Structure-Based Reaction Descriptors for Predicting Rate Constants by Machine Learning: Application to Hydrogen...

Area of Science:

• Computational chemistry
• Chemical kinetics
• Machine learning applications

Background:

• Determining thermal rate constants for combustion reactions is experimentally and theoretically challenging.
• Machine learning (ML) offers a powerful approach for predicting reaction rate constants.

Purpose of the Study:

• To develop quantitative structure-property relationship (QSPR) models for estimating rate constants of hydrogen abstraction reactions.
• To evaluate the performance of XGB, FNN, and XGB-FNN algorithms using novel molecular descriptors.

Main Methods:

• Employed three supervised ML algorithms: XGBoost (XGB), Feedforward Neural Network (FNN), and a hybrid XGB-FNN.
• Utilized Morgan molecular fingerprints and topological indices to represent chemical reactions.
• Developed QSPR models for hydrogen abstraction reactions involving alkanes and free radicals (CH3, H, O).

Main Results:

• The hybrid XGB-FNN model achieved average deviations of 65.4% (alkanes + CH3), 12.1% (alkanes + H), and 64.5% (alkanes + O).
• Performance was comparable or superior to models using activation energy as a descriptor, avoiding computationally expensive ab initio calculations.
• The XGB-FNN models demonstrated good generalization ability, accurately predicting rate constants for different abstraction sites and isomers.

Conclusions:

• Novel molecular descriptors combined with ML, particularly XGB-FNN, provide an accurate and efficient method for predicting combustion reaction rate constants.
• The proposed descriptors show potential for broader application in ML modeling of various chemical reactions.
• This approach offers a viable alternative to traditional methods requiring complex calculations.