Medicane Zorbas: Origin and impact of an uncertain potential vorticity streamer

Abstract. Mediterranean tropical-like cyclones (Medicanes) can have high societal impact and their accurate forecast remains a challenge for numerical weather prediction models. They are often triggered by upper-level potential vorticity (PV) anomalies, such as PV streamers and cut-offs. But knowledge is incomplete about their detailed formation processes and factors limiting their predictability. This study exploits a European Centre for Medium-Range Weather Forecast (ECMWF) operational ensemble forecast with an uncertain PV streamer over the Mediterranean, which, three days after initialisation, resulted in an uncertain 5 development of Medicane Zorbas in September 2018. Using an ad-hoc clustering of the ensemble members according to the PV streamer position, it is demonstrated that uncertainty in the initial conditions near an upper-level jet streak over the Gulf of Saint Lawrence is the dominant source of the subsequent uncertainty in the position of the PV streamer over the Mediterranean. The initial condition uncertainty strongly amplifies baroclinically after 18 h in a region of strong quasi-geostrophic forcing for ascent in the left exit of a jet streak over the North Atlantic. The further amplification and downstream propagation of the 10 tropopause-level PV uncertainty leads to a large spread in the position of the PV streamer over the Mediterranean after three days, directly limiting the predictability of the position, thermal structure and evolution of Zorbas. Two low-level airstreams possibly play a key role in linking the uncertainties of the large-scale upper-level flow with meso-scale uncertainties in the cyclone structure. Overall, this study is an illustrative example that uncertainties in large-scale initial conditions can determine the practical predictability limits of a high-impact weather event. 15


Medicanes: high impact, limited understanding, and uncertain forecasts
In the last 15 years, there has been increasing research on cyclones placed at the interface between the the classical concepts of tropical and extratropical cyclones. Several studies focused on so-called subtropical cyclones (STCs; e.g. Guishard et al., 2009;González-Alemán et al., 2015). STCs are low pressure systems that initially form from baroclinic processes, but they later 20 acquire tropical characteristics due to convective processes, while fronts dissolve. In the second phase convection begins to build a low-level warm core and the environment becomes more barotropic. STCs have gained attention because they can have substantial impacts on society (Guishard et al., 2007), but also because they potentially convert into fully-tropical cyclones, a process called tropical transition (Davis and Bosart, 2004). Davies, 1997;Dirren et al., 2003;Davies and Didone, 2013;Gray et al., 2014;Baumgart et al., 2018). They can originate from errors in initial conditions and have shown to result in so called "forecast busts" over Europe, i.e. periods of anomalously low predictability (Rodwell et al., 2013;Magnusson, 2017;Grams et al., 2018). Also, the misrepresentation of diabatic processes in the forecast model can induce PV errors, as for example in warm conveyor belts (Gray et al., 2014;Martinez-Alvarado et al., 2016). Warm conveyor belts affect the Rossby wave guide via their low-PV outflow in the upper troposphere. PV errors near 70 the tropopause, due to errors in initial conditions and/or model physics, translate into Rossby wave errors that then propagate downstream.
To become relevant, forecast errors in Rossby waves must amplify. Different case studies have shown that this can be due to a range of processes. Baumgart et al. (2018) point out the importance of non-linear (barotropic) dynamics near the tropopause and to a lesser extent baroclinic interaction and effects of upper-tropospheric divergent winds. The contribution of non-linear 75 tropopause dynamics to amplification (and downstream propagation) of forecast errors can be understood by the mutual interaction of negative and positive PV errors near the tropopause (Davies and Didone, 2013). Grams et al. (2018) identified warm conveyor belts in a forecast bust case as key for amplifying forecast errors in the tropopause region. This process includes error amplification in the baroclinic growth of a cyclone, possibly enhanced by model error in the representation of diabatic processes in the warm conveyor belt, resulting in errors in the divergent wind and size and amplitude of the negative PV anomaly 80 in the upper troposphere. Together, these studies indicate that forecast errors in near-tropopause Rossby waves can grow just due to their internal non-linear dynamics, but baroclinic coupling can, via rapid baroclinic growth or warm conveyor belts and their associated low-PV outflow, enhance this amplification.
The relevance of these processes to the amplification of forecast errors and uncertainty is likely very case dependent and their interplay potentially complex. However, only few case studies exist that investigate the origin, amplification and propagation 85 of forecast errors in Rossby waves and their impact on the uncertainty in the prediction of high-impact weather events.

