Abstract
Propeller-induced ship wakes modulate corrosion and cathodic-protection currents and thereby generate low-frequency underwater electromagnetic (EM) signatures. Sea-trial measurements show that, after propagation in a conductive marine environment with multiple transmission paths, the dominant received frequency is systematically lower than the propeller shaft frequency. To describe this behavior, we present a closed-loop data-driven framework for forward modeling and inverse recovery of ship-wake-induced underwater EM propagation. The framework combines SpectralDriftNet for received-frequency prediction, a conditional diffusion model for receiver-spectrogram generation, and a Swin-Transformer-based inverse model for source-trajectory and propeller-frequency reconstruction. A reconstruction-consistency constraint links the forward and inverse modules to improve stability and internal consistency. Experiments on real-sea EM data show that the framework reproduces the observed frequency downshift, preserves the dominant line-spectrum structures in generated spectrograms, and supports stable inversion under limited-data conditions. These results indicate that closed-loop learning is a practical route for studying complex underwater EM propagation when exact analytical modeling remains difficult.