Prediction of Large-Depth Transmission Loss Matrices Using Conditional Diffusion Models

The transmission loss (TL) matrix is vital in underwater acoustics, ocean exploration, [1] and sonar design, yet traditional numerical methods like ray tracing, finite element methods, and parabolic equations struggle with the high computational complexity and reliance on prior environmental parameters in complex marine settings, especially during large-scale parameter sweeps or Monte Carlo simulations where costs escalate exponentially, limiting their practicality. Diffusion models, an emerging generative approach excelling in image and speech synthesis via progressive de noising, adeptly capture dynamic variations in high-dimensional data. [2]–[3] This study introduces a diffusion-based method to predict TL matrices in large-depth regions, using Bellhop-generated TL data to train the model on implicit distributions tied to source depth, frequency, and seabed topography, enhanced by condition embeddings and attention mechanisms. Experimental results show it efficiently generates accurate TL matrices across multiple source depths with significantly reduced computational cost, providing an innovative solution for rapid acoustic field computation, design, and optimization.