The role of North Atlantic-European weather regimes in the surface impact of sudden stratospheric warming events

Sudden stratospheric warming (SSW) events can significantly impact tropospheric weather for a period of several weeks, in particular in the North Atlantic European (NAE) region. While the stratospheric forcing often projects onto the North Atlantic Oscillation (NAO), the tropospheric response to SSW events, if any, is highly variable and it remains an open question what determines the existence, location, timing, and strength of the downward impact. We here explore how the variable tropospheric response to SSW events in the NAE region can be characterised in terms of a refined set of seven weather 5 regimes and if the tropospheric flow in the North Atlantic region at the onset of SSW events is an indicator of the subsequent downward impact. The weather regime analysis reveals the Greenland blocking (GL) and Atlantic Trough (AT) regimes as the most frequent large-scale patterns in the weeks following an SSW. While the GL regime is dominated by high pressure over Greenland, AT is dominated by a southeastward shifted storm track in the North Atlantic. The flow evolution associated with GL and the associated cold conditions over Europe in the weeks following an SSW occur most frequently if a blocking 10 situation over western Europe and the North Sea (European Blocking) prevailed at the time of the SSW onset. In contrast, an AT regime associated with mild conditions is more likely if GL occurs already at SSW onset. For the remaining tropospheric flow regimes during SSW onset we cannot identify a dominant flow evolution. Although it remains unclear what causes these relationships, the results suggest that specific tropospheric states during the onset of the SSW are an indicator of the subsequent tropospheric flow evolution in the aftermath of an SSW, which could provide crucial guidance for subseasonal prediction. 15 Copyright statement. TEXT

