Can stratospheric nudging improve surface predictability? Insights from the 2019 Southern Hemisphere sudden stratospheric warming

Abstract

This study isolates the role of the 2019 Southern Hemisphere (SH) sudden stratospheric warming (SSW) for surface climate in eight Stratospheric Nudging And Predictable Surface Impacts project (SNAPSI) models. The novel nudging experiments allow for a clearer disentanglement of the SSW’s impact on surface climate than the standard free-running forecasts, in which the entire atmospheric system evolves freely and deviates from observed reality. SNAPSI models capture the downward propagation of the negative Southern Annular Mode (SAM) from the stratosphere when the zonally symmetric stratospheric circulation is nudged toward the observations. In addition, zonally asymmetric stratospheric variations amplify warm anomalies and dry conditions over eastern Australia, enhancing the Australian local high-pressure system. The composite analysis reveals that the zonally asymmetric stratospheric variations driven by the westward-shifted vortex influence the eastern Australian precipitation forecast by strengthening the downward coupling of the SAM and enhancing the meridional ridge-trough structure over Australia. Westward-shifted vortex events are identified in over 60% more ensemble members in the nudged run than in the free run, associated with precipitation closer to observed values in the nudged run. The wildfire weather potential risk in Australia is assessed using the Fraction Attributable Risk (FAR) of the Hot Dry Windy (HDW) wildfire weather index. The positive FAR of the HDW wildfire weather index indicates that the nudged stratosphere contributes up to ~30% to the increased wildfire weather risk along eastern and southern Australia, highlighting the contribution of the 2019 SSW event to Australian extreme weather from mid-October to mid-November.

Introduction

The extremely cold winter pole is enclosed by fast circumpolar westerlies from the troposphere to the stratosphere, and within the stratosphere, this singular feature of rotational cold air mass comprises the stratospheric polar vortex^1,2. While tropospheric variability typically evolves on shorter timescales, the stratospheric polar vortex changes more gradually. Stratospheric perturbations originating from a disrupted polar vortex can descend into the troposphere and modulate subseasonal forecast skill^3,4,5. Given the differences in the land-sea distribution and topography between the Southern and Northern Hemispheres, the Southern Hemisphere (SH) stratospheric polar vortex tends to be more circular, stronger, and is less frequently disturbed by planetary waves than the Northern Hemisphere (NH) counterpart^6,7. When an SSW occurs, the stratosphere over the polar region experiences a temperature increase of >30 K, and the polar westerly winds suddenly decelerate and even reverse to easterlies^8,9. Major SSWs refer to those events with a reversal of the westerly zonal winds at 10 hPa in the subpolar region, while minor ones are merely accompanied by rapid westerly deceleration without a complete reversal of the wind from westerly to easterly¹⁰.

Compared with the NH^11,12, major sudden stratospheric warmings (SSWs) are much more rarely observed in the SH¹³. A sudden decrease of 60 m/s in the SH circumpolar westerly winds at 10 hPa was observed in mid-September 2019 (without a wind reversal), and the wind deceleration magnitude for this “minor” SSW is even greater than that of some NH major SSWs^14,15,16. Since the Subseasonal to Seasonal (S2S) project launched in 2013^17,18, only one minor SSW occurred in the SH – during September/October 2019 – when an extreme Indian Ocean Dipole (IOD) event also occurred¹⁵. The stratospheric anomalies associated with the minor SSW in 2019, coupled with record-high Antarctic stratospheric ozone, propagated into the troposphere, likely contributing to the extreme drought and wildfires in Australia during late spring^14,15,19. These downward impacts were similar to those of the only observed SH major SSW in 2002^20,21. A study even showed that the impacts of this minor SSW were found in the NH tropics through an enhanced Brewer-Dobson circulation²².

In the NH and SH alike, stratospheric variability during SSW events can propagate downward and impact the troposphere and even the surface²³. Existing S2S models capture the tropospheric responses to some extent in free-running forecast experiments^24,25,26. However, the inconsistencies in ensemble size and in initialization and output strategies across models limit the ability to perform a fair model-by-model comparison. Further, the free experiments are limited by the models’ inherent bias in forecasting the SSW and thus may not accurately represent the observed stratospheric state during the event, which impacts the attribution of predictability sources for surface forecasts. Separating the stratospheric impact on surface climate in S2S models continues to be a complex problem. To push forward a better understanding of the stratospheric role in surface predictability, the Stratospheric Nudging And Predictable Surface Impacts (SNAPSI) project was initiated²⁷. The SNAPSI project required a uniform initialization time and a comparably large ensemble size of at least 50 members for three recent SSW events, one of which was the September 2019 minor SSW event in the SH. Further, in one experiment, the zonal-mean component of the stratospheric wind and temperature is nudged to reanalysis, allowing for the isolation of its role in the surface forecast. The large ensemble size in the SNAPSI project, compared to the original S2S project in which some models had fewer than 10–15 members, helps to reduce the influence of unrelated tropospheric variability and sampling noise, thereby enhancing the robustness and statistical significance of the stratospheric signal^28,29.

Forecasts for the September 2019 SH SSW were available in eight SNAPSI models, and each model provides free runs, nudged runs, and control runs (for details, see “Methods” section). Previous studies have emphasized the stratospheric role in the Australia hot and dry weather (and therefore wildfires) in the SH spring–early summer of 2019^14,15, yet we still do not clearly understand the origins of near-surface predictability, and whether nudging the stratosphere improves the surface forecast. In this study, we seek to answer the following question: To what extent do the stratospheric circulation variations contribute to the near-surface forecast of Australia's hot and dry weather?

