A deep equivariant neural network approach for efficient hybrid density functional calculations.

Area of Science:

• Computational materials science
• Quantum chemistry
• Machine learning in physics

Background:

• Hybrid density functional calculations are crucial for accurate electronic structure but computationally expensive, limiting their application to large systems.
• Current methods require time-consuming self-consistent field (SCF) iterations, posing a bottleneck for materials discovery.
• Deep learning approaches offer potential to accelerate these calculations.

Purpose of the Study:

• To develop a deep equivariant neural network, DeepH-hybrid, for learning hybrid-functional Hamiltonians.
• To enable accurate and efficient electronic structure calculations for large-scale materials.
• To apply the method to study complex systems like Moiré-twisted materials.

Main Methods:

• Development of DeepH-hybrid, a deep equivariant neural network.
• Training the network to learn the hybrid-functional Hamiltonian directly from material structure.
• Bypassing traditional self-consistent field iterations.
• Extensive experimental validation and testing for reliability, transferability, and efficiency.

Main Results:

• DeepH-hybrid demonstrates high reliability, transferability, and computational efficiency.
• The method successfully predicts electronic structures with hybrid-functional accuracy.
• Application to large-supercell Moiré-twisted materials, including magic-angle twisted bilayer graphene.
• First study on the impact of exact exchange on flat bands in these systems.

Conclusions:

• DeepH-hybrid significantly reduces the computational cost of hybrid-functional calculations.
• The method enables the study of large and complex materials previously inaccessible.
• This work extends deep learning to electronic structure beyond conventional density functional theory.
• Facilitates the development of advanced deep learning-based ab initio methods for materials science.