Zorbas: An uncertain Medicane
In this context, we exploit the recent Medicane Zorbas, which rapidly acquired fully tropical-like characteristics. Operational ensemble forecasts by the European Centre for Medium Range Weather Forecast (ECWMF) did not agree on the development of Zorbas, even at short lead times. Zorbas, from the moment of its formation until it acquired full tropical-like characteristics, 90 led to considerable damage through severe winds, torrential rainfall, major flooding and even tornadoes. The main affected region was Southern Greece, especially Crete, Peloponnese, Evia, and the region around Athens.
We investigate the origin and amplification of the forecast uncertainty associated with the PV streamer that triggered Zorbas (see Section 3) and identify the relevant large-scale conditions limiting the Medicane's predictability. The key questions of our study are: What is the origin of the forecast uncertainty in the PV streamer and which sequence of dynamical processes lead 95 to its amplification and propagation into the Mediterranean? Was the uncertainty in the PV streamer the direct cause of the uncertain Medicane prediction and why? This study does not aim to analyze the details of the Medicane dynamics but rather focuses on the large-scale processes prior to the Medicane formation.
The remainder of this article is structured as follows. After describing the data used in this study we give an overview on the synoptic evolution of Zorbas and introduce the main method that builds the basis for all subsequent analyses. Then, the next 100 Sections are organized as a journey from the origin of the large-scale forecast uncertainty to its effect on the uncertainty of the cyclone development and precipitation impacts. This is then connected to uncertainties in the low-level moisture and warm-air advection prior to the Medicane formation. Finally, uncertainties in the evolution of the vertical thermal structure of Zorbas are diagnosed and links to the relevant precursors discussed. We close with highlighting the main conclusions and discuss implications for further research on Medicane dynamics and on uncertainty in Rossby wave forecasts. 105 2 Operational ECWMF products The basic data for this study is from the ECMWF Integrated Forecasting System (IFS, Cycle 45r1; ECMWF, 2018). We use the operational ensemble forecasts with 50 perturbed members initialized at 0000 UTC 24 Sep and 0000 UTC 27 Sep 2018 (46 perturbed members available), the operational analysis, and operational short-term forecast of 6-hourly accumulated precipitation, are used. The spectral resolution of the operational ensemble is TCO639 (about 18 km) on 91 model levels, and 110 the resolution of the operational analysis TCO1279 (about 9 km) on 137 model levels. The data is available every 6 h and has been interpolated to a regular grid with a horizontal resolution of 1 • x 1 • . From the standard variables we additionally compute PV, equivalent potential temperature and quasi-geostrophic omega (QG ω) on 850 hPa as forced by levels above 550 hPa [computed the same way as in Graf et al. (2017) ]. As a measure for forecast skill, anomaly correlation coefficients (ACC) are calculated for geopotential height at 500 hPa for each ensemble member of the forecasts initialized at 0000 UTC 24 115 Sep and 0000 UTC 27 Sep 2018.

Synoptic overview
Figures 1, 2 and 3 provide an overview of the atmospheric processes before and during the formation of Zorbas based on ECMWF operational analysis and infrared images from the European Organisation for the Exploitation of Meteorological Satellites (EUMETSAT). At 0000 UTC 26 Sep 2018, about a day before cyclogenesis, a few convective clouds (white patches) 120 and some precipitation (red contours) can be identified in the Southern Mediterranean and Lybia (Fig. 1a). Convective activity is confirmed by the occurrence of lightning in this region (as inferred from www.lightningmaps.org, not shown). At this time, a large-scale trough is present over Eastern Europe (Fig. 2a). In the Mediterranean region, the low-level air masses (850 hPa) 4 https://doi.org/10.5194/wcd-2019-1 Preprint. Discussion started: 28 August 2019 c Author(s) 2019. CC BY 4.0 License. have very high equivalent potential temperatures whereas in Eastern Europe the values are much lower (Fig. 3a) indicating substantial baroclinicity and moisture gradients between these two regions.

