Machine learning-based activity prediction of phenoxy-imine catalysts and its structure-activity relationship study

Area of Science:

• Catalysis Science and Engineering
• Computational Chemistry
• Materials Science

Background:

• Structure-activity relationships (SAR) are crucial for designing efficient catalysts.
• Ti-phenoxy-imine (FI-Ti) catalysts show promise in polymerization reactions.
• Predictive modeling can accelerate catalyst discovery and optimization.

Purpose of the Study:

• To systematically investigate the SAR of 30 FI-Ti catalysts using machine learning (ML).
• To identify key descriptors governing catalyst performance in ethylene polymerization.
• To develop a data-driven framework for designing novel FI-Ti catalysts.

Main Methods:

• Employed machine learning algorithms, with XGBoost showing superior predictive performance (R²=0.998 training, R²=0.859 test).
• Utilized feature importance analysis (SHAP, ICE) to identify critical descriptors and understand nonlinear interactions.
• Applied polynomial feature expansion to capture complex descriptor relationships.

Main Results:

• Identified three composite descriptors (ODI_HOMO_1_Neg_Average GGI2, ALIEmax GATS8d, Mol_Size_L) accounting for over 63% of predictive power.
• Achieved high predictive accuracy on the training set and reasonable performance on the test set.
• Enhanced model interpretability through SHAP and ICE analyses, revealing threshold effects.

Conclusions:

• The study provides a robust ML-based framework for predicting FI-Ti catalyst performance.
• Identified key descriptors offer insights for rational catalyst design.
• Experimental validation and larger datasets are recommended to improve generalizability and confirm findings.