Predicting Materials Properties with Little Data Using Shotgun Transfer Learning.

Area of Science:

• Materials Science
• Machine Learning
• Computational Chemistry

Background:

• Growing demand for machine learning (ML) surrogate models for materials properties.
• Digital transformation in materials science yields databases, but insufficient data volume and diversity limit ML advancements.
• Transfer learning (TL) offers a solution for limited materials data by leveraging physically interrelated property types.

Purpose of the Study:

• To develop a pretrained model library, XenonPy.MDL, to facilitate widespread adoption of transfer learning in materials science.
• To demonstrate the effectiveness of transfer learning in building predictive models with limited datasets.
• To explore cross-disciplinary transferability of learned features in materials science.

Main Methods:

• Development of the XenonPy.MDL library, comprising over 140,000 pretrained models for diverse material types and properties.
• Application of transfer learning by pretraining models on related proxy properties using abundant data.
• Repurposing machine-acquired features from pretrained models for target property prediction with limited data.

Main Results:

• XenonPy.MDL release with over 140,000 pretrained models for small molecules, polymers, and inorganic crystalline materials.
• Successful model building with as few as dozens of data points using transfer learning.
• Enhanced extrapolative prediction capabilities through strategic model transfer.
• Discovery of non-trivial transferability across different material types and scientific disciplines (e.g., small molecules to polymers, organic to inorganic chemistry).

Conclusions:

• Transfer learning significantly enhances predictive modeling in materials science, especially with limited data.
• The XenonPy.MDL library provides a valuable resource for accelerating materials discovery and design.
• Transfer learning reveals fundamental interconnections between different material classes and chemical domains.