125
In the evening of the same day, at 1800 UTC 26 Sep 2018, cloud formation and precipitation over the Mediterranean is enhanced (Fig. 1b). The upper-level trough has elongated into a narrow PV streamer on 325 K that extends towards the Central Mediterranean ( Fig. 2b). At the streamer's downstream side (to the east), a large area of enhanced QG forcing for ascent by the upper levels (QG ω, red contour in Fig. 2b) provides favorable conditions for the relatively strong convective precipitation off the Lybian coast [local maximum of 24 mm (6 h) −1 indicated by the blue contours in Fig. 1b]. High wind speeds on 850 hPa 130 over the Aegean Sea (arrows in Fig. 3b) indicate strong low-level cold air advection from Eastern Europe and the Black Sea region towards Greece and the Central Mediterranean. Cold and dry air masses that are advected over a warm ocean surface can enhance sea-surface latent heat fluxes and become relevant for Medicane formation (Miglietta and Rotunno, 2019).
A day later, at 1800 UTC 27 Sep 2018, the clouds have formed a spiral-like structure with a weak frontal cloud band extending from the Mediterranean over Lybia (Fig. 1c). Surface cyclogenesis has taken place off the Lybian coast close to Benghazi 135 (closed yellow contours in Fig. 3c) and increased gradients of equivalent potential temperature at this location ( Fig. 3c) confirm the presence of a weak surface cold front. The PV streamer has broken up into segments (Appenzeller and Davies, 1992). Its tip has formed a C-shaped PV cut-off (shading in Fig. 2). To the east of it, a large region of enhanced QG ω (red contours) is present and strongly ascending air masses (black crosses, ascent rate larger than 600 hPa in 24 h) as identified from trajectory calculations using the Lagrangian analysis tool LAGRANTO (Wernli and Davies, 1997;Sprenger and Wernli, 2015) are located 140 in the dent of the cut-off, above the cyclone centre. This is where also the precipitation maximum occurs [88 mm (6 h) −1 , red contours in Fig.1c]. The presence of rapidly ascending and strongly precipitating air masses are indicative for substantial diabatic effects [cf. the concept of warm conveyor belts; Joos and Wernli (2012); Binder et al. (2016)]. They not only heat the cyclone centre and potentially form a low-level positive PV anomaly but favor the direct erosion of the PV cut-off due to diabatic PV modification and entrainment (Portmann et al., 2018). Furthermore, they transport low-PV air to the 325 K level, 145 which causes the dent structure in the upper-level PV and further affects the evolution of the cut-off. The ascent seems to be strongly convective, as confirmed by the high lightning activity in this region (as inferred from www.lightningmaps.org, not shown).
At 1800 UTC 28 Sep 2018, almost no strongly ascending air masses are present on 325 K, the PV cut-off has almost completely decayed on 325 K, and the enhanced QG ω has vanished (Fig. 2d). The surface cyclone has further intensified and moved over 150 the Central Mediterranean. The satellite data shows a cyclone without clear frontal cloud bands and still substantial precipitation in its centre (Fig. 1d), which is, however, not associated with strong QG ω but likely due to conditional instability. The region with enhanced gradients of equivalent potential temperature previously identified as a weak cold front has moved eastward and become disconnected from the cyclone centre. A local maximum of low-level equivalent potential temperature is present in the cyclone centre, indicating the formation of a warm seclusion (Shapiro and Keyser, 1990). Warm seclusion events have 155 been previously linked to Medicane formation (Mazza et al., 2017). All this suggests that the cyclone underwent the transition from an extratropical to a subtropical-like or even tropical-like system between 1800 UTC 27 Sep and 1800 UTC 28 Sep 2018. As diagnosed from the cyclone phase space [CPS; Hart (2003), for more details see Sect.6.3], a low-level warm core is 5 https://doi.org/10.5194/wcd-2019-1 Preprint. Discussion started: 28 August 2019 c Author(s) 2019. CC BY 4.0 License. already present at 1800 UTC 27 Sep and a deep warm core is reached only 12 hours later at 0600 UTC 28 Sep (see Fig 4a).
Consequently, Zorbas also visually acquires tropical characteristics including the formation of an eye on 29 September 2018 160 as it moves over Greece (not shown). However, this later period of the cyclone evolution is not in the focus of this study.
In summary, Zorbas forms via extratropical cyclogenesis forced by a breaking-up PV streamer over a baroclinic zone in the Central Mediterranean. As low-level moisture and temperature over the Mediterranean are very high and north-easterly cold air advection at low levels occurs, cyclogenesis is accompanied by strong convection, cloud formation and intense precipitation, first at the PV streamers downstream flank and later in the cyclone centre. Zorbas then transitions into a subtropical and later 165 even tropical-like system: It loses its frontal structures and becomes a circular cyclone with a warm seclusion, which is where most precipitation occurs.

