The neural adjoint method and its application to sound speed learning

Abstract

Ray-based acoustical tomography typically provides estimates of local ocean sound speed profile (SSP) fluctuations from precise measurements of acoustic travel times (AT) fluctuations (with respect to a reference environment) between multiple sources and receiver arrays along with a model of the ray propagation in the reference (i.e., fluctuation-free) environment. Classically, inverting the forward model (if available) yielding SSPs from ATs can be done by computing the pseudo-inverse (expensive) or using an iterative-method (e.g., gradient descent); however, including any acoustic forward model (e.g., ray-tracing) in an optimization loop is non-trivial because the nonlinear mapping between SPP and AT is computationally expensive to differentiate. Instead, we use a Neural Adjoint (NA) approach [Ren et al., NeurIPS (2020)] to circumvent this problem by replacing the physics-based forward model with a deep neural network approximation, allowing for an inexpensive way to compute a gradient through backpropagation. The iterative nature of the NA method permits thorough exploration of the solution space depending on the initialization, as opposed to outputting a single point estimate. Here, we continue the discussion of data-driven methods presented in the companion presentation by Saha et al. and show that NA has the potential to further refine existing data-driven SSP estimation techniques.