Do changing circulation types raise the frequency of summertime thunderstorms and large hail in Europe?

We study the role of changes in circulation type frequency on the evolution of summertime thunderstorm and large hail frequency across Europe since 1950 until 2020 to find out if they are responsible for the changes that an additive regression convective hazard models model (AR-CHaMo) predicts to have happened. To define circulation types, the 500 hPa geopotential height anomaly field on each day was clustered into 14 distinct patterns using principal component analysis and k-means clustering. We show that lightning and hail occurrence, both observed and modeled by AR-CHaMo, strongly depend on the circulation type, with a higher frequency observed in poleward flow downstream of a trough and on the lee side of mountains. AR-CHaMo predicts strong increases in hail frequency across central parts of Europe to have occurred in the 1950–2020 period. During this period, changes in circulation type frequency are small and only significant for 2 of the 14 clusters. The trends in both lightning and hail frequency to be expected if they were solely determined by circulation patterns, are small, with typical values of 1%–3% per decade relative to the mean, whereas the trends expected by AR-CHaMo are on the order of 4%–10% in most areas. Across many regions, the sign of the changes does not agree in sign, in particular across European Russia where circulation types have become more favorable for lightning and hail, but a decreasing probability was modeled by AR-CHaMo. We conclude that changing circulation types are, in general, not responsible for changes in thunderstorm and hail frequency, which included the strong increase of conditions favorable for large hail in central Europe.


Introduction
The frequency of lightning and severe thunderstorms across Europe is changing (Rädler et al 2019, Taszarek et al 2020, Battaglioli et al 2023).Although such changes cannot yet be inferred from lightning detection networks because they are still too short to detect multi-decadal trends (Taszarek et al 2019, Enno et al 2020), proxies from reanalyzes indicate that conditions supportive of thunderstorms have become more frequent across central and northern Europe since 1979 (Rädler et al 2019, Taszarek et al 2020, Battaglioli et al 2023).These studies agree that convective available potential energy (CAPE) to support thunderstorms is increasing, an observation corroborated by radiosonde measurements (Mohr andKunz 2013, Taszarek et al 2021a).This, in turn, is a consequence of rising moisture content in the lower troposphere caused by increased evapotranspiration as temperatures rise (Rädler et al 2018).Increases in lightning are counteracted in some regions by higher convective inhibition (CIN), i.e. the presence of a warm, dry layer of air that acts as a lid for initiating convective clouds, while the low humidity causes dilution of the buoyancy of the cloud, effectively preventing them from growing into thunderstorms (Westermayer et al 2017, Kahraman et al 2022, Battaglioli et al 2023).This effect dominates across parts of southern and southwestern Europe (Taszarek et al 2021b).
Large hail, the costliest weather hazard associated with thunderstorms, has also increased in frequency across central Europe (Púc ˇik et al 2019).In central Europe and specifically around the Alps, conditions favoring large (⩾2 cm) and very large (⩾5 cm) hail have locally become 100% to even 200% more frequent in the decade 2010-2019 compared to 1950-1959 (Battaglioli et al 2023).This change is driven by increases in CAPE on thunderstorm days, which ever more often reaches levels supportive of very large hail, while other prerequisites for (very) large hail, such as sufficient vertical wind shear, have changed very little.
Changes in CAPE and CIN are the result of changes in tropospheric moisture and static stability at different altitudes in the troposphere (Púc ˇik et al 2017), which can have both local and remote causes.Local causes include diabatic effects, such as changed heating or evaporation of moisture at the earth's surface, or changes in radiative flux to and from higher layers of the atmosphere or free space.Alternatively, they may result from remote causes, i.e. changes in transport of air-masses with a certain moisture content and stability from elsewhere.Climate change may manifest itself both by local changes and by changes in the large-scale flow.Coumou et al (2018), for example, detected changes occurring in association with Arctic amplification, i.e. the weakening pole-to-equator temperature gradient, on the mid-latitude circulation in summer, the season of highest (severe) thunderstorms frequency across Europe.They concluded that there is evidence that summertime circulation types are becoming more persistent.In this study, we investigate the role of changes in the large-scale flow on severe thunderstorms and clarify their relative importance.
Prior studies have found that some circulation types indeed favor the occurrence of thunderstorms while others suppress them (Wapler andJames 2015, Piper et al 2019).This is not surprising, since a circulation type associated with thunderstorms must simultaneously bring together the ingredients required for deep, moist, convection: low-level moisture, instability, and a lifting mechanism (Doswell et al 1996).As an example, thunderstorms in Western Europe are often characterized by a circulation type that features an upstream synoptic-scale trough that brings low-level moisture northward (van Delden 2001).This layer of air is initially capped by an elevated mixed layer with high instability until a front or a prefrontal convergence line produces the required lift to initiate storms (van Delden 1998, Dahl and Fischer 2016, Schemm et al 2016).In studies of severe weather events across Europe, the synoptic scale was found to be important in providing the required atmospheric moisture, e.g. by ensuring low-level moisture from subtropical regions by an atmospheric river from distant upwind sources (Dayan et al 2015, Davolio et al 2020, Ibebuchi 2022).Piper et al (2019) found a strong link between local-scale thunderstorm activity and large-scale flow as well as teleconnection patterns such as the North Atlantic Oscillation.Kapsch et al (2012), who used an objective weather type classification criterion for synoptic scale patterns, found that the frequency of circulation types which were associated with larger hail would increase, leading to a slight increase in hail occurrence in future climate projections in Germany for the period 2030-2045.In this study, we will use an objective classification algorithm and both observations and (severe) thunderstorm probability as modeled by the additive regression convective hazard models (AR-CHaMo) developed by Battaglioli et al (2023) to investigate trends across Europe as a function of changes in tropospheric circulation types for the period 1950-2020.

