Random Noise in the Ice: How Stochastic Perturbations Sharpen Seasonal Forecasts

Every winter, forecasters at the European Centre for Medium-Range Weather Forecasts (ECMWF) attempt one of the hardest problems in Earth science: predicting how the Arctic’s sea ice will evolve over the coming months, and how that evolution will ripple into the weather of northern Europe. A new study published in the journal Weather and Climate Dynamics suggests that the missing ingredient in these forecasts is, counterintuitively, randomness itself. Kristian Strommen of ECMWF and his colleagues show that injecting carefully constructed random perturbations into the sea ice component of the centre’s flagship forecast model makes sea ice predictions more reliable — and, remarkably, appears to improve seasonal forecasts of the winter atmospheric circulation over Europe and the North Atlantic.

The problem the team set out to solve is known in the forecasting community as underdispersion. Ensemble forecasts work by running a model many times with slightly different starting conditions and perturbed physics, producing a cloud of possible futures whose spread is meant to mirror the true uncertainty of the forecast. When the ensemble spread is much smaller than the average forecast error, the ensemble is underdispersive: it is overconfident, presenting a range of outcomes that is too narrow to encompass what actually happens. Sea ice forecasts, the authors note, are systematically and often severely underdispersive — even at the moment they are initialised, before any forecast error has had time to grow. This is likely tied to the large biases that models exhibit in sea ice and to the poor representation of unresolved processes such as ice fracturing, ridging, and the turbulent exchange of heat between ice, ocean, and atmosphere.

The remedy tested in the study is a stochastically perturbed parameterisation, or SPP, scheme applied to SI³, the state-of-the-art sea ice model now coupled to ECMWF’s Integrated Forecast System. Rather than perturbing the ice directly, the scheme perturbs nine physical parameters inside the model — including the ice strength parameter P*, the ice–ocean drag coefficient, and snow-related properties — multiplying each by a log-normally distributed random field that evolves in space and time. The random field is generated by repeatedly applying a Shapiro filter to white noise, producing smooth spatial correlations on scales of roughly 750 kilometres, and it decorrelates over about ten days. Because the perturbations act on parameters rather than on tendencies or energies, the scheme conservatively respects tracers, momentum, and sea ice volume, keeping it physically consistent with the stochastic schemes already used in the atmospheric component of the model since the late 1990s.

To evaluate the scheme, the team ran 50-member ensemble forecasts initialised on 1 November and 1 May for every year from 1993 to 2023, each running six months to cover full winter and summer seasons. The experiments used a prototype of cycle 49R2 of the IFS, the configuration that will underpin the next-generation seasonal forecast system SEAS6 and the ERA6 reanalysis. Two ensembles were compared: a control run and an otherwise identical run with sea ice SPP switched on. The only difference between them was the stochastic noise in the ice, which means any divergence between the two ensembles must ultimately trace back to the sea ice itself.

The headline result is a robust increase of roughly 10 percent in ensemble spread for both sea ice concentration and sea ice thickness, in both summer and winter, bringing the spread–error ratio closer to the ideal value of one that defines a well-calibrated probabilistic forecast. But the noise did something unexpected as well. Despite being mean-zero by construction, the perturbations systematically shifted the mean state: sea ice concentration decreased on average, thickness increased, and in winter the ice was redistributed outward from the central Arctic pack toward the ice edge, with drift changes on the order of 20 percent of baseline velocities. The authors explain this through a simple but powerful conceptual model built on two asymmetries. First, sea ice concentration is bounded between zero and 100 percent, and its distribution is strongly bimodal: perturbations can only push near-total ice cover downward, while gridpoints knocked to zero effectively vanish from the scheme’s view, so the mean tends to fall. Second, in winter the dense ice pack resists convergence — piling ice up costs energy — whereas there is little resistance to moving ice toward the open edge, so random perturbations to advection preferentially push ice outward along the path of least resistance. A toy model based on repeatedly perturbing a truncated bimodal distribution reproduces exactly this signature of a declining mean.

These ice changes did not stay confined to the cryosphere. Because fluctuations in ice concentration and thickness modulate surface heat fluxes and albedo, the perturbations altered air temperature up to at least the 850 hectopascal level, with changes of order half a degree Celsius. In winter, the central Arctic warmed where ice cover thinned, allowing more heat to escape the ocean, while regions along the Greenland, Barents, and Kara sea ice edges cooled where the perturbations increased ice concentration. Circulation anomalies, including a cyclonic feature near the Ural Mountains that advected cold Siberian air over the Barents and Kara seas, reinforced these thermodynamic responses. Ensemble spread in air temperature also increased by around 10 percent, both over the Arctic and to some extent at midlatitudes. In summer, by contrast, the atmospheric response was much smaller, as Arctic surface temperatures are pinned near the freezing point and the ice has little leverage on the air above it.