1 of the troposphere itself is high, such that it can mask a stratospheric influence. Predicting if, when, and where a downward impact from SSW events will occur is therefore not straightforward, yet a better prediction of the type and timing of a downward impact would significantly benefit a wide range of users.
The tropospheric impact of SSW events is communicated by a range of mechanisms including synoptic and planetary-scale 25 waves (e.g. Song and Robinson, 2004;Domeisen et al., 2013;Hitchcock and Simpson, 2014;Smith and Scott, 2016). In particular, after SSW events the North Atlantic-European region (NAE) tends to exhibit more persistent states of the negative phase of the North Atlantic Oscillation (NAO-, Domeisen, 2019), as well as more frequent transitions towards NAO-and fewer away from NAO- (Charlton-Perez et al., 2018). NAO-is associated with enhanced meridional air mass exchanges, in particular, more cold air outbreaks in Northern Europe but fewer over the Nordic Seas (Kolstad et al., 2010;Kretschmer 30 et al., 2018b;Papritz and Grams, 2018;Huang and Tian, 2019), as well as increased precipitation in Southern Europe (Butler et al., 2017;Ayarzagüena et al., 2018). The Pacific sector tends to be less strongly affected in the aftermath of SSW events (Greatbatch et al., 2012;Butler et al., 2017), though the occurrence of wave reflection in the stratosphere can be associated with Pacific blocking (Kodera et al., 2016) and cold spells over North America (Kretschmer et al., 2018a). Given the preferred occurrence and the increased persistence of certain surface signatures after SSW events as compared to climatology, medium-35 to long-range predictability over Europe has been suggested to increase after SSW events (Sigmond et al., 2013;Domeisen et al., 2015;Karpechko, 2015;Butler et al., 2016;Scaife et al., 2016;Jia et al., 2017;Beerli et al., 2017;Butler et al., 2019;Domeisen et al., 2020a), although SSW events themselves are often not predictable beyond deterministic lead times (Taguchi, 2014(Taguchi, , 2016Domeisen et al., 2020b).
Despite the preferred occurrence of the negative phase of the NAO, the downward influence of an SSW event on the evolution 40 of the tropospheric flow can be highly variable between events. This issue is further complicated by the fact that there exists a range of different metrics for characterizing the downward impact, with each definition yielding a different set and number of SSW events with apparent surface impacts. In particular, the occurrence and type of downward impact has been investigated with respect to the SSW geometry (Charlton and Polvani, 2007;Mitchell et al., 2013;Maycock and Hitchcock, 2015;Seviour et al., 2016), though no statistically robust differences with respect to wave geometry emerge in the tropospheric response. 45 More promising pathways include SSW precursors (Nakagawa and Yamazaki, 2006;White et al., 2019), the evolution of the stratosphere -troposphere system following the SSW (Kodera et al., 2016), and in particular the persistence of the lower stratospheric response (Hitchcock et al., 2013a;Karpechko et al., 2017;Runde et al., 2016;Polichtchouk et al., 2018) after the SSW event. These studies use indices for the downward effect that are based on exclusively stratospheric or a combination of stratospheric and tropospheric indicators. In this study we will investigate purely tropospheric indicators of downward impact 50 for SSW events. Definitions of a downward impact using tropospheric indicators are generally based on large-scale circulation indices such as the NAO (Charlton-Perez et al., 2018;Domeisen, 2019) or tropospheric jet location (Garfinkel et al., 2013;Afargan-Gerstman and Domeisen, 2020;Maycock et al., 2020).
While a causal downward link from the stratosphere after SSW events has been confirmed in idealized experiments (e.g. Gerber et al., 2009), remote forcings can affect both the stratosphere and the troposphere, and thereby either mask or strengthen 55 the downward response from the stratosphere. Indeed, a range of tropical remote connections can impact the NAE region through both a tropospheric and a stratospheric pathway (Attard et al., 2019), such as the Quasi-Biennial Oscillation (QBO) (Gray et al., 2018;Andrews et al., 2019), the MJO (Garfinkel et al., 2014;Barnes et al., 2019), and the El Niño Southern Oscillation (ENSO) (Jiménez-Esteve and Domeisen, 2018;Domeisen et al., 2019), in addition to extratropical tropospheric forcing in the North Pacific (Honda and Nakamura, 2001;Sun and Tan, 2013;Drouard et al., 2013), Arctic sea ice (Sun et al., 60 2015), and snow cover in Eurasia (Cohen et al., 2014). It therefore has to be kept in mind that the stratosphere is often only one possible forcing of the troposphere. In addition, it has recently been suggested that the precursors to SSW events with a downward influence differ from those without such a tropospheric impact in terms of strength and location (Domeisen, 2019;Zhang et al., 2019), in particular with respect to forcing over Eurasia (White et al., 2019;Tyrrell et al., 2019;Peings, 2019).
Given the large variability of the tropospheric flow evolution following SSW events and the influence of other remote factors 65 mentioned above, the prediction of the SSW response in the troposphere is difficult for an individual event, despite the general shift towards NAO negative conditions in a statistical sense. The goal of this study is to investigate if tropospheric flow regimes in the NAE region help to understand the variability of the SSW response in the observational record. More specifically, we here address the question if the tropospheric flow evolution in the NAE region after an SSW is statistically different from that without an SSW using seven weather regimes in the NAE region. Weather regimes are quasi-stationary, recurrent, and 70 persistent patterns of the large-scale extratropical circulation (e.g. Michelangeli et al., 1995). While many studies showed that there are preferred transitions between different regimes, internal tropospheric variability is high and a regime onset often occurs on short timescales (e.g. Vautard, 1990;Michel and Rivière, 2011). Therefore predictability due to regimes arises from regime persistence on time scales of several days rather than typical regime sequences over several weeks. However, recent work revealed important shifts of regime occurrence and transition probabilities between regimes on subseasonal time scales 75 of several weeks dependent on the external forcing such as the stratospheric polar vortex state (Charlton-Perez et al., 2018;Papritz and Grams, 2018;Beerli and Grams, 2019). This motivates the study at hand aiming at investigating if the variability in the tropospheric flow evolution following SSW events can be characterised in terms of the weather regime at the time of SSW onset.   Table 2 in Butler et al. (2017). The SSW central dates for the remaining years are 12-Feb-2018 and 02-Jan-2019.

