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

Area of Science:

• Polymer Science
• Materials Informatics
• Machine Learning

Background:

• Current machine learning models in polymer science often fail due to simplistic molecular representations.
• Macromolecules possess inherent hierarchical and statistical characteristics that are typically overlooked.

Purpose of the Study:

• To develop a novel structure-aware graph convolutional network (GCN) framework for polymer informatics.
• To enhance machine learning efficiency by incorporating the statistical nature and molecular mass distribution (MMD) of polymers.

Main Methods:

• Developed a GCN framework treating polymer samples as statistical ensembles with a hierarchical graph representation.
• Integrated molecular mass distribution (MMD) data to account for sample dispersity.
• Employed an ensemble-based training strategy with topologically realistic graphs generated via kinetic Monte Carlo simulations.

Main Results:

• Achieved over 98% accuracy in classifying complex polymer architectures on synthetic data.
• Predicted glass transition temperatures (T_g) with high accuracy (R² = 0.89 ± 0.01) on experimental data.
• Demonstrated successful learning of the T_g-molar mass relationship by integrating MMD information.

Conclusions:

• The proposed GCN framework offers a physically realistic paradigm for polymer informatics.
• This approach enables more accurate polymer property predictions and accelerates in silico material design.
• The integration of MMD data is crucial for capturing the statistical nature of polymer samples.