Inverse design of polarization-insensitive all-dielectric BIC metasurface with dual Fano-resonances by deep learning

The quasi-bound states in the continuum (quasi-BIC) based on all-dielectric metasurfaces have been widely applied to enhance nonlinear optical processes. To this end, a polarization-insensitive all-dielectric tetrameric metasurface with dual-quasi-BIC is demonstrated for the third-harmonic generation (THG) in this paper. In practical applications, the structural parameters of the metasurface need to be adjusted according to the target excitation wavelength to produce a wavelength-matched quasi-BIC. However, structural adjustment is time-consuming by forward simulation. Therefore, a rapid inverse design method based on deep learning is proposed. This method consists of dual-channel convolutional neural network (DCCNN), multi-head attention (MHA) and Gramian angular field (GAF) transformation. GAF can effectively encode the relationship between transmission values at different wavelengths, which is conducive to obtaining complete transmission spectrum characteristics. Finally, the proposed all-dielectric metasurface can easily achieve frequency conversion from fundamental waves (1300 nm–1600 nm) to third-harmonic waves (433 nm–533 nm) and the inverse design can reach the mean relative deviation of 0.8 %.