Pangu-Weather underestimates the 2019 European summer heatwave: A storyline comparison with ICON

Although it has been demonstrated that data-driven models can outperform physical models in simulating mean states, their capabilities to simulate extremes and their generalizability across climate scenarios remain questionable. This study compares Pangu-Weather to the physics-based ICON-CLM model in simulating peak of the July 2019 European summer heatwave and its unfolding in +2, +3, and +4 K future warmer scenarios, following a storyline approach. We initialize both Pangu-Weather and ICON-CLM using global storyline simulations from the coupled climate model AWI-CM1, which were spectrally nudged towards ERA5 reanalysis to maintain the large-scale flow and to accurately assess the impacts on heatwave severity due to thermodynamic changes only. To quantify uncertainties in Pangu-Weather’s predictions, the random field perturbation (RFP) technique is used to generate probabilistic forecasts. The study reveals that Pangu-Weather shows a systematic cold bias of about 2 – 3° C over the heatwave core region in both present and future warmer scenarios. Pangu-Weather fails to capture maximum 2-m temperature (T2m ) exceeding 45° C in +3 and over a limited area in +4 K global warming scenarios compared to ICON-CLM, indicating its limitations to accurately predict out-of-sample extremes even in an ensemble framework. ... mehrFurther, the maximum T2m response is linearly dependent on global warming levels consistent with ICON-CLM simulations but with reduced warming amplification. Additionally, it shows that Pangu-Weather needs enough time to leave the trajectory specified by forcing data. Our study emphasizes that the storyline approach and comparison with physics-based models show a promising way to test data-driven models for their applicability to warmer climate conditions.