Data-Driven Differentiable Simulations for Three-Dimensional Range-Dependent Underwater Acoustics

Abstract Artificial intelligence is transforming the shipping industry by providing cutting-edge solutions that improve ship efficiency and safety. The combination of these technologies, as well as the unparalleled power of sensor technologies and data analysis, has the potential to transform shipping by effectively processing real-time data, identifying potential hazards, reducing the impact on marine life and assisting in decision making. There is a critical need for accurate and reliable prediction of far-field noise emanating from shipping vessels. Traditional full-order models relying on high-dimensional PDEs can be inefficient for real-time far-field noise prediction. Recent advancements in deep learning-based reduced-order models have demonstrated speeds several orders of magnitude faster than full-order simulations. However, existing models lack uncertainty quantification for decision-making and safety-critical tasks. This research addresses this gap by focusing on quantifying uncertainty in data-driven models for underwater acoustics. Using stochastic variational Gaussian process regression, we develop a data-driven model to predict underwater transmission loss (TL, expressed in dB) for regions surrounding Vancouver Island and the Port of Vancouver. The intrinsic ability of Gaussian process regression to predict variance facilitates robust uncertainty quantification in TL predictions. Furthermore, the model is implemented in PyTorch to ensure differentiability, enabling seamless integration with ship route optimization frameworks to minimize the impact on marine mammals. We also developed an algorithm utilizing radial basis function interpolation to reconstruct TL data from scattered points.