Ensemble clustering according to position of PV streamer
The position of the PV streamer at 0000 UTC 27 Sep (shown in Fig. 2 six hours earlier) varies strongly in the ensemble forecast initialized at 0000 UTC 24 Sep 2018, with about equal shares of ensemble members where the PV streamer is roughly correct, 170 too far west and too far east, respectively. This offers the opportunity to use this ensemble forecast to study the dynamical processes that lead to this significant forecast uncertainty in the upper troposphere, which subsequently affected Medicane formation (see Sect.6). Therefore a pragmatic clustering method was designed to separate the strongly diverging PV streamer evolutions in these ensemble members into clusters with a similar evolution. The three identified PV streamer scenarios, i.e.  Fig. 4). The clustering uses PV vertically averaged from 320 to 330 K, where all tropospheric PV values (<2 PVU) are set to zero, hereafter called PV av . Hence, PV av is high in areas where the PV streamer is strong and deep, and low where it is weak and shallow. The clustering is based on two different steps: First, all ensemble members are identified for which the region with PV av ≥2 PVU has more than 75% overlap with the corresponding area in the analysis. In these 19 members (cluster 1), the streamer has 180 a similar shape and location as in the analysis (see blue shading in Fig. 4). The remaining members are separated into two clusters depending on whether the maximum PV av is shifted to the west (cluster 2, 12 members, green shading in Fig. 4) or east (cluster 3, 18 members, green shading in Fig. 4) relative to the analysis. There is one ensemble member that can not be attributed to one of the three clusters because its overlap is less than 75% and the maximum of PV av is located at the same longitude as in the analysis.

185
The meaningfulness of this clustering for understanding predictability is supported by the fact that it helps explaining the temporal development of the ACC in the Mediterranean box. To this aim, Fig. 5a summarizes the synoptic sequence of the case, i.e. the formation of the PV streamer over the Mediterranean, the break-up resulting in a PV cut-off (grey boxes), the minimum sea-level pressure of Zorbas (solid line) and its thermal structure as diagnosed from the CPS (colors, for details see Sect.6.3). As shown in Fig. 5b, the ACC of geopotential height at 500 hPa in the Mediterranean starts to drop (median and We conclude that errors in the position and shape of the PV streamer limited the large-scale predictability as measured by the ACC of geopotential height on 500 hPa in the Mediterranean, and that the clustering incorporates the relevant characteristics of the PV streamer well. However, the ACC of geopotential height on 500 hPa does not fully account for errors in the vertical structure of the cyclone and its intensity and exact position, aspects that are potentially relevant for predicting the cyclone's 200 impact. Hence, even if the large-scale predictability related to the PV streamer is high, still there can be relevant errors in the details of the cyclone evolution and its interaction with the upper levels, which may severly limit meso-scale predictability.

