Predicting Rate Constants of Alkane Cracking Reactions Using Machine Learning.

Area of Science:

• Theoretical chemistry
• Computational chemistry
• Chemical kinetics

Background:

• Calculating thermal rate constants is crucial for understanding combustion reactions.
• Existing machine learning methods using molecular similarity are limited to specific reaction types like alkane hydrogen abstraction.
• Alkane cracking reactions, involving C-C bond cleavage, require different predictive approaches.

Purpose of the Study:

• To develop a novel feature selection scheme applicable to both bimolecular and unimolecular alkane cracking reactions.
• To enhance machine learning models for accurate prediction of thermal rate constants in combustion.
• To extend the applicability of machine learning in theoretical combustion chemistry.

Main Methods:

• Utilizing molecular descriptors generated by the RDKit software.
• Implementing a new feature selection scheme tailored for reactants and products in cracking reactions.
• Employing machine learning models, including XGB-FNN, to predict rate constants.

Main Results:

• The proposed feature selection scheme accurately predicts rate constants for both alkane hydrogen abstraction and cracking reactions.
• The XGB-FNN model achieved an average deviation of approximately 60% for hydrogen abstraction and 100% for cracking reactions.
• Demonstrated the capability of selected molecular descriptors to represent complex reaction pathways.

Conclusions:

• The developed feature selection method provides a robust approach for predicting rate constants in diverse combustion reactions.
• This work expands the utility of machine learning in theoretical combustion chemistry.
• The proposed descriptors are expected to be applicable to a broader range of chemical reactions.