Directed message passing neural networks enhanced graph convolutional learning for accurate polymer density prediction

Polymer density is a critical factor influencing material performance and industrial applications, and it can be tailored by modifying the chemical structure of repeating units. Traditional polymer density characterization methods rely heavily on domain expertise; however, the vast chemical space comprising over one million potential polymer structures makes conventional experimental screening inefficient and costly. In this study, we proposed a machine learning framework for polymer density prediction, rigorously evaluating four models: neural networks (NNs), random forest (RF), XGBoost, and graph convolutional neural networks (GCNNs). Using a curated dataset of 1432 homopolymers from the PoLyInfo database, our comparative analysis indicated that the GCNN model enhanced by a directed message passing neural network (D-MPNN) for feature extraction achieves superior predictive accuracy (mean absolute error, MAE = 0.0497 g/cm3; coefficient of determination, R2 = 0.8097) relative to the other models. Experimental validation on six polymers demonstrated strong agreement between measured densities and GCNN predictions with relative errors not exceeding 4.8%. In-depth error analysis combined with SHapley additive exPlanations (SHAP) and subgroup evaluations reveals fundamental relationships between specific functional groups and polymer density, improving the model's interpretability. Furthermore, t-SNE visualization and kernel density estimation (KDE) analysis revealed that the D-MPNN framework effectively captures critical chemical space features governing density, explaining the GCNN's superior performance. This work establishes a reliable and scalable computational framework for high-throughput polymer screening, providing scientific insights into structure-property relationships while positioning density prediction as a prerequisite for incorporating density-dependent properties into polymer informatics and accelerating the discovery of novel polymer materials.