Latent-space correction for transmission loss prediction in deep water under sound speed profile perturbations

Accurate prediction of underwater acoustic transmission loss (TL) for each forecast sound speed profile (SSP) is important for acoustic applications that rely on forecast ocean environments. Rather than predicting TL directly from each forecast SSP, we represent the forecast SSP as a perturbation relative to a historical monthly mean SSP. In deep water, the relationship between SSP perturbations and TL changes is nonlinear, with SSP perturbations shifting the convergence zones (CZs) rather than causing pointwise TL fluctuations. A latent-space correction method is therefore proposed, in which SSP perturbations are mapped to latent features and used to correct the encoded reference TL. As the TL latent features show different levels of variability across dimensions, we introduce a channel-attention module to assign dimension-specific weights during correction. To increase the diversity of SSP perturbations, the training data are augmented with different reference SSPs and corresponding perturbations. Within-year and cross-year evaluations show that the proposed method has smaller TL prediction and CZ leading-edge position errors, compared with the uncorrected reference TL, while requiring much less computation time than conventional acoustic propagation modeling.