85
The tropospheric flow over the NAE region is described in terms of quasi-stationary large-scale flow patterns, given by seven year-round weather regimes defined in Grams et al. (2017) based on six-hourly data for the period 1979-2015. We use this weather regime classification to stratify SSWs according to the large-scale tropospheric flow conditions at their onset. To do so, we select for each SSW the dominant weather regime that is active during at least one 6-hourly time step throughout the onset day of the SSW (see Table 1). As for the canonical seasonal definition using four regimes (e.g. Michelangeli et al., 90 1995;Michel and Rivière, 2011;Ferranti et al., 2015;Charlton-Perez et al., 2018), the mean patterns of the seven regimes are based on a k-means clustering in the phase space spanned by the leading seven empirical orthogonal functions (EOFs; explaining 76% of the variance) of 10-day low-pass filtered 500hPa geopotential height anomalies. In addition, the normalized projection (weather regime index I W R ) following Michel and Rivière (2011) for each of the seven regimes is employed for defining objective weather regime life cycles and for a filtering of time steps without a clear regime structure ("no regime" 95 category). In essence, an active life cycle requires an I W R above a certain threshold for at least 5 consecutive days (minimum duration for an active regime life cycle) and a continuous increase/decrease during the onset/decay phases (see methods of Grams et al., 2017, for details). As different life cycles can be active simultaneously, in particular during the onset and decay phases, individual days are attributed to a specific regime life cycle only if I W R is also the maximum of all I W R . The life cycle definition allows for a continuous extension of the weather regime attribution to more recent data without repeating the EOF A potential modulation of the frequency of occurrence of the seven regimes can be understood in terms of the link between the respective regimes and the NAO (Beerli and Grams, 2019, their Figs. 2,6) and the link between the stratospheric polar vortex and the NAO (Charlton-Perez et al., 2018). Since ZO and ScTr project onto NAO+, they are suppressed after a weakening of the stratospheric polar vortex, and vice versa after a strengthening. In contrast, GL strongly projects onto NAO-and is enhanced 110 following a weak stratospheric polar vortex, while it is suppressed in the aftermath of a strong vortex. EuBL and AT do not project strongly onto either NAO phase and are, thus, only weakly modulated by the strength of the stratospheric polar vortex.

Statistical testing
The rare occurrence of SSWs (26 events during 1979-2019) and the subsequent stratification according to tropospheric flow conditions requires careful statistical testing to extract significant and robust results that are distinct from sampling uncertainty.

115
The overarching questions we address in this study are whether after SSWs the tropospheric flow evolution is different from situations without an SSW and to what extent this depends on the tropospheric state at the time of the SSW. To investigate the latter, we will consider subsamples of all SSWs. Hence, in all cases the relevant null hypothesis is that the flow evolution after SSWs is indistinguishable from that occurring in the absence of an SSW. The testing procedure, thus, comprises the following two steps: 120 1. First, we assess the robustness of the samples by performing a Monte Carlo resampling. For that purpose, we resample the original samples 100 times with repetitions. The number of random samples is chosen according to the maximum  1979, 24.02.1984, 01.01.1985, 11.02.2001, 21.01.2006, 07.01.2013 Scandinavian Blocking (ScBL) -Greenland Blocking (GL) 04. 03.1981, 08.12.1987, 30.12.2001, 24.02.2007, 09.02.2010 no regime at onset (no) 29.02.1980, 18.01.2003, 05.01.2004 number of possible combinations with repetitions of the smallest subset of SSW events that will be considered in this study (N = 5 events corresponding to 126 independent combinations). This yields confidence intervals, estimating the uncertainty inherent in each sample. Due to the small sample size, these confidence intervals are relatively large. Applying this method to geopotential height anomalies, we consider anomalies as robust if the width of the confidence interval is smaller than the amplitude of the anomaly. In addition, the sample mean is significant at, e.g., the 10% level, if the confidence intervals overlap by less than 10% with the Monte Carlo distribution. A similar procedure is applied to test significance of lagged weather regime occurrence.

