Deep-Learning Density Functional Theory Hamiltonian in Real Space

Area of Science:

• Computational physics
• Materials science
• Artificial intelligence

Background:

• Integrating physical principles into deep learning for first-principles calculations is crucial.
• Existing deep learning density functional theory Hamiltonian (DeepH) methods have limitations.

Purpose of the Study:

• To enhance the DeepH method by modifying its learning objective.
• To improve prediction accuracy and generalization ability in electronic structure calculations.
• To facilitate AI-driven materials discovery.

Main Methods:

• Developed DeepH-R, a novel deep learning approach using a rotation-invariant, basis-free learning objective: the real-space Kohn-Sham potential.
• Trained foundation models for electronic structure with sub-meV accuracy.

Main Results:

• DeepH-R demonstrated substantially improved prediction accuracy and generalization compared to prior DeepH methods.
• The new method offers a more accurate and direct pathway to deep-learning density functional perturbation theory.
• Enabled the training of highly accurate electronic structure foundation models.

Conclusions:

• DeepH-R represents a significant advancement in integrating physical priors into deep learning for electronic structure.
• The method enhances computational efficiency and accuracy, paving the way for accelerated AI-driven materials discovery.