Title:Efficient prediction of density functional theory Hamiltonian with graph neural network

Abstract:Despite the past decades have witnessed many successes of machine learning methods in predicting physical properties, the prediction of Hamiltonian and thus electronic properties is still out of satisfactions. Here based on the graph neural network architecture, we present an extendible neural network model to learn the Hamiltonian from ab initio data with only local structures as inputs. The local coordinates, encoded using the convolutional neural network and designed to reserve the Hermitian symmetry, is used to map hopping parameters onto local structures. We demonstrate the performance of our model using SiGe random alloys as an example. We show that, our neural network model, although trained using small-size systems, can predict Hamiltonian as well as electronic properties such as band structures and densities of states (DOS) for large-size systems within the ab initio accuracy, justifying its extendibility. In combination with the high efficiency of our model, i.e., it takes only seconds to get the Hamiltonian of a 1728-atom system, our work provides a general framework to efficiently and accurately predict the electronic properties, which provides new insights into computational physics and will accelerate the research for large-scale materials.