Results

Observed evolution of the September 2019 SH SSW

We begin with the evolutions of the circumpolar winds and polar height in Fig. 1a, b. The climatological zonal-mean zonal wind at 10 hPa and 60°S in SH late winter is ~80 m/s, peaking in late August. Afterward, the climatological westerlies gradually weaken due to increased solar radiation reaching the Antarctic. Westerlies decelerate by ~60 m/s in the following two months, reaching ~20 m/s at the end of October. After that, westerlies reverse in late November and early December when the climatological final warming typically happens³⁰.

Fig. 1: Stratospheric circulation anomalies during the 2019 minor SSW event.

a Evolution of zonal-mean zonal winds at 10 hPa and 60°S from 29 August to 31 October for every year between 1979 and 2018 (black dashed curves). The climatology of the zonal-mean zonal winds is shown in the thickened green curve. The evolution of zonal winds in 2019 is highlighted in the red curve, with the 2002 major SSW case in the yellow curve as a comparison. The horizontal and vertical reference lines denote the threshold for a minor SSW event and the onset date (15 September) of the September 2019 minor event, respectively. b As in (a) but for polar area-averaged geopotential height at 10 hPa. c–h Synoptic charts of geopotential height at 10 hPa for the weekly mean in September and October 2019.

The westerly wind at the end of August 2019 was ~70 m/s, nearly ~10 m/s slower than the climatology (Fig. 1a). In the following 20 days, the westerlies weakened dramatically to 20 m/s on 15 September 2019 when a minor SSW occurred¹⁵. The weakest wind speed was ~10 m/s on 17 September. In the subsequent two weeks, the wind slightly recovered but remained below 28 m/s. It then decreased again to around 20 m/s, where it persisted for another half month before decelerating further. Finally, the westerly winds reversed to easterly winds on 30 October 2019 without any recovery until the following autumn, denoting an earlier final warming. As a comparison, the westerly wind during the 2002 major SSW event experienced a sharp decrease of ~80 m/s within 20 days and dropped below −20 m/s at the end of September (Fig. 1a).

The 2019 minor SSW impacted other aspects of the stratospheric circulation¹⁵. The polar cap geopotential height experienced extremely large positive anomalies during the SSW event. Polar cap height at 10 hPa around the SSW onset was ~30,280 gpm (compared with the climatology of ~28,600 gpm). The height soon further increased to ~30,500 gpm for a few days, and oscillated around the 30,280 gpm in the following two months (Fig. 1b).

While the increase in polar cap height captures the overall weakening of the vortex, further insight into the zonal structure of the vortex anomaly is evident in Fig. 1c–h, which shows maps of height anomalies for every 7-day interval before and after the SSW. Before the SSW onset, the vortex strength gradually weakened, manifesting as a zonal wavenumber 1 (wave 1) pattern as the vortex center moved toward the Antarctic Peninsula and a high developed over the Southern Indo-Pacific Oceans. The vortex center was displaced toward ~75°S in the Western Hemisphere (Fig. 1c). During 15–21 September, the high shifted eastward toward New Zealand and the daughter vortex nearly collapsed with a clear elliptical shape (Fig. 1d). During the following week, the daughter vortex continued to weaken (Fig. 1e) and moved back toward the South Pole. However, the residual daughter vortex never reaches the pole and remains displaced toward the Atlantic sector, as a ridge is evident over the Pacific sector (Fig. 1f–h); the small-sized vortex remnant and Australian ridge persisted in the following months until it completely collapsed when the final warming occurred. The key point is that the observed vortex stayed predominantly in the Western Hemisphere and extended toward the Antarctic Peninsula during the minor SSW, which maintains a distinct zonally asymmetric component well into October, and we will next demonstrate that this zonal structure matters for the surface impact over populated regions.

Forecasts of downward impacts of the 2019 SH minor SSW event in SNAPSI models

Previous studies reported that the historical hot and dry weather during the Australian spring 2019 was primarily caused by the weakened Antarctic polar vortex, which propelled an equatorward shift of the descending branch of the Hadley cell and the subsequent high-pressure system over Australia^14,19. Following the 2019 SH minor SSW, Australia suffered from prolonged and devastating wildfires peaking in December, which garnered widespread attention. An anomalously strong negative SAM propagated downward into the troposphere, which was the primary driver for the near-surface extremes^19,31,32. Rao et al.¹⁵ examined the near-surface predictability in S2S models, initialized at 29 August and 5 September, after two weeks relative to the 2019 SH minor SSW onset, and suggested that attribution of Australian climate extremes was complex due to the co-occurrence of the SSW with a moderate Central Pacific El Niño and a strong IOD. Other studies indicated that the impacts of this minor SSW event propagated much more slowly into the troposphere than other NH major SSW events^27,33. However, few studies have investigated the pure tropospheric and surface responses to this event using forecast models. We focus here on the initialization on 1 October 2019 (at a lag of 16 days after the SSW onset), as it allows us to quantify the role of the stratospheric signals for subsequent surface impacts.

Figure 2a illustrates the pressure-time evolution of the normalized SAM index derived from ERA5. The negative SAM index was −3.5 standard deviations in the austral spring (September–November). The minor SSW did not excite an instant downward propagation of negative SAM into the troposphere in September. Rather, the negative SAM in the stratosphere following the SSW descended to the troposphere only until 18 October, more than a month after this SSW event. Tropospheric negative SAM formed in late austral spring.

Fig. 2: Observed and forecasted SAM evolution during the 2019 minor SSW event.

a Pressure-time evolution of the SAM index in ERA5. The horizontal reference line marks 500 hPa and 10 hPa, and the vertical reference lines denote the SSW onset and the days with the largest tropospheric negative SAM. b–i Forecasts of the SAM evolution at 500 hPa (SAM500; bottom) and 10 hPa (SAM10; top) for control, free, nudged, and nudged-full run experiments. The black solid curve denotes the SAM index from the observation. The green solid curve denotes the control run ensemble mean (note that the stratospheric state is nudged gradually (not instantly) toward the climatology to maintain computational stability, so the stratospheric SAM in the starting days is not zero), the red solid curve denotes the free run ensemble mean, the light blue solid curve denotes the nudged run ensemble mean, and the deep blue dashed curve denotes the nudged-full run ensemble mean (only available for ECMWF and UKMO). The legend is displayed at the top of the figure.

