Fluid-based reservoir computing for distributed signal processing

Abstract

Physical Reservoir Computing (RC) leverages the interactions of a physical medium to efficiently accomplish the tasks of an artificial neural network with substantially reduced training cost. In this work, a subset of physical reservoirs focused on fluid-based media are compared, examining their uses of nonlinear phenomena to accomplish their computational objectives. The RCs are shown to depend heavily upon the nonlinearities inherent to the Navier-Stokes equations, while sampling only a subdomain of the entire fluid. The phenomena are chosen based on their desired end-functionality, indicating a wide application space including: surfaces waves, submerged structures, and both stationary and moving fluids. To further grow this nascent field, new multiphysics design frameworks and low-cost metrics will be needed as these systems are tuned for criticality. These techniques, along with more cost-effective, high-speed, non-intrusive fluid sampling, will unlock the potential of the field to radically reduce training and deployment costs for distributed flow signal sensing and processing.