Presenter: Ryan McCarthy (Scripps Institute of Oceanography)

Description:
Next generation environmentally aware autonomous platforms hold promise for the ocean science community and will be reliant on underwater sound to communicate and sense the ocean. Our goal for this research is to enable more efficient autonomous operation through sound-awareness. Underwater acoustic propagation simulations through high-resolution environmental models are computationally expensive, often with low return on investment due to uncertainty in the environment. In-situ measurements provide limited spatial and temporal information and require assumptions for use in simulations (e.g. range independence). To address limited computational resources onboard low power autonomous platforms, we present a machine learning approach to enhance acoustic communications through adaptive repositioning within the environment. Specifically, a physics informed conditional generative adversarial network (PI-CGAN) is explored to improve underwater acoustic sensing through acoustic transmission loss field realizations. Inputs into the PI-CGAN include ray paths from a source, average sound speed profile, time of year (i.e., summer, winter, etc.), and depth of the transmitter. The PI-CGAN is trained and tested through generated transmission loss fields from a ray tracing model (BELLHOP). Transmission loss fields are calculated for environments of 100m depth at a frequency of 25kHz and range of 3km. Sound speed profiles collected across 3 years off the coast of Southern California are incorporated in the BELLHOP simulations. The PI-CGAN is trained by including a mean square loss error to enhance the generator’s predicted transmission loss. The PI-CGAN results are also compared to in-situ measurements of communication ranges from a UUV for validation.