The Evolution of AI in Sea Ice Prediction: Bridging Gaps and Building Trust

A decade ago, dedicated artificial intelligence (AI) systems for predicting sea ice were virtually non-existent. Fast forward to today, and we see the emergence of numerous cutting-edge technologies across both polar regions capable of forecasting sea ice conditions over time frames that range from mere days to entire seasons. The computational speed at which these models can operate is remarkable; tasks that previously took days on supercomputers can now be executed in minutes on a standard laptop.

However, advancements in speed and accuracy do not equate to trustworthiness. The vital observational data required to construct and rigorously test these systems remains limited and unevenly distributed between the poles. This leads us to the core of the conversation: observational requirements. Addressing this gap is the central aim of the Observational Requirements in the Context of AI prediction systems for Sea ice (ORCAS) initiative.

Understanding ORCAS

Established in 2025, ORCAS has emerged as a vital international scientific community. It’s affiliated with the Scientific Committee on Oceanic Research (SCOR) as a working group and serves as a task team within the World Weather Research Programme’s (WWRP) Polar Coupled Analysis and Prediction for Services (PCAPS) project. The initiative combines efforts from observational scientists, AI developers, and physical modelers worldwide, focusing on determining the observational needs of AI-based sea ice prediction systems.

Notably, this community is heavily driven by early-career researchers eager to make a significant impact in the field of climate prediction and modeling.

The Importance of Observational Data

Sea ice prediction is a critical component of climate services that impact operational decision-making and the safety of high-latitude communities. Indigenous groups in the Arctic rely on reliable forecasts for safe hunting and travel. Vessels navigating these treacherous waters need accurate, near-real-time predictions to avoid ice floes. A variety of sectors, from fisheries to search and rescue teams, depend on these forecasts for effective planning and operations.

Initially, sea ice prediction systems arose from research efforts aimed at understanding the predictability of seasonal sea ice conditions. As the field matured, these systems evolved into a significant enterprise, underscored by innovative methodologies and scientific advances that now enable near-real-time forecasts at both poles. This rapid evolution is exemplified by initiatives like the Sub-seasonal to Seasonal (S2S) project and the Sea Ice Prediction Network (SIPN) South, which highlight the burgeoning capabilities of seasonal forecasting.

Diverse Approaches to Sea Ice Prediction

The approaches to sea ice prediction are increasingly varied, with ORCAS working to incorporate many methodologies. Traditional process-based dynamical climate models offer detailed simulations of interactions among the atmosphere, ocean, and ice. Despite their reliance on extensive historical datasets for testing and large ensembles for estimating uncertainty, these models can be resource-intensive and exhibit systematic biases.

In contrast, emerging AI-based methods, including data-driven models and physics-informed systems, present formidable alternatives. Running faster while maintaining predictive skill, these models often focus on short-range regional forecasts for navigation or long-term overviews of ice conditions across the Arctic. Nevertheless, few have honed in on Antarctica, and even fewer cater to specific local scales that are crucial for community safety and navigation.

The Observation Gap

As climate change progresses, polar regions are increasingly faced with unprecedented conditions. Unfortunately, this complicates the landscape for observational data collection, which remains limited and disparities in data availability raise significant challenges. Recent shifts in ice behavior—often not reflected in the historical data used for AI training—pose operational risks to prediction systems.

Addressing this issue involves both AI-ready and process-focused observations. AI-ready data are typically gridded, standardized, and accessible for use in machine learning. However, process-focused observations—rich in detail about ice thickness, drift, and deformation—are essential for evaluating the physical credibility of predictions. Yet, they tend to be localized and irregular, posing challenges for harmonization.

Building Trust in AI Predictions

The rapid proliferation of AI systems in earth sciences brings the pressing concern of ensuring their reliability and physical integrity. For sea ice prediction, trust relies not just on accuracy but also on how physically coherent the models are and their ability to provide useful insights across various conditions.

Through ORCAS, researchers aim to establish frameworks that allow for comprehensive evaluation of sea ice prediction systems. This includes using shared validation scenarios, standardized datasets, and metrics linking model performance to real-world ice processes. The emphasis is not just on prediction accuracy; rather, there’s a crucial need to ensure that AI systems can reliably predict conditions outside their training datasets, responding to the rapidly changing climatic landscape.

Observations for the New Era

ORCAS’s efforts represent a paradigm shift in prioritizing observations pertinent to AI applications. There’s a dual focus on ensuring that crucial observational data are adaptable for machine learning ingestion while also incorporating high-value process-focused datasets for benchmarking AI predictions against established physical laws. This strategic balancing act aims to enhance the physical understanding inherent to sea ice predictions.

As initiatives like SIPN South have proven effective in intercomparison of forecast skills, ORCAS zeroes in on another important question: Are AI predictions physically realistic? The curation of campaign and process study datasets for validation provides a pathway for addressing this need.

Applications Beyond Sea Ice

Sea ice presents one of the most challenging arenas for AI prediction, characterized by sparse data and rapidly changing conditions. Yet, the challenges faced here resonate across the broader Earth sciences, as advancements in AI prediction technologies often outpace the methodologies used to validate them.

By forging stronger links between AI forecasting systems and observational requirements, ORCAS not only aims to enhance its validation mechanisms in sea ice prediction but also aspires to contribute to a more trustworthy AI landscape across various fields. The lessons learned from sea ice could very well inform a wider array of scientific endeavors aimed at harnessing AI for environmental prediction.

In navigating this complex landscape, ORCAS remains committed to building a community that encourages collaboration, innovation, and shared goals in the pursuit of reliable and effective sea ice prediction systems.