Predicting transmission loss in underwater acoustics using continual learning with range-dependent conditional convolutional neural networks

Efficient and accurate prediction of underwater acoustic transmission loss (TL) is important for minimizing noise impacts on marine ecosystems and supporting naval operations. Traditional wave-based solvers are computationally expensive, especially for range-dependent bathymetry, rendering them unsuitable for real-time applications. Recent advances in data-driven models, particularly convolutional and recurrent neural networks, provide a more efficient alternative by substantially reducing the dimensionality of the data. However, these deep-learning models struggle with long-range wave forecasts as they often rely on auto-regressive predictions and lack far-field bathymetry information. This research aims to improve the accuracy of deep learning models for forecasting underwater radiated noise in far-field scenarios. We introduce a range-dependent conditional convolutional neural network that predicts TL fields in a single step by conditioning directly on input bathymetry. The model is trained using a replay-based continual learning strategy, which allows generalization across sequential bathymetric changes without retraining. We evaluate our model using multiple test cases and a benchmark scenario that involves predictions over the Dickins Seamount. Our architecture effectively captures transmission loss over range-dependent bathymetry profiles. The proposed framework provides an efficient deep learning model for digital twins of the ocean soundscape, enabling real-time decision-making for underwater radiated noise.