Multi-Task Learning in Homogeneous Catalysis: A Case Study for Predicting the Catalytic Performance in Ethylene Polymerization

Abstract

This study focuses on training a multi-task learning (MTL) type machine learning (ML) model to predict diverse catalytic performance of 195 bis(imino)pyridine transition metal complexes toward ethylene polymerization, with comparison to their single-task learning (STL) counterparts. The CatBoost MTL model outperforms all other models, showing predictions and generalization errors for the properties of catalytic activity (R_t²=0.741, R² = 0.985, Q² = 0.600), molecular weight (R_t²=0.873, R² = 0.997, Q² = 0.846), molecular weight distribution (R_t²=0.831, R² = 0.999, Q² = 0.839), and melting temperature (R_t²=0.813, R² = 0.992, Q² = 0.625) of the produced polymer. The interpretation of the model reveals that complexes with electron-donating groups, simple alkyl groups (such as methyl groups etc.), and a higher degree of unsaturation (presence of double or triple bonds) positively influence the predicted properties. Subsequently, providing insights into the underlying mechanisms of variation in catalytic performance, new complexes are designed with superior catalytic performances.