Forecasts of the SAM evolution at 500 hPa and at 10 hPa are examined for SNAPSI models in Fig. 2b–i. The tropospheric negative SAM index exceeded 2 standard deviations around 25 October. However, all models underestimate the negative tropospheric SAM in both the free (red) and nudged (light blue) run experiments. The free runs generally capture the timing of the downward-propagating SAM as indicated by forecasts trending toward a tropospheric negative SAM around 10 October. The persistent stratospheric negative SAM is captured by most models before 25 October, albeit with relatively weaker intensity. However, SNAPSI models miss the stratospheric signals in the free run initialized on 1 October, and the negative SAM is underestimated for both the stratosphere and troposphere. CMA only captures a transient tropospheric negative SAM during mid-October (Fig. 2b). NCAR even forecasts a tropospheric positive SAM, which is mainly due to the too early reversal of the stratospheric SAM sign around 20 October (Fig. 2g). By nudging the stratospheric zonal-mean circulation, forecasts of negative SAM at 500 hPa improve compared to those from the free run for most models. The most negative SAM in the nudged run falls below −1 standard deviation for most models. The free run of Meteo-France is the only one that consistently remains below −1 standard deviation throughout the 10-day period centered on the peak negative SAM at 500 hPa (25 October), and the free run and nudged run share comparable skill in this model, although the tropospheric negative SAM on 25 October is underestimated (Fig. 2f). The difference between nudged (light blue) and nudged-full (deep blue) runs in ECMWF and UKMO is not significant (Fig. 2d, i), since the SAM mainly reflects the atmospheric variability of the zonal-mean symmetric circulation. In addition, the stratospheric SAM in the control (green) run remains close to zero across all models after 5 days following initialization, consistent with the nudging of the stratospheric zonal-mean circulation toward the natural evolution of climatology. However, the tropospheric SAM exhibits divergence among the control runs. Most control runs simulate a neutral tropospheric SAM after 25 October except for CMA, CNR-ISAC, and NCAR (Fig. 2b, c, g). Notably, the control runs of CMA and NCAR tend to produce a tropospheric positive SAM, even when the initialization is gradually adjusted toward a stratospheric neutral SAM, which partially explains the tropospheric positive SAM in their free runs (Fig. 2b, g). In contrast, the control run of CNR-ISAC simulates a tropospheric negative SAM (around -1 standard deviation) under a stratospheric neutral SAM (Fig. 2c). Note that the control run includes the IOD forcing, and hence there is little agreement across models as to whether the IOD induced a tropospheric SAM response in October and November 2019. In contrast, there is agreement across the models that the stratospheric nudging induced a negative tropospheric SAM in this period.

The negative SAM anomalies propagated into the troposphere around 18 October, intensifying and persisting throughout October and November (Fig. 2a). A zonally elongated anomalous high-pressure system subsequently formed over the southwest of Australia (Fig. 3a1, a2), corresponding to subsequent hot and dry weather. The Australian anomalous high-pressure system developed with a clear negative SAM pattern in the extratropics (Fig. 3a). The Australian local high was associated with the development of a wave train emanating from the tropics to the SH extratropics^34,35. All models forecast the positive geopotential height anomalies over Australia in the free run compared to the ERA5 climatology, although the height anomaly pattern diverges among models (Fig. 3b1–i1). CMA and NCAR forecast overly broad and strong positive height anomalies over Australia and fail to capture the southern trough (Fig. 3b1, g1). The remaining models perform better than these two in representing the spatial extent of the ridge-trough structure.

Fig. 3: Observed and forecasted 500 hPa geopotential height anomalies.

a1 ERA5 geopotential height anomalies averaged from 18 October to 14 November at 500 hPa for Australia and a2 the SH, respectively. b1–i1 Forecasts of geopotential height anomalies averaged from 18 October to 14 November at 500 hPa for Australia in the free run relative to the ERA5 climatology from eight SNAPSI models, and b2–i2 forecast height anomaly differences averaged from 18 October to 14 November at 500 hPa for the SH between the nudged and control run. d3, e3 The differences between nudged-full and control run for ECMWF and UKMO. Dots in the free runs mark the model anomalies relative to the simultaneous ERA5 climatology at the 95% confidence level according to a one-sample t-test. Dots in the nudged/nudged-full minus control run mark the difference between the two ensemble means at the 95% confidence level according to a two-sample t-test. The sample size refers to the number of ensemble members for each model (see Table 1 in the “Methods” section).

The differences between the nudged run and control run denote the contribution from stratospheric zonally symmetric variations during the 2019 SSW to the tropospheric prediction (Fig. 3b2–i2). The stratospheric nudging leads to a clear negative SAM for all models, characterized by positive geopotential height anomalies over the polar region and negative anomalies over the mid-latitudes, as seen in the differences between the nudged and control runs. However, the SAM signal is weaker in nearly all forecast models relative to the observed SAM in ERA5. The nudged run minus control run difference pattern is fairly zonally symmetric in all models: positive height anomalies cover the Antarctic, while negative height anomalies cover the midlatitude oceans. One exception is CMA (Fig. 3b2): it successfully reproduces the Australian local high, continuously connected with the Antarctic high anomalies. However, the CMA control run reproduces a tropospheric positive SAM during the focused period (18 October–14 November), which partially enlarges the positive difference in midlatitude height anomalies between the nudged and control runs (Fig. 3b). In the remaining models, the observed anomalous high over Australia, South Africa and southern South America at 500 hPa are missing in the nudged minus control difference field.