135
3 Weather regimes during SSW events As a first step, we evaluate the sequence of weather regimes from 60 days before to 60 days after an SSW for all 26 SSW cases during 1979-2019 ( Fig. 1, cf. Tab. 1). This figure suggests a preferred occurrence of AT (purple) and GL (blue) during the weeks after an SSW compared to the weeks before. This is further emphasized by the 5-day running mean of the anomalous frequency of weather regimes around SSW events, which provides a more complete overview over the modulation of regime 140 frequencies after SSWs (Figure 2). We show the 5-day running mean frequency anomaly to account for the 5-day minimum duration of an active regime life cycle. Different from the testing procedure outlined in Section 2.2, we here consider the distribution of lagged 5-day mean frequencies by selecting for each day in the original sample a random day ±15 days around the original day of year but from a different winter. In addition, the random day must exhibit the same weather regime as the original day to replicate potential regime-dependence. We then compute the mean lagged weather regime frequency for each 145 random sample as for the original sample and test for significance at the 10% level (bold). For reference, we show the absolute frequencies of weather regimes in Fig. A2.
GL and EuBL are the most prominent regimes at the onset of SSW events with a relative 5-day mean frequency of more than 18% and 20%, respectively ( Figure A2a). The frequency of EuBL is significantly enhanced from 5 days prior until the onset of the SSW (Figure 2a), in agreement with Woollings et al. (2010) and Nishii et al. (2011). The cyclonic regimes ZO, ScTr, as 150 well as the blocked regimes AR and ScBL tend to be suppressed at the time of SSW events. This is consistent with the strong projection of the ZO and ScTr regimes onto NAO+, which also tends to be suppressed after SSW events (Charlton-Perez et al., 2018). On the other hand, AR (yellow, significant peak around lag -20 to -10) and the related ScTr (orange, significant range at lag -10) regimes are more frequent in the period 1-3 weeks before the SSW onset. The prominence of AR around 15 days before the onset of an SSW event agrees with the suggested precursor role of blocking over the Atlantic before SSW events 155 (Martius et al., 2009). After the SSW onset, AT frequencies are significantly enhanced, peaking at more than 23% after 7 days (Figure A2a) corresponding to a frequency anomaly of 15% for the same lag (Figure 2a). Thereafter, GL (lag 12 to 42 days) 6 and AT (lag 17-35 days) are the most likely weather regimes with enhanced frequency anomalies of up to 15% (Fig. A2a), but in absolute terms frequencies for both are only around 20-25% and none of the two clearly dominates (Figure 2a). This behaviour obscures the potential tropospheric impact of an SSW in a composite as AT and GL trigger contrasting large-scale 160 weather conditions (rather mild and windy for AT, cold and calm for GL) for large parts of Europe (Beerli and Grams, 2019). We now sub-divide the 26 SSW events with respect to the weather regime that dominates during the SSW onset: GL (5 cases), EuBL (6 cases), and the cyclonic regimes (ZO, ScTr, AT; 9 cases). The remaining 6 cases either have no clear regime signature (no-regime, 3 events) or are associated with AR (3 events) at their onset. Because of the small sample size, we do not consider these cases here. For the GL subset ( Fig. 2b / A2b), all other regimes are subsequently suppressed except for AT 165 and EuBL. The frequency of GL itself drops immediately after the SSW to below 10% around a lag of 20 days -far below its climatological mean frequency. AT, and to a lesser degree also EuBL, become significantly more frequent immediately after the SSW until about a lag of 10 days, reaching absolute frequencies of 35% and 20%, respectively (Fig. A2b). After a period with no clear regime assignment, AT becomes the dominant regime starting at lag 18 days with anomalous frequencies above 40% (Fig. 2b), peaking above 50% absolute frequency about 23 days after the SSW and remaining significantly enhanced until 170 7 a lag of 33 days (Fig. A2b). From lag 25 days until lag 40 days, EuBL becomes significantly enhanced peaking at 40% absolute frequency around lag 30 days.
For the EuBL subset ( Fig. 2c / A2c), the subsequent regime frequencies are quite different to GL at the onset of an SSW.
First, the frequencies of ScTr and AR are significantly enhanced directly after the SSW, with peaks at 20% and 30% absolute frequency at lag 10 days. This is then followed by a period of preferred occurrence of GL (around lag 20 days) and AT (around 175 lag 30 days) with a frequency of about 25% each. The dominance of GL from lag 35 to 45 days (50% peak frequency) is particularly striking. At the same time, also ScBL is enhanced with a frequency of 20%, while all other regimes are suppressed.
Cyclonic regimes at the time of the SSW (Fig. 2d / A2d) exhibit a less prominent regime frequency modulation after an SSW compared to the EuBL and GL subsets. Still, GL (lag 10-35 days, lag 45 days), AR (lag 20-30 days), and AT (lag 25-35 days) are significantly enhanced, but absolute frequencies barely exceed 20%. Note that this corresponds to significantly increased 180 frequencies of 10-20% for these regimes in the considered time windows. However, most often no single regime dominates after an SSW event with a cyclonic regime at lag 0, hinting at cases with a "missing" response after the SSW event.
Despite the large tropospheric variability in the aftermath of SSW events, the investigation of lagged regime frequencies reveals that (1) the AT and GL regimes are more likely to follow an SSW (as compared to other weather regimes) and (2)

