Deep-learning density functional theory Hamiltonian for efficient ab initio electronic-structure calculation

The marriage of density functional theory (DFT) and deep-learning methods has the potential to revolutionize modern computational materials science. Here we develop a deep neural network approach represent DFT Hamiltonian (DeepH) crystalline materials, aiming bypass computationally demanding self-consistent field iterations substantially improve efficiency ab initio electronic-structure calculations. A general framework is proposed deal with large dimensionality gauge (or rotation) covariance matrix by virtue locality, this realized message-passing for learning. High accuracy, high good transferability DeepH method are generally demonstrated various kinds material system physical property. provides solution accuracy–efficiency dilemma opens opportunities explore large-scale systems, as evidenced promising application in study twisted van der Waals materials. developed learn mapping function from atomic structure Hamiltonian, which helps address useful studying