Polymer informatics with multi-task learning.

Area of Science:

• Polymer Science
• Materials Informatics
• Machine Learning

Background:

• Data-driven methods are revolutionizing polymer development.
• Current surrogate models often underutilize available data, ignoring correlations between properties.
• Sparse datasets limit the predictive power of conventional models.

Purpose of the Study:

• To demonstrate the effectiveness of multi-task learning (MTL) for polymer property prediction.
• To exploit inherent correlations within polymer datasets.
• To develop interpretable models for rational polymer design.

Main Methods:

• Deep learning multi-task architectures were employed.
• Data for 36 properties of over 13,000 polymers were utilized.
• MTL models were compared against conventional single-task learning models.

Main Results:

• The multi-task approach significantly improved accuracy and efficiency compared to single-task models.
• The developed models demonstrated scalability and amenability to transfer learning.
• Interpretable chemical rules explaining property trends were successfully extracted.

Conclusions:

• Multi-task learning offers a potent strategy for enhancing polymer property prediction.
• This approach facilitates the rational design of application-specific polymers.
• The interpretability of MTL models paves the way for deeper scientific understanding and targeted material development.