Evaluation of a stochastically perturbed parametrisations scheme for sea ice in sub‐seasonal forecasts

Abstract

Within the European Centre for Medium‐Range Weather Forecasts (ECMWF) sub‐seasonal ensemble prediction system, sea ice forecasts exhibit pronounced underdispersion, with ensemble spread being systematically smaller than the mean squared error of the ensemble mean. A contributing source of this underdispersion is that key sea ice model parameters are poorly constrained and, in many cases, not directly measurable, yet these parameters are typically assigned a single constant value in the model. This neglects both their intrinsic uncertainty and their possible variability in space and time, leading to overconfident forecasts. To address this challenge, we explore the impact of a stochastically perturbed parametrisations (SPP) scheme for sea ice on sub‐seasonal forecasts, where nine key sea ice model parameters are perturbed stochastically to represent their uncertainty better. Using a prototype configuration of the coupled IFS‐NEMO4‐SI3 forecasting system for Integrated Forecasting System (IFS) cycle CY49R2, this study extends previous assessments of sea ice SPP in seasonal forecasting. The SPP scheme effectively increases the ensemble spread of sea ice concentration and thickness forecasts on sub‐seasonal time‐scales, while changing the mean state only negligibly. The increased spread establishes within a few days and saturates after about three weeks from initialisation. The increase in spread is most pronounced along the ice edge during boreal winter and over the central Arctic during summer, with relative increases of up to 30%. This leads to improved ensemble calibration and probabilistic skill for sea ice forecasts, though forecasts remain underdispersive in ways that likely reflect limitations beyond subgrid‐scale variability. Atmospheric impacts are small at sub‐seasonal lead times, likely reflecting the limited mean‐state adjustments introduced by SPP on these time‐scales. This also makes the sub‐seasonal range a particularly well‐suited testbed for sea ice SPP, as results are not strongly confounded by mean‐state changes that may emerge at longer lead times, providing a useful framework for further experimentation and evaluation.