190
We focus in the following on the modulation of the stratosphere-troposphere coupling for the previously discussed sets of SSWs. For that purpose, we evaluate the temporal evolution of standardized geopotential height anomalies averaged over the NAE sector (-80 • E to 40 • E / 60 • N to 90 • N) by compositing a given set of SSW events. Using the full hemisphere, that is, the full longitude range instead of the here used sectorial view over the North Atlantic, yields the same qualitative results due to the strong imprint of the anomalies induced by the SSW in the NAE sector (Fig. A3).

195
Compositing all SSW events (Fig. 3a)   . When compositing all SSW events, the downward impact between 10 to 60 days after the SSW onset is robust at the 25 % but not the 10 % level (see Fig. A4a). Together with the relatively weak amplitude of the anomalies, this reflects the 200 large case-to-case variability in the tropospheric impact of SSWs. A more robust anomaly (10 % level) can only be observed at around 15 days after an SSW. This anomaly around a lag of 15 days is unlikely to be obtained from a random sampling as evident from the less than 10 % overlap between the confidence and random distributions (Fig. 3a). This suggests that in the immediate aftermath of an SSW (lag 15 -25 days), indeed positive geopotential height anomalies over the NAE sector are significantly more likely than in the absence of an SSW.  SSW events that occur during GL (Fig. 3b) are associated with an immediate, strongly positive anomaly in the troposphere. Fig. A2b, when GL is present at the onset of the SSW, GL or AR are often already present before the SSW event, which is likely the cause of the positive tropospheric geopotential height anomalies several days prior to the event.

Consistent with
Notably, there are no significant and robust (cf. Figs. 3b and A4b) anomalies after 10 days of the onset of the SSW except for a weak negative geopotential height anomaly after 20 days, indicating a cyclonic flow regime in the NAE region. This is 210 consistent with the significantly enhanced likelihood for the occurrence of the AT regime at this lag (Fig. A2b). Note that both the immediate positive geopotential height anomalies and the weak tropospheric anomalies in the aftermath of the event are not the result of cancellations in the composites but are rather typical across cases.
For EuBL at the onset of the SSW event, only a weakly significant positive anomaly can be observed at the time of the SSW, but highly robust, significant, and strongly positive geopotential height anomalies are present in the troposphere at lags 215 of 15 -20 and 30 -55 days after the SSW event (Fig. 3c). These positive anomalies are consistent with the finding that first AR and then GL are much more likely in the aftermath of an SSW with EuBL at lag 0 (compare to Fig. A2c). Furthermore, comparing to the panel for all SSW events (Fig. 3a) indicates that the EuBL cases dominate the perceived downward response in the canonical response for SSW events.
During cyclonic regimes at the onset of the SSW, there is no substantial tropospheric anomaly in the NAE region at the 220 time of the SSW, but a positive albeit weak anomaly can be observed around days 15 -20 after the SSW event (Fig. 3d). This anomaly is not robust at the 10 % level, but it is significantly different from a random sample at least at the 25 % level. Several SSWs with a cyclonic regime at the onset are followed by GL at a longer lag (Fig. A2d), thus, likely causing these anomalies.
Still, the GL frequencies only reach 25% at most and also other regimes occur more often albeit with low frequencies around 25%. These findings and the small amplitude of the anomalies suggest that the variability in the tropospheric flow evolution 225 after SSWs is large after a cyclonic regime at lag 0, which is also confirmed by the inspection of individual cases (not shown).
While not the focus of this study, the question arises whether other factors might contribute to the differing surface evolution in the aftermath of the SSW event. In fact, it is certainly the case that tropospheric variability at SSW onset alone is not responsible for the response to SSWs. In particular, differing values of the amplitude and persistence of the lower stratospheric anomaly can be observed in Fig. 3 between the different composites. In general, the events with EuBL at the onset also tend 230 to have a longer stratospheric persistence, although a long (but slightly weaker) persistence can also be observed for cyclonic regimes at the onset of the SSW. The five SSW events associated with GL have a more short-lived lower stratospheric response.
As events with a persistent lower stratospheric response are often associated with so-called polar jet oscillation (PJO) events, a comparison with Table 1 in Karpechko et al. (2017) reveals that 2 out of 5 SSW events with a GL regime (and, respectively, 4 out of 6 EuBl events) at the onset have a persistent lower stratospheric temperature response characterized by a polar jet 235 oscillation (PJO) event (Kuroda and Kodera, 2004;Hitchcock et al., 2013b). While this is not a clear result, it indicates that the shorter (longer) persistence in the lower stratosphere for the SSWs associated with GL (EuBL) may add some support to the persistence of the tropospheric response, but the statistics are too small to provide a clear result. Similarly, 4 out of 6 EuBL events are split events (rather than displacements), while 2 out of 5 GL events are split events.

