Physics-informed and machine learning-enabled retrieval of ocean current speed from flow noisea)

Episodes of exceptionally strong near-bottom currents were encountered at depths of over 2500 m in a 52-day deployment of moored autonomous acoustic noise recorders (MANRs) over the Atlantis II Seamounts in the Northwest Atlantic. A strong correlation is found between the current speed and acoustic noise intensity, especially at infrasonic frequencies below 20 Hz. Flow noise and ambient sound, including shipping noise, made comparable contributions to the measured acoustic intensity but had distinct spectral properties. This paper explores a way to identify and quantify the differences between flow noise and ambient sound in the pressure fluctuations measured by a hydrophone and find statistical characteristics of the fluctuations which contain robust information about the flow speed. A regression tree machine learning model was developed to relate the acoustic features of flow noise to directly measured current speeds. By training the model using data from a MANR equipped with a hydrophone and current meter, the time series of current speed was obtained with 1-min resolution at another MANR, where only acoustic data were available. Accuracy of the inferred current speeds was confirmed by comparing the dependence of flow noise spectra on the current speed at the two MANRs.