Abstract

Sea ice plays a key role in Earth’s climate system and exhibits significant seasonal variability as it advances and retreats across the Arctic and Antarctic every year. The production of sea ice forecasts provides great scientific and practical value to stakeholders across the polar regions, informing shipping, conservation, logistics, and the daily lives of inhabitants of local communities. Machine learning offers a promising means by which to develop such forecasts, capturing the nonlinear dynamics and subtle spatiotemporal patterns at play as effectively—if not more effectively—than conventional physics-based models. In particular, the ability of deep generative models to produce probabilistic forecasts which acknowledge the inherent stochasticity of sea ice processes and represent uncertainty by design make them a sensible choice for the task of sea ice forecasting. Diffusion models, a class of deep generative models, present a strong option given their state-of-the-art performance on computer vision tasks and their strong track record when adapted to spatiotemporal modelling tasks in weather and climate domains. In this talk, I will present preliminary results from a IceNet-like [1] diffusion model trained to autoregressively forecast daily, 6.25 km resolution sea ice concentration in the Bellingshausen Sea along the Antarctic Peninsula. I will also touch on the downstream applications for these forecasts, from conservation to marine route planning, which are under development at the British Antarctic Survey (BAS). I welcome ideas and suggestions for improvement and look forward to discussing opportunities for collaboration within and beyond BAS .

[1] Andersson, Tom R., et al. “Seasonal Arctic sea ice forecasting with probabilistic deep learning.” Nature communications 12.1 (2021): 5124. https://www.nature.com/articles/s41467-021-25257-4