Physics Informed Generative Surrogate Learning for Full Wave Validated Real Time Beamforming in Programmable Metasurfaces

ABSTRACT Programmable metasurfaces enable agile electromagnetic beam shaping, yet the inverse design of dense binary apertures is complicated by feed illumination, local coupling, finite aperture effects, and mode dependent state statistics. A physics informed generative surrogate framework is developed for real time beamforming with full wave validation. The framework integrates a conditional state matrix generator, a parametric aperture physics surrogate with hard routed residual experts, two stage NSGA‐II optimization, and CST verification. The generator achieves 96.67% binary accuracy and a 94.47% F1 score. The best predictor gives a test RMSE of 1.108 dBi, an MAE of 0.549 dBi, and a peak gain MAE of 0.114 dB. Across 137 optimized full wave validation cases, the mean peak gain bias remains close to ‐0.13 dB. For 35 paired real time/final cases, the refined search improves the full wave target region gain by 2.943 dB on average and the peak gain by 0.573 dB, indicating that the surrogate can support practical metasurface beam optimization.