Physics-Based Feature Makes Machine Learning Cognizing Crystal Properties Simple.

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Machine learning (ML) accelerates materials discovery, with feature engineering being crucial.
• Existing descriptors can be computationally expensive or lack physical insight.

Purpose of the Study:

• To introduce a novel, low-cost descriptor, electron probability waves (EPW), for ML in materials science.
• To evaluate EPW's efficacy in predicting magnetic and electronic properties of materials.

Main Methods:

• Extracting EPW descriptors from high-symmetry points in the Brillouin zone based on electronic structures.
• Utilizing 10-fold cross-validation for model training and performance assessment.
• Comparing EPW-based models against established crystal graph features-based deep learning models.

Main Results:

• Achieved 0.92 accuracy (ACC) and 0.83 area under the receiver operating characteristic curve (AUC) in distinguishing ferromagnetic/antiferromagnetic materials.
• Demonstrated EPW's effectiveness in classifying metals/semiconductors and determining semiconductor bandgap types (direct/indirect).
• EPW-based models showed performance comparable to deep learning models using crystal graph features.

Conclusions:

• EPW is a computationally efficient and physically meaningful descriptor for ML-driven materials design.
• EPW captures essential electronic structure information relevant to material properties like magnetism and conductivity.
• The EPW descriptor offers a promising alternative for accelerating the discovery of novel functional materials.