240
Since each weather regime is associated with characteristic surface weather, the modulation of regime successions in the aftermath of an SSW by the tropospheric state at the time of an SSW might contribute to the marked variability in the surface impact. Hence, we here consider spatial composites of 2m temperature anomalies and anomalies of 500 hPa geopotential height (Z500') for the three groups of SSW events discussed in the previous sections (Fig. 4a-c) and for all SSW events (Fig. 4d) for days 0 to 25 after the SSW (cf. Fig. A5 for days 25 -50).

245
During SSWs dominated by GL at the onset, initially strong warm anomalies prevail over Greenland and the Canadian Archipelago, whereas western Russia and Scandinavia are anomalously cold, consistent with the anomalous ridge over Greenland and the low geopotential height anomalies over Scandinavia (Fig. 4a). With the subsequent progression of weather regimes -typically towards the cyclonic AT regime or EuBL -mild conditions are established throughout central Europe from a lag of 20 days onwards. This is in stark contrast to the negative NAO and the associated cold conditions that are commonly expected 250 as the canonical response to SSWs over Europe (Butler et al., 2017;Kolstad et al., 2010;Domeisen et al., 2020a).
For SSWs that are dominated by EuBL at the onset, cold anomalies prevail over Northern Europe, albeit also extending over large parts of central Europe (Fig. 4b). They peak at -4 K to -6 K around lags beyond 20 days, which corresponds well with the occurrence of the GL regime. Note that cold anomalies in the composite for all SSWs are much weaker (cf. Fig. 4d) reaching -1 K to -3 K in Central Europe. The retrogression of initial positive Z500' over the eastern North Atlantic to Greenland along 255 with a strengthening of negative Z500' over the southeastern North Atlantic around lag 15-25 days is striking. Furthermore, GL is associated with warm anomalies over Greenland and North America.
Finally, as expected by the varied regime succession for the SSWs with cyclonic regimes at their onset, composite temperature and Z500' are weaker for these events (Fig. 4c). Thus, the canonical response of surface temperature (i.e., the composite for all SSWs, Figure 4d) is the result of averaging over -in important regions opposing -temperature anomalies for SSWs 260 with GL, EuBL, or a cyclonic regime at the onset.

Summary and Discussion
This study aimed to shed light on the large case-to-case variability of the tropospheric response to SSW events and their associated surface impacts, as well as the dependence on the tropospheric weather regime at the onset of the SSW. To that end, we have exploited in a statistical framework the observational record of the satellite era (1979 -2019) as represented in the 265 ERA-Interim reanalysis. Our conclusions are as follows: 1. In the aftermath of an SSW event, the tropospheric flow in the NAE region exhibits an evolution that is unlikely to occur in the absence of an SSW. Specifically, positive geopotential height anomalies related to Greenland blocking are statistically more likely to occur after the onset of the SSW than in the absence of an SSW. This is consistent with the expected (canonical) negative NAO response of the troposphere to SSWs (e.g. Charlton-Perez et al., 2018). 270 2. The significant and robust positive geopotential height anomalies found in the period 10-60 days after SSWs are predominantly the result of SSWs with European blocking at their onset. This is manifest for this subset of events in a transition from EuBL to GL that then dominates at lags of 15-20 and 30-55 days after the SSW onset, which is statistically significantly different from the natural progression from EuBL to GL. For other tropospheric regimes at SSW onset the tropospheric response is weaker and less robust and significant. 3. For Greenland blocking at the SSW onset, a weak preference for cyclonic flow regimes around 20-30 days after the SSW is apparent, with an opposite surface response in the aftermath of the SSW as compared to SSW onsets dominated by EuBL.
4. SSWs that occur during cyclonic weather regimes exhibit a considerably weaker and less significant response with a modestly enhanced likelihood for GL.

