Thunderstorm Environments in Europe

. Meteorological environments favorable for thunderstorms are studied across Europe, including rare thunderstorm conditions from seasons with climatologically few thunderstorms. Using cluster analysis on ERA5 reanalysis data and EUCLID lightning data, two major thunderstorm environments are found: Wind-ﬁeld thunderstorms, characterized by increased wind speeds, high shear, strong large-scale vertical velocities, and low CAPE, occurring mainly in winter; and mass-ﬁeld thunderstorms, characterized by increased mass-ﬁeld variables such as large CAPE values, high dew point temperatures, and elevated 5 isotherm heights, occurring mostly in summer. Several sub-environments of these two major thunderstorm environments exist. Principal component analysis is used to identify four topographically distinct regions in Europe that share similar thunderstorm characteristics: The mediterranean, alpine-central, continental, and coastal regions, respectively. Based on these results it is possible to differentiate lightning conditions in different seasons from coarse reanalysis data without a static threshold or a seasonal criterion.


Introduction
Lightning, the defining characteristic of thunderstorms, can originate in a variety of meteorological settings. Some conditions that lead to lightning occur more frequently and are better understood than others. Numerous lightning climatologies are available but many focus on the dominant characteristics and seasons while infrequent thunderstorm conditions are often neglected.
Thunderstorms during the cold season are generally rare but pose a serious threat to wind turbines and other tall structures 15 because it has been observed that lightning strikes to tall infrastructure have no or only a weak annual cycle whereas lightning in general has a pronounced annual cycle Matsui et al., 2020;Vogel et al., 2016). This study describes thunderstorm environments occurring in Europe using a balanced view on all four seasons to include also seasonally infrequent thunderstorm conditions. A comparison between different regions provides a comprehensive overview of the lightning characteristics in Europe. 20 The general lightning pattern in Europe is well described in various climatologies (e.g., Taszarek et al., 2019;Enno et al., 2020;Poelman et al., 2016;Wapler, 2013;Taszarek et al., 2020a, b;Mäkelä et al., 2014;Vogel et al., 2016;Ukkonen and Mäkelä, 2019;Simon et al., 2017;Kotroni and Lagouvardos, 2016;Piper and Kunz, 2017;Anderson and Klugmann, 2014;Hayward et al., 2022;Holt et al., 2001;Enno et al., 2013;Poelman, 2014;Taszarek et al., 2015;Schulz et al., 2005;Coquillat et al., 2022;Manzato et al., 2022;Simon and Mayr, 2022). There is a north-south gradient of lightning frequency with a 25 maximum in northern Italy and the Mediterranean. Lightning in central Europe follows a clear annual cycle with a maximum over land in summer (MJJA) and a secondary peak in fall and early winter (SONDJ) in the Mediterranean (Taszarek et al., 2019;Poelman et al., 2016;Enno et al., 2020). In south-central Europe, the annual cycle is less pronounced and has sometimes two lightning maxima along with a local minimum in summer (Taszarek et al., 2019). There are differences in the annual lightning cycle: Offshore and coastal areas have a lower amplitude and a later maximum compared to inland or mountainous 30 locations (Wapler, 2013;Enno et al., 2013). In the northern Atlantic region, occasional intense thunderstorms are possible, even though the climatological thunderstorm activity is low (Enno et al., 2020). There, lightning in the cold season (Oct.-Apr.) occurs predominantly over the seas (North Sea, Baltic Sea, Atlantic) and less so over the land (Mäkelä et al., 2014). The question remains whether there are meteorologically different thunderstorm conditions at work, resulting in these spatial and temporal differences in lightning characteristics across Europe. 35 Many processes influencing lightning occurrence are known: The diurnal lightning cycle in Europe peaks in the afternoon over land and at night over the sea (Taszarek et al., 2020a;Enno et al., 2020;Manzato et al., 2022). Nighttime offshore lightning (Bay of Biscay, the North Sea, and the Baltic Sea) is explained by convection initiated over land and advected out to sea, where lightning activity endures longer as the sea surface temperatures are unaffected by nighttime cooling (Enno et al., 2020). The most pronounced diurnal cycle is found over mountainous areas and commonly explained by the topography. Complex terrain 40 favors more unstable environments (less CIN when the surface is close to the level of free convection), mechanical forcing (forced lifting), and thermal forcing (elevated heating leads to positive buoyancy and up-mountain flow; Manzato et al., 2022). This is particularly relevant after the snow has melted at higher elevations . Most of continental Europe experiences 20 − 40 thunderstorm days annually, but the mountain ranges in southern Europe have thunderstorm frequencies of > 60 thunderstorm days per year (e.g., northern Italy). 