To identify the contribution from stratospheric zonally asymmetric variations, results from the nudged-full run minus the control run are displayed for ECMWF and UKMO (Fig. 3d3, e3). Even though the polar cap antinode of the tropospheric SAM is nearly the same in nudged-full runs as compared to nudged runs (Fig. 2d, i), the midlatitude antinode of the tropospheric SAM pattern is stronger in both models, especially in the Indo-Pacific sector. Furthermore, the anomalous high-pressure system over subtropical Australia is only evident in nudged-full, though its location is shifted from that observed. Therefore, the tropospheric response over Australia is more robust when the stratospheric zonally asymmetric variations are included.

There are several possible explanations as to why the nudged minus control difference does not explain much of the tropospheric circulation anomalies over the Australian region in both the ERA5 reanalysis and free runs. First, the predictability of high pressure over land could arise due to other forcings such as the positive IOD SST forcing and the Central Pacific El Niño. As shown in Fig. 2b–i, the free run tends to lose the stratospheric SAM signal quickly within 20 days of the 1 October initialization. Second, the stratosphere–troposphere coupling in SNAPSI models may be too weak to affect subtropical changes, and SNAPSI models could have limited skill in forecasting the stratospheric signals over land (especially over Australia) in the SH³⁶. Third, the zonally symmetric nudging technique does not account for zonal shifts in the vortex center, thus masking the potential contribution from the stratospheric zonally asymmetric variations.

Forecasts of Australian climate extremes in SNAPSI models

Forecasts of the 2-meter temperature (t2m) anomalies over Australia and the SH by eight SNAPSI models are shown in Fig. 4. Due to the long-lasting anomalous high over Australia, this continent experienced anomalously hot weather with the peak warming in this period along the western coast of Australia (Fig. 4a1). The temperature anomaly reached 3 K over western Australia and around 2 K over the eastern coast. As the negative SAM descended to the near surface, broad warm anomalies prevailed over the Antarctic with a magnitude of 3–5 K, while cold anomalies emerged around east of the Ross Sea (Fig. 4a2). In contrast, the midlatitude t2m anomalies were relatively weak. Further, South Africa and southern South America also underwent extremely hot weather¹⁴. In the free run, almost all models simulate hotter weather to varying degrees over Australia than in ERA5. For example, in the eastern half of Australia, ERA5 shows only a narrow band of anomalous warming along the east coast, with magnitudes of approximately 0.5–1 K, and only a few areas reaching 1–2 K. In contrast, all models exhibit anomalous warming of above 1–2 K across the entire eastern half of Australia, with CMA, CNR-ISAC, and Meteo-France generally showing anomalous warming of 2–3 K, and up to 3–4 K over central-eastern Australia (Fig. 4b1, c1, f1). ERA5 shows the peak warming over the southern half of Western Australia, while models forecast the peak warming over the eastern coast of Australia (Fig. 4b1–i1), though note that the observed peak warming shifted east after this period¹⁴.

Fig. 4: Observed and forecasted 2-meter temperature anomalies.

a As in Fig. 3a, but for ERA5 2-m temperature (t2m) anomalies averaged from 18 October to 14 November for Australia and the SH, respectively. b1–i1 Forecasts of t2m anomalies averaged from 18 October to 14 November for Australia in the free run relative to the ERA5 climatology from eight SNAPSI models, and b2–i2 forecast t2m anomaly differences averaged from 18 October to 14 November for the SH between the nudged and control run. d3, e3 The differences between nudged-full and control runs for ECMWF and UKMO. Dots in the free runs mark the model anomalies relative to the simultaneous ERA5 climatology at the 95% confidence level according to a one-sample t-test. Dots in the nudged/nudged-full minus control run mark the difference between the two ensemble means at the 95% confidence level according to a two-sample t-test. The sample size refers to the number of ensemble members for each model (see Table 1 in the “Methods” section).

The stratospheric contribution to the t2m predictability (nudged minus control) is largest in CMA (Fig. 4b2), indicating that the stratosphere accounts for 1–2 K of the surface warming over the southwest of Australia. However, it also overestimates the cold and warm anomalies in the Antarctic continent. These strong surface temperature impacts in CMA over Australia are consistent with the collocated height response (Fig. 3b2). The stratospheric signals shown by ECMWF and UKMO are second in strength only to CMA, reaching about 0.2–0.5 K over central-eastern Australia (Fig. 4d2, e2). The stratospheric signal in the remaining models is limited to the Antarctic with a maximum t2m anomaly magnitude of 1–3 K (Fig. 4c2–i2, d3, e3), relatively weaker than the observed t2m anomalies, 3–5 K (Fig. 4a2). The stratospheric zonally asymmetric variation in nudged-full runs partially contributed to the warm anomalies in central-eastern Australia (0.5–1 K, much higher than their nudged runs) and the Antarctic region (1–2 K), reaffirming the influence of stratospheric zonally asymmetric variation on near-surface predictability (Fig. 4d3, e3). The contribution of the zonally symmetric stratospheric signals to the t2m forecast is relatively weak over Australia across eight SNAPSI models (Figs. 4b1–i1). This might imply that the stratosphere–troposphere coupling in SNAPSI models is not fully captured, or that the downward propagation of the zonally symmetric stratospheric SAM contributed minimally to the predictability of Australian hot weather in October to nearly November. The predictability of the Australian hot weather in 2019 was sourced from both the zonally asymmetric component of the minor SSW event and the surrounding ocean forcing.

Forecasts of the precipitation anomalies over Australia and the SH by eight SNAPSI models are shown in Fig. 5. The persistent drought conditions over Australia after this minor SSW event are consistent with anomalously decreased precipitation (Fig. 5a1). Dry conditions were also prevalent across South Africa, Southeast Asian archipelago, and southern South America (Fig. 5a2). We mainly focus on dry conditions over Australia, which are forecast by six models except CNR-ISAC and NCAR (Fig. 5c1, g1) in the free run. Only SNU forecasts the drought anomalies that exceed −2 mm/day along the east coast of Australia in the free run (Fig. 5h1). Other models forecast the dry center along the northern coast but miss the dry center along the eastern coast.

