Abstract: To address the significant accuracy degradation in traditional evaporation duct inversion under non-ideal observational conditions—arising from “one-dimensional assumption distortion” due to atmospheric horizontal inhomogeneity and extreme sensitivity to meteorological parameter perturbations—this study proposes a Naval Postgraduate School(NPS)-Transformer inversion algorithm integrating physical principles with deep learning. The method constructs a high-fidelity electromagnetic dataset incorporating gravity wave perturbations and multi-scale parameter drift characteristics based on the wide-angle parabolic equation (WAPE). The algorithm follows a “physics-guided + residual-corrected” strategy: first, designing a cross-modal attention module to deeply fuse meteorological priors with radar features; second, utilizing the Transformer network to learn nonlinear residuals between physical benchmarks and true states; finally, introducing physical gradient constraints to suppress error amplification. Experiments demonstrate exceptional robustness in simulations featuring a 30% systematic relative humidity bias and high atmospheric turbulence uncertainty. Compared to the traditional NPS model, the mean absolute error (MAE) of modified refractivity decreases by 86.7%, the effective detection rate (i.e., probability of detection) for complex evaporation ducts increases to 96.8%., and electromagnetic blind zones are accurately reconstructed. This approach effectively overcomes the performance limitations of purely physical models when input data contains biases and disturbances. It achieves deep integration between meteorological-oceanographic information and the physical mechanisms of evaporation ducts, providing a reliable solution for high-precision electromagnetic sensing in complex environments.