A universal machine learning model for the electronic density of states

Area of Science:

• Computational materials science
• Machine learning in chemistry
• Electronic structure theory

Background:

• Machine learning interatomic potentials predict atomic configurations with high accuracy.
• Existing models focus on energy and forces, not explicit electronic structure.
• Predicting electronic properties is crucial for understanding material behavior.

Purpose of the Study:

• To develop a universal machine learning model for predicting electronic structure.
• To focus on predicting the electronic density of states (DOS) and bandgaps.
• To assess the model's performance in finite-temperature simulations.

Main Methods:

• Developed PET-MAD-DOS, a transformer model based on the Point Edge Transformer (PET) architecture.
• Trained the model on the Massive Atomic Diversity (MAD) dataset.
• Evaluated predictions on diverse external datasets and specific material systems (LPS, GaAs, HEA).

Main Results:

• PET-MAD-DOS accurately predicts electronic density of states (DOS) and bandgaps.
• The model achieves semi-quantitative agreement for ensemble-averaged DOS and electronic heat capacity.
• Fine-tuning with limited data yields performance comparable or superior to bespoke models.

Conclusions:

• Universal machine learning models can predict explicit electronic structure.
• PET-MAD-DOS offers a fast and accurate method for electronic property prediction.
• Fine-tuning enhances model performance for specific material systems, demonstrating broad applicability.