Reproduce “Machine learning Hubbard parameters with equivariant neural networks” on OpenResearch

About this paper

Martin Uhrin, Austin Zadoks, Luca Binci, Nicola Marzari, Iurii Timrov

An equivariant neural network model predicts self-consistent on-site U and inter-site V Hubbard parameters for DFT+U+V calculations by learning from computationally expensive first-principles Density-Functional Perturbation Theory (DFPT) data. The model achieves high accuracy and transferability across various materials and oxidation states, significantly reducing the computational cost of obtaining these crucial parameters.

Key insights

• Equivariant neural networks can accurately learn complex relationships between electronic structure (via occupation matrices) and Hubbard parameters, making input U/V parameters largely redundant.
• The machine learning model can accelerate the self-consistent determination of Hubbard parameters by improving upon early-terminated DFPT calculations, reducing the required number of expensive iterations.
• The model exhibits robust transferability across different oxidation states, compositions, and crystal structures, suggesting its applicability to new materials not seen during training.

Summary by alphaXiv.