Identification of High-Reliability Regions of Machine Learning Predictions Based on Materials Chemistry

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning

Background:

• Machine learning (ML) applications in materials design are advancing rapidly.
• A key challenge is understanding the reliability of ML predictions, particularly with small datasets common in materials science.

Purpose of the Study:

• To develop methods for assessing the reliability of ML predictions in materials science.
• To demonstrate how to identify reliable prediction regions and extract physical understanding from ML models.

Main Methods:

• Utilized ML predictions for formation energy and band gap of transparent conductor oxides, dilute solute diffusion, and perovskite properties.
• Constructed a convex hull in feature space to delineate regions of reliable ML predictions.
• Analyzed systems within the convex hull to extract physical insights.

Main Results:

• A convex hull in feature space effectively identifies regions with highly reliable ML predictions.
• Analysis of enclosed systems yields valuable physical understanding.
• Materials obeying physical principles are likely similar and show strong feature-property relationships.
• Including diverse material classes in training data does not improve accuracy; narrow, similar material classes yield reliable ML results.

Conclusions:

• The convex hull method is a powerful tool for validating ML predictions in materials science.
• Focusing ML models on narrow classes of similar materials enhances prediction reliability.
• Integrating physical principles with ML can improve materials design and discovery.