Scientists have developed research systems to test new ideas in data assimilation, but these often lack the efficiency and robustness needed for operational use. We addressed this gap with key innovations: a flexible observation database, first guess at the appropriate time, and modular, parallelised software enabling the assimilation of millions of observations.

We tested a new way to improve Arctic sea ice forecasts by adding satellite-based surface brightness, or albedo, into a sea ice model. This approach captures key surface changes like melting and snowfall that affect ice loss. We found it often gives better results than using standard data like ice coverage or thickness, especially during the melt season. This method offers a powerful tool for tracking Arctic sea ice in a changing climate.

Integrating observations into complex sea ice models improves model estimates, but the impact of specific kinds of observations may vary in space and time. By modeling sea ice at single locations, this work quantifies the impact of four different observation kinds on sea ice at three characteristic locations in the Arctic. The results indicate that this simplified experimental framework is a useful tool for developing methods to meld new and existing observations with modern sea ice models.

Statistically combining models and observations with data assimilation (DA) can improve sea ice forecasts but must address several challenges, including irregularity in ice thickness and coverage over the ocean. Using a sea ice column model, we show that novel, bounds-aware DA methods outperform traditional methods for sea ice. Additionally, thickness observations at sub-grid scales improve modeled ice estimates of both thick and thin ice, a finding relevant for forecasting applications.

Accurate sea ice conditions are crucial for quality sea ice projections, which have been connected to rapid warming over the Arctic. Knowing which observations to assimilate into models will help produce more accurate sea ice conditions. We found that not assimilating sea ice concentration led to more accurate sea ice states. The methods typically used to assimilate observations in our models apply assumptions to variables that are not well suited for sea ice because they are bounded variables.