45 The sea has an effect on lightning as the number of lightning strokes and the sea-surface temperature are positively correlated in fall (Kotroni and Lagouvardos, 2016). Mallick et al. (2022) even suggests the use of sea surface temperature as a proxy for seasonal lightning forecasts. Warm oceanic currents are known to increase lightning densities in each season and particularly so in winter (Iwasaki, 2014;Holle et al., 2016). Wintertime lightning occurs usually in mid-latitudinal cyclones (Bentley et al., 2019) and lightning bands are found in wintertime storm tracks (Zhang et al., 2018;Virts et al., 2013). In general, the European 50 lightning patterns are well described (e.g., Wapler and James, 2015;Enno et al., 2014), but the meteorological drivers leading to lightning in the winter compared to summer are less understood. High structures such as wind turbines or radio towers increase the occurrence of lightning , especially in the cold season (Vogel et al., 2016;Pineda et al., 2018) so that lightning damage to infrastructure is evenly distributed over the year even though lightning occurrence in the surroundings has a strong annual cycle . 55 The 2018 update of the lightning protection standard for wind turbines introduces different lightning threats in winter and in summer (Méndez et al., 2018;IEC 61400-24, 2019). Using the maps from March et al. (2016), the environmental factor in the standard now includes the local threat of winter lightning. While considering winter lightning is a good first step, the quality of the risk assessment could be improved because the maps are very coarse, underestimate winter lightning, and use a static threshold (< 5 • C at 900 hPa). The reasons for the insufficient consideration of lightning in winter in the standard are 60 the different processes leading to upward lightning and the limited meteorological knowledge of lightning in the cold season (Becerra et al., 2018).
One approach to investigate these differences would be to numerically simulate individual thunderstorms, which may require horizontal resolutions of O(100 m) (Bryan et al., 2003) and still fail to make thunderstorms appear at the correct times and places. Our approach takes advantage of already knowing where and when lightning occurs from measurements. Thunderstorm cles. Almost one hundred such variables were computed and exploratively analyzed. By eliminating highly correlated variables that provide limited additional information, a set of 25 variables remains ( Table 2). The set is indicative of substantial clouds (e.g., moisture, large-scale vertical velocity, precipitation, cloud size), charge transfer (e.g., ice, snow, supercooled liquids), and charge separation (e.g., shear, wind, CAPE, CIN). As some required variables are not directly available at ERA5 95 (https://www.doi.org/10.24381/cds.adbb2d47, accessed 2023-02-15), they are derived using also the model level data. These additionally derived variables include variables such as the height of the −10 • C isotherm, cloud mass between −10 and −40 • C (ice and snow), and the product of maximum large-scale vertical velocity and liquid particles between −8 and −12 • C (vertical liquids flux). Further derived variables are cloud size, cloud shear, wind speed at cloud base, maximum upward vertical velocity, and the temperature difference between the air mass at 1000 m a.g.l. and the surface (sea surface temperature or skin 100 temperature). All 25 variables are listed in Table 1 and details about them are provided in the online supplement (Morgenstern et al., 2023).
To ease interpretation, physical-based categories group the variables: Mass-field variables refer to temperature, pressure, and humidity. Surface-exchange variables include atmospheric fluxes interacting with the surface. Wind-field variables cover everything related to wind. Cloud-physics variables refer to measures directly related to clouds. Topographic variables consist 105 of the surface geopotential height (orography) and a binary land-sea mask.

