Design and Experimental Validation of a Photocatalyst Recommender based on a Large Language Model

Utilising an extensive library of literature on photocatalytic transformations, we disclose the development of a Machine Learning model for the recommendation of photocatalysts most suitable for reactions of interest. The model is trained on >36,000 such literature examples and uses an architecture inspired by the BERT large language model. Under cross-validation, it can suggest the “correct” photocatalysts with ~90% accuracy. When experimentally tested on four types of out-of-box reactions, this algorithm consistently suggests photocatalysts that give yields competitive to those chosen by human experts, frequently suggesting alternative photocatalysts that are more appealing than the original selected photocatalyst. Altogether, this platform serves as a valuable tool for researchers undertaking reaction optimization programs. The model is free to use at http://photocatals.grzybowskigroup.pl/predict/.