Combining DeepH with HONPAS for accurate and efficient hybrid functional electronic structure calculations with ten thousand atoms

Abstract

Density functional theory (DFT) calculations of hybrid functionals have traditionally been limited to small systems containing hundreds of atoms due to substantial computational constraints. In this work, we introduce an interface between DeepH, a machine learning-based Hamiltonian approach, and HONPAS, a density functional theory (DFT) software package. By leveraging DeepH’s ability to bypass self-consistent field (SCF) iterations, DFT calculations in HONPAS become significantly more efficient, including computationally intensive hybrid functional calculations. This combined approach is particularly advantageous for twisted van der Waals systems, as demonstrated through examples of twisted bilayer graphene and twisted bilayer $\mathrm{MoS_2}$. The substantial reduction in computation time for the HSE06 functional suggests that our method effectively addresses the efficiency-accuracy trade-off in DFT calculations, making high-accuracy calculations feasible for large systems containing more than ten thousand atoms.