The most striking finding concerns forecast skill beyond the ice itself. In winter, the anomaly correlation coefficient of 500 hectopascal geopotential height — the standard yardstick for seasonal circulation forecasts — increased substantially over the subpolar North Atlantic and extended into Scandinavia, with the difference peaking near 0.5. Averaged over a broad subpolar North Atlantic box, the correlation rose from about 0.3 in the control to about 0.6 with the perturbations switched on, and the winter North Atlantic Oscillation, the dominant mode of European winter variability, was forecast somewhat better as well. The authors are careful to stress the statistical caveats: with only about 30 winters in the sample, field significance was not achieved, and Monte Carlo resampling suggests a roughly 40 percent chance that the pattern of significant gridpoints could arise by chance. Yet the same improvement appeared in several additional hindcasts with different configurations, and the region of enhanced skill coincides with where the ice was most strongly affected, lending credibility to the signal.

The proposed mechanism is as intriguing as the result. The control model carries a uniform warm bias across the Arctic, which weakens the meridional temperature gradient in the subpolar North Atlantic. By increasing ice concentration along the Greenland, Barents, and Kara sea edges, the stochastic perturbations cool the overlying atmosphere and partially correct that warm bias, potentially strengthening the temperature gradient that steers the North Atlantic jet stream. Bootstrap resampling of the perturbed ensemble revealed a statistically significant correlation of −0.28 between the degree of cooling along the ice edge and the skill of the North Atlantic circulation forecast, consistent with this pathway. The authors caution that this accounts for only about 8 percent of the variability, that causality is difficult to establish in a system this noisy, and that some of the improvement may reflect bias compensation rather than a genuinely better simulation of the coupled system. Whether the benefit transfers to other models will likely depend on their own baseline biases.

The study also uncovered a subtlety with practical consequences for how such schemes should be deployed. In uncoupled ocean-and-ice-only simulations, the perturbations were dramatically more effective at generating spread than in fully coupled forecasts, where the atmosphere dominates the error growth and may already saturate the achievable spread; switching to independent random fields for each parameter boosted spread by up to 60 percent in the uncoupled case but made little difference when coupled. This suggests the scheme could be especially valuable in data assimilation, where enhancing the spread of ocean and ice initial conditions is a recognised priority, and work is already under way to apply a variant of the scheme in that context. The team also found no evidence that the perturbations strengthened the much-debated teleconnection between Barents–Kara sea ice and the North Atlantic Oscillation reported in earlier climate-model work, hinting that in these forecasts the relevant pathway may run instead through Greenland sea ice.

For operational forecasting, the message is that a modest dose of well-designed randomness in the ice can pay dividends far beyond the polar cap. The scheme achieved its primary goal of better-calibrated sea ice probabilities, and it did so while introducing mean state changes small enough — around 4 percent in concentration and 10 centimetres in thickness — that modest retuning could offset them if desired. As ECMWF prepares its next generation of seasonal and sub-seasonal systems, and as climate change continues to make Arctic sea ice both more variable and more consequential for midlatitude weather, the idea that noise, properly structured, is a form of signal may prove one of the more quietly important lessons in modern numerical weather prediction.

Subject of Research: Stochastic sea ice perturbation schemes for improving seasonal ensemble forecasts of sea ice and the midlatitude atmosphere

Article Title: The impact of stochastic sea ice perturbations on seasonal forecasts

Article References: Strommen, K., Mayer, M., Storto, A., Spaeth, J., & Tietsche, S. (2026). The impact of stochastic sea ice perturbations on seasonal forecasts. Weather and Climate Dynamics, 7(3), 1593-1618. https://doi.org/10.5194/wcd-7-1593-2026

Image Credits: AI Generated

DOI: 10.5194/wcd-7-1593-2026

Keywords: sea ice, seasonal forecasting, stochastic parameterisation, ensemble forecasts, ECMWF, Arctic, North Atlantic Oscillation, underdispersion, sea ice concentration, jet stream, weather and climate dynamics, coupled modelling