Representing Polymers as Periodic Graphs with Learned Descriptors for Accurate Polymer Property Predictions.

Area of Science:

• Polymer Science
• Machine Learning
• Computational Chemistry

Background:

• Predicting polymer properties computationally accelerates materials discovery.
• Current machine learning models struggle to represent polymers' periodic structures and require manual feature engineering.
• Developing automated descriptor generation for polymers is a key challenge in polymer informatics.

Purpose of the Study:

• To develop an advanced machine learning framework for accurate *a priori* prediction of polymer properties.
• To address the limitations in capturing polymer periodicity and automate descriptor generation.
• To enhance the efficiency and accuracy of polymer property prediction for accelerated materials development.

Main Methods:

• Utilized a novel periodic polymer graph representation to encode structural periodicity.
• Coupled this representation with a message-passing neural network (MPNN) for deep learning.
• Employed graph deep learning to automatically learn chemically relevant polymer descriptors, eliminating manual feature design.

Main Results:

• Achieved state-of-the-art performance on 8 out of 10 diverse polymer property prediction tasks.
• Demonstrated the effectiveness of the periodic graph representation in capturing polymer structure.
• Showcased the ability of the MPNN to automatically learn optimal polymer descriptors.

Conclusions:

• The proposed method significantly advances predictive capabilities for polymer properties.
• Learned descriptors optimized for polymer structures lead to superior prediction accuracy.
• This approach offers a powerful tool for accelerating polymer discovery and development through machine learning.