Challenges with\nLiterature-Derived Data in Machine\nLearning for Yield Prediction: A Case Study on Pd-Catalyzed Carbonylation\nReactions

The application of machine learning (ML) to predict reaction\nyields\nhas shown remarkable accuracy when based on high-throughput computational\nand experimental data. However, the accuracy significantly diminishes\nwhen leveraging literature-derived data, highlighting a gap in the\npredictive capability of the current ML models. This study, focusing\non Pd-catalyzed carbonylation reactions, reveals that even with a\ndata set of 2512 reactions, the best-performing model reaches only\nan <i>R</i><sup>2</sup> of 0.51. Further investigations\nshow that the models’ effectiveness is predominantly confined\nto predictions within narrow subsets of data, closely related and\nfrom the same literature sources, rather than across the broader,\nheterogeneous data sets available in the literature. The reliance\non data similarity, coupled with small sample sizes from the same\nsources, makes the model highly sensitive to inherent fluctuations\ntypical of small data sets, adversely impacting stability, accuracy,\nand generalizability. The findings underscore the inherent limitations\nof current ML techniques in leveraging literature-derived data for\npredicting chemical reaction yields, highlighting the need for more\nsophisticated approaches to handle the complexity and diversity of\nchemical data.