Accelerating Polymer Discovery with Uncertainty-Guided PGCNN: Explainable AI for Predicting Properties and...

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Predicting polymer properties from monomer composition is crucial for materials discovery but faces challenges like limited data and poor explainability.
• Existing methods struggle with accurate property prediction and understanding the physical basis of material behavior.
• The vast chemical space of polymers necessitates advanced computational tools for efficient exploration.

Purpose of the Study:

• To develop an interpretable deep learning model for accurate polymer property prediction.
• To address challenges of insufficient data, ineffective representation, and lack of explainability in polymer informatics.
• To enable high-throughput screening and accelerate the discovery of novel polymer materials.

Main Methods:

• Proposed the Polymer Graph Convolutional Neural Network (PGCNN), an interpretable model integrating evidential deep learning.
• Trained the PGCNN using the RadonPy dataset and validated with experimental data.
• Employed uncertainty-guided active learning and global attention mechanisms for enhanced training and interpretability.

Main Results:

• Achieved accurate prediction of various polymer properties.
• Quantified prediction uncertainty and enabled sample-efficient training.
• Identified key functional groups influencing material attributes, aiding mechanistic understanding.
• Successfully screened one million hypothetical polymers for thermal conductivity.

Conclusions:

• The PGCNN model offers a trustworthy and explainable approach for data-driven polymer discovery.
• This work advances mechanistic understanding of polymers through explainable AI.
• The model facilitates high-throughput screening for identifying promising polymer candidates with desired properties.