Lightning climatology
We used lightning data from the arrival time differencing network (ATDnet) operated by the Met Office (Enno et al 2020) from the 2008-2020 period to investigate the climatology of lightning across Europe.The study area is defined as the area between 34 • N and 71.25 • N latitude and 13 • E and 48 • W longitude and includes a number of subregions (figure 1).The detections of lightning strikes were binned onto a three-dimensional grid with 1 hourly temporal resolution and 0.25 • × 0.25 • degrees spatial resolution.Each bin with one or more lightning strikes is defined as a lightning hour.
The monthly pattern for a region (defined in figure 1) and each year (figure 2(a)), shows that lightning hours occur preferentially in the summer months, especially in June and July, while lightning is least frequent in the winter months (December, January and February; DJF), which is due to a lack of solar heating and atmospheric water vapor which limits instability (Brooks et al 2007, Taszarek et al 2021b).The differences between individual years are small when averaged across all of Europe.Between regions large differences exist (figure 2(b)): In January and February, lightning occurs most often in the Southeast, South, and Southwest, where convection is generated by heat and water vapor from the Mediterranean Sea.From April through June, lightning frequency dramatically increases across Europe in response to solar heating and the subsequent steepening of lapse rates.The maximum lightning activity is in June across the East and Southeast, and in July in the Centre, Northeast, West, and Northwest.In the Northwest, lightning activity remains the lowest of all regions.The Southwest and South feature maxima in September, when latent instability is maximized as the Mediterranean Sea that has reached relatively high temperatures after the summer and subtropical subsidence retreats equatorward (Taszarek et al 2021b).In the remainder of this study, we will consider the occurrence of lightning and hail only in the months May-August, which accounts  for the main thunderstorm season in Europe, but excludes the autumn Mediterranean thunderstorm season when the relation between circulation patterns and thunderstorms may be quite different than in summer.

