Prediction and Interpretation of Polymer Properties Using the Graph Convolutional Network.

Area of Science:

• Materials Science
• Computational Chemistry
• Polymer Science

Background:

• Predicting polymer properties is crucial for materials design.
• Machine learning offers a promising avenue for accelerating materials discovery.
• Graph convolutional networks (GCNs) are emerging as powerful tools for molecular property prediction.

Purpose of the Study:

• To develop and evaluate GCN-based machine learning models for predicting thermal and mechanical properties of polymers.
• To compare the performance of GCN representations against traditional fingerprints like ECFP.
• To investigate how GCNs capture polymer structural features relevant to property prediction.

Main Methods:

• Development of GCN models for predicting glass transition temperature (Tg), melting temperature (Tm), density (ρ), and elastic modulus (E).
• Utilizing dimensionality reduction to analyze polymer organization within GCN representation spaces.
• Comparison of GCN performance with extended-connectivity circular fingerprints (ECFP).

Main Results:

• GCN models achieved reliable predictions for Tg, Tm, ρ, and E, with performance varying by property (R² up to ~0.9 for Tg).
• GCN representations demonstrated comparable or superior predictive performance to ECFP, especially when combined with neural network regression (GCN-NN).
• GCNs effectively capture polymer backbone rigidity, a key factor correlated with Tg, and show potential for transferability to other rigidity-dependent properties.

Conclusions:

• GCN-based machine learning models are effective for predicting diverse polymer properties.
• GCNs offer advantages in automatically extracting relevant structural features, such as backbone rigidity.
• The study highlights the capabilities and limitations of GCNs in modeling polymer systems, emphasizing property-dependent performance.