Data-Driven Exploration of the Polyethylene Catalyst Chemical Space via Machine Learning

Area of Science:

• Catalysis science
• Materials science
• Computational chemistry

Background:

• Developing highly active polyethylene (PE) catalysts requires understanding complex structure-activity relationships.
• The vast chemical space for catalyst design presents a significant challenge for traditional methods.

Purpose of the Study:

• To develop a data-driven framework for predicting and discovering novel PE catalysts.
• To establish structure-condition-activity relationships using explainable machine learning.
• To generate a large virtual library of potential catalyst candidates.

Main Methods:

• Curated dataset of 507 catalysts (bis(phenoxyimine) and bis(imino)pyridine ligands, seven metals).
• Gradient Boosting Regression (GBR) model for activity prediction (test R² = 0.91).
• SHAP analysis for descriptor importance and combinatorial fragment assembly for virtual library generation.

Main Results:

• GBR model outperformed neural networks in predicting catalyst activity.
• Key descriptors identified: topological (Chi2v), electronic (EState_VSA), and hydrophobic (SlogP_VSA).
• Identified 1090 synthetically accessible candidates with high predicted activity (> 2 × 10⁷ g mol⁻¹ h⁻¹).
• Uncovered metal-dependent design rules for catalyst optimization.

Conclusions:

• A data-driven approach combining ML and virtual screening is effective for catalyst discovery.
• Explainable AI provides insights into catalyst design principles.
• This framework enables the rapid generation of actionable catalyst designs for polyethylene production.