Fig. 5: Observed and forecasted precipitation anomalies.

a As in Fig. 3a, but for ERA5 precipitation anomalies (units: mm/day) averaged from 18 October to 14 November for Australia and the SH, respectively. b1–i1 Forecasts of precipitation anomalies averaged from 18 October to 14 November for Australia in the free run relative to the ERA5 climatology from eight SNAPSI models, and b2–i2 forecast precipitation anomaly differences averaged from 18 October to 14 November at 500 hPa for the SH between the nudged and control run. d3, e3 The differences between nudged-full and control run for ECMWF and UKMO. Dots in the free runs mark the model anomalies relative to the simultaneous ERA5 climatology at the 95% confidence level according to a one-sample t-test. Dots in the nudged/nudged-full minus control run mark the difference between the two ensemble means at the 95% confidence level according to a two-sample t-test. The sample size refers to the number of ensemble members for each model (see Table 1 in the “Methods” section).

To quantify the stratospheric role in precipitation forecasts, the nudged minus control run difference is shown in Fig. 5b2–i2. CMA, UKMO, and Meteo-France capture a small area of negative precipitation anomalies over eastern Australia (Fig. 5b2, e2, f2). The dry conditions over South America are limitedly forecast by CMA, CNR-ISAC, and NCAR (Fig. 5b2, c2, g2), while only Meteo-France reproduces the dry conditions over South Africa (Fig. 5f2). By comparison, the stratospheric zonally asymmetric variations slightly better explain the dry condition along the eastern coast of Australia (Fig. 5d3, e3).

Thus far, we have demonstrated that midlatitude surface impacts from the SSW rely on successfully capturing the zonally asymmetric component of the vortex anomalies. This motivates the next part of our study: to what extent do the SNAPSI experiments employing either zonal-mean nudging or no nudging capture stratospheric zonally asymmetric variations and their subsequent impacts on surface weather?

The role of vortex center shifts in tropospheric and precipitation forecasts

The Southern Hemisphere stratospheric polar vortex remained in the Atlantic sector throughout this SSW event (Fig. 1c–h), and this zonally asymmetric stratospheric circulation variation is simulated by the nudged-full runs. Previous studies have confirmed the synoptic-scale and decadal-scale shifts in the NH polar vortex and its underlying effects on surface climate regionally^37,38. Therefore, to what extent the SNAPSI models capture the vortex center shifts, and what impact the center shifts of the polar vortex lead to, will be explored in this section.

Using elliptical diagnostics, the vortex center longitude is calculated for the 10 hPa daily geopotential height field (see “Methods” section). The Pearson correlation coefficient is calculated between the vortex center longitude and the Eastern Australia (EA; 20°S–40°S, 145°E–155°E; see also yellow boxes in Fig. 7) precipitation anomalies, averaged from 18 October to 14 November, across the multimodel multi-ensemble members (totaling 352) for control, free, and nudged run experiments (see Fig. S1 in Supplementary Information). The EA precipitation anomaly is around −1.2 mm/day during this period, and the vortex center longitude is around 40°W in the ERA5 reanalysis. The control runs exhibit a correlation coefficient of ~0.2, and the free and nudged runs both exhibit a coefficient above 0.3 (Fig. S1). This indicates that a westward shift of the polar vortex is moderately but statistically significantly linked with more decrease in EA precipitation. Both the control and nudged runs employed zonally symmetric nudging, which preserves the model’s intrinsic variability in vortex position. By contrast, the nudged-full run included zonally asymmetric nudging, which strictly constrains the vortex center to observed positions, as shown in Fig. S1. This explains the minimal variability in nudged-full runs, and so the coefficient is small and not significant.

Generally, the center longitudes in the multiple model ensemble members spread from 60°W to 30°E. We equally divide the forecasts into three groups: westward-shifted vortices, eastward-shifted vortices, and the remaining (the neutral). A westward-shifted vortex is one with a center west of 30°W (toward the Antarctic Peninsula) and an eastward-shifted vortex with the center east of 0°E (toward southern Africa). This classification aligns well with observations: during the 2019 SSW event, the vortex was westward-shifted (~40°W; gray dashed vertical line in Fig. S1), while the climatological mean vortex lies near the central longitudes (~15°W; black dashed vertical line in Fig. S1), representing the neutral state. The composite analysis reveals that the mean westward-shifted vortex centers in the control, free, and nudged runs are situated around 45°W, 43°W, and 43°W, respectively, and the sample sizes are 145, 78, and 126, respectively (see Fig. S2 in Supplementary Information). In contrast, the mean eastward-shifted vortex centers are around 14°E, 14°E, and 24°E, respectively, and the sample sizes are 54, 116, and 98, respectively. It seems that SNAPSI models tend to simulate a westward-shifted vortex in the control runs (compared to the observed climatological position of ~15°W), and the number is nearly three times that of the eastward-shifted vortices. The zonally symmetric nudging to observations leads to an increase in westward-shifted centers as compared to the free runs (126 compared to 78), and a decrease in eastward-shifted centers (98 compared to 116). This suggests that nudging the vortex strength zonally symmetrically can, to some extent, improve the vortex location simulation.

Figure 6 displays the composite tropospheric anomalies for the two types of vortex center shifts. For the control run, a westward-shifted vortex tends to generate larger and broader positive geopotential height anomalies within the lower latitudes as compared to an eastward-shifted vortex (Fig. 6a). In contrast, when the 2019 SH SSW occurs, free and nudged run members that capture a westward-shifted vortex also forecast a stronger negative SAM pattern (Fig. 6b, c). This might indicate that a westward-shifted vortex can strengthen stratosphere–troposphere coupling during the 2019 SH SSW event.

