Atmospheric circulation compounds anthropogenic warming and impacts of climate extremes in Europe

Significance

We address the key question of whether and how the dynamics of the atmosphere may enhance the impacts of anthropogenic climate change. We specifically focus on temperature and surface wind extremes in Europe, two classes of events which have caused a high death toll and large insured losses over the continent in the recent decades. We find that large-scale atmospheric patterns which favor summertime heatwaves and wintertime windstorms over large parts of the continent are becoming increasingly frequent. This effect sums to that of the average global climate change. A key implication of our work is that circulation changes modulate extreme events already in the present climate.

Abstract

Diagnosing dynamical changes in the climate system, such as those in atmospheric circulation patterns, remains challenging. Here, we study 1950 to 2021 trends in the frequency of occurrence of atmospheric circulation patterns over the North Atlantic. Roughly 7% of atmospheric circulation patterns display significant occurrence trends, yet they have major impacts on surface climate. Increasingly frequent patterns drive heatwaves across Europe and enhanced wintertime storminess in the northern part of the continent. Over 91% of recent heatwave-related deaths and 33% of high-impact windstorms in Europe were concurrent with increasingly frequent atmospheric circulation patterns. While the trends identified are statistically significant, they are not necessarily anthropogenic. Atmospheric patterns which are becoming rarer correspond instead to wet, cool summer conditions over northern Europe and wet winter conditions over continental Europe. The combined effect of these circulation changes is that of a strong, dynamically driven year-round warming over most of the continent and large regional and seasonal changes in precipitation and surface wind.

Extreme weather events exact a heavy and steadily increasing socioeconomic toll on Europe (1), eliciting both scientific and media interest in the atmospheric circulation patterns favoring the occurrence of heatwaves (2), cold spells (3), heavy precipitation (4), and windstorms (5, 6). Such circulation patterns may be understood as spatial patterns in a given atmospheric variable which repeatedly occur in conjunction with specific classes of extreme events. The question of whether, how, and why these circulation patterns have been changing under anthropogenic forcing, including the role of Arctic Amplification, has been the source of a lively discussion (7–25).

Of particular interest is whether atmospheric circulation patterns favoring specific extreme events have become more persistent. There are arguments supporting an increasingly persistent summer atmospheric circulation (10, 16, 18), which is highly relevant for the occurrence of heatwaves (19, 20). Nonetheless, (23) did not find a systematic slowdown of planetary waves over the Northern Hemisphere summers in the historical period, and (26) found no significant changes in their amplitude. There are also studies arguing against any projected changes in summertime atmospheric persistence, albeit on a regional scale (27). No consensus has therefore been reached on the topic. The picture for the winter season is equally debated, with some studies finding recent increases in the sinuosity of the midlatitude flow (11) and others not finding any trends in planetary wave phase speeds (23) or amplitude (26) or arguing for a decrease in high-amplitude waves and blocking—both regarded as persistent atmospheric patterns associated with surface extremes (8). A complementary line of work has analyzed the frequency of occurrence of specific atmospheric circulation patterns associated with extreme events (9, 13–15, 24). Heatwaves have again been a focus, on account of their intimate link with the large-scale atmospheric circulation and rapidly increasing frequency and duration (25). For example, (9) found that well over half of the total trend in hot extremes over Europe may be linked to the increased occurrence of anticyclonic circulation patterns over the eastern part of the continent. Similarly, (15) argues that circulation patterns like the ones associated with the 2003 European heatwave may become increasingly frequent. Along a similar line, (24) identified observed changes in hemispheric-scale, wavy atmospheric circulation patterns that increase the risk of concurrent heatwaves across eastern North America, eastern and northern Europe, and other downstream regions in Eurasia. Specific atmospheric patterns associated with other extremes, including drought (7) and heavy precipitation (21), have also been studied. The two perspectives of persistence and frequency of occurrence are intimately related since the recurrence of a specific pattern likely translates to persistent weather and vice-versa (20).

