Machine learning-based design of pincer catalysts for polymerization reaction

We present a generic machine learning (ML) workflow for rapid screening of 3d-transition metal pincer catalysts utilizing bis(imino)pyridine ligands for homogeneous polymerization reactions. In this study, we compiled a dataset comprising over 503 catalysts. We identified the best quantitative structure–activity relationship (QSAR) models among hundreds, including both categorical and continuous, to predict catalytic activity, combining molecular fingerprints and descriptors. Top-performing models achieve high accuracy with coefficient of determination (R²) of nearly 80% for test sets; further validated by 5-fold cross-validation. We devised a mechanism to filter the outliers, such as ligands with meta-substituted groups. Steric map and QM-based descriptors like NBO charge, spin, HOMO-LUMO gap, and Fukui indices are identified as important features. In silico catalysts were generated by enumeration at ortho positions with bulky ligands like CHPh₂, CH(p-F-Ph)₂; directly linked with buried volume. This work supports increased innovation and accelerated development of new pincer catalysts using ML-based predictive modeling.