Abstract This study develops an ensemble forecast system for Arctic sea ice with a horizontal resolution of ∼2 km to resolve linear kinematic features (LKFs). The system is designed to run operationally on the Sunway computing architecture, which employs many‐core processors with a distinct programming logic. The Parallel Data Assimilation Framework has been optimized for Sunway's architecture by leveraging fine‐grained thread‐level parallelism and restructuring memory hierarchies. Results from multiple process configurations show that the optimized analysis step achieves an average speedup of 9.19, while overall performance improves by an average of 3.66 compared to CPU‐based architectures. To address the strong nonlinearity of LKFs, we employ ice strength parameter perturbations and a novel localized observation error function that preserves valuable LKF information from both the model and sea ice concentration observations. Compared to Synthetic Aperture Radar observations at 1–3‐day lead times, the hindcast experiment yields a bidirectionally averaged minimum Hausdorff distance of 39.7–41.9 km for large‐scale LKFs, and outperforms persistence forecasts in spatial maximum cross‐correlation error within the 48 hr lead time. The probability density functions of LKF orientation and length largely follow the observed distributions. Results indicate that LKF predictability can be enhanced through model dynamics inherently from well‐initialized sea ice states like concentration or thickness, despite only limited details of LKFs in the observations used during assimilation. With higher‐frequency data assimilation, this system shows potential for operational LKF forecasting in support of practical Arctic navigation.
Mu et al. (Sun,) studied this question.