[Paper Review] Efficient hybrid density functional calculation by deep learning

Zechen Tang, He Li+8 more|arXiv (Cornell University)|Feb 16, 2023

Machine Learning in Materials ScienceMaterials Science3 citations

TL;DR

This paper introduces DeepH-hybrid, a deep equivariant neural network method that learns hybrid-functional Hamiltonians from small-scale self-consistent field calculations and enables efficient, accurate electronic-structure simulations for large systems. The method achieves orders-of-magnitude speedup over conventional DFT while maintaining high accuracy, demonstrated by predicting HSE band structures of magic-angle twisted bilayer graphene with 11,164 atoms, revealing a 10-fold increase in flat band width due to exact exchange.

ABSTRACT

Hybrid density functional calculation is indispensable to accurate description of electronic structure, whereas the formidable computational cost restricts its broad application. Here we develop a deep equivariant neural network method (named DeepH-hybrid) to learn the hybrid-functional Hamiltonian from self-consistent field calculations of small structures, and apply the trained neural networks for efficient electronic-structure calculation by passing the self-consistent iterations. The method is systematically checked to show high efficiency and accuracy, making the study of large-scale materials with hybrid-functional accuracy feasible. As an important application, the DeepH-hybrid method is applied to study large-supercell Moiré twisted materials, offering the first case study on how the inclusion of exact exchange affects flat bands in the magic-angle twisted bilayer graphene.

Motivation & Objective

• To overcome the high computational cost of hybrid density functional theory (DFT) in large-scale materials simulations.
• To extend deep learning-based electronic structure methods (DeepH) beyond local and semilocal functionals to non-local hybrid functionals within the generalized Kohn-Sham framework.
• To enable efficient, accurate prediction of electronic properties—especially band structures and band gaps—of complex superstructures like twisted bilayer graphene using hybrid functionals.
• To investigate the impact of exact exchange on flat band physics in magic-angle twisted bilayer graphene (MATBG) at scale.
• To validate the generalizability and transferability of the learned Hamiltonian across different twist angles and atomic configurations.

Proposed method

• A deep E(3)-equivariant neural network is trained to predict the hybrid-functional Kohn-Sham Hamiltonian $ H_{\text{DFT}}^{\text{hyb}} $ as a function of atomic structure, using self-consistent field data from small reference systems.
• The method leverages the nearsightedness principle in localized atomic basis sets, enabling effective representation of non-local exchange-correlation effects in hybrid functionals via learned neural network parameters.
• The trained model bypasses iterative self-consistent field (SCF) procedures by directly outputting the Hamiltonian, drastically reducing computational cost compared to standard DFT.
• The model is first trained on perturbed untwisted bilayer graphene with randomized in-plane shifts and interlayer distances to simulate diverse stacking configurations.
• Generalization to twisted bilayer graphene (TBG) with various twist angles (e.g., 21.79°, 13.17°, 9.43°) and large supercells (up to 11,164 atoms) is achieved through data augmentation and architectural invariance.
• The method is validated by comparing predicted band structures and Fermi velocities against reference DFT calculations, with quantitative metrics reported for accuracy and efficiency.

Experimental results

Research questions

• RQ1 Can deep learning effectively represent the non-local exchange potential in hybrid functionals within the generalized Kohn-Sham framework?
• RQ2 To what extent can a deep neural network trained on small, simple systems generalize to large, complex superstructures like twisted bilayer graphene with over 10,000 atoms?
• RQ3 How does the inclusion of exact exchange via the HSE functional affect the flat band properties in magic-angle twisted bilayer graphene compared to PBE functionals?
• RQ4 What is the computational efficiency gain of the DeepH-hybrid method over conventional DFT for large-scale hybrid-functional calculations?
• RQ5 Can the model accurately predict electronic properties such as band gaps and Fermi velocities with ab initio accuracy at a fraction of the cost?

Key findings

• DeepH-hybrid reduces computational cost by orders of magnitude compared to conventional DFT, with lower scaling with respect to system size.
• The model achieves high accuracy in predicting band structures of twisted bilayer graphene across multiple twist angles, with excellent agreement to reference DFT results.
• For magic-angle twisted bilayer graphene (11,164 atoms), the inclusion of exact exchange via HSE increases the bandwidth of flat bands from 4.1 meV (PBE) to 41.1 meV, indicating a significant reduction in flatness.
• Fermi velocity increases from 0.9 m/s (PBE) to 36.2 m/s (HSE), suggesting a qualitative change in electronic dynamics due to exact exchange.
• The method successfully generalizes from untwisted bilayer graphene to various twisted superstructures, demonstrating robustness to structural variations and stacking configurations.
• DeepH-hybrid enables the first large-scale study of hybrid-functional electronic structure in Moiré twisted materials, opening the door to accurate, efficient simulations of complex correlated systems.