Fig. 1. Schematic diagram of CAHM monolithic design via REDCNN model

Fig. 2. REDCNN model design and feature extraction. (a) The architecture and dimension of REDCNN model. (b) The downsampling process of feature encoder. (c) The upsampling process of feature decoder. (d) Feature transfer of different channels in encoding process. (e) Feature transfer of different channels in decoding process

Fig. 3. Training and test of the REDCNN model. (a) The variation of MAE loss value in deep learning process. (b) The variation of MAE loss value in transfer learning process. (c) The error histogram of deep learning in training set. (d) The error histogram of deep learning in test set. (e) The error histogram of transfer learning in training set. (f) The error histogram of transfer learning in test set.

Fig. 4. The comparison of predicted metasurface and real metasurface with error distributions. (a) Input images. (b) Phase profiles of metasurface. (c) Amplitude profiles of metasurface. (d) Theoretical electric field distributions calculated by diffraction theory. (e) Simulated electric field distributions.

Fig. 5. Measurement verification and comparation of the metasurfaces. (a) Photograph of fabricated metasurface prototype. (b, c) Photograph of orthogonal metal gratings. (d) Photograph of real metasurface pattern. (e) Photograph of predicted metasurface pattern. (f) Electric-field measurement environment in microwave anechoic chamber. (g) Measured electric field distribution of real metasurface. (h) Measured electric field distribution of predicted metasurface. (i) The error of measured electric field distribution between the real and predicted metasurfaces