The structure–property relationship of metallocene-based ethylene oligomerization catalysts using DFT and graph neural networks

Ethylene oligomerization catalysts have been extensively studied in both experimental and simulation contexts, yet a molecular-level understanding of structure-property relationship remains far from full understanding. Herein, an applicable strategy for the design of ligands of ethylene oligomerization catalysts is proposed. Density functional theory (DFT) and 3D graph neural networks (3D GNNs) have been combined to establish the relationship between the catalyst structure and its property. A series of titanium-based metallocene catalysts with different ligands were designed and calculated using DFT to establish a dataset. The catalyst prediction model was constructed using 3D GNNs, and a weighted removal approach was used to compare output results and study the impact of different ligand structures on the oligomerization selectivity represented by the energy barrier difference between β-hydrogen transfer and the fourth ethylene insertion. The R² values of the energy barrier difference predictions by four 3D GNNs were 0.93-0.96, indicating good predictive accuracy of the graph network models. Using the graph neural network explanation algorithms, we investigated the influence of different substructures within the ligands on trimerization selectivity. Based on the training and explanation results of the model, an external validation set is designed, and the R² is 0.92, suggesting the generalization ability of the model. This enabled a molecular-level study of the relationship between the structure of the titanocene catalyst and its properties, providing guidance for the design of new catalyst structures.