Models for lightning and hail occurrence
To study the evolution of lightning and hail outside the 2008-2020 temporal window, we use the AR-CHaMo models developed by Battaglioli et al (2023) based on the fifth-generation atmospheric reanalysis (ERA5; Hersbach et al 2020) of the European Centre for Medium-range Weather Forecasts (ECMWF), lightning detection data, and severe weather reports from the European Severe Weather Database (Dotzek et al 2009).The models use an additive logistic regression which yields the probability of lightning or large hail occurrence as a function of certain reanalysis-derived predictor parameters.The model for large hail uses two regressions: one for the probability of a thunderstorm and one for the conditional probability of hail given a thunderstorm forms (Rädler et al 2019).The probability of large hail is given by the product of these probabilities.
The modeled probabilities are available on the same 0.25 • × 0.25 • degree and one hourly grid to which the lightning data were mapped, but for a much longer period than observations allow, backward to 1950.The predictors for the lightning and hail models were chosen following a procedure based on an ingredient-based approach.For lightning the latent instability, mid-tropospheric relative humidity, precipitation, mixing ratio of the convective source layer and a land-sea mask were used.For large hail, CAPE above the −10 • C level, effective deep-layer vertical wind shear, mixing ratio of the source layer and the height of the 0 • isotherm were identified to be jointly very skillful.For more details, we refer to Battaglioli et al (2023).
The observed and modeled frequency of lightning in the period 2008-2020 resemble each other to a large extent.In figure 3(a) Scandinavia, Finland, the British Isles, and northwestern Russia exhibited low values of up to 15 h with lightning per year, while southern European mountain ranges such as the Alps, the Pyrenees, the Apennines, the Carpathians, and the Caucasus had much higher numbers of lightning hours.The highest  Compared to climatology based on hail reports, the modeled distribution represents a more objective map that overcomes the inhomogeneities in large hail reporting caused by non-meteorological factors such as those described by Allen and Tippet (2015).
The most prominent differences between the spatial distribution of large hail and lightning occurrence are the local minima across the central parts of mountain ranges, such as the Alps, Pyrenees, and Caucasus, where thunderstorms are very likely, but not storms with large hail.The cause is that high CAPE required for hail rarely occurs there because of a lack of low-level moisture (Taszarek et al 2020).Instead, maxima of hailstorm occurrence are found in the forelands of these mountain ranges, such as the Po Valley, and the lowlands north of the Caucasus.In addition, a relatively high probability of large hail is found across the Atlas Mountains and eastern Spain, where a lack of CAPE is usually not the limiting factor, but rather the warm and dry air that limits convective initiation as a result of subsidence in the subtropical downward branch of the Hadley circulation (Taszarek et al 2021b): once a storm forms, CAPE is comparatively often sufficient to sustain storms powerful enough to produce large hail.

Circulation type classification
We performed a cluster analysis of all 8856 summer days (May-August) using the geopotential height at 500 hPa extracted from the ERA5 archive at 12 UTC, for a 72 year period , over Europe.In our data analysis, we employed the scikit-learn library (Pedregosa et al 2011) for data processing tasks, facilitating the implementation of various clustering and classification algorithms.We first calculated daily geopotential anomalies by subtracting the time-averaged daily mean.These deviations from the geopotential height are scaled by the cosine of latitude to account for the latitude-dependent surface area represented by each grid point and finally the data is rescaled to fall between 0 and 1 (Cassou 2008).We then performed a principal component analysis and identified the first 12 modes that explained 95% of the total variance.The associated principal component time series were then subjected to a cluster analysis to search for commonly occurring weather patterns.The clustering analysis was performed using the k-means procedure of MacQueen (1967).Weather patterns of single days were assigned to the cluster to which the Euclidean distance was lowest (Zhang and Villarini 2019, Zhang et al 2022, Saenz et al 2023).To accurately determine the final number of regimes, we employed a two-step procedure.First, we used the elbow method, which plots the within-cluster sum of squares (WCSSs) against the number of clusters (Bholowalia and Kumar 2014).The WCSS quantifies the variability within clusters; as the number of clusters increases, the WCSS tends to decrease.The 'elbow point' in the graph signifies a point at which yield decreases and further cluster partitioning results in a small decrease in WCSS.This point represents an optimal balance between the number of clusters and their homogeneity.To complement this, we also used the silhouette score, a metric that measures the quality of clustering (Rousseeuw 1987).It evaluates how well each data point lies within its assigned cluster compared to other clusters.Combining insights from the elbow method and the silhouette score, we rigorously selected 14 as the final number of regimes, as this marked the point of balanced clustering and high-quality assignments for our weather patterns.We chose k-means as our clustering method because it is widely used for similar approaches and is relatively simple.However, it has a number of limitations that we have tried to mitigate.We addressed the limitation of k-means that outliers are not handled well by performing a principal component analysis.The limitation that k-means are sensitive to the initial position of the clusters was mitigated by choosing a high number of 500 iterations.Furthermore, robustness to the choice of a specific number of clusters was demonstrated by a sensitivity analyses showing that a wide range of cluster numbers leads to similar results (see appendix).An anonymous reviewer pointed out that we could have chosen other methods, such as fuzzy c-means, which, in their opinion, would have been the optimal choice.To acknowledge their reservations, we added this discussion of limitations of k-means to the article and hereby pass on their general message of 'caveat emptor' to the reader.
Each of the 14 identified clusters of synoptic-scale flow over Europe in summer (figure 5) occurred between 5.0% (Cluster 5) and 9.5% (Cluster 13) of the time.The locations of troughs and ridges in the geopotential pattern differ strongly among clusters, as does the temperature at 850 hPa.The distribution of CAPE (for the most unstable parcel) indicates that, in the summer, regardless of the synoptic pattern, there is CAPE of at least 100 J kg −1 over the Mediterranean region.CAPE tends to be higher when the 850 hPa temperature is above 15 • C because the higher temperature is associated with higher rates of evapotranspiration and, subsequently, more low-level moisture in the lower troposphere.Second, the 850 hPa level is often near the bottom of an elevated-mixed layer, a layer with a steep mid-tropospheric lapse rate that originates over elevated and arid areas such as the Iberian Peninsula, North Africa, or Turkey.Both low-level moisture (figure 6) and steep mid-tropospheric lapse rates are ingredients to CAPE.Across Southeast Europe, across the Mediterranean, the presence of sizable CAPE often occurs independently of the synoptic scale pattern.A much stronger dependence between CAPE and the synoptic-scale pattern is found over the western half of Europe, especially France, the Benelux countries, and western Germany, which see sizable CAPE only in case of a southwesterly flow in the mid-troposphere, for example in Clusters 4, 5, 9, and 11, where isohypses across France are southwest-to-northeast oriented (figure 5).Southwesterly flow advects steep lapse rates northward from the Iberian Peninsula, a pattern known among forecasters in western Europe as the 'Spanish Plume' (van Delden 2001).Over Central Europe (Germany, the Alpine region, the Czech Republic, Slovakia, and Poland), non-zero CAPE occurs either with a warm air mass and southwesterly flow aloft for the same reason as over Western Europe, or within a deep trough.CAPE occurs within a deep trough (cluster 3 over Germany and western Poland or cluster 14 over the Baltic states) because of cold air aloft in the core of the trough, which implies a strong vertical temperature gradient.