Fig. 6: Composite 500 hPa geopotential height anomalies for different vortex center shifts.

a1–c1 Composite 500 hPa geopotential height anomalies averaged from 18 October to 14 November, for westward-shifted vortex centers (west of 30°W) for control, free, and nudged runs, respectively, using multimodel multi-ensemble members (totaling 352, with CMA excluded). Dots mark the composite anomalies relative to the simultaneous ERA5 climatology at the 95% confidence level according to a one-sample t-test. a2–c2 As in (a1–c1), but for eastward-shifted vortex centers (east of 0°E). a3–c3 Composite difference between westward-shifted vortex centers. Dots mark the difference between two composite means at the 95% confidence level according to a two-sample t-test. N represents the sample size in each composite. The Meridional Ridge-Trough Index (MRTI) is calculated as the height difference between two boxes (20°S–40°S, 120°E–180°; 40°S–60°S, 120°E–180°).

Previous research primarily links the Australian late spring hot and dry conditions with the anomalously intensified Australian local high^14,19. The meridional circulation dipole has been emphasized repeatedly to account for EA summer heatwaves – a subtropical ridge over Australia accompanied by a midlatitude trough to its south^39,40 (red and blue boxes in Fig. 6). The presence of a meridional ridge-trough dipole structure is reported to enhance offshore flow, thus channeling dry air from the continental interior to cooler coastal locations³⁹. This highlights the importance of both the Australian local high and the southern trough. In our composite analysis, such a meridional ridge-trough dipole structure also emerges and is significantly different between the two types of vortex center shifts. We calculate the height difference between the ridge (20°S–40°S, 120°E–180°) and trough (40°S–60°S, 120°E–180°) marked in two boxes as the Meridional Ridge-Trough Index (MRTI) to quantify the intensity of this dipole structure. It is revealed that the MRTI dipole is enhanced under the westward-shifted vortex condition in all experiments (Fig. 6a–c). Comparing the three runs, the MRTI dipole is enhanced under the westward-shifted vortex conditions, with the dipole structure strongest in the nudged run (left column in Fig. 6). These findings also align with Gibson et al.³⁹ that the negative SAM condition is favorable for this meridional ridge-trough structure.

In short, it is inferred that the vortex center shifts can significantly impact the tropospheric circulations. Namely, zonally asymmetric stratospheric variations might influence the EA precipitation forecast by shifting the vortex toward the Antarctic Peninsula, projecting onto stronger downward propagation of the negative SAM signals into the troposphere, and enhancing the MRTI dipole.

A shifted polar vortex can also influence precipitation. We demonstrate this by repeating the analysis of Fig. 6, but now for the total precipitation anomalies (Fig. 7). A westward-shifted vortex is consistently associated with larger negative EA precipitation anomalies compared to its eastward-shifted counterpart across the three experiments, and this difference even reaches over −0.5 mm/day when the SSW occurs (Fig. 7a3–c3). For the westward-shifted vortex centers (left column in Fig. 7), the negative EA precipitation anomalies intensify progressively across the three experiments from control, free to nudged run, consistent with the corresponding strengthening of the MRTI dipole (Fig. 6a1–c1). The nudged EA precipitation anomalies are strengthened to −1.25 mm/day if only the westward-shifted vortex cases are selected, comparable to the observations (−1.2 mm/day in ERA5).

Fig. 7: Composite precipitation anomalies for different vortex center shifts.

As in Fig. 6, but for total precipitation anomalies. a1–c1 Composite precipitation anomalies averaged from 18 October to 14 November, for westward-shifted vortex centers (west of 30°W) for control, free, and nudged runs, respectively, using multimodel multi-ensemble members (totaling 352, with CMA excluded). Dots mark the composite anomalies relative to the simultaneous ERA5 climatology at the 95% confidence level according to a one-sample t-test. a2–c2 As in (a1–c1), but for eastward-shifted vortex centers (east of 0°E). a3–c3 Composite difference between westward-shifted vortex centers. Dots mark the difference between two composite means at the 95% confidence level according to a two-sample t-test. N represents the sample size in each composite. The precipitation anomalies within Eastern Australia (EA) are calculated by averaging the yellow box (20°S–40°S, 145°E–155°E).

The modeling evidence demonstrates that EA precipitation is significantly impacted by the stratospheric vortex center shifts (and therefore the zonally asymmetric stratospheric variations). The westward-shifted vortex during the 2019 SH SSW event enhanced the Australian dry conditions. The increase in the sample size of westward-shifted vortex cases from the free to the nudged experiments (from 78 to 126; Fig. 7b1–c1) indicates that the nudging technique also helps to improve the simulation of the vortex center. The nudged run members reproduce the observed EA negative precipitation anomalies if they also capture the westward-shifted vortex center.

To support the robustness of the above conclusions, we tested two compositing methods. The first, used in the main text, averages across all samples that meet the respective criteria, regardless of model origin—meaning that models contributing more west- or east-shifted members may have a greater influence on the composite. The second method computes composites within each model and then takes the multimodel mean (see Figs. S7–S9 in Supplementary Information). While the latter yields slightly weaker differences between the west- and east-shifted composites, the overall conclusions remain consistent. Given that the aim of this section is to examine the relationship between vortex shifts and precipitation responses, we chose to retain the first approach, as its larger sample size helps improve the statistical robustness of the results.

Attribution of Australian wildfire conditions in 2019 to the stratosphere

In late spring 2019, Australia experienced exceptionally hot and dry conditions, and consequent severe wildfires, causing enormous societal and environmental damage^14,41. To assess the Australian wildfire risk and to quantify the stratospheric contribution from a probabilistic forecast perspective, we employed the Hot Dry Windy (HDW) wildfire weather index and the Fraction of Attributable Risk (FAR). The HDW index is derived from near-surface wind speed and 2-meter air temperature, along with specific humidity at 925 hPa. A positive FAR indicates a higher probability of extreme wildfire weather events in the nudged run compared to the control run (see the “Methods” section).