PV streamer scenarios emerge from initial condition uncertainties and baroclinic amplification
We now investigate how the diverging PV streamer scenarios identified in Sect.4 emerge from differences in initial conditions that amplify baroclinically at the left exit of a jet streak over the North Atlantic. To this end, we analyze the differences of the 205 means of clusters 2 and 3 as these are the clusters that deviate the most in terms of the PV streamer evolution. We investigate the upper-level development using differences in PV and winds on 325 K ( Fig. 7a-d), and the baroclinic interaction with the lower levels using QG ω and geopotential height on 850 hPa ( Fig. 7e-h). For upper-level PV and winds, we use normalized cluster-mean differences (Torn et al., 2015): where σ P V is the standard deviation of all ensemble members. Hence, ∆P V becomes large when the cluster-mean difference of PV at a given location is much larger than the ensemble standard deviation at the same location, i.e. when the two clusters contain the members of the ensemble that are most different from each other. Large absolute differences in regions of strong gradients, particularly at the tropopause, are given less weight. Additionally, it allows us to easily compare different lead times. For example, if ∆P V increases with lead time, the cluster differences grow faster than the ensemble standard deviation, 215 which means that the clusters become increasingly distinct from each other, relative to the full ensemble. Further, in order to make statistically robust statements, regions where cluster-mean PV values significantly differ are identified using a two-sided Wilcoxon rank-sum test (see supplementary material). By design, significant PV differences tend to be co-located with high values of ∆P V .
Already at 0000 UTC 24 Sep 2018, i.e. the initialization time of the forecast, a relatively large area of significantly positive PV 220 differences is discernible on the stratospheric side of a jet streak over the Gulf of Saint Lawrence (teal contour in Fig. 6). The normalized PV differences in this area are between 0.5 and 1.5 standard deviations (red shading in Fig. 6). With increasing lead time this PV difference in the initial conditions moves eastward along the 2-PVU contour, amplifies over the North Atlantic and produces a dipole of negative and positive PV differences further downstream that ultimately result in the differences in the PV streamer formation over the Mediterranean (Figs. 7a-d). In the following, this development is discussed in more detail.

225
The initial PV difference as shown in Fig. 6 moves eastward and after 6 h is located northeast of Newfoundland, with an amplitude still between 0.5 to 1.5 standard deviations (Fig. 7a). After 18 h (Fig. 7b), the PV difference is largest over the central North Atlantic, has amplified, reaching values above 1.5 standard deviations, and covers a larger area. Also, the mean 2-PVU contours of clusters 2 (dashed) and 3 (solid) start to separate at this location and a clear cyclonic difference wind field (arrows) is present. Additionally, downstream of the positive PV difference, a negative PV difference north of the British Isles emerges. faster downstream than the low-level wave. They reach the Mediterranean region while the low-pressure system resulting from this interaction becomes stationary east of Iceland.

Diverging synoptic development impacts Medicane predictability
Having elucidated the reason for the uncertainty in the position of the PV streamer, we now have a closer look at how this uncertainty affects the subsequent cyclone development in the Mediterranean. In addition, we analyze how the PV streamer 265 scenarios affect the potential precursors for the formation of a low-level warm core, which is crucial for the meso-scale dynamics of the transition of an extratropical to a Medicane-like cyclone. Finally, we diagnose the predicted number of Medicane-like systems in each cluster and discuss the relevant large-scale precursors.