Flow-dependent lightning distributions
Like CAPE, the observed distribution of lightning (figure 7) and its anomaly (figure 8) are strongly influenced by synoptic patterns.Downstream of the geopotential trough, where differential advection of mid-level vorticity implies upward vertical motion and poleward advection of lower tropospheric moisture and steep lapse rates, the development of thunderstorms is favored.This is particularly evident in cluster 13, where a trough axis over central Europe corresponds to high lightning activity over the Alpine region and eastern Europe, and in cluster 6, where lightning frequency is maximized over the Balkans.In contrast,  clusters with a strong zonal tropospheric flow as cluster 3 (3.0% of all lightning) and cluster 8 (6.8%) are associated with the lowest frequency of lightning hours.A number of clusters are associated with an upper-level ridge (e.g.clusters 1, 4, 9, 11) and widespread convective initiation over the continent.In these constellations, convective initiation is caused by local processes rather than synoptic features, for example, by mountain-range induced mesoscale circulations, which is supported by the observation that most storms in those patterns occur across mountain ranges, such as the Alps, Pyrenees, Carpathians, and the Caucasus.Indeed, although different tropospheric circulation types can strongly modulate the likelihood of lightning in some areas, it occurs somewhat independently of the tropospheric flow in others.This is especially true for the southern Alps, which exhibit a high number of lightning hours in all clusters.Indeed, the presence of nearby sources of low-level moisture, i.e. the Mediterranean Sea and adjacent Po river Plain, and the diurnal mesoscale circulation creating upslope flow lead to the frequent initiation of thunderstorms, regardless of the circulation type.

