Guest
Iuri Gorenstein, Universidade de São Paulo The Ocean circulation is a highly nonlinear system, which requires some of the world’s largest computers to accurately simulate its turbulent activity and mesoscale eddies formation. Recently, the usage of neural-networks (NN) data-driven methodology in climate sciences has proven the ability to learn complex nonlinear system dynamics and assist numerical models in creating more realistic simulations. Recently, NNs have been able to accurately simulate the atmospheric dynamics of the whole globe or at regional configurations in a closed domain. However, these networks have not resolved primitive equations per se, they have been trained to evolve the dynamics of strict systems at specific resolutions and, although very successful at their task, they need to be retrained every time they are applied using different constraints. Here, we present a physics-informed NN auto-encoder able to embed the dynamics from the regional numerical model CROCO (Coastal and Regional Ocean Community Model) Barotropic mode at any ocean domain. Using one GPU, the NN was trained to evolve the ocean free surface with the same applicability as the numerical model, which uses the primitive equations to evolve the Barotropic free surface of the ocean pixel by pixel from any grid given a set of initial conditions. The NN’s auto-encoder is able to accurately encode the Ocean surface circulation system at any resolution and ocean domain. The NN dynamics embedding of the Barotropic mode was achieved at multiple domains and different resolutions (1/3 and 1/12 degree) for the training data-sets, however, the network is currently failing to generalize outside of it.
