NeuralGeoFlow

NeuralGeoFlow explores a novel neural solver for the fast simulation of geophysical flows. Specifically, we propose a data-free training paradigm based on the minimization of physical energies, that allows a neural network to self-correct on-the-fly during a simulation. We expect this methodology, inspired by visual computing, to bring significant speed-ups at a marginal precision cost for the simulation of geophysical flows (tectonics, glaciers, rivers, oceans, climate, lava, etc.).