Neural network and genetic algorithm-driven inverse design of a full-phase metasurface for wavefront manipulation

This work proposes an inverse design framework integrating deep learning and genetic algorithms (GA) to address the inefficiency and computational limitations of traditional metasurface design methods. A fully connected neural network (FCN) model is constructed to predict the forward electromagnetic response (real and imaginary parts of the S parameters) of the 6×6 binary-coded metasurface units, achieving a test set mean squared error (MSE) of 0.05. The GA-driven inverse optimization enables rapid generation of eight phase-modulated units covering the complete 0−2 π phase range at 13 GHz. As a proof-of-concept, based on the inverse method, we successfully design deflection and focused vortex metadevices at 13 GHz, with prediction speeds significantly faster than traditional full-wave simulations, which means that eight corresponding metasurface units can be found in 70 s. Two metasurfaces for holographic imaging at 13 GHz are designed from the same network showing different images in orthogonal polarization states, which are validated by full-wave simulation. Compared to traditional iterative full-wave optimization, the pre-trained FCN-GA framework can generate target-compliant binary codes in about 10 s, which significantly reduces the time it takes to design metasurfaces for both experienced and inexperienced metasurface designers.