EUCLID lightning data and geographical domains
Lightning data are provided by the European Cooperation for Lightning Detection (EUCLID Schulz et al., 2016;Poelman et al., 2016), a cooperation of several local lightning location systems (LLS) in Europe. Only cloud-to-ground lightning flashes between 2010-2020 are considered, as this period is most stable regarding the hardware and software configuration of the 110 network. If at least one lightning flash occurred within an ERA5 cell in a given hour, the whole cell-hour is regarded as one lightning observation.
The EUCLID territory is separated into 12 domains with rather homogeneous topography and lightning detection efficiency

Methods
To investigate spatio-temporal lightning characteristics, lightning data sets for the 12 domains are constructed that have the same number of observations from each season. The lightning data sets are then combined with 25 ERA5 variables representing the atmospheric conditions at the hour of the lightning observations. Using the domain means, a spatial lightning analysis for Europe is performed with the help of a principal component analysis. Then, thunderstorm environments are found indi-130 vidually on each domain by a cluster analysis with k = 3 clusters. A seasonal lightning analysis follows by analyzing how many observations from each season have been classified into which thunderstorm environment. Finally, the thunderstorm environments are compared to one another using again a principal component analysis.

Composition of data
EUCLID lightning data is aggregated to the spatio-temporal resolution of ERA5 resulting in binary cell-hours indicative of 135 lightning. For each lightning cell-hour, ERA5 data at the respective cell and from the last full hour is taken to capture the build-up of the thunderstorms. Accumulated variables, such as precipitation, are taken from the next full hour to capture everything within the hour in which lightning was observed. Only cell-hours with lightning are considered. To investigate seasonal differences, the available data are reduced to contain the same number of lightning cell-hours from each season (winter = DJF, spring = MAM, summer = JJA, fall = SON). Therefore a random sample without replacement is drawn from 140 the seasons with more lightning cell-hours. Depending on the domain size and general lightning frequency, the data set in each domain consist of 5 320 − 40 000 observations (Table 1). For robustness, the whole analysis is performed on 50 different samples in each domain. A visual comparison of the resulting figures reveals qualitatively the same results between these repetitions. Hence, the samples are representative and it is sufficient to discuss only one sample in the following.
k-means clustering requires scaled input variables that follow rather similar distributions. Therefore all ERA5 variables are 145 square root transformed and scaled to a mean of zero and a standard deviation of one.
with x being the original ERA5 value and x t its transformation.
µ and σ are the empirical mean and standard deviation and x s is the scaled value. The applied algorithm is supplied in the

Statistical methods
Principal component analysis (Mardia et al., 1995) is an approach for dimension reduction that computes several linear com-155 binations of projected input data (principal components, PC) aiming to capture as much variability from the data as possible.
The first PC explains the most variance and each following PC is oriented perpendicular to the previous PC explaining less and less variance. Omitting the later PC's results in the intended dimension reduction. In this study, the first two PC's are used as axes for a so-called biplot to visualize the variance in the 25-dimensional data.
k-means cluster analysis (MacQueen, 1967) is a data-driven approach to find groups in data, aiming at maximum similarity 160 within and minimum similarity between the groups. The similarity is measured with the squared euclidean distance between each observation and the cluster means. Starting with k random cluster means, new cluster means are calculated iteratively to which the observations are assigned forming the clusters. The optimal number of clusters k for the data used in this study is derived from the sum of the squared residuals and ranges from 2 to 4. The results for k = 3 are presented in detail, and the results for k = 2 and ≥ 4 are also described. Cluster analysis is used to identify different thunderstorm environments. To 165 account for possible regional differences, clustering is performed separately on each of the 12 topographically homogeneous domains.
The online supplement provides the R code to replicate the cluster analysis and the principal component analysis (Morgenstern et al., 2023). thunderstorm environments are found by k-means clustering and a decision tree is presented to differentiate them (Sect. 4.2).
Finally, the thunderstorm environments are seasonally analyzed and compared in Sections 4.3 and 4.4.

