Machine-Learning-Based Design of Metallocene Catalysts for Controlled Olefin Copolymerization

Polyolefins are versatile materials for various purposes, but their functionality should be fine-tuned for target applications including the mitigation of adverse environmental impacts. Producing such polymers with desired properties requires catalysts that can control polymerization at an atomistic level. However, complex reaction mechanisms and very limited experimental data make it difficult to design new efficient catalysts using conventional computational and data-driven approaches. Here, we present a pragmatic strategy based on data-efficient predictive models combined with a genetic algorithm to design new catalysts for controlled ethylene/hexene copolymerization. By deriving the chemically intuitive descriptors from the mechanistic analysis of the polymerization, we achieved the promising predictive models with small data applicable to various core structures and different experimental conditions, respectively. We screened catalysts through a virtual screening scheme combining a genetic algorithm and predictive models using chemically intuitive descriptors and considered their synthesizability through the manual inspections of experts. As a result, we successfully designed nine catalysts with desired comonomer ratios and diverse core structures.