Generalized microwave network-driven physics-informed neural operator surrogate model for metasurfaces network modeling

To address the computational bottlenecks in the optimized design of reconfigurable electromagnetic metasurfaces and achieve a transition from full-wave simulations to efficient intelligent operator inference, this paper proposes a metasurface physics-informed neural operator surrogate model driven by generalized microwave network modeling. First, a unified near/far-field physical modeling method based on generalized microwave network theory is proposed, enabling accurate field prediction through scattering parameter matrices, effective excitation, and angular spectrum propagation. Then, a physics-guided Fourier neural operator intelligent surrogate framework is constructed to learn the operator mapping from port configurations to field distributions. Preliminary validation based on an 8-focus focusing scenario using a 20×20 PIN diode metasurface shows that the proposed generalized microwave network method aligns well with HFSS full-wave simulations, providing a reliable data foundation for AI training. Under the case studied in this paper, the trained Fourier neural operator intelligent surrogate achieves a storage reduction of 10,000 times and a computational speedup of 2 million times compared to HFSS, with a near-field prediction error below 3% and a power conservation error less than 10%. Experimental results demonstrate that the generalized microwave network theory enables a learnable operator representation of complex electromagnetic processes through effective excitation and port-field decomposition. The Fourier neural operator trained on this basis achieves efficient approximation of operator mappings via spectral convolutions in the frequency domain, making the proposed framework suitable for real-time optimization scenarios of metasurfaces.