Previous efforts in the detection of atmospheric circulation shifts have often focused on the average behavior (11, 16, 23), for example, presenting hemispheric or seasonal-mean results, or on a specific set of extreme events or reference circulation patterns (9, 15, 19, 24, 28, 29), for example, the canonical North Atlantic weather regimes. Here, we consider in turn each daily atmospheric circulation pattern in 72 y of sea-level pressure reanalysis data (30), over the period 1950 to 2021 and the Euro-Atlantic sector. This results in 26,280 daily sea-level pressure latitude–longitude maps. The same analysis is repeated using 500 hPa geopotential height data (SI Appendix). Our approach is thus distinct from the above-cited studies since we look at individual atmospheric circulation patterns without needing to select a priori a set of reference patterns. We then select those days in the winter (December–February) and summer (June–August) seasons displaying significant trends in the occurrence of their analogs (hereafter referred to simply as “occurrence trends” Methods and SI Appendix Figs. S1 and S2). The vast majority (92.7%) of circulation patterns show no significant occurrence trend in the historical period; 5.1% show increasing trends and 2.2% show decreasing trends. Notwithstanding their rarity, the circulation patterns with significant occurrence trends have major implications for surface climate. To isolate the effect of circulation changes, we also consider daily reanalysis data for 10-m horizontal winds, 2-m temperatures, and precipitation rates, over the same time-period and region. All the data shown in the composites are detrended and deseasonalized (Methods).

During both the boreal summer and winter seasons, the atmospheric circulation patterns with positive occurrence trends show highly coherent surface climate anomalies over Europe. Specifically, patterns occurring more frequently in winter (Fig. 1A) show a north–south cyclonic–anticyclonic dipole. This is associated with anomalously windy, warm, and wet conditions in Northern and Eastern Europe and moderately warm, dry conditions in Southern Europe (Fig. 1C, E, and G). Wintertime patterns with a negative occurrence trend (Fig. 1B) have a cyclonic structure in the central North Atlantic and an anticyclone to the Northeast, leading to anomalously windy, cool, and dry conditions over Northern Europe and windy, wet, and moderately warm conditions over Central and Southern Europe (Fig. 1D, F, and H). We next analyze the average wintertime temperature and precipitation associated with patterns displaying either decreasing or increasing occurrence trends over Europe (Fig. 1 I and J; see mask in SI Appendix Fig. S3). We do not detect additional significant circulation-related trends, besides the thermodynamics ones that have been already removed from the datasets (Methods). For 2-m temperatures associated with patterns with increasing occurrence trends, the average value of 1.8 to 1.9 ^°C (black solid line, Fig. 1I) is coherent with the large areas displaying positive 2-m temperature anomalies in Fig. 1E).

Fig. 1.

Sea-level pressure wintertime atmospheric circulation patterns with significant occurrence trends and associated surface anomalies: Composite anomalies of DJF sea-level pressure (A and B), 10-m horizontal wind speed (C and D), 2-m temperatures (E and F), and precipitation rates (G and H) for days with increasing (A, C, E, and G) or decreasing (B, D, F, and H) occurrence trends. In the composites (A–H), contours indicate regions with changes significant at the one-sided 5% level, computed with a bootstrap sample size of 500. Spatial averages of seasonal temperature anomalies (black) and precipitation rates (blue) during the days with increasing (I) or decreasing (J) occurrence trends and count of days displaying the corresponding occurrence trend (orange stems) during DJF. Solid lines represent linear trends of the spatial averages with the 95% confidence intervals of the two linear fits in each panel shown in the legends. The averages in (I) and (J) are computed on all European land points (SI Appendix Fig. S3).

In summer, the atmospheric circulation patterns with positive occurrence trends are associated with an anticyclonic anomaly over the Labrador Sea and Greenland and a cyclonic anomaly to the East of the British Isles (Fig. 2A). Such circulation patterns drive anomalously calm and dry conditions over much of the continent and anomalously warm conditions over Western Europe (Fig. 2C, E, and G). Circulation patterns with decreasing occurrence trends show a ridge of high-pressure anomalies over the Atlantic and a large cyclonic structure centered over Northern Europe. These are associated with anomalously calm conditions over most of Europe and cool and wet conditions over Northern Europe (Fig. 2D , F, and H). Again, no significant trends appear in summertime European precipitation or temperature associated with any of the above patterns (Figs. 2 I and J).

