Transfer Learning Approach to Multitarget Temperature-Dependent Reaction Rate Prediction.

Area of Science:

• Chemical kinetics
• Computational chemistry
• Atmospheric chemistry and combustion science

Background:

• Accurate prediction of temperature-dependent reaction rate constants is vital for atmospheric chemistry and combustion.
• Automated mechanism generation is hindered by a lack of quality reaction rate data.
• Machine learning shows promise for predicting kinetic data, addressing data gaps.

Purpose of the Study:

• To develop easily accessible, general-purpose, temperature-dependent, and multitarget machine learning models for predicting reaction rates.
• To overcome the bottleneck of limited reaction rate data in detailed chemical mechanisms.

Main Methods:

• Utilized Morgan fingerprints and learned representations from the QM9 dataset.
• Developed a model predicting modified-Arrhenius parameters (A, n, B) instead of direct rate constants.
• Employed an Arrhenius-based loss function for model training.

Main Results:

• The proposed model demonstrated over 35% greater accuracy compared to a baseline feed-forward network (FFN) using Morgan fingerprints.
• Successfully predicted temperature-dependent reaction rate constants for organic compounds.

Conclusions:

• The developed machine learning model offers a significant improvement in predicting reaction rate constants.
• This approach can alleviate data scarcity issues in computational chemistry, benefiting atmospheric and combustion science.