A posteriori learning of quasi-geostrophic turbulence parametrization: an experiment on integration steps

In this paper, Hugo Frezat, Julien Le Sommer, and their co-authors, train a NN-based closures for QG dynamics with an a posteriori learning criteria based on model integrations over several time-steps. While this strategy yields closures which outperform existing baselines, it requires the coarse resolution model to be differentiable in order to minimize the training loss.