Synoptic development over the Mediterranean
To examine the diverging synoptic development we analyze for all three clusters the evolution of mean upper-level PV as well 270 as geopotential height on 850 hPa and surface precipitation. Regions where these fields significantly differ between two clusters are identified using a two-sided Wilcoxon rank-sum test (see supplementary material). Figure  the 2-PVU contour of cluster 1 is very similar to the analysis, the contour is clearly shifted to the west compared to the analysis in cluster 2 and to the east in cluster 3. At 1200 UTC 26 Sep 2018, after the narrow PV streamer has formed, the differences between the clusters become more obvious (Figs. 8b,f,j). The shape and position of the streamer in cluster 1 is still very close to the analysis, whereas in cluster 2 the tip of the streamer is thinner and extends more to the west, and in cluster 3 it is shifted to 280 the east. In these regions, clusters 2 and 3 significantly differ from cluster 1 (see supplementary material). This is not surprising as the clustering was specifically designed to focus on these differences.
After the PV cut-off formation, at 1200 UTC 27 Sep 2018 (Figs. 8c,g,k) the differences in the scenarios over the Mediterranean are very prominent. While in cluster 1, the cut-off is located south of Italy in the Central Mediterranean (as in the analysis), cluster 2 exhibits a much weaker cut-off shifted to the west over Tunisia and cluster 3 a stronger cut-off shifted to the east 285 over the Eastern Mediterranean. In all clusters surface cyclogenesis occurs slightly east of the cut-off (closed purple contours and black circles), which is where we expect the strongest QG forcing for ascent (as visible in Fig. 2). Hence, in cluster 1 the cyclone forms in northeastern Lybia (as in the analysis, indicated by the teal star), in cluster 2 in northwestern Lybia close to Tripoli, and in cluster 3 in the Eastern Mediterranean over Crete. At this stage, cluster 3 exhibits the strongest, most developed surface cyclone. One day later, at 1200 UTC 28 Sep 2018 (Figs. 8d,h,l), the cut-off in the analysis has decayed already into smaller patches due to the effects of strong latent heat release in rapidly ascending and precipitating air masses, as discussed in Sect.3. In cluster 1 the cut-off has clearly weakened (PV values < 3 PVU), in cluster 2 it fully decayed, and in cluster 3 it is still very prominent and strong (PV values > 6 PVU), indicating substantial differences in latent heat release. In both clusters 1 and 3 the vertical structure of the system has become more barotropic, i.e. the cut-off and the surface cyclone are vertically aligned. In cluster 1 295 the cyclone has further intensified with the minimum cluster-mean sea-level pressure dropping from 1010 hPa to 1007 hPa (not shown), whereas in clusters 2 and 3 it has weakened from 1011 hPa to 1013 hPa and from 1009 hPa to 1011 hPa, respectively.
The accumulated cluster-mean precipitation during the period when most members exhibit a cyclone track (1800 UTC  there is no cyclone track in the ensemble that closely follows the observed one in the second part of the life cycle, when the cyclone moves over Greece and leads to substantial precipitation over the Peleponnes, the Athens region, and Evia. We conclude that the uncertainties in the PV streamer's zonal position over the Mediterranean, as a consequence of the processes discussed in Sect.5, directly lead to uncertainties in the location of cyclogenesis and the amount and location of precipitation. For cluster 1, where the PV streamer location most closely matches with the analysis, the predicted cyclogenesis, 310 cyclone tracks, PV cut-off evolution, and precipitation patterns also compare most favourable with the operational analysis and short-term forecasts. The uncertainty in the evolution of the PV cut-off likely directly affects the development of the cyclone and might be a crucial factor determining if a Medicane finally forms or not: If the cut-off is weak and decays early (as in cluster 2), the cyclonic circulation, destabilization, and quasi-geostrophic forcing for ascent are also weak and might not be sufficient to produce a strong cyclone with a deep warm core. We will elaborate further on this hypothesis in Sect.6.3.

Low-level airstreams relevant for Medicane formation
To investigate potential precursors of a low-level warm core and subsequent development of a Medicane-like system, we now focus on two different low-level air masses in the cyclone around cyclogenesis time: The air directly located in the cyclone centre and the air with the highest equivalent potential temperature in the warm sector of the cyclone, which could potentially be advected into the cyclone centre to form a warm seclusion. Note that we do not identify low-level warm cores directly and do 320 not investigate their formation in detail. However, we hint on two airstreams that are potential precursors of a low-level warm core and are dominated by the large-scale situation in each cluster. We highlight the important differences between the clusters and use them to hypothesize about the relevance of the large-scale situation for the probability of transition into Medicane-like systems in each cluster. This probability is then quantified in the subsequent section.
(Wernli and Davies, 1997;Sprenger and Wernli, 2015) (i) from all grid points on 850 hPa within a radius of 200 km around the cyclone centre, and (ii) from the 15% of grid points with the highest equivalent potential