Generalizing deep learning electronic structure calculation to the plane-wave basis.

Area of Science:

• Computational Materials Science
• Quantum Chemistry
• Artificial Intelligence in Science

Background:

• Deep neural networks (DNNs) show potential for electronic structure calculations by learning the density functional theory (DFT) Hamiltonian.
• Existing DNNs are limited to the atomic-orbital (AO) basis, excluding the widely used plane-wave (PW) basis.

Purpose of the Study:

• To develop a method enabling DNNs to utilize plane-wave (PW) basis DFT results for electronic structure calculations.
• To bridge the gap between AO-based deep learning and PW-based DFT methods.

Main Methods:

• Proposed a real-space reconstruction method to compute AO Hamiltonian matrices directly from PW DFT results.
• Demonstrated the efficiency and accuracy of the reconstruction method compared to traditional projection-based techniques.

Main Results:

• The reconstruction method is significantly faster (orders of magnitude) than conventional projection methods.
• Reconstructed Hamiltonian matrices accurately reproduce electronic structures calculated using the PW basis.
• Successfully integrated the advantages of PW methods (accuracy, flexibility, applicability) into deep learning electronic structure approaches.

Conclusions:

• The developed method overcomes the basis limitation of previous DNNs for DFT.
• Enables the creation of large-scale, high-fidelity training datasets using PW DFT.
• Paves the way for developing more precise and broadly applicable deep learning models for electronic structure.