Fig. 2.

Sea-level pressure summertime atmospheric circulation patterns with significant occurrence trends and associated surface anomalies: Composite anomalies of DJF sea-level pressure (A and B), 10-m horizontal wind speed (C and D), 2-m temperatures (E and F), and precipitation rates (G and H) for days with increasing (A, C, E, and G) or decreasing (B, D, F, and H) occurrence trends. In the composites (A − H), contours indicate regions with changes significant at the one-sided 5% level, computed with a bootstrap sample size of 500. Spatial averages of seasonal temperature anomalies (black) and precipitation rates (blue) during the days with increasing (I) or decreasing (J) occurrence trends and count of days displaying the corresponding occurrence trend (orange stems) during DJF. Solid lines represent linear trends of the spatial averages with the 95% confidence intervals of the two linear fits in each panel shown in the legends. The averages in (I) and (J) are computed on all European land points (SI Appendix, Fig. S3).

We underline that these results are obtained for detrended and deseasonalized datasets, meaning that the observed signals are chiefly related to the atmospheric circulation. Analyses conducted on NCEP/NCAR reanalysis data (SI Appendix Figs. S4 and S5) and E-OBS gridded observational data (SI Appendix Figs. S6 and S7) provide similar conclusions. SI Appendix Figs. S8–S11 further show that the atmospheric circulation patterns with positive or negative occurrence trends are generally similar at the beginning and end of the analysis period. In other words, the atmospheric circulation patterns identified as having significant occurrence trends stay roughly constant over time. We further find that several features of the identified spatial patterns are highly coherent across the individual days in the composite plots, and that trends are robust to changes in the details of our trend calculation (SI Appendix Figs. S12–S17).

We repeat the same analysis for the detrended 500 hPa geopotential height data (SI Appendix Figs. S18 and S19), to verify whether our results depend on the choice of observable for the atmospheric circulation. There is a strong resemblance of the spatial anomaly composites for atmospheric circulation patterns with increasing frequency of occurrence in sea-level pressure and 500 hPa geopotential height, as well as the associated impacts on temperature, winds, and precipitation (cf. Figs. 1A, C, E, G, 2A, C, E, G and SI Appendix Figs. S18 A, C, E, and G, and S19 A, C, E, and G). Patterns with decreasing frequency of occurrence show larger differences in both seasons (cf. Figs. 1B, D, F, and H, 2B, D, F, and H and SI Appendix Figs. S18 B, D, F, and H and S19 B, D, F, and H). This may be partly attributable to the smaller sample size for patterns with decreasing occurrence trends compared to patterns with increasing occurrence trends (approximately 43% of the sample size, computed jointly over winter and summer). It further suggests that our results for these sets of days should be interpreted with care. We conclude that the qualitative large-scale circulation patterns for days showing increasing occurrence trends and the associated surface anomalies are robust to the choice of the observable, while some notable differences emerge for days with decreasing occurrence trends. We estimate the magnitude of the trends for days showing significant positive or negative trends as number of analogs/decade for the different datasets and variables we consider here (SI Appendix, Table S1). Overall, we find that increasing (resp. decreasing) trends induce between 2 and 4 analog occurrences more (resp. less) every decade.

