Surface symmetry determination by a deep learning application of optical second-harmonic generation results

Abstract

Optical second-harmonic generation (SHG) technique is widely used to characterize the structural symmetry of condensed matters not only for the bulk but also for the surface states. Since experimental results of the SHG anisotropy patterns often contain multiple contributions from bulk and surface states, a conventional regression process may not provide a unique solution about the symmetry of each part. With a given symmetry for the bulk state, we here develop a discriminator for the surface symmetry by exploiting neural architectures based on 1D convolution layers where simulation results of SHG anisotropy patterns are taken as learning sequence data. Since SHG experimental results have limited information for determining a unique symmetry out of several symmetry candidates, our discriminator consists of tens of neural architectures optimized with different hyperparameters. As a final output, the surface symmetry discriminator gives a weighted sum of results suggested from all the individual neural architectures. We demonstrate that the discrimination results of the deep learning approach agree well with those of the conventional regression method for GaAs and Bi2Se3.