Figure 8 presents the FAR of the HDW wildfire index, averaged over 15 October to 14 November, from six SNAPSI models (variables used to calculate HDW are only available for those six models). Five of the six models (except CNR-ISAC; Fig. 8a) simulate a positive FAR across most of Australia, with the 2019 SSW event explaining 20%–30% of the increased wildfire weather risk along the eastern coast. In the Meteo-France (Fig. 8d), stratospheric variations account for 30%–50% of the increased wildfire weather risk over a large portion of eastern Australia, with the attributable fraction exceeding 50% near Brisbane. This increase is primarily associated with a higher vapor pressure deficit, rather than changes in wind speed. The enhanced vapor pressure deficit results from a combination of higher temperatures and lower humidity (for details, see the diagnostic analysis of relevant physical variables in Figs. S3–S6 of the Supplementary Information). In contrast, the CNR-ISAC model fails to forecast the positive FAR and increased wildfire weather risk along the eastern coast and northern Australia, likely due to its unrealistic near-surface temperature and humidity forecasts (Figs. S3–S5).

Fig. 8: Forecasts of wildfire risk.

a–f Fraction of Attributable Risk (FAR) of the Hot Dry Windy (HDW) wildfire weather index averaged from 15 October to 14 November for Australia from six SNAPSI models. See the “Methods” section for the calculation of the FAR of the HDW wildfire index.

Overall, most models confirm the stratospheric contribution to the increased wildfire weather risk during 15 October–14 November, highlighting the potential role of the 2019 SSW event in explaining the extreme hot and dry conditions, and thus wildfire weather onset across Australia in late spring 2019.

Discussion

Previous studies have examined the predictability of the rare minor SSW occurring in September 2019 in the SH^14,15,16. However, those studies diverged in the relative role of the stratosphere for forecasts of Australia's climate extremes. This research provides new evidence to quantify the contribution of the stratospheric signals to the predictability of near-surface weather extremes in the SH under the SNAPSI protocol. In the experiments initialized on 1st October 2019, the nudged run from most models fails to capture a strong stratospheric contribution to the surface predictability over midlatitude continents during 18 October–14 November 2019. Instead, most SNAPSI models present a much more zonally symmetric response in the troposphere in the nudged run relative to the control run, with positive height anomalies over Antarctica and negative height anomalies over SH midlatitude oceans. Using nudged-full run minus control run from two models (ECMWF and UKMO), the role of stratospheric zonally asymmetric variation is also investigated. Compared with the impact of the stratosphere nudged toward the zonal-mean circulation, the zonal asymmetry of the stratospheric circulation indeed amplifies the warm anomalies (0.5–1 K; Fig. 4d3, e3) and dry conditions (−0.2 to −0.5 mm/d; Fig. 5d3, e3) over EA by enhancing the Australian local high at 500 hPa. Such an effect is evident if we subselect ensemble members with a westward-shifted vortex as was observed. Namely, the observed vortex center was located around 40°W during the focused period (Fig. S1), and forecasts with the vortex center shifted westward toward the Antarctic Peninsula produce a stronger tropospheric SAM response than forecasts with the vortex center shifted eastward toward South Africa. This is due to a more realistic prediction of the Australian meridional ridge-trough dipole. As a consequence, the forecast precipitation from the SNAPSI models is sensitive to the vortex shifts. The EA precipitation difference between a westward-shifted and an eastward-shifted vortex center is up to −0.5 mm/day in the free and nudged run experiments (Fig. 7b3, c3), consistent with the differences in their tropospheric responses (Fig. 6b3, c3). Further, the zonally symmetric nudged stratosphere reproduces over 60% more occurrences of westward-shifted vortex cases than the free runs (The samples of westward-shifted vortex cases in the nudged and free runs are 126 and 78, respectively, and therefore the samples from nudged runs are (126/78−1)*100% = 60% more; Fig. 6b1, c1). Therefore, the EA precipitation is better forecast in the nudged run members with a westward-shifted vortex center (−1.25 mm/day in nudged runs and −1.2 mm/day in ERA5). During late spring 2019, Australia also suffered from severe wildfires. The positive FAR of the HDW wildfire weather index along eastern and southern Australia reveals that the nudged stratosphere accounts for up to 30% and even higher ratios of the increase in wildfire weather risk.

It is acknowledged that stratospheric nudging improves surface forecasts from both ensemble and probabilistic perspectives, particularly when the nudged stratosphere captures a westward-shifted vortex center. While the nudging technique provides valuable insights into the contribution of the stratosphere to surface forecasts, we acknowledge that its impact on predictability is not uniform across all surface variables or regions. For instance, the improvement in near-surface winds over Australia is relatively modest compared with that in other diagnosed surface variables (Figs. S3–S6). Other factors could still influence our assessment of the isolated stratospheric signal. For example, the diversity within the nudged run minus control run difference in some models might be owing to systematic biases in the tropospheric SAM response in the control run (Fig. 2b, c, g). These biases might amplify/suppress the difference between the nudged run and the control run, potentially masking the stratospheric signal. In addition, the SAM-linked Australian local high is missing in the nudged minus control difference for some models, although the free run forecasts it (Fig. 3c1–i1). Finally, while the stratosphere–troposphere coupling during the 2019 SSW event has been widely investigated, the interference of oceanic forcings in the stratospheric signals still remains uncertain. The Australian local high and the extreme hot and dry weather are not only linked with the negative SAM associated with the downward propagation of the stratospheric disturbance, but are also explained by other known and unknown drivers^14,33,42.

