A machine learning approach to predict tight-binding parameters for point defects via the projected density of states

Calculating the impact of point defects on the macroscopic properties of technologically relevant semiconductors remains a considerable challenge. Semi-empirical approaches, such as the tight-binding method, are very efficient in calculating the electronic structure of large supercells containing one or several defects. However, the accuracy of these calculations depends on the quality of the parameters. Obtaining reliable parameters by fitting to the large number of entangled bands in defective supercells is a demanding task. We therefore present an alternative way by fitting to the atom and orbital projected densities of states. Starting with a tight-binding fit of the pristine material, we only need a few physically motivated parameters for the fitting of defects. The training is done on data sets generated purely with parameter variations of tight-binding Hamiltonians. We demonstrate the efficiency of our approach for the calculation of the carbon monomer and the carbon dimer substitutions in hexagonal boron nitride. The method opens a path towards understanding complicated defect landscapes using a computationally affordable semi-empirical approach without sacrificing accuracy.