The surface climate anomalies associated with increasingly or decreasingly frequent circulation patterns can be directly related to the occurrence of high-impact summertime heatwaves and wintertime stormy weather in Europe. We draw high-impact heatwaves and the associated excess deaths from the EM-DAT disaster database (31) (Methods). We examine 228 heatwave days over the analysis period. Here, 9.7% of these days correspond to atmospheric circulation patterns increasing in occurrence (versus a climatological summertime occurrence of these patterns of 4.5% and a 97.5th percentile from random sampling of 7.0%), while none correspond to decreasing occurrence trends. As a term of comparison, only 3.1% of 480 cold spell days match atmospheric circulation patterns increasing in occurrence, in line with climatology. The 228 heatwave days occur during 10 major heatwave episodes, associated with 83,462 deaths. Four of these heatwaves include an above-average fraction of days with positive occurrence trends (Methods). The latter heatwaves are responsible for 91.4% of the total heatwave-related excess deaths (Fig. 3A). Excluding from the analysis the summer 2003 heatwave, which is associated with days with positive occurrence trends and alone accounts for the bulk of the total heat-related deaths in Europe in the period considered, we still find that heatwaves associated with circulation patterns with positive occurrence trends are responsible for 43.2% of total excess deaths. Although this heuristic argument is based on a limited sample of events, it nonetheless serves to illustrate the potential societal impact that circulation patterns with increasing occurrence trends may have, and motivates a future, more systematic impact-based analysis.

Fig. 3.

Impacts of changing atmospheric circulation patterns in terms of heatwave casualties and European Windstorms. (A) The size of the pie chart for each country shows the total number of heatwave excess deaths as recorded in the EM-DAT database for the whole year. The purple slices show the fraction of excess deaths associated with heatwaves showing an above-average frequency of circulation patterns with a positive occurrence trend. The yellow slices show the excess deaths associated with heatwaves that we excluded from our analysis (Methods). Finally, the blue slices show the corresponding fractions for heatwaves showing circulation patterns with no or negative occurrence trends. The shading on the geographical map shows the temperature anomalies (°C) during the 228 heatwave days retained for analysis. (B) The size of the pie chart for each country shows the total number of destructive windstorms in the storm database for the whole year (Methods). The purple slices show the fraction of windstorms showing an above-average frequency of circulation patterns with a positive occurrence trend. The blue slices show the corresponding fraction for windstorms showing analogs with no or negative occurrence trends. The shading on the geographical map shows the precipitation anomalies (mm day⁻¹) during the 438 windstorm days retained for analysis.

We conduct a similar analysis for 90 European windstorms which resulted in a high number of casualties and/or large insured losses, extending the storm database from ref. 32 (Methods and SI Appendix Fig. S20). Over a total of 438 windstorm days, we find that 66 (15%) are associated with circulation patterns with increasing occurrence trends and only 4 (0.9%) with patterns with decreasing trends. We find that 30 windstorms (33%) show an above-average fraction of days with positive occurrence trends while 4 events (4.4%) are linked to negative trends (Fig. 3B and SI Appendix Fig. S20b). These statistics aggregate values on a continental level. A geographically resolved picture highlights that the proportion of windstorms associated with patterns showing positive occurrence trends is larger in Continental/Northern Europe than in Southern Europe (Fig. 3B, e.g., 40% in United Kingdom versus 0% in Italy). This is consistent with the pattern shown in Fig. 1 C and G), which is reminiscent of the anomalies associated with destructive windstorms over Continental Europe found by ref. 5.

In this paper, we identified the atmospheric circulation patterns over the North Atlantic and Europe that have become less or more frequent in the historical period. We based our analysis on all daily patterns in the data, rather than on aggregated patterns, average behavior, or a specific set of extreme events or reference circulation patterns (9, 11, 15, 16, 19, 23, 24, 28, 29). In other words, we do not constraint our analysis to a fixed number of reference spatial patterns of some atmospheric field but rather view each atmospheric pattern as unique and characterized by rare recurrences (a few % of all data in our analysis). This approach does not require introducing any specific decomposition bases and the associated hyperparameters. We are thus able to isolate robust linear trends in the frequency of occurrence of specific, observed circulation patterns. Our results show minimal differences if a quadratic or cubic fit is applied instead of a linear one. The difference from previous studies is not only methodological but also interpretative. For example, we find increasing occurrence of zonal flow patterns and decreasing occurrence of antizonal patterns over the North Atlantic during winter, which do not emerge from a conventional weather regime-based perspective (33, 34).

