The accuracy of simulation models for natural phenomena like tsunamis is heavily dependent on its input parameter space. We evaluate a deep feature consistent variational autoencoder that aims to reduce the input parameter space of bathymetric data while preserving the accuracy of simulations run on said data. We compare cutouts of the GEBCO datasets and their reconstructions based on multiple criteria to ascertain the overall accuracy of the reconstruction process before running simple wave simulations on our datasets to determine the impact of discrepancies between the original and reconstructed datasets. For this purpose we plot the water height over time at a buoy point and use the time frequency misfit method to more accurately quantify differences in our plots. Since both these procedures produce reasonable results we conclude that our models reconstructions are accurate for the purpose of simulation.

Keywords: Parameter Reduction, Bathymetry, Topography, ExaHyPE, GEBCO, Autoencoder