Construction of an Interpretable Regression Model for Yield Prediction and Mechanistic Insight Enabled by Automated...

Area of Science:

• Computational chemistry
• Machine learning in chemistry
• Chemical reaction prediction

Background:

• Machine learning (ML) models are increasingly used for predicting chemical reaction outcomes.
• Current ML models often lack mechanistic interpretability due to arbitrary descriptors.
• Developing interpretable ML models is crucial for advancing chemical research.

Purpose of the Study:

• To develop mechanistically interpretable regression models for predicting chemical reaction yields.
• To demonstrate the utility of a physically motivated energy descriptor for ML models.
• To provide insights into the relationship between reaction intermediates and outcomes.

Main Methods:

• Utilized an energy descriptor based on the energies of potential reaction intermediates.
• Employed the single-component artificial force induced reaction (SC-AFIR) method to autonomously identify reaction intermediates.
• Trained regularized linear regression models using the energy descriptor for reaction yield prediction.

Main Results:

• The energy descriptor enabled the construction of interpretable regression models.
• The developed models achieved good predictive performance for reaction yield (RMSE < 7%).
• Model coefficients provided insights into how intermediate energies influence reaction outcomes.

Conclusions:

• Energy descriptors offer a physically grounded approach for building interpretable ML models in chemistry.
• This method enhances mechanistic understanding of chemical reactions.
• The approach is valuable for predictive tasks and gaining mechanistic insights in chemical research.