Deep-Learning Density Functional Theory Hamiltonian in Real Space

Integrating essential physical priors into deep-learning first-principles methods is a critical fundamental problem. Here we demonstrate that the deep learning density functional theory Hamiltonian (DeepH) method can be substantially improved by changing the learning objective to a rotation-invariant and basis-free quantity-the real-space Kohn-Sham potential (named DeepH-R). Benefiting from the enhanced physical priors, DeepH-R achieves substantially improved prediction accuracy and generalization ability compared to previous DeepH approaches. Moreover, DeepH-R provides a more accurate and straightforward route to deep-learning density functional perturbation theory, and enables the training of foundation models of electronic structure with sub-meV accuracy, opening new opportunities for AI-driven materials discovery.