2023_programme: Approximating normal mode propagation modeling using a dense neural network

• Session: 07. Modeling techniques for underwater acoustic scattering and propagation (including 3D effects)
  Organiser(s): Boris Katsnelson and Pavel S. Petrov
• Lecture: Approximating normal mode propagation modeling using a dense neural network
  Paper ID: 1855
  Author(s): Varon Arthur, Mars Jérome, Bonnel Julien
  Presenter: Varon Arthur
  Abstract: The underwater environment acts as a dispersive acoustic waveguide in which the propagation can be described by a sum of modal components. However, the simulation of this propagation can be computationally expensive, especially when considering multimodal wideband signals. Here, the objective is to accelerate the resolution of normal mode propagation models using machine learning. More specifically, a Dense Neural Network (DNN) is trained to predict the wavenumbers and group velocities for a variety of possible environments (with a known water column, an unknown sediment layer and an unknown basement). Predicted wavenumbers are then used to evaluate modal depth functions (using known water column and inverse iteration method) and transmission losses. Replacing only part of the physical simulation by a DNN allows the reduction of the computational cost of at least an order of magnitude with only a minor loss in physical explainability.\nThe whole method is used in a simulated benchmark reproducing geoacoustic inversion of Shallow Water 2006 data, illustrating the capacity to perform fast inversion using DNN instead of traditional propagation models.
    Download the full paper
• Corresponding author: Mr Arthur Varon
  Affiliation: Université Grenoble Alpes, CNRS, Grenoble-INP, GIPSA-Lab, Grenoble, France
  Country: France
  e-mail: