Leveraging Theory for Enhanced Machine Learning.

Area of Science:

• Materials Science
• Computational Chemistry
• Polymer Science

Background:

• Machine learning (ML) in materials science faces challenges with limited data and lack of physical understanding.
• The polymer domain is particularly affected by the scarcity of large, curated datasets for ML model training.

Purpose of the Study:

• To address data scarcity and improve the explainability of ML models in the polymer domain.
• To investigate methods for integrating scientific knowledge, specifically imperfect theories, into ML workflows.
• To enhance ML predictions for polymers under both interpolation and extrapolation scenarios.

Main Methods:

• Explored various techniques for incorporating scientific theories into ML models.
• Utilized a polymer system in different solvent qualities as a test case.
• Employed diverse ML models, including Gaussian process regression.
• Compared encoding the functional form of a theory versus its numeric values.

Main Results:

• Encoding the functional form of scientific theories into ML models yielded the best performance.
• Encoding the numeric values of theories also improved predictions compared to theory-agnostic ML.
• The approach enhanced predictive accuracy for smaller datasets and provided a degree of explainability.

Conclusions:

• Integrating scientific knowledge, particularly imperfect theories, is a viable strategy to overcome data limitations in materials ML.
• Encoding the theoretical functional form offers a superior method for knowledge integration, enhancing both prediction accuracy and interpretability.
• This approach holds promise for advancing ML applications in polymer science and other data-scarce materials domains.