Our method is flexible and could, for example, be used as a complement to extreme event attribution studies conditioned on the circulation (35, 36), or to the storyline approach (37). The latter approach has been criticized for not taking into account the role of climate change-induced circulation changes (38), and we provide a readily applicable toolkit to address this problem. Our approach is not data or location-specific and may be applied to different regions or datasets. This makes it well-suited as a tool to evaluate the consistency of dynamical trends in numerical simulations with those observed in reanalysis data. In the future, our approach could be complemented by the use of machine learning techniques for identifying atmospheric analogs, as opposed to using the Euclidean distance, e.g., ref. 39.

Our analysis does not enable to make a robust statement as to the potential role of natural variability versus anthropogenically driven climate change in modulating the trends in atmospheric circulation pattern occurrence over the recent decades. An analysis of climate modes of low-frequency variability highlights a possible modulation of the Altantic Multidecadal Oscillation, El Niño–Southern Oscillation and the North Atlantic Oscillation variability on trends in circulation pattern occurrence (SI Appendix and SI Appendix Figs. S21–S23), which should be taken into consideration when interpreting our results. Indeed, recent work has suggested that the variability of the North Atlantic Oscillation can modulate decadal temperature trends (33). In future work, one may also leverage initial condition large ensembles to separate forced signals from internal variability in the context of our analog analysis, e.g., ref. 40. A further limitation of our study is that it assumes that the circulation patterns of interest have good analogs in the dataset being used (Methods). This assumption is problematic in the presence of strong nonstationarities, which may lead to unprecedented atmospheric states. A theoretical advance would be to complement the analysis with a systematic investigation of analog quality, based on the distance between a day and its analogs. Preliminary analyses to this effect show no long-term trends in analog quality, yet highlight some interannual variability which may relate to the above-discussed modulation by climate modes of low-frequency variability. A second caveat is that the analysis of surface impacts through the EM-DAT and the windstorms database likely suffers from temporal and spatial inconsistencies in reporting—a common limitation of impacts data (41). In our specific case, the partial reliance of the windstorms database on catalogs compiled for the scientific literature (Methods), could alleviate the issue, but by no means completely resolves it.

We underscore that the temperature anomalies associated with increasingly frequent circulation patterns (Figs. 1E and 2E) are regionally comparable to or larger than the average global climate change best estimates reported by the IPCC (+1,07 °C, (42)) with winter (summer) temperatures anomalies up to +5.7 °C (+1.1 °C). These results are obtained after detrending and deseasonalizing the data and are consistent with those obtained for the 500 hPa geopotential height in the SI Appendix.

The circulation types we study, while rare (approximately 7% of all winter and summer days in our dataset), are increasing and decreasing at a pace of about 2 to 4 analogs per decade, depending on the variable and dataset examined (SI Appendix Table S1). If these trends are due to anthropogenic forcing, we expect the circulation patterns with positive occurrence trends to dominate the respective seasons in 2 to 4 decades. If, instead, they are due to internal variability, we expect to continue seeing their footprint for the next few decades. Regardless of their origin, we stress that these circulation patterns are associated with impactful extreme events, both deadly heatwaves and destructive windstorms. While these events are not exclusively caused by circulations changes, our results show that circulation changes cannot be neglected when evaluating the consequences of anthropogenic climate change or other sources of climatic variability on extreme weather events and their impacts.

Materials and Methods

Computing Trends in the Occurrence of Atmospheric Circulation Patterns.

For the computation of the trends in the occurrence of atmospheric circulation pattern analogs (which we hereafter refer to as “occurrence trends”), we use daily sea level pressure and 500 hPa geopotential height data from the ERA5 reanalysis (30) over the period 01/01/1950 to 31/12/2021 (26,280 d). This is the sample size for all analyses presented in the paper. The data have a horizontal resolution of 0.25 ° ×0.25°, and we restrict our analysis to 22.5N–70N and 80W–50E. This corresponds to the North Atlantic and Europe, with a size of 200 × 530 grid cells. Part of the analysis was repeated on additional datasets and subperiods of the full dataset (SI Appendix). For both the sea-level pressure and 500 hPa geopotential height data, we remove a grid point by grid point linear trend for the whole analysis period. This ensures that the selection of circulation analogs is not affected by long-term thermodynamic trends. We further deseasonalize the data using a mean seasonal cycle computed by averaging over the same calendar days. The robust trends in the occurrence of atmospheric circulation patterns are computed as follows:

