Structure-Aware Machine Learning for Polymers: A Hierarchical Graph Network for Predicting Properties From...

Machine learning applications in polymer science are often inefficient due to molecular representations that neglect the inherent hierarchical and statistical nature of macromolecules. This work introduces a structure-aware graph convolutional network (GCN) framework that addresses this limitation by treating polymer samples as statistical ensembles. The approach utilizes a hierarchical graph representation where nodes correspond to monomer units and explicitly integrates molecular mass distribution (MMD) data to account for sample dispersity. A key innovation is an ensemble-based training strategy using topologically realistic graphs generated on-demand via an optimized kinetic Monte Carlo simulation. The model's efficacy was validated on a broad range of tasks. On synthetic data, it achieved more than 98% accuracy in classifying complex polymer architectures. When applied to a large experimental dataset, the model predicts glass transition temperatures (T_g) with high accuracy (R² = 0.89 ± 0.01). Crucially, a fine-tuning experiment demonstrated that the model could successfully learn the physically / chemically grounded relationship between T_g and molar mass by integrating MMD information. This work establishes a robust and physically realistic paradigm for polymer informatics, enabling more accurate property predictions and paving the way for accelerated in silico material design.