Materials Prediction via Classification Learning.

Area of Science:

• Materials Informatics
• Computational Materials Science
• Machine Learning in Materials Discovery

Background:

• Feature set selection is crucial for accelerated materials discovery in materials informatics.
• Existing feature sets often rely on heuristic justifications for their functional forms.
• Orbital radii offer potential for physically-based feature sets but lack rigorous mathematical grounding.

Purpose of the Study:

• To demonstrate how machine learning can mathematically derive optimal feature sets for materials informatics.
• To provide a rigorous, assumption-free basis for constructing predictive material features.
• To apply these principles to identify novel ductile materials within specific compound classes.

Main Methods:

• Utilized machine learning algorithms to uncover functional forms for feature sets from data.
• Applied the derived feature sets to analyze wide band gap AB compounds and rare earth-main group RM intermetallics.
• Developed predictive models based on the machine learning-derived features.

Main Results:

• Machine learning naturally identified functional forms mirroring commonly used literature features, establishing a mathematical basis.
• Successfully applied the approach to AB compounds as a model system.
• Identified specific rare earth-main group intermetallics (ScCo, ScIr, YCd) predicted to be ductile, contrary to prior predictions.

Conclusions:

• Machine learning provides a powerful, data-driven method for constructing robust feature sets in materials informatics.
• This approach validates and mathematically grounds previously heuristic features.
• The study successfully predicts new ductile materials, advancing accelerated materials design.