB-FNN hybrid model demonstrated the most robust performance.
• Achieved average deviations of 89.13% for alkanes + OH reactions and 190.93% for alkanes + HO2 reactions.
• The model exhibited extrapolation capabilities for larger alkanes and could distinguish reactivity based on hydrogen atom abstraction sites.

Conclusions:

• The XGB-FNN model is effective for predicting thermal rate constants of hydrogen abstraction reactions in fuel oxidation.
• The developed model shows promise for understanding complex combustion chemistry and guiding experimental efforts.