The SuperIce Webinars are a series of scientific online presentation and discussion on the topic of sea ice and machine learning. They are organized by the SuperIce project, which is led by the Nansen Center. The webinars are open to everyone. For this first webinar, we will have two presentations.
Charlotte Durand – Deep learning for surrogate modeling of new sea-ice models
A novel generation of sea-ice models can represent the drift and deformation of sea ice with an unprecedented accuracy at the mesoscale. Since these models are computationally heavy, we investigate supervised deep learning techniques for surrogate modeling of neXtSIM from its Arctic-wide simulations.
We train a convolutional neural network architecture, namely U-Nets, to emulate the sea-ice thickness for a forecast lead time of twelve hours. The U-Net extracts information at multiple scales and correctly predicts the advection of thickness given several atmospheric forcings. In general, autoregressively applying the neural network performs 36% better than a persistence forecast on a daily timescale, and this gain persists on monthly timescales with improvements of up to 40%. Still representing the advection of sea-ice, these surrogate models can also handle seasonal forecast of sea-ice thickness with a forecast lead time of up to 6 months. The obtained neural network models allow us to speed-up sea-ice simulations by several orders of magnitude compared to physical models like neXtSIM.
Julien Brajard – Super-resolution of satellite observations of sea ice thickness using diffusion models and physical modeling
This work introduces a simulator of high-resolution sea ice thickness in the Arctic based on diffusion models, which is a type of artificial intelligence (AI) generative model. Diffusion models have already shown impressive skill in generating realistic high-resolution images (e.g., DALL-E, Midjourney, Stable Diffusion).
Current satellite-based observations or climate model simulations of sea ice thickness provide valuable data but are limited by their coarse spatial resolution. High-resolution information is crucial for useful predictions and understanding small-scale features such as leads and thin ice, which significantly impact seasonal forecasting and heat flux calculations.
To increase the resolution of sea ice thickness products, we propose the following. First, a physically-based sea ice model, neXtSIM, is employed to generate a synthetic but realistic high-resolution sea ice thickness dataset. This synthetic dataset is then filtered to mimic the resolution of present satellite products or climate model outputs. An AI-based diffusion model is then trained to enhance the low-resolution SIT data.
By comparing the field produced by the simulator and a high-resolution test image, we will demonstrate that the simulator is able to produce accurate and realistic high-resolution sea ice thickness fields.