LiNbO3 domain structures have been widely applied in nonlinear beam shaping, quantum light generation, and nonvolatile ferroelectric memory. The recent developments in nanoscale domain engineering techniques make it possible to fabricate sub-diffracted nanodomains in LiNbO3 crystal for high-speed modulation and high-capacity storage. However, it still lacks a feasible and efficient way to characterize these nanoscale domains. In this work, we propose and experimentally demonstrate a deep-learning-assisted identification of sub-diffraction LiNbO3 nanodomain lines. In the experiment, we record the second-harmonic (SH) images of nanodomain lines by using a confocal microscope. The domain linewidths range from 200 nm to 600 nm, which are beyond the spatial resolution of the used microscope (∼800 nm). After training a neural network with 1568 SH images, it is capable of recognizing different nanodomain lines at an accuracy of 81.25%. Our approach leverages the exceptional recognition capability of the neural network, which provides an efficient method to identify sub-diffraction nanodomains from diffraction-limited images.
Read full abstract