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

Area of Science:

• Materials Science
• Computational Chemistry
• Polymer Science

Background:

• Polymer density is crucial for material performance and applications.
• Experimental characterization of polymer density is time-consuming and expensive due to the vast chemical space.
• Predictive models are needed to efficiently screen potential polymer structures.

Purpose of the Study:

• To develop and evaluate a machine learning framework for accurate polymer density prediction.
• To compare the performance of various machine learning models, including neural networks, random forest, XGBoost, and GCNNs.
• To enhance model interpretability and understand structure-property relationships governing polymer density.

Main Methods:

• Utilized a dataset of 1432 homopolymers from the PoLyInfo database.
• Implemented and compared four machine learning models: NNs, RF, XGBoost, and GCNNs.
• Employed a directed message passing neural network (D-MPNN) for feature extraction within the GCNN framework.
• Performed experimental validation and utilized SHapley additive exPlanations (SHAP) for interpretability.

Main Results:

• The GCNN model enhanced with D-MPNN achieved superior predictive accuracy (MAE = 0.0497 g/cm3, R2 = 0.8097).
• Experimental validation showed strong agreement with GCNN predictions (relative errors ≤ 4.8%).
• SHAP analysis revealed key functional group contributions to polymer density, enhancing model interpretability.

Conclusions:

• The developed GCNN-D-MPNN framework provides a reliable and scalable computational tool for high-throughput polymer screening.
• The study offers insights into structure-property relationships, specifically concerning polymer density.
• This predictive capability is essential for advancing polymer informatics and accelerating the discovery of new materials.