The lag between this SH minor SSW event and its maximum impacts on the near surface is longer (more than a month until 18 October 2019) as compared to NH major SSW events²⁷. Although this period partially overlaps with the climatological early final warming, the average stratospheric wind speed difference between the nudged run and the control run during 18 October to 14 November still reaches ~20 m/s (Fig. S10 in Supplementary Information), indicating a persistent and non-negligible influence. Since the SNAPSI project does not include ocean-targeted control sensitivity experiments, it is difficult to quantitatively assess the relative roles of the IOD and the El Niño in Australian weather forecasts, in comparison to those of the SSW and the associated negative SAM during austral spring in 2019. To disentangle the respective contributions of the oceanic and stratospheric drivers and to gain deeper insights into forecast predictability in the free experiment, additional targeted experiments will be essential in future research.

Methods

SNAPSI experiments

The SNAPSI project aims to explore forecasts of extreme stratospheric polar vortex events and the stratospheric role in surface weather variations. For the September 2019 SH minor SSW event, two initialization dates are set, and three experiments are commonly run²⁷. Two dates (29 August and 1 October 2019) were chosen for initialization as the SNAPSI protocol requires. The primary purpose of the first initialization is to investigate SSW forecasts, while for the second initialization, the primary purpose is to evaluate the following downward impacts of the SSW. The data output interval is every 6 h, and the total length of the forecasts is 45 days. Three experiments, namely free, nudged, and control runs, are generated for different purposes. Specifically, free runs can be used to examine the original forecast skill in each individual model without any factitious intervention after initialization. Nudged runs represent the “perfect” stratospheric simulation with the stratosphere relaxed toward observations using a zonal-mean stratospheric nudging technique. Similar to nudged runs, the stratosphere (ua, va, and ta) in the control run was relaxed toward the zonal-mean climatology to examine the tropospheric predictability under the natural evolution of the stratosphere. Based on the nudged and control runs, where zonally symmetric nudging is adopted, the nudged-full run additionally contains the zonally asymmetric components, which help us understand the contribution from zonally asymmetric stratospheric variation.

Models and data

Eight models from the SNAPSI project provided the forecast data at the very beginning of this study (May 2024). Those eight models are CMA (BCC-CSM2-HR⁾⁴³, CNR-ISAC (GLOBO)⁴⁴, ECMWF (IFS)⁴⁵, KMA (GloSea6-GC32)⁴⁶, Meteo-France (CNRM-CM6-1)⁴⁷, NCAR (CESM2-CAM6)⁴⁸, SNU (GRIMs)⁴⁹, and UKMO (GloSea6)⁵⁰. They are provided by eight model developing institutes from different countries. Models have a varying horizontal resolution: ECMWF has the highest horizontal resolution, TCo639-319 (about 16–31 km), while NCAR has the coarsest resolution, 0.9° × 1.25° in latitude × longitude (about 95 km). The stratosphere is included in all eight models: the highest model top is 85 km (about 0.005 hPa) in KMA, and the lowest model top is 2 hPa in NCAR and SNU. Nearly all the available models provide at least 50 members for each initialization and each experiment, except for CMA, which currently provides only 4 members. Models provided 6-hourly outputs, and their average on a calendar day is computed as the daily mean for easy handling. For full expansion of each model, please refer to Table 1.

Table 1 Brief information for the forecast models available in this study

Since the SNAPSI project uses the ERA5 reanalysis^27,51 as the reference line for its nudged experiment, this reanalysis is also used to denote the realistic conditions. The monthly and daily climatology are computed as the mean for each calendar month and day during 1979–2018 at a horizontal resolution of 1.5° × 1.5° and 37 vertical levels (the local model resolution is 0.25° × 0.25° and model top is 1 hPa).

Calculation of indices and statistical methods

The Southern Annular Mode (SAM) index is calculated to track the downward propagation of the stratospheric signals associated with the SSW. The SAM index is defined as Eq. (1):

$${\rm{SAM}}={\bar{{Z}^{* }}}^{\text{GM}}-{\bar{{Z}^{* }}}^{\text{SP}}$$

(1)

where ${Z}^{* }$ represents the seasonally detrended geopotential height. The overbar plus superscript ${GM}$ denotes the global area-averaged mean, while the overbar plus superscript ${SP}$ denotes the polar area-averaged mean (South of 60°S). The SAM index is standardized by dividing the raw SAM index by its standard deviation during September–November⁵².

The vortex center is calculated through elliptical diagnostics in the planar coordinate using the stereographic projection, and then transformed back into the spherical coordinate^7,53,54.

The hot dry windy (HDW) wildfire weather index⁵⁵ is calculated through Eq. (2):

$${HDW}=U\times {\text{VPD}}\left(T,q\right)$$

(2)

where $U$ represents the near-surface total wind speed, and $\text{VPD}$ represents the vapor pressure deficit. In Eq. (2):

$${VPD}={e}_{s}\left(T\right)-e\left(q\right)$$

(3)

$${e}_{s}\left(T\right)=6.112\times \exp \left(\frac{17.67T}{T+243.5}\right)$$

(4)

$$e\left(q\right)=q\times \frac{P}{0.622+0.378q}$$

(5)

where ${e}_{s}$ is the saturation vapor pressure, $e$ is the actual vapor pressure, $T$ is the near-surface air temperature, $q$ is the specific humidity at 925 hPa, and $P$ is 925 hPa.

The fraction of attributable risk (FAR)⁵⁶ is calculated through Eq. (6):

$${FAR}=1-\frac{10 \% }{{p}_{90{th}}}$$

(6)

The 90th percentile in the control run was chosen as the threshold to indicate dangerous conditions. The ${p}_{90{th}}$ represents the probability of a nudged run exceeding the threshold, and 10% corresponds to the probability of a control run exceeding this threshold. ${FAR} > 0$ represents an increased chance of dangerous conditions.