Flow-dependent hail distributions
Using modeled hail probabilities by AR-CHaMo, we next investigate the dependence of the modeled occurrence of large hail on the circulation type (figures 9 and 10).A strong dependence similar to that for observed lightning exists across most regions: positive anomalies are found downstream (east) of troughs and negative anomalies upstream (west) of troughs, near trough axes and in zonal flow across northern Europe.This means that above-average hail activity can be expected with a southerly wind component at 500 hPa (inferred from south-to-north oriented isohypses with lower heights to the west), while below-average hail activity occurs with a northerly wind component (figure 10).An important contributor to hail probability in southerly flow is arguably the advection of steep mid-tropospheric lapse rates from arid and high-altitude  areas, such as Iberia, North Africa, and Turkey, which are required for the large CAPE that is needed for the formation of large hail (Púc ˇik et al 2015, Taszarek et al 2020).The only exceptions are northern Italy and Spain, where above-average hail activity occurs with westerly to northwesterly prevailing flow.Indeed, near the Alps, the relative location of most modeled hailfall is particularly dependent on the flow: in northern Italy, the highest number of hours of large hail is simulated in northwesterly flow (clusters 13 and 14).In this situation, warm and humid air masses can be trapped south of the Alps while colder air wraps around the mountain range.Furthermore, vertical wind shear on the southern flank of the Alps is typically enhanced because of lee cyclogenesis that leads to easterly flow across the Po Plain opposing the mid-level flow.North of the Alps, synoptic-scale patterns with a southwesterly flow show the highest number of hours with large hail (clusters 5 and 9).The situation is similar in southern France, where large hail is most frequent when the prevailing flow is southwesterly (clusters 9, 11 and 12).Over the Balkans, large hail is uncommon with northerly flow in the mid-troposphere while it is most frequent ahead of a large-scale trough (cluster 13).

Lightning frequency trends
Circulation type frequency is highly variable from one year to the next (figure 11), with frequencies ranging from absent (0%) to over 20% in some years for most clusters.Statistically significant trends (p < 0.05) are found only for two clusters: cluster 2 and cluster 11.Cluster 2 has become less frequent since 1950.This cluster is characterized by a ridge over the Atlantic and northwesterly flow over northern Europe (figures 5 and 6), with comparatively small anomalies for lightning and hail occurrence, except for a pronounced negative anomaly over Northeast Europe (figures 7-10).Cluster 11 has become more frequent and is characterized by a ridge over central and northern Europe and a relatively high frequency of lightning across west-central and northwestern Europe.Trends in other clusters are not significant at the 5% level.The closest significant is the downward trend of cluster 8 (p = 0.08) with a mostly zonal flow and a weak trough west of Iberia and France and a negative lightning and hail frequency over northern Europe, and the increase of cluster 6 (p = 0.10) with low geopotential over the Alps and Italy and an attendant low hail and lightning frequency across West Europe and high frequency across eastern Europe.
We next quantified the total effect of circulation type changes on the frequency of lightning and large hail in the period 1950-2020 and compare them with the changes modeled by AR-CHaMo over the same period.First, we calculated the average lightning frequency distribution for each cluster.Then changes in lightning occurrence exclusively due to circulation type changes were simulated by assuming that lightning occurrence would, at all times, match that of the average of the respective cluster.These averages were calculated in two ways: (1) using modeled lightning probability between 1950 and 2020 using AR-CHaMo and (2) using ATDnet lightning observations from 2008 to 2020.
Mathematically, we do the following.Let N y be the number of events, i.e. lightning or large hail, in year y.Then AR-CHaMo gives predictions of the number of events N ARCHaMo (y) for y = 1950, 1951, …, 2020.Let p c be the probability (0 ⩽ p c ⩽ 1) of an event occurring on a day with circulation pattern c.These probabilities p c can be estimated by the observed frequency of events on days with circulation pattern c.For lightning we can use the observed event frequency during the period 2008-2020, i.e. p c,observed_lightning , since such observations are indeed available.Alternatively, we can take the modeled frequency of events by AR-CHaMo for the period 1950-2020 for both lightning, p c,modeled_lightning , and hail, p c,modeled_hail .If in a given year, a circulation pattern c occurs on f c days, then the expected number of events in year y, based on the relative frequencies of circulation patterns in that year N expected (y) = ∑ c=n clusters c=1 p c • f c (y).We next compare the trends of N ARCHaMo (y) on the one hand, with those of N expected (y) using either observations or AR-CHaMo, on the other hand.The sign, magnitude, and significance of the simulated trends N expected,modeled_lightning (figure 12(a)) and N expected , observed_lightning (figure 12(b)) agree fairly well with each other over most of Europe.The use of lightning observations yields a higher coverage of statistically significant trends in lightning.The use of the AR-CHaMo-based lightning probability results in more statistically significant trends across the northern Atlantic Ocean.The effect of the changed occurrence of circulation types should be to slightly increase the lightning frequency across Europe except the Southwest (figure 12), typically by 0%-4% per decade.This increase is statistically significant across parts of the North Atlantic, and across northern, and eastern Europe.Note that the magnitude of the relative changes in figure 12 is high and noisy where thunderstorms are very rare, for example across the Northeast Atlantic Ocean.These large relative changes are in fact caused by very small absolute changes of lightning frequency.
Changes in the occurrence of synoptic-scale patterns cannot alone explain the changes in the lightning frequency found by AR-CHaMo over Europe.The trend in lightning frequency directly from AR-CHaMo has different sign and magnitude in many regions (figure 13) than the expected trends due to circulation pattern changes in figure 12.In most areas, AR-CHaMo simulates much stronger changes than can be expected on the basis of circulation type changes.The strongest contrast is observed over Russia, where synoptic-scale circulation types favorable for thunderstorms have become more common, and on that basis, an upward  trend of 1%-3% per decade could be expected, while AR-CHaMo shows a statistically significant decrease in the lightning frequency of −2% to −6% per decade.In the Alpine region, synoptic-scale patterns indicate a small and insignificant increase of about 0%-1.5% per decade, while AR-CHaMo shows a statistically significant increase in lightning hours of up to 5%-12% per decade, which, over the entire 71 year period equals an increase of 35%-85%.Over northwestern Europe, AR-CHaMo shows an increase which has the same sign but is much larger than that which can be attributed to changing cluster frequency: AR-CHaMo shows increases of more than 10%, but the cluster frequency can be held accountable for only about 3% of the increase.
To ensure that these results are not sensitive to the number of 14 clusters that we have chosen, this analysis was repeated for different numbers, i.e. 3, 5, 10, and 20 (figure A1).The modeled trends for the 5, 10, and 20 are qualitatively similar to those obtained for 14 clusters, both using observations and AR-CHaMo for the per-cluster lightning frequency.For the 3 member analysis, the trends deviate strongly from those found for other cluster numbers, signaling that using 3 clusters is insufficient.

