Direct prediction of electrical properties of grain boundaries from photoluminescence profiles using machine learning

Abstract

We present a machine learning model to directly predict the carrier recombination velocity, vGB, at the grain boundary (GB) from the measured photoluminescence (PL) intensity profile by training it with numerical simulation results. As the training dataset, 1800 PL profiles were calculated with a combination of random values of four material properties—vGB, the GB inclination angle, and the carrier diffusion lengths in the grains on both sides of the GB. In addition, the measured noise was modeled artificially and applied to the simulated profiles. A neural network was constructed with the inputs of the PL profile and the outputs of the four properties. This served as the solver of the reverse problem of the computational simulation. The coefficient of determination and the root mean squared error of vlog, which is the common logarithm of vGB, for the test dataset were 0.97 and 0.245, respectively. This prediction error was sufficiently low for the practical estimation of vGB. Moreover, the calculation time was reduced by a factor of 198 000 compared to conventional numerical optimization of repeating the computational simulations. By utilizing this fast prediction method, continuous evaluation of vGB along a GB was demonstrated. The finding is expected to advance scientific investigation of the electrical properties of local defects.