Dataset of Sensitivity Experiments Using the Second-Order Sampling Method within the IAP-CAS System

An English version consistent with atmospheric-science manuscript style is: To evaluate the impact of ensemble initialization strategies on MJO forecast performance and to optimize the ensemble generation method, a sensitivity experiment dataset targeting individual MJO events was constructed based on the IAP-CAS S2S forecasting system. The experiments focus on representative MJO events occurring during the boreal winters of 2019–2023 (November to April). MJO events were identified using the Real-time Multivariate MJO (RMM) index, and cases with relatively strong amplitudes and clear propagation characteristics were selected as forecast events. For each selected MJO event, ensemble forecasts were initialized from the corresponding start dates. The forecast length was set to 35 days to cover the full life cycle of the MJO from development to decay. All experiments employed the same atmospheric model configuration, with differences only in the ensemble initialization strategy, allowing a direct comparison of different ensemble generation approaches. For ensemble perturbation generation, the Second-Order Exact Sampling (SOES) method was introduced. A large historical sample of model states was used as a statistical dataset to extract the dominant modes of error covariance. The historical samples were obtained from long-term hindcast experiments of the model. By applying singular value decomposition to these historical samples, the leading structures associated with error growth were identified, and physically consistent initial perturbations were generated accordingly. Compared with traditional random perturbation methods, this approach can more realistically represent model uncertainty while maintaining relatively low computational cost.