A lightweight conditional mean-field network for predicting azimuth-range acoustic transmission loss in dynamic ocean environments

Prediction of underwater acoustic transmission loss (TL) is critical for sonar performance evaluation. Traditional physical models entail high computational costs, while existing data-driven methods are mostly limited to static fixed environments and cannot adapt to the spatiotemporal dynamics of the real marine environment. To address this, this paper proposes a Lightweight Conditional Mean-Field Network (LCMF-Net) for efficient azimuth-range TL prediction in dynamic ocean environments. Adopting an encoder-modulator-decoder architecture, LCMF-Net fuses source positions with time-varying sound speed profiles via a conditional encoding mechanism to capture the spatiotemporal environmental variations. It integrates the historical-mean field as physical before drastically reducing learning complexity, and employs FiLM-based conditional residual blocks to dynamically calibrate TL features with environmental information. With only 1.56 × 106 parameters, LCMF-Net outperforms the U-Net-2D and GAN-2D benchmarks in all prediction accuracy metrics in tests on two typical sea areas: the Northwestern Pacific and the northern South China Sea. Its computational efficiency is also improved by more than three orders of magnitude compared with the Bellhop3D model. Ablation experiments further validate the effectiveness of core designs including the historical-mean field, FiLM modulation, and dilated convolution. This study provides a lightweight and efficient deep learning solution for azimuth-range TL prediction, and exhibits significant potential for edge computing and real-time sonar applications.