A Physics-Informed Framework Linking Deep Learning Force Fields and Electronic Structure Calculations

Recent developments in machine learning techniques demonstrate promising potential for addressing physical problems. In this presentation, we introduce a multifunctional computational method based on deep learning. The multifunctional model (WANDER) can both simulate atomic forces and electronic energy bands in solid-state materials, making it a powerful tool for studying large-scale systems. This model uses the Wannier function as the basis and streamlines its design by categorizing Wannier Hamiltonian elements through physics-informed principles to improve its effectiveness. This approach is demonstrated using twisted two-dimensional materials, showing its ability to address large-scale systems with decent accuracy. By endowing traditional machine-learning force fields with electronic structure simulation capabilities, WANDER bridges the gap between force field and electronic structure methods. The study highlights the potential of deep-learning computational methods that can achieve multiple functionalities in simulating large-scale systems, which are challenging for conventional first-principles calculations.