A data-driven approach for tight-binding parametrization

Abstract

To determine how the Hamiltonian of a system manifests in experimental observation is, in general, a highly complex task. In certain systems, the Hamiltonian can be inferenced via symmetry analysis, and the problem reduces to determine a finite set of parameters, which is typically done by fitting through experiment or more complete numerical simulations. Artificial neural networks are a tool that allows to perform this fitting in a model agnostic way, as long as the input is provided in a standardized manner. These networks have the added benefit that they can handle and suppress noisy inputs. In this work, we investigate the efficiency of artificial neural networks to determine the hopping and on-site energies of a graphene Hamiltonian based on a noisy dataset of energies within the Brillouin zone. We have designed a standard to submit the input and optimize the network, ensuring that the minimal error is below 5%.