The discovery of general catalysts has historically relied on determining the most widely applicable catalyst from ex-tensive experimentation. However, certain catalysts with potentially broad applicability remain underexplored due to limited reporting and small, biased datasets. In this study, we apply machine learning (ML) techniques tailored to these data-constrained environments to uncover general catalysts from limited historical data. Our approach surfaced several high-performing structures, including a rarely studied imidazolidinone bearing a benzyl-protected indole substituent. Despite minimal precedent, this scaffold showed broad generality upon experimental validation involving selected experiments from eight different reaction types. Moreover, retrospective modeling using only pre-2005 data further demonstrated that ML could have predicted the rise of now-prominent catalysts, such as diarylprolinol silyl ethers and imidazolidinones, well before their widespread adoption. These findings highlight the potential of ML to accelerate catalyst discovery, prioritize overlooked scaffolds, and compress timelines traditionally spanning decades. By integrating bias-aware modeling with focused experimentation, this strategy offers a more predictive and efficient pathway to identifying broadly effective catalysts.
Li et al. (Mon,) studied this question.