AI-based sea ice prediction systems have multiplied over the past decade, now covering both poles from days to seasons ahead. Calculations that once required supercomputers can run on a laptop in minutes. But speed and skill, the authors note, are not the same as trustworthiness. The observations needed to build and stress-test these systems remain sparse across both polar regions, and climate change is pushing conditions beyond historical precedent.
To address this gap, a new international group called ORCAS was established in 2025 under the Scientific Committee on Oceanic Research and the World Weather Research Programme. It brings together observational scientists, AI developers, and physical modelers to assess what emerging AI-based sea ice prediction systems actually need. The group emphasizes that both process-based dynamical models and AI approaches share a common dependence on observational data—for tuning, initialization, training, and independent validation.
A key challenge is making datasets more accessible and interoperable, in line with FAIR principles. Satellite records, such as passive microwave sea ice concentration data dating to 1978, are plentiful but fragmented. AI-ready gridded datasets and process-focused local measurements offer different benefits, and both are needed to ensure forecasts are physically credible. The authors stress that sustained, well-documented observations remain the keystone of reliable sea ice prediction, regardless of the modeling approach. As the field grows, ORCAS aims to ensure that observation strategies keep pace with AI's demands. The article is based on a single source, so no conflicting perspectives are presented.}```markdown? Wait, the user asked for strict JSON with exactly these keys. I provided JSON. But I accidentally included a trailing