Polymer Informatics at Scale with Multitask Graph Neural Networks.

Area of Science:

• Polymer Science
• Materials Informatics
• Computational Chemistry

Background:

• Current polymer screening relies on handcrafted features, which is time-consuming and difficult to scale with growing polymer libraries.
• Manual feature extraction from polymer repeat units is a bottleneck in polymer informatics.

Purpose of the Study:

• To develop and validate a machine learning approach for direct feature extraction from polymer repeat units.
• To demonstrate that deep learning methods can replace handcrafted features for efficient polymer screening.

Main Methods:

• Utilized graph neural networks and multitask learning, advanced deep learning techniques.
• Implemented a method for direct feature learning from polymer repeat units, bypassing manual feature engineering.
• Applied the approach to various polymer property prediction tasks.

Main Results:

• Achieved 1-2 orders of magnitude speedup in feature extraction compared to handcrafted methods.
• Maintained model accuracy for polymer property prediction tasks.
• Demonstrated a viable and cost-effective alternative to manual feature extraction.

Conclusions:

• Directly learning features using machine learning is an efficient alternative to handcrafted features in polymer informatics.
• This approach enables the screening of significantly larger polymer libraries, advancing polymer discovery.
• Facilitates more sophisticated and large-scale screening technologies in the field of polymer informatics.