Deep learning surrogate model of the Arctic sea ice

As a world first, my colleague Charlotte Durand at CEREA, École des Pont and EdF R&D, has built a purely data-driven model of the sea-ice thichness of the Arctic. She learns this surrogate model of the Arctic from the nEXtSIM physical model, using a deep learning U-net architecture. With a stronger contraint on the total thickness of the Arctic sea ice, and using atmsopheric forcings, she obtains a year-long stable model which convincingly forecasts the advection of ice shelves, and the thermodynamics. She evidences and quantifies the long-term loss of the fine-scale geometry of sea ice due to the inherent diffusion of the surrogate model.

The paper, a contribution to the Schmidt Sciences SASIP project, is entitled Data-driven surrogate modeling of high-resolution sea-ice thickness in the Arctic, and is published in The Cryosphere.