A physically-informed deep-learning model using time-reversal for locating a source from sparse and highly noisy sensors data

We approximate the underwater acoustic wave problem for locating sources in that medium. We create a time dependent synthetic data-set of sensor recorded pressures, based on a small set of sensors placed in the domain, and perturb this data with high random multiplicative noise. We show that reference time-reversal based method struggles with high noise, and a naive deep-learning method also fails. We propose a method, based on physically-informed neural-networks and time-reversal, for approximating the source location even in the presence of high sensors noise. •Solving the wave problem with a numerical scheme.•Formulating the problem of locating sources as a data driven problem.•Creating a 4 blocks method: Noise level inference; Time-reversal; Deep-learning sharpening; Physically informed loss.•Testing the method for many scenarios.•Comparing to reference methods and achieving satisfactory results.