Fully connected-transformer synergistic neural network for customized design of high-Q electromagnetically induced transparency

High quality-factor (high-Q) electromagnetically induced transparency (EIT) greatly enhances light-matter interactions but is empirically designed and time-consuming, requiring precise resonant mode tuning. While deep learning aids photonic device design, conventional methods fail for high-Q EIT due to its nonlinear complexity and extreme parameter sensitivity. To address these challenges, we present a fully connected-transformer synergistic neural network (FTSNN) based on a decomposition-prediction-combination strategy, which combines the advantages of the fully connected neural network (FCNN) and transformer neural network (TNN), for the rapid and accurate prediction of transmission responses and inverse design of structural parameters with high-Q EIT. We decompose the EIT response into a narrow high-Q resonance and a broad smooth background. Analytical parameters are extracted by fitting the EIT peak with a Lorentzian model, which directly recognizes the narrow high-Q response without higher-order complexity by using the FCNN. Concurrently, the TNN captures the long-range correlations of the broad background spectra; thus, accurate prediction and inverse design in a customized way can be realized by combining the results of the FCNN and TNN. The mean squared error (MSE) losses for both the FCNN and TNN are smaller than 0.01 after 1000 iterations. Excellent forward prediction for high-Q EIT can be achieved with Q factor up to 1.2 × 10 4 and a peak transmission of 0.9. Customized high-Q EIT spectra can be generated by the inverse design with a Q factor reaching 8,170, as confirmed by rigorous coupled wave analysis (RCWA). Our findings offer unique opportunities for the design and applications of high-Q resonators with enhanced performance.