Large hail frequency trends
The occurrence of large hail simulated by AR-ChaMo N ARCHaMo, hail (figure 14(a)) shows trends that, in the same way as for lightning, cannot be explained by trends in circulation type occurrence across Europe N expected,modeled_hail (figure 14(b)).Based on the trend in the circulation types, only very small changes of mostly less than 2% would be expected over Europe.This is in strong contrast to the trend derived from the AR-CHaMo hail model, which shows a statistically significant increase of large hail days by more than 10% per decade over the Alps, northern Italy, northwestern France, the Benelux countries, northwestern Germany, and southwestern Norway.This implies that any changed occurrence of the circulation types has not contributed much to this increase.As for lightning, the results for 5, 10 or 20 clusters are qualitatively similar (figure A2).

Summary and conclusion
We found the conditions associated with thunderstorms in European summer to have become more frequent in the period 1950-2020, with the exception of a large portion of Russia and the Iberian Peninsula, but these increases cannot be attributed to the fact that circulation types with a high (low) lightning probability occurred more (less) often.Although two of the 14 circulation types that were distinguished by k-means clustering exhibited significant trends, others did not and their combined effect on lightning probability leads to only small decadal trends in the lightning probability that do not generally correspond to the changes derived from the AR-CHaMo.Only across Northwest Europe the frequency of the flow-pattern changes and the total changes in lightning agree on a significant increase.For that region, it can be said that circulation type changes have indeed contributed to the increase.This cannot be said for other regions, and across large parts of European Russia, where a strong inhibiting effect on lightning occurred from pre-convective environments that counteracted the increases to be expected from flow-pattern changes to result in a net decreasing trend.
For large hail changes to be expected from changing circulation types are much smaller than the total modeled change from AR-CHaMo.The AR-CHaMo derived changes are maximized across western and central Europe and are generally 4%-10% per decade, relative to the mean of the period.Across European Russia, areas with a decrease were modeled.Across the entire study area there is little agreement in sign and magnitude between the changes to be expected from changes in the frequency of circulation types (mostly less than 2% per decade) compared to those which are attributable to other factors.
The answer to the question posed in the title of this study is therefore a no: the thunderstorm and large hail frequency in Europe is mostly determined by other factors than changes in how often the circulation types occur over a year.The exception is Northwest Europe, where lightning has become significantly more frequent as a result of the increased occurrence an omega-blocking pattern centered across the Baltic Sea and an upstream southerly flow (cluster 11) that is associated with lightning.
It is well possible that changes in circulation types will become a more important driver of severe weather potential in the future.For this to occur, changes in the properties of the mid-latitude circulation must become larger than have been the case until the year 2020.Any projections for the future similar to that of Kapsch et al (2012) should take into account that the magnitude of increases in CAPE and the low-level humidity that causes will likely dominate the effects of circulation type changes, at least in the near future, since they have dominated in the recent past as we have shown.
An aspect that we have not studied is how the frequency of lightning and hail changes within a given circulation type cluster.After all, the increases in thunderstorm and hail probability may have not occurred with the same magnitude in each cluster.Furthermore, we have not studied the aspect of changes in circulation type persistence that has been highlighted by Coumou et al (2018) and others.These are as of now questions for future investigation.

