Speaker
Description
Ocean models have to balance efficiency and accuracy. The general assumption is that the higher the resolution, the more realistic the simulated flow. However, the actual grid spacing is not the physically realistic and reliable resolution of a model configuration. Numerical artifacts due to choices of solvers, numerical schemes or parameterizations can affect the accuracy and reliability of the simulated flow. At intermediate model resolutions of a few to tens of kilometers, viscosity parameterizations play a critical role in maintaining model stability and in dampening numerical modes that may otherwise result in numerical artifacts. In this study, we present a machine learning method to identify numerical noise in configurations of the FESOM2 unstructured grid ocean model. We train a vector-quantized variational autoencoder (vqAE) on surface vorticity data from an idealized channel configuration, interpolated onto a regular grid using nearest neighbour interpolation. The reference configuration employs sufficiently high viscosity to produce a smooth, largely noise-free vorticity field optimized for the setup. After training, we then test the autoencoder on simulations with (temporally) reduced viscosity. Thus, we identify regions of intensified noise that are revealed after reconstruction. To target noise at specific scales, particularly after interpolation to a finer regular grid, we conduct sensitivity tests with the image Euclidean distance metric (IMED) and the feature similarity index (FSIM). These methods aim to enhance spatial robustness and apply a targeted edge detection algorithm, respectively, improving the detection of anomalies. The vqAE robustly identifies and flags intensified noise, therefore enabling automated detection of potential numerical instabilities that may degrade the simulated flow without causing obvious model blow-ups. It can ultimately act as a warning flag to modelers for model tuning and before analyzing or releasing model data.