Fig. 1. Schematic diagram of the framework used for the bidirectional neural network design process. (a) The unit cell structure diagram of ASRR; (b) the forward neural network diagram; (c) transmission spectrum of the forward prediction output; (d) optimal parameters of the reverse design output; (e) inverse neural network diagram; (f) transmission spectrum of the inverse design input

Fig. 2. Neural network model

Fig. 3. Sigmoid activation function

Fig. 4. Structural parameters of DNN

Fig. 5. Forward neural network loss

Fig. 6. Comparison of the influence of different network layers on the loss function

Fig. 7. Forward prediction results and numerical simulation results

Fig. 8. (a) Evolution of model training loss for an inverse neural network. (b) Comparison of inverse neural network output，CST simulation results and target spectrum

Fig. 9. Top view of the optimized model. Incident along the y-axis, periodicity P=90 µm. Aluminum ring thickness H=14 μm, aluminum ring arm width W=6 μm, gap G=3 μm, asymmetric tuning factor D=3 μm

Fig. 10. The near-field total electric and magnetic field amplitude, |E| and |B| of the Fano resonance under the conditions of the structural parameters of [70, 14, 6, 3, 3] lattice mismatched and [90, 14, 6, 3, 3] lattice matched. All maps of the same field share a color scale with the same range