Figure 2 .
Figure 2. Seasonal cycle of the total number of lighting hours by month and year in Europe: (a) the seasonality from 2008 to 2020 for the whole domain is shown in gray and the mean seasonality of 13 years is shown in black.In (b), the seasonality of the individual subregions over 13 years is shown in colors.

Figure 5 .
Figure 5.The 14 clusters with geopotential height at 500 hPa plotted each 50 geopotential meters intervals (lines) and with, (a) the temperature of ERA5 shaded, and (b) the most unstable CAPE shaded.The headings list the relative frequency of occurrence of each cluster.

Figure 6 .
Figure 6.The distribution of 0-1 km average water vapor mixing ratio and 100 m AGL wind vectors for each of the 14 clusters.

Figure 7 .
Figure 7. lightning hours (2008-2020) per cluster (shaded) per year and geopotential heights at 500 hPa (with contours every 100 geopotential meters) for each of the 14 clusters.

Figure 9 .
Figure 9.As figure 6, but for hours with large (⩾2 cm) hail as modeled by AR-CHaMo for the period 1950-2020.

Figure 10 .
Figure 10.As figure 8, but for the relative anomaly (climatology = 1) of hours with large (⩾2 cm) hail as modeled by AR-CHaMo for the period 1950-2020.

Figure 11 .
Figure 11.Frequency of occurrence of circulation types per cluster (%) and linear regression lines (1950-2020).The coefficient of determination R2, p-value, and slope are given for each cluster.

Figure 12 .
Figure 12.Trend (1950-2020) of change per decade (relative to the average) of lightning days in the months May-August, expected due to circulation type changes, (a) based on per cluster lightning frequency modeled by AR-CHaMo (1950-2022) N expected,modeled_lightning , (b) based on per cluster lightning frequency observed by ATDnet (2008-2020) N expected,observed_lightning .Hatching indicates areas where trends are statistically significant.

Figure 13 .
Figure 13.Trend (1950-2020) of lightning days (days with at least one lightning hour), expressed as changes relative to the average lightning frequency in the period, in the months May-August, as modeled by AR-ChaMo N ARCHaMo,lightning .Hatching indicates areas where trends are statistically significant (p < 0.05).

Figure 14 .
Figure 14.Trend (1950-2020) of the annual number of hail days in the months May-August, expressed as changes relative to the average hail frequency in the period, (a) as modeled by AR-CHaMo N ARCHaMo,hail and (b) expected due to circulation type changes N expected,modeled_hail .Hatching indicates areas where trends are statistically significant.In panel (a), values exceeding 15 and −15 which occur in regions where hail is extremely rare exceed the colormap and are shown as white.

Figure A1 .
Figure A1.Trend (1950Trend ( -2020) )  of the annual number of lightning days in the months May-August, expected due to circulation type changes, expressed as changes relative to the average lightning frequency in the entire period in % per decade.Hatching indicates statistical significance of the trends.

Figure A2 .
Figure A2.Trend (1950Trend ( -2020) )  of the annual number of hail days in the months May-August, expected due to circulation type changes, expressed as changes relative to the average hail frequency in the entire period in % per decade.Hatching indicates statistical significance of the trends.