Deep-learning prediction of ab initio parameters

A pipeline, from an MLP to a graph neural network, that infers Hubbard U and related parameters from atomic and electronic descriptors.

Earlier work · 2025

Built at C2N as an extension of BMach: automated data ingestion, training and validation with Python and HPC workflows, aimed at fast prediction of parameters inside DFT+U simulations.

AI/ML Physics Graph neural networks Machine learning DFT+U Hybrid functionals

Research