Predicting reaction conditions from limited data through active transfer learning.

Area of Science:

• Organic Chemistry
• Machine Learning
• Computational Chemistry

Background:

• Accelerating the discovery of new chemical reactions is crucial for scientific advancement.
• Machine learning, particularly transfer and active learning, offers a promising approach by leveraging existing data and new experiments.

Purpose of the Study:

• To demonstrate how specifically tuned machine learning models can expand the scope of palladium-catalyzed cross-coupling reactions.
• To investigate the effectiveness of model transfer and active learning for predicting reactions with novel nucleophiles and reagent combinations.

Main Methods:

• Utilized random forest classifiers for machine learning models.
• Employed model transfer learning with closely related reaction mechanisms and substrates.
• Implemented a model simplification scheme for enhanced predictivity.
• Introduced an active transfer learning strategy for challenging reaction targets.

Main Results:

• Model transfer proved effective for related reaction mechanisms and substrates, even with limited data.
• A simplified model scheme achieved comparable predictivity for new nucleophiles and unseen reagent combinations.
• Active transfer learning significantly improved predictions in challenging cases where standard model transfer offered minimal benefit.

Conclusions:

• Tuned machine learning models, especially random forest classifiers, can successfully expand the applicability of Pd-catalyzed cross-coupling reactions.
• Model simplification and active transfer learning are effective strategies for improving predictive performance with novel chemical entities.
• Simple, interpretable models are key for generalizability and performance in active transfer learning for chemical synthesis.