Predicting reaction conditions for cross-coupling reactions is challenging when a new class of substrate is involved. To predict couplings for a new nucleophile, for example, there is insufficient data (i.e. little to none) that can be used to model plausible reaction conditions. Leveraging data of a related transformation, however, can boost the predictive power of models transferred into the target reaction space. This talk considers Pd-catalyzed cross-coupling reactions of one nucleophile class (source) to identify working conditions for a distinct nucleophile (target) under a classification problem formulation. First, we will show how source model complexity reduction can lead to higher predictivity on target reactions, though ideal performance may not be reached with transfer learning alone. To further accelerate discovery active transfer learning is considered, where the source model is adapted in the target domain. Moreover, the simplicity of models allows full interpretation of the algorithm’s behavior, particularly emphasizing the importance of catalysts for C-N coupling reactions. The qualitative aspects learned from this study will be used to propose generalizable characteristics that will be needed in other transfer learning problems in chemistry, where acquiring data is costly.

Reaction condition exploration for new substrate classes guided by active transfer learning

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.