• 1.We select daily sea-level pressure or 500 hPa geopotential height latitude–longitude maps, which we interpret as atmospheric circulation patterns.
• 2.We compute the Euclidean distance between daily maps, taking each daily map in turn as reference state and computing its distance from all other maps in the dataset. We then define a high quantile q to select the analogs. We chose q = 0.98, meaning that we take as analogs the 2% closest fields to the target. We describe below how the sensitivity to the choice of q is tested.
• 3.We divide the time interval of 73 y into 9 periods of roughly 8 y. We then count how many analogs N fall in each period t, obtaining N(t) with 0 < t ≤ 9. Shortening these periods has no qualitative impact on our results (SI Appendix Figs. S12 and S13).
• 4.We perform a linear fit of N(t) of the type at + b. Using a cubic fit does not qualitatively affect the results (SI Appendix Figs. S14 and S15).
• 5.We estimate the upper and lower 95% confidence intervals (CI) of the a parameter of the fit using the Wald method (43). If the lower and the upper bounds of the CI for a are positive (negative), we interpret this as a significant positive (negative) trend for the selected daily sea-level pressure or 500 hPa geopotential height map and quantile q. If the CI contains zero, the trend is nonsignificant.
• 6.We repeat the above steps for q = 0.99 and q = 0.995. We retain as daily maps with significant increasing (decreasing) occurrence trends only those having consistent (same sign AND significant) occurrence trends for all three quantiles. These are the robust circulation patterns that are analyzed in this paper. We additionally verify that the quality (i.e., distance) of analogs for these patterns is comparable to that for all other days in our dataset.

The 95% CIs that we compute indicate that there is a 5% chance that the real trend is out of the CI interval. Since the CIs are defined as symmetrical intervals, this implies that the chance that the “real” trend is zero or of opposite sign to that of our sample is ≤2.5%. In practice, since we use three different quantiles q simultaneously, this chance is very small, even taking into account that these are not independent samples. While analogs are a relatively common analysis approach in the atmospheric sciences, our methodology and its focus on trends is distinct from previous work. It is further fundamentally different from conventional decompositions of the atmospheric variability, such as self-organizing maps (44), k-means clustering (45), Empirical Orthogonal Functions, or others (46). Indeed, we do not constrain our analysis to a fixed number of circulation patterns. The occurrence trends are thus identified for daily patterns rather than for a fixed pattern associated with, e.g., a self-organizing map or weather regime. Indeed, the computation of the analogs does not require introducing any specific decomposition bases, which in turn introduce arbitrary hyperparameters such as the number of components into which the atmospheric variability is partitioned. Moreover, we do not “create” reference fields that may have never been observed in the data by orthogonal decomposition, averaging, or centroid computations. Our approach thus relies directly on the atmospheric variability present in the data, without adding an intermediate projection step.

Computation of Significant Circulation and Surface Anomalies.

We next produce composite anomalies for the days displaying robust increasing or decreasing trends within the winter (DJF) or summer (JJA) months (without including their analogs) for several daily variables: sea-level pressure, 500 hPa geopotential height, 2m temperature, horizontal 10-m wind speed, and precipitation rate. All these data are detrended and deseasonalized using the same procedure as for sea-level pressure and 500 hPa geopotential height. Significance for the geographical anomalies is estimated by using a bootstrap procedure (sample size m = 500) consisting in randomly drawing from the whole dataset a number of days equal to the number of days in each composite, regardless of their trend. Significant anomalies are those below the fifth or above the 95th percentiles of the bootstrap distribution at each grid point. We have also tested gridpoint by gridpoint sign agreement of the days within each composite, to ensure that the composite anomalies result from a spatially coherent set of daily anomalies (SI Appendix, Figs. S16 and S17). Long-term trends shown in the composites of Figs. 1–2 and SI Appendix Figs. S4–S19 are computed by averaging, for each season, the spatial averages over Europe (SI Appendix Fig. S3) for the patterns displaying occurrence trends. Years without patterns displaying occurrence trends are treated as NaNs. Significance for trends is estimated by using the Wald method and looking at the sign of both the upper and lower bounds of the CI. CIs of trends are displayed in the legends of the relevant figure panels.

