General framework for E(3)-equivariant neural network representation of density functional theory Hamiltonian

Area of Science:

• Computational materials science
• Artificial intelligence in scientific research
• Quantum mechanics and condensed matter physics

Background:

• Deep learning and ab initio calculations offer revolutionary potential for scientific research.
• Designing neural networks with prior knowledge and symmetry is a significant challenge.
• Representing material structures and their properties computationally requires advanced methods.

Purpose of the Study:

• To propose an E(3)-equivariant deep-learning framework for representing density functional theory (DFT) Hamiltonians.
• To develop a method that naturally preserves Euclidean symmetry, even with spin-orbit coupling.
• To enable efficient and accurate electronic structure calculations for large material systems.

Main Methods:

• Developed an E(3)-equivariant deep-learning framework (DeepH-E3) to model DFT Hamiltonians.
• Trained the model on DFT data from small material structures.
• Ensured preservation of Euclidean symmetry and spin-orbit coupling effects.

Main Results:

• Achieved ab initio accuracy in electronic structure calculations.
• Enabled efficient calculations for large-scale supercells (>10^4 atoms).
• Demonstrated sub-meV prediction accuracy with high training efficiency, achieving state-of-the-art performance.

Conclusions:

• The DeepH-E3 method significantly advances deep learning for scientific research.
• This framework makes routine studies of large material supercells feasible.
• Opens new avenues for materials research, including database creation for Moiré-twisted materials.