280
Depending on the tropospheric weather regime at SSW onset different surface signatures result. Specifically, the canonical signature in 2m temperature, e.g., cold conditions prevailing over much of northern Europe, occurs for the EuBL cases. In  contrast, mild temperatures in large parts of Europe are found for SSWs with GL at their onset. It is important to distinguish these cases, since although EuBL and GL frequently occur at the onset of SSW events, they lead to a different subsequent evolution and different associated surface temperatures. In particular, the most common transitions are from EuBL to GL and 285 from GL to AT around 3-4 weeks after the SSW, respectively, along with their contrasting large-scale weather impacts (Beerli and Grams, 2019). These findings corroborate that the presence of either a EuBL or GL regime at SSW onset will allow us to disentangle the difference in surface weather, and hence to determine if and when a "downward impact" of the SSW is expected. This is highly relevant for subseasonal forecasting.
While these findings are limited by the small sample size of available SSW events, the rigorous statistical testing for sig-290 nificance and robustness performed here suggests that the large case-to-case variability in the tropospheric response to SSWs can be described in terms of NAE weather regimes and depends on the regime at the onset of the SSW. Our findings confirm that while the stratosphere does not represent the sole forcing of the tropospheric state, for many events it may be able to nudge the tropospheric flow into a particular direction by suppressing some weather regimes and by favoring others, as found in Charlton-Perez et al. (2018). We here in addition show that the susceptibility of the troposphere to the stratospheric nudging 295 depends on the tropospheric state at the time of the SSW. Other factors that can modulate the tropospheric response are the persistence of the temperature anomaly in the lower stratosphere (Hitchcock et al., 2013a;Karpechko et al., 2017;Runde et al., 2016;Polichtchouk et al., 2018), as well as upstream effects in the eastern North Pacific (Afargan-Gerstman and Domeisen, 2020). An analysis of differences in the lower stratospheric persistence for the here considered weather regimes did not yield conclusive results, which warrants further studies. In particular, a model analysis to quantify the respective contributions to the 300 tropospheric impact of different remote factors in comparison to the role of local North Atlantic variability might shed further light onto the complex role of stratosphere -troposphere coupling in surface weather.
However, it is currently not sufficiently known to what extent complex prediction models are able to represent the diversity of tropospheric responses to stratospheric forcing, as this has not been sufficiently tested in models beyond the canonical response and selected case studies. From a preliminary analysis of S2S prediction model data we anticipate large biases and 305 a very complex role of the representation of stratosphere -troposphere coupling in prediction models that will be difficult to disentangle. Hence, while state-of-the-art subseasonal prediction systems are often unable to forecast at the time of occurrence of the SSW event if a surface response is to be expected, our findings suggests that the presence or absence -and in fact the timing -of a surface impact following SSW events might in some cases be predictable based on the weather regime at the onset of the SSW event. This could significantly improve the subseasonal prediction of tropospheric winter weather over Europe.

310
Data availability. The ERA-interim reanalysis data (Dee et al., 2011) is available from ECMWF at https://apps.ecmwf.int/datasets/data/interimfull-daily/.  sub-divided by the weather regime that is dominant at the onset of the SSW as indicated by the titles of the panels. Robustness is assessed using confidence intervals by resampling the SSW events 100 times with repetition. If the magnitude of the anomaly exceeds the interquartile or the 10 th -90 th percentile ranges the anomaly is highlighted by stippling or hatching, respectively. See Section 2.2 for details.