Molecular Topological Deep Learning for Polymer Property Prediction.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Predicting polymer properties is crucial for efficient polymer design.
• Traditional experimental and density functional theory (DFT) methods are time-consuming and costly.
• Existing deep learning models often overlook high-order and multiscale information in molecular data.

Purpose of the Study:

• To develop a novel deep learning framework, molecular topological deep learning (Mol-TDL), for accurate polymer property prediction.
• To incorporate high-order interactions and multiscale properties into a topological deep learning architecture for enhanced analysis.
• To demonstrate the potential of Mol-TDL in accelerating polymer design and discovery.

Main Methods:

• Representing polymer molecules as multi-scale simplicial complexes.
• Building simplicial neural networks to process topological information.
• Employing a multiscale topological contrastive learning model for pretraining simplex-based message passing.

Main Results:

• The Mol-TDL model significantly outperforms existing learning models on DFT-based and experimental polymer datasets.
• Mol-TDL achieved highly accurate predictions for the glass transition temperatures of eight synthesized advanced polymers, with a mean error of approximately 45 °C.
• The model effectively captures high-order and multiscale information, leading to superior predictive performance.

Conclusions:

• Mol-TDL offers a powerful and efficient approach for predicting polymer properties, overcoming limitations of traditional methods.
• The integration of topological features and multiscale analysis enhances the accuracy and applicability of deep learning in polymer science.
• Mol-TDL shows significant promise for accelerating the discovery and design of novel polymers with desired properties.