Recent deep learning weather prediction (DLWP) models have continued to improve in forecast skill. Since the publication of Pangu-Weather in 2022, DLWP models have showed better performance than ECMWF’s Integrated Forecasting System (IFS), while req...
Recent deep learning weather prediction (DLWP) models have continued to improve in forecast skill. Since the publication of Pangu-Weather in 2022, DLWP models have showed better performance than ECMWF’s Integrated Forecasting System (IFS), while requiring far less computational time. Despite these advances, evaluation of DLWP models for extreme weather events remains limited.
In this study, I assessed whether DLWP models can realistically simulate an extreme wintertime polar stratospheric event, a Stratospheric Sudden Warming (SSW), through its dynamical process. SSWs are triggered when planetary waves in the upper troposphere and lower stratosphere propagate upward into the stratosphere, converge, and deposit momentum and energy. This wave forcing disrupts the polar vortex, often through displacement or splitting, alters the mean stratospheric circulation, and produces a rapid temperature increase over a short period. Following SSWs, their influence can extend downward to the troposphere and surface. In the polar troposphere, geopotential height anomalies (GPHA) tend to increase. At the surface, sea-level pressure (SLP) anomalies also increase over high latitudes, and cold-air outbreaks can develop over Eurasia. Because these impacts can persist on subseasonal-to-seasonal (S2S) timescales, SSWs provide a demanding dynamical test for forecast models. I focused on the 2018 major SSW, a case with a clear and well-documented downward influence, to examine DLWP model behavior from a dynamical perspective.
My experimental design followed the SNAPSI project methodology. I conducted three sets of experiments: Free, Nudged, and Control. In the Free experiment, forecasts were produced without any additional constraints. In the Nudged and Control experiments, forecasts were initialized after nudging stratospheric zonal wind (U) and temperature (T) toward two different reference states: ERA5 (Nudged) and the ERA5 climatology (Control). I analyzed five DLWP models—Pangu-Weather, GraphCast, FuXi, FengWu, and FourCastNet v2—and used the operational numerical model ECMWF IFS (cycle 48r1) as a reference. All experiments were initialized four days before the onset of the 2018 major SSW and were run with 6-hourly output for a 30-day lead time.
All models captured the key elements of the 2018 major SSW, including the upward propagation of planetary waves into the stratosphere and the resulting splitting of the polar vortex. Among the DLWP models, GraphCast most clearly reproduced the polar vortex split into two distinct vortices at the event’s onset. After the SSW, GraphCast, FuXi, and FengWu also reproduced tropospheric and surface responses comparable to those in the dynamical reference, including enhanced polar-cap GPHA and increased SLP, with broadly similar spatial patterns. GraphCast further captured the amplitude of these anomalies more realistically and reproduced signals consistent with Eurasian cold-air outbreaks.
To evaluate whether the DLWP models maintained hydrostatic balance, I additionally decomposed GPHA into contributions from temperature and surface pressure (SP). GraphCast showed temperature and SP contributions comparable to those in IFS. For all DLWP models except FengWu, the summed temperature and SP contributions closely matched the predicted GPHA. This agreement suggests that these models produce forecasts that account for hydrostatic balance.
Overall, this study evaluated DLWP models using the dynamical process of an extreme event. I showed that GraphCast reproduced not only the development of the 2018 major SSW but also its downward influence, in broad agreement with the dynamical reference model (IFS).