Global Ensemble Weather Prediction from a Deep Learning–Based Model (Pangu-Weather) with the CMA-GEPS Initial Condition Perturbations

Pangu-Weather (PGW) trained with deep learning–based methods (DL-based model) shows significant potential for global medium-range weather forecasting. However, the interpretability and trustworthiness of global medium-range DL-based models raise many concerns. This study uses singular vector (SV) initial condition (IC) perturbations of the China Meteorological Administration Global Ensemble Prediction System (CMA-GEPS) as inputs of PGW for global ensemble prediction (PGW-GEPS) to investigate the ensemble forecast sensitivity of DL-based models to the IC errors. Meanwhile, the CMA-GEPS forecasts serve as benchmarks for comparison and verification. The spatial structures and prediction performance of PGW-GEPS are discussed and compared to CMA-GEPS based on seasonal ensemble experiments. The results show that the ensemble mean and dispersion of PGW-GEPS are similar to those of CMA-GEPS in the medium range but with smoother forecasts. Meanwhile, PGW-GEPS is sensitive to the SV IC perturbations. Specifically, PGW-GEPS can generate realistic ensemble spread beyond the sub-synoptic scale (wavenumbers ≤ 64) with SV IC perturbations. However, PGW’s kinetic energy is significantly reduced under the sub-synoptic scale, leading to error growth behavior inconsistent with CMA-GEPS in the sub-synoptic scale. Thus, this behavior indicates that the effective resolution of PGW-GEPS is beyond the sub-synoptic scale and is limited to predicting mesoscale atmospheric motions. In terms of the global medium-range ensemble prediction performance, the probability prediction skill of PGW-GEPS is comparable to CMA-GEPS in the extratropics when they use the same IC perturbations. That means that PGW has a general ability to provide skillful global medium-range forecasts with different ICs from NWP.