Representation and temporal prediction of ocean states with fully complex autoencoders

The one-directional oceanic channel is very dynamic and complicated. Changing temperatures affect sound speed, the tidal effects of the lunar cycle change depth and pressure, and the physical aperturing from undersea terrain filters the ever-changing multipath scattering rays. All of these effect the impulse response of the channel. This is a time-variant system with a very high number of degrees of freedom, and thus difficult to model and make predictions for. Autoencoders allow for the projection of the space of the varying channel transfer functions to a much lower dimensional “encoding space” from which physical interpretations may be gleaned, and constructing an autoencoder with fully complex neuron layers allows a better representation of the complex-valued transfer function space. Within the encoding space, temporal trends can be observed at scales of minutes or days, and ongoing research explores the predictive capacity of trends at these scales. [Work is done in collaboration with the Naval Postgraduate School and funded by US Department of Defense NEEC grant N00174-20-1-0016.]