Machine learning Hubbard parameters with equivariant neural networks.

Area of Science:

• Computational Materials Science
• Quantum Chemistry
• Machine Learning in Physics

Background:

• Density-functional theory with extended Hubbard functionals (DFT+U+V) accurately describes complex materials with transition-metal or rare-earth elements.
• Accurate prediction of DFT+U+V requires precise on-site (U) and inter-site (V) Hubbard parameters, often determined by computationally expensive first-principles calculations.
• Existing methods for parameter determination include semi-empirical tuning or rigorous but time-consuming ab initio calculations.

Purpose of the Study:

• To develop a machine learning model for rapid and accurate prediction of Hubbard U and V parameters.
• To circumvent the computational cost associated with traditional DFT+U+V parameter calculations.
• To accelerate materials discovery and design through high-throughput computational screening.

Main Methods:

• Utilized equivariant neural networks trained on atomic occupation matrices as descriptors.
• Descriptors capture electronic structure, local chemical environment, and oxidation states.
• Targeted prediction of Hubbard parameters computed via self-consistent linear-response calculations within density-functional perturbation theory (DFPT).

Main Results:

• The machine learning model achieved mean absolute relative errors of 3% for Hubbard U and 5% for Hubbard V parameters.
• The model was trained on data from 12 diverse materials, demonstrating robustness across different crystal structures and compositions.
• Predictions closely approach the accuracy of DFPT calculations but with significantly reduced computational overhead.

Conclusions:

• The developed machine learning model offers a computationally efficient and accurate alternative for predicting Hubbard parameters.
• Its high transferability facilitates accelerated materials discovery and design for technological applications.
• This approach bypasses expensive self-consistent DFT or DFPT protocols, enabling faster materials design.