175
This section investigates how the meteorological thunderstorm conditions vary regionally in Europe and whether some of the 12 domains can be grouped together based on their meteorological similarities during lightning throughout the year.    Accumulated variables considered at the next full hour after lightning observation. All other variables are considered at the last full hour.
CAPE: convective available potential energy, CIN: convective inhibition, msl: mean sea level, a.g.l.: above ground level, Temp. diff. sfc (sst or skt) -1000 m a.g.l.: Temperature difference between the surface and 1000 m a.g.l., where sfc is either sea surface temperature or skin temperature.
The domains A, B, C, and D (blue triangles) are all located in the top-left of Fig. 2. The labeled arrows indicate that these domains are physically characterized by increased boundary layer heights (~800 m) and increased wind speeds at 10 m (~6 m s −1 ) and at cloud base (~12 m s −1 ) relative to all other domains (Table 1). Long arrows pointing in opposite directions of domains A-D indicate decreased values. For example, decreased values in CAPE, CIN, pressure, and cloud size. The temperature difference between the ocean (or skin temperature over land) and the air at 1000 m altitude is on average 6.4 • C indicating a rather cool ground (Table 1). The regional characteristics of domains A-D are their large ocean areas including 190 coastlines, hence they are grouped together as the 'coastal' region. The domains J, K, and L (turquois squares) gather in the lower part of Fig. 2. Their common physical characteristics relative to the other domains are high 2 m dew point temperatures above 13 • C, elevated −10 • C isotherm heights of more than 4100 m, large CAPE values (> 400 J kg −1 ), the presence of CIN, and high amounts of total column water vapor (~25 kg m −2 ). The temperature difference between the surface and at 1000 m altitude is more than 8 K, indicating a warm ground (Table 1). Regionally, all these domains are located in the Mediterranean,195 and are hence grouped together as the 'mediterranean' region. The domains H and I are located on the right or top-right of   The abbreviations in the sub-environments stand for: CP = additionally increased cloud-physics variables, noMF = decreased mass-field variables, and SX = increased surface-exchange variables.

Thunderstorm environments
After finding four regions where thunderstorms have similar characteristics throughout the year, the next goal is to investigate whether individual thunderstorms occur under similar larger-scale meteorological conditions, i.e. whether different thunder-215 storm environments exist.
Cluster analysis with k = 3 is performed separately on every domain to find thunderstorm environments (clusters) relative to the overall lightning characteristics in that domain. Each found cluster from each domain is then described by its driving meteorological characteristics using the average values of the 25 input variables (cluster means). Then the average values within the physically-based categories (mass field, wind field, cloud physics, surface exchange, and topography) are computed for each 220 cluster to yield an overall characterization. Two major thunderstorm environments emerge as the wind-field category and the mass-field category always deviate substantially. The decision tree in Fig. 4 distinguishes between these two thunderstorm environments and helps to identify further sub-environments. Wind-field thunderstorms are characterized by increased windfield variables and sometimes decreased mass-field variables and are indicated by bluish colors. There are two wind-field sub-environments: Wind-field CP thunderstorms (dark blue) that have additionally enhanced cloud-physics variables (CP) and 225 wind-field noMF thunderstorms (light blue) that have decreased mass-field variables (no MF) while wind-field variables and cloud-physics variables are at their average values. The other major thunderstorm environment, mass-field thunderstorms, is characterized by average or increased mass-field variables plus often decreased surface-exchange variables and is indicated by reddish colors. There is one mass-field sub-environment, mass-field SX thunderstorms (dark red) with increased surfaceexchange variables (SX) and sometimes average mass-field values.   Table 2, and in the online supplement (Morgenstern et al., 2023). For robustness, each cluster analysis is repeated 50 times but only one representative result is shown. Figure 5 shows, how the wind-field thunderstorms (middle-blue triangles) in domain B are driven by enhanced wind speeds, enhanced boundary layer dissipation, lower −10 • C isotherm heights, little water vapor, small CAPE, and large boundary layer 240 heights of more than 1200 m (Table 2). Different from this, the wind-field CP thunderstorms (dark-blue circles) in domains I and sub-environments wind-field noMF and mass-field SX both occur in conditions where the (sea) surface is hot relative to the air at 1000 m altitude with an average temperature difference of 10.5 K, while in mass-field thunderstorms (without sub-environment) the air mass at 1000 m is only about 3.9 K colder than the surface. Regarding the topographical influences, mass-field thunderstorms occur more often over land (higher land-sea mask values), and wind-field thunderstorms more often over the sea.
In each domain, at least one wind-field related thunderstorm environment and one mass-field related thunderstorm environ-260 ment are found. The two major thunderstorm environments clearly separate from one another. Varying the number of clusters k robustly finds similar results. With k = 2 only the two major thunderstorm environments are found. With k > 3 more and more clusters are found referring to an already existing thunderstorm environment revealing no additional meteorological insights.
In summary, there are two major thunderstorm environments in Europe (wind-field thunderstorms and mass-field thunderstorms) and three sub-environments thereof. Thunderstorm environments are found by applying cluster analysis on 12 domains.

