Grid-Agnostic Sea Ice Downscaling with Neural Operators

Mon, 26 Oct 2026 00:00:00 +0000

Background

Downscaling sea ice concentration (SIC) maps from coarse to fine spatial resolution is an important task in polar climate science, relevant to both operational forecasting and the analysis of observational records. A recurring practical challenge is that sea ice data products and climate model outputs live on a variety of grids, including polar stereographic projections and latitude-longitude grids. U-Nets and diffusion models require all data to be on the same regular grid before training, which means regridding is a mandatory preprocessing step. This regridding is not harmless: bilinear or nearest-neighbour interpolation introduces smoothing artefacts, can distort the sharpness of the ice edge, and discards the native resolution structure of each data source.

SwinIR: A Deterministic Baseline for Sea Ice Downscaling

Mon, 26 Oct 2026 00:00:00 +0000

Background

Accurate forecasting of Arctic sea ice is a critical challenge in climate science, with implications for shipping routes, ecosystem monitoring, and climate modelling. Recent work has shown that deep learning models can produce skillful sea ice forecasts and downscaled sea ice concentration (SIC) maps, often outperforming classical numerical methods at certain lead times and spatial scales. A central open question in this space is what constitutes the strongest deterministic baseline before introducing probabilistic or generative approaches.

Downscaling on CogSys Project Descriptions

Grid-Agnostic Sea Ice Downscaling with Neural Operators

Background

SwinIR: A Deterministic Baseline for Sea Ice Downscaling

Background