Dynamic multi-task neural network for end-to-end prediction of deep-sea acoustic transmission loss

Accurate and efficient prediction of deep-sea acoustic transmission loss is essential for underwater applications, such as sonar design and underwater communication. In practical acoustic field computation tasks, it is often necessary to analyze the spatial distribution characteristics of the acoustic field in different regions. However, traditional numerical models require large-scale simulations on dense grids to predict acoustic fields over multiple spatial ranges, resulting in high computational cost. To address these limitations, this study proposes a dynamic multi-task U-Net (DMT-UNet) neural network model. Built on a multi-task learning framework, the model can dynamically adjust its network structure for end-to-end joint modeling of acoustic transmission loss at different computation ranges. Incorporating source information and deep-sea sound speed profiles as inputs enhances adaptability to model complex environments. DMT-UNet achieves an average root mean square error of approximately 1.6 dB on simulated deep-sea acoustic field datasets, with computational efficiency improved by more than 98.8% compared to traditional numerical models. Gradient-based visualization reveals how the model reconstructs acoustic field distributions during decoding, demonstrating interpretability. Experimental results show that DMT-UNet yields high prediction accuracy and computational efficiency, while maintaining consistency with physical laws and generalization capability. Thus, DMT-UNet enables real-time, multi-task acoustic modeling in complex deep-sea environments.