Accurate Prediction of Polymerization Performance for Metallocene Catalysts via a Dual-Path Neural Network and Local Feature Learning.

Metallocene catalysts, distinguished by their well-defined active centers and tunable coordination geometries, are pivotal in the homopolymerization of propylene to produce polypropylene with tailored properties. However, the rational design of such catalysts remains challenging due to the complex coupling between ligand structures and polymerization conditions. Conventional trial-and-error approaches are inefficient, while existing machine learning (ML) models often overlook critical ligand descriptors, limiting their generalization for industrial use. To address this, we developed a hybrid ML framework that integrates both reaction parameters and catalyst structural features. A dual-path neural network processes numerical and categorical inputs separately to avoid feature semantic distortion, enabling accurate predictions of catalyst activity (R2 = 0.9201) and number-average molecular weight (R2 = 0.9133). For the narrow molecular weight distribution typical of metallocene-derived polypropylene─a characteristic leading to compact, locally correlated data─a k-nearest neighbor regression model achieved superior performance (R2 = 0.9766) by effectively capturing local sample relationships. Both models outperformed eight other benchmark ML algorithms across all metrics. This work provides a robust, interpretable computational strategy for linking catalyst chemistry to polymer properties, offering a practical tool for the targeted design and scalable application of high-performance polypropylene materials.