2025_programme: deep SSP: Enhanced Split Step Pade methods using deep learning

Back


  • Day: June 16, Monday
      Location / Time: C. THALIA at 12:40-13:00
  • Last minutes changes: -
  • Session: 09. Modeling techniques for underwater acoustic scattering and propagation (including 3D effects)
    Organiser(s): Boris Katsnelson, Pavel S. Petrov
    Chairperson(s): Boris Katsnelson, Sven Ivansson, Pavel Petrov
  • Lecture: deep SSP: Enhanced Split Step Pade methods using deep learning
    Paper ID: 2221
    Author(s): Daniel Walsken
    Presenter: Daniel Walsken
    Abstract: This work aims to improve the accuracy of the numerical solution of the parabolic wave equation by Split-Step Pade (SSP) techniques, developed by Collins, using deep learning. Based on the recently developed deep finite difference method (deep FDM) by Kossaczka et al., an artificial neural network is trained to approximate the truncation error of the method, thus improving the accuracy without sacrificing the consistency properties of the split-step method. To train the model, a PINN loss function is used to learn the dynamics of the parabolic wave equation directly, without relying on higher-order solutions as training data. Currently, the idea is only applied to the simplest splitting type, but can in principle be modified to improve the accuracy of any splitting method.
  • Corresponding author: Mr Daniel Walsken
    Affiliation: Bergische Universität Wuppertal
    Country: Germany

Back