(a) Schematic drawing of CTINN with the circuit-theory-derived capability of generalization for metamaterial design [81]. (b) Design generalization of sample space and wavelength range for PSM with meta-atoms of single MDM stack [81]. (c) Design generalization of sample space and wavelength range for PSMs with meta-atoms of double MDM stacks [81]. (d) Conceptual diagram of highly intelligent metamaterial design empowered by circuit-physics-driven deep learning [81]. (e) Schematic illustration of the CNN–LSTM neural network structure [80]. (f) Comparison of the point-to-point mean square error with the expected value and predicted value of η, PCR, using or not using data augmentation methods [80].