Computing Impacts of Atmospheric Circulation Patterns with Positive Occurrence Trends.

We take heat waves and cold spells from the EM-DAT disaster database (31). We focus on events over Western-Central Europe (excluding, e.g., Western Russia, outlying islands such as the Canary Islands etc.), excluding those heat waves and cold spells where no start and/or end day was provided, or which lasted only for one day. Dates included in several events were only counted once. Following these criteria, from a total of 34 heatwave and 66 cold spell episodes included in EM-DAT, we identified 228 d over 10 heatwaves and 480 d over 16 cold spells. Of the original 34 heatwave events, 3 were excluded because they were entirely outside Western-Central Europe, 11 because they were missing start and/or end dates, 9 because a single day was assigned to the heatwave. Finally, 2 heatwaves were merged into 1 because, even though they had different “disaster numbers” (i.e., unique identifiers of each disaster in EM-DAT), they had overlapping dates and occurred in geographically contiguous regions. Of the 66 cold spells, a relatively large number was excluded due to being outside of Western-Central Europe (13 events were only registered in Russia). An additional 35 were excluded due to missing or single-day dates. Finally, four events were merged into 2 because even though they had different “disaster numbers”, they had overlapping dates and occurred in geographically contiguous regions.

European Windstorms data are taken from an updated version of the storm database of ref. 32, which originally covered the period 1948 to 2015. This database is largely based on the catalogs by ref. 47 (ending 1971) and (48) (ending 2012). Additional windstorms (2013 to 2020) have been integrated from the Wikipedia web page https://en.wikipedia.org/wiki/List_of_European_windstorms because of their relevance in terms of human losses, damages, or their profile in the media. We have only selected windstorms that have been analyzed by a meteorological office or research institute: a link to the documentation is provided together with each entry in the database. The database itself is available as SI Appendix. Using publicly contributed databases as sources for scientific information is becoming a common practice in citizen science projects, which are gaining momentum in the geosciences (49, 50). Our database includes a total of 438 windstorm days and 90 distinct windstorms or windstorm clusters. SI Appendix Fig. S20 presents the database entries per country a) and per time of occurrence b).

The database is. organized in four columns:

• 1.The day of occurrence in the format yyyymmdd;
• 2.The name(s) of the windstorm;
• 3.The country/countries or the region(s) affected;
• 4.A reference to a peer-reviewed article, a report or a press article describing the importance of the windstorm

As a caveat to our methodology, we note that the increasing coverage of both meteorological instruments and technological means of information results in an increasing number of windstorms with time.

To determine whether a given extreme event is associated with circulation patterns with positive or negative occurrence trends, we first compute the fraction of days with positive occurrence trends within each extreme event. We then compare these fractions to the average frequency of occurrence of circulation patterns with positive or negative occurrence trends within the 20 y centered on the extreme. This is because the frequency of occurrence of atmospheric circulation patterns with positive occurrence trends is by definition higher in the later part of the data than in the earlier years, such that each extreme should be compared to the period within which it occurs. Based on this comparison, we then separate extreme events into groups which have an above or below-average fraction of daily analogs with either trend.

Supplementary Material

Appendix 01 (PDF)

Dataset S01 (XLSX)

Data, Materials, and Software Availability

The code to compute the dynamical analysis is availabe at https://fr.mathworks.com/matlabcentral/fileexchange/95768-attractor-local-dimension-and-local-persistence-computation. ERA5, NCEP, ENSO, AMO and NAO data are publicly available at climate explorer (http://climexp.knmi.nl/). E-OBS data are available from the EU-FP6 project UERRA (https://www.uerra.eu) and the Copernicus Climate Change Service, and the data providers in the ECA&D project (https://www.ecad.eu). Other study data are included in the article and/or SI Appendix.