Transfer learning across different photocatalytic organic reactions

Area of Science:

• Organic Chemistry
• Machine Learning
• Catalysis

Background:

• Predicting catalysts for new reactions is challenging and time-consuming for organic chemists.
• Machine learning (ML) offers a potential solution to replicate expert chemical intuition.

Purpose of the Study:

• To apply a domain-adaptation-based transfer-learning (TL) approach to photocatalysis.
• To improve the prediction of photocatalytic activity for [2+2] cycloaddition reactions.

Main Methods:

• Utilizing transfer learning (TL) to adapt knowledge from photocatalytic cross-coupling reactions.
• Applying ML models to predict the activity of organic photosensitizers (OPSs).

Main Results:

• Successfully transferred knowledge of OPS behavior from cross-coupling to [2+2] cycloaddition reactions.
• Achieved improved prediction of photocatalytic activity compared to conventional ML.
• Demonstrated satisfactory predictive performance with only ten training data points.

Conclusions:

• Transfer learning (TL) significantly enhances ML-based prediction of photocatalytic activity.
• A small dataset is sufficient for effective catalyst exploration using TL.
• This approach shows promise for identifying effective organic photosensitizers (OPSs) for various reactions, including alkene photoisomerization.