Seasonal differences between thunderstorm environments in Europe
The stacked barplots in Fig. 6 show how many lightning observations from each season belong to a given thunderstorm environment. As the data set is built to have the same number of observations from each season, the bars are equally high.
The absolute numbers of observations per domain are given in Table 1. In all domains, winter (DJF) is dominated by wind-270 field thunderstorm environments (blues) and summer (JJA) by mass-field thunderstorm environments (reds). Spring and fall are transitional seasons with varying proportions. If a domain has two wind-field thunderstorm environments (e.g., domain K), there is often a dominant thunderstorm environment with a more pronounced annual cycle (wind-field noMF ) and a smaller environment (wind-field CP ) with less seasonality.
The map in Fig. 7 spatially compares barplots that are estimated individually on each domain using local mean and standard deviations for scaling. The polygon colors indicate which domains are similar to one another (Sect. 4.1) and hence more comparable as they are scaled with similar values (baselines, Table 1). In every domain, wind-field thunderstorms (blues) dominate in winter (first bar) and contribute to a varying fraction of thunderstorms in spring and fall, which is higher the more maritime a domain gets ( In summary, wind-field thunderstorms dominate the cold season and are more important over the sea while mass-field 290 thunderstorms dominate the warm season and are more important over the mainland.

Comparability of the thunderstorm environments
The thunderstorm environments are identified relative to the general meteorological conditions during lightning in each domain, and the question remains how similar the thunderstorm environments of the same name from different domains are.
To make the thunderstorm environments more comparable, a principal component analysis is estimated on all cluster means 295 from every domain using the same scaling (Fig. 8). Again, the first two principal components are displayed on the axes explaining together about 70 % of the variance (PC 3 explains additionally 13.4 %) and the labeled arrows (loadings) indicate the contribution of each variable to the variance in the respective direction. Each domain (letters) is represented by three colored circles, the thunderstorm environments found there. First of all, the figure shows that the two major thunderstorm environments, wind-field thunderstorms and mass-field thunderstorms, clearly separate as the bluish and reddish circles are located in different 300 parts of the figure. The difference between the mass-field thunderstorms is small as the reddish circles gather close to one another. Their major difference is in the surface-exchange variables that separate the light red mass-field environment from the dark red mass-field SX sub-environment, which becomes more relevant in PC 3. Wind-field thunderstorms are more diverse as the bluish circles spread widely.   Table 2 and taking the first two principal components (PC) as axes. Each point represents a cluster and is colored and labeled according to its thunderstorm environment and domain. The labeled arrows (loadings) indicate the contribution of each variable to the variance in the respective direction.

Discussion
Regional lightning differences are described by four distinct regions: coastal, continental, mediterranean, and alpine-central.
Thunderstorm characteristics in different meteorological conditions are provided by the thunderstorm environments (wind-field thunderstorms and mass-field thunderstorms plus sub-environments).

310
Other authors have also investigated thunderstorm conditions. The variables important for our wind-field thunderstorms are similar to Mäkelä et al. (2013)'s investigations of winter lightning in Finland revealing the importance of vertical temperature difference between the surface and mid-troposphere (700/500 hPa) and shear, but not CAPE. Another classification was performed by Fujii et al. (2013) in Japan, who found that the number of winter lightning strokes and the probability of high-current lightning strokes, group into the storm type and inactive type dependent on the −10 • C isotherm height. Market environments, especially the sub-environment with reduced mass-field values (wind-field noMF ). Because high shear results in tilted clouds, charge separation in wind-field thunderstorm environments occurs along a slanted path within the cloud, resulting in charge centers that are also horizontally separated. This is known as the tilted charge hypothesis (Takeuti et al., 1978;Brook et al., 1982;Engholm et al., 1990;Williams, 2018;Takahashi et al., 2019;Wang et al., 2021), and is often described for cold season thunderstorms. Those thunderstorms do not require the release of CAPE to explain charge separation, so CAPE is often 325 low and consequently a poor predictor of such thunderstorms. Stucke et al. (2022) relate our two major thunderstorm environments as described in Morgenstern et al. (2022) to upward lightning at two alpine towers and find that most upward lightning occurs in wind-field thunderstorm conditions. Thus, windfield thunderstorms pose a particular risk to tall infrastructure and should be considered when determining the lightning threat to wind farms. If Stucke et al. (2022)'s relation between wind-field thunderstorms and upward lightning (i.e., lightning to tall 330 structures) holds also for flat terrain and for the thunderstorm sub-environment noMF, then offshore wind farms are at particular risk from wind-field lightning. The lightning protection standard IEC 61400-24 (2019) could be improved by additionally including the proportion of wind-field thunderstorms at sites considered for wind farms. Currently, only the local lightning density, the height of the structure, and an environmental factor (i.e., factors for winter lightning, terrain slope angle, and elevation) are taken into account (IEC 61400-24, 2019). The concept of thunderstorm environments introduced here is superior 335 to the idea of winter lightning versus summer lightning as it takes into account regional and seasonal differences. The concept is easily transferable to many locations as it is independent of static thresholds as they are used for example by March et al. (2016), Montanyà et al. (2016), or Sherburn and Parker (2014).
In general, thunderstorm frequencies under different synoptic conditions are often described (e.g., Wapler and James, 2015;Enno et al., 2014;Bielec, 2001;Kolendowicz, 2006) and regional thunderstorm differences are often subject of classical 340 climatologies as mentioned in the introduction. For the Baltic countries, Enno et al. (2013) found three distinct thunderstorm regions (continental, transitional, maritime) similar to some of our thunderstorm regions (continental, coastal). The Baltic countries probably have similar proportions of wind-field thunderstorms and mass-field thunderstorms as domain C because this domain covers parts of Lithuania and Latvia. However, Enno et al. (2013)'s thunderstorm regions indicate that the transition between maritime and continental thunderstorms is just a few dozen kilometers inland. Hence, a higher proportion of wind-field 345 related thunderstorm environments compared to domain C is expected in the Baltic countries, as most parts of these countries are close to the coast.

355
This study is limited by the resolution of the data used. Finer distinctions in the thunderstorm sub-environments are expected with higher resolution in the reanalysis data. More details in the model topography might lead to a more precise thunderstorm differentiation in complex terrain and a convection-resolving resolution could reveal more details about the meteorological characteristics of the thunderstorms themselves, not just the environments in which they occur. But this requires reanalysis on a scale finer than currently available. Further, a longer time series in the EUCLID data would allow more regions to be analyzed.

360
Scandinavia has been excluded from this investigation because the scarcity of lightning in winter caused the sample size over the investigated 11 years to be too small for an unbiased statistical analysis. These limitations affect only the thunderstorm sub-environments; the main results (wind-field thunderstorms and mass-field thunderstorms) are expected to remain the same.
Based on the results presented, several new research questions arise that are beyond the scope of this paper. How often are wind turbines or other tall structures struck by which thunderstorm environment? What is the relationship between thunder-365 storm environments and lightning properties such as the lightning duration, transferred charge, polarity, or channel length?
Is the decision tree (Fig. 4) valid for other extratropical regions? Are there other thunderstorm (sub-) environments in other climate zones such as the tropics? It would also be interesting to model lightning probability maps for each thunderstorm environment in each season, and to investigate the proportion of all thunderstorms in a given region that occur in a particular thunderstorm environment.

Conclusions
This study investigates seasonal and regional differences of meteorological environments in which lightning occur in Europe.
Highly destructive lightning damages often occur in seasons and regions where lightning is climatologically unlikely. They pose a challenge for lightning risk assessments because time series of lightning observations are often short and the meteorological conditions for lightning in the cold season are not well understood. This study explicitly includes infrequent lightning 375 conditions by considering an equal number of lightning observations from each season. EUCLID lightning data are combined with meteorological ERA5 data to answer two research questions: "Are there regions in Europe, where thunderstorms occur under similar meteorological conditions?" and "What characterizes thunderstorms in different meteorological environments and how do they vary seasonally across Europe?" Using coarse but consistent reanalysis data, this study paves the way for lightning reconstructions by providing tools to diagnose favorable lightning conditions.

380
Using principal component analysis, the European territory can be divided into four regions where the atmospheric conditions for thunderstorms are similar throughout the year: The alpine-central region with thick clouds, large cloud particle concentrations, and strong large-scale vertical velocities relative to the other regions; the mediterranean region with increased mass-field variables; the coastal region with increased wind speeds; and the continental region with in general average conditions and increased solar radiation relative to the other regions. Cluster analysis is performed individually on 12 domains 385 in Europe to find and describe different thunderstorm environments and to name them according to their characteristics compared to other thunderstorms in that domain. A decision tree is developed to easily distinguish the thunderstorm environments ( Fig. 4).
There are two major thunderstorm environments -wind-field thunderstorms and mass-field thunderstorms -and three sub-

395
The other major thunderstorm environment, wind-field thunderstorms, originate in more diverse weather conditions, but share the characteristics of average or reduced values in mass-field variables and elevated or average values in wind-field variables (high wind speeds at different heights, strong large-scale vertical velocities, large cloud shear, and increased boundary layer dissipation) relative to other thunderstorms in that domain. In this environment, CAPE is a poor predictor of whether lightning will occur because it is generally small. High wind speeds and shear cause the charged particles to be separated 400 along slanted paths. Wind-field thunderstorms dominate the cold season, especially winter, and are more important over the sea. Sometimes the cloud-physics (CP) variables are additionally enhanced leading to the wind-field CP thunderstorm subenvironment with large cloud sizes, increased concentrations of cloud particles (snow, ice, supercooled liquids), and large amounts of precipitation. Another sub-environment, wind-field noMF , is characterized by decreased mass-field variables (no MF) and often occurs over the sea.

405
The process-oriented view of different thunderstorm environments challenges the traditional idea of winter lightning versus summer lightning and makes it easier to compare similar processes with different magnitudes in different regions. In summary, this study shows that lightning in Europe originate in different meteorological environments, that winter lightning is not just a rarer sibling of summer lightning, and provides a decision tree to easily differentiate thunderstorm environments in Europe independent of a seasonal criterion or static thresholds.
Author contributions. DM performed the investigation, wrote the software, visualized the results, and wrote the paper. IS, TS, and DM performed the data curation, built the data set, and derived variables based on ERA5 data. TS contributed coding concepts. GJM provided 430 support for the meteorological analysis, data organization, and funding acquisition. AZ supervised the formal analysis and interpretation of the statistical methods. AZ, GJM, and TS are the project administrators and supervisors. All authors contributed to the conceptualization of this paper, discussed the methodology, evaluated the results, and commented on the paper.
Competing interests. The authors declare no competing interests.