Recent advances in artificial intelligence (AI) are transforming sea-ice forecasting. AI models now offer performance comparable to or better than traditional physics-based systems, while requiring far fewer computational resources. These efficiencies allow for more frequent, timely predictions that could benefit a wide range of stakeholders. However, the effective development and validation of these systems rely on high-quality observational data. While AI models are often trained on analysis and reanalysis products, observational campaign data—vital for understanding physical processes—remains underused. Such data is essential for evaluating the realism of AI outputs and building trust in their predictions.
The ORCAS Working Group (WG) will systematically assess the observational requirements for emerging AI-based sea-ice prediction systems up to the seasonal timescale. Our approach combines the analysis of historical field campaign data with collaborative model evaluation, focusing on sea ice, ocean, and near-surface atmospheric conditions observed during campaigns such as MOSAiC, Antarctica InSync. We will investigate how different data types contribute to training, initialisation, and validation, and evaluate how physically consistent AI forecasts are compared to traditional methods.
This work will identify key gaps in observing systems and inform the design of future campaigns and networks, including those planned for the fifth International Polar Year (IPY5). Our recommendations will target the specific needs of AI-driven forecasting to support reliable, actionable Arctic and Antarctic predictions. Establishing the ORCAS WG is needed to guide the integration of AI models and observations, helping ensure these new tools are physically grounded and trusted by the polar prediction community.
The working groups selected at the 2025 SCOR annual meeting are still taking feedback from the reviews under consideration, which may result in additions to the membership and other small revisions.
- Chair(s)
- Clare Eayrs (USA), Lorenzo Zampieri (Germany), Malte Müller (Norway)
- Other Full Members
- Luisa von Albedyll (Germany), Sandra Barreira (Argentina), David Bromwich (USA), Petra Heil (UK), Zachary Labe (USA), Yafei Nie (China-Beijing), Luciano Ponzi Pezzi (Brazil)
- Associate Members
- David Clemens-Sewell (USA), Yonghan Choi (South Korea), Wayne De Jager (South Africa), Simon Driscoll (UK), Lauren Hoffman (Belgium), François Massonnet (Belgium)
- Reporter
- Paul Myers
- Terms of Reference
Design validation scenarios by identifying key historical observational campaigns (e.g MOSAiC, ASPeCt) and datasets suited for evaluating data-driven polar prediction models up to the seasonal timescale. Prioritise events and variables of interest for consistent benchmarking across AI architectures.
Evaluate AI model performance by designing and implementing standardised methods to test AI predictions against observational data. Quantify model accuracy and uncertainties in extreme and representative polar conditions.
Assess the physical consistency of AI systems by comparing model outputs with known physical constraints and conservation laws. Identify failure modes and assess the interpretability and limits of emerging AI approaches.
Recommend future observational priorities by identifying which data types, temporal/spatial resolutions, and variables are most effective for training, initialising and validating AI prediction systems. Generate guidance for upcoming international campaigns (e.g. Antarctic InSync, IPY5) so that observation systems can be strengthened and developed to support AI models.
Foster collaboration throughout all activities by facilitating inclusive engagement among AI developers, model users and observational scientists. Co-design methodologies and benchmarks through community workshops and interactions with key programs (SOOS, ASPeCt, CliC Arctic Sea-ice Working Group, SIPN, and with the full suite of PCAPS Task Teams).
- Approved
- October 2025
- Financial Sponsors
- SCOR, NSF