The low frequency Kauai Beacon (KB) source, situated at 1000 m depth on the north shore of Kauai, provides a unique opportunity to investigate basin-scale ocean dynamics and underwater acoustic propagation through the lens of controlled, long range transmissions (3,500 + km) of broadband m-sequence signals to fixed receivers, such as those of the Comprehensive Nuclear Test Ban Organization (CTBTO) hydroacoustic monitoring station at Wake Island. This presentation focuses on an extensive dataset of received and simulated signals, spanning more than a year, and explores how travel time fluctuations on the order of a second could be related to significant oceanographic events or changes to the environment. Simulated data are synthesized using physics-based acoustic propagation software and ocean state estimates from reanalysis data, providing machine learning frameworks an opportunity to directly train on the mapping between the ocean state and the received signal. We then leverage these simulated datasets to train neural networks, enabling comparisons with models trained on the measured receptions. This approach elucidates the ability of machine learning to uncover unmodeled physical effects while maintaining coherence with the underlying physics.
Durofchalk et al. (Tue,) studied this question.