Machine learning-based forecasting of ocean channel impulse responses over basin-scale distances motivated by prior knowledge of the waveguide propagation physics

The underwater acoustic channel is challenging to predict given the uncertainty of the oceanic environment. This is particularly difficult across basic-scale distances due to the intricacies and interactions between the time-varying oceanic processes that influence the ocean state and, therefore, the channel impulse response. From an application perspective, such channel forecasting knowledge can vastly improve the quality of interpretation of sensor measurements as the impact of the channel delay and delay-Doppler spread can be accurately compensated for in sensor network data. Similar gains may be expected for target-specific interpretation of sonar pings by harnessing the accurate forecasting of channel impulse response from the sonar transceiver to the target and back. In this work, we will focus on ideas to improve machine learning-based forecasting of ocean channel impulse responses over basin-scale distances. Specifically, we will investigate the use of constraints during training on the ML predictions motivated by prior knowledge of the waveguide propagation physics. These include penalty functions that force causality, constraining ray path structures, expected channel spreading etc. The approaches are demonstrated on simulated and real measurements of channel impulse responses using M-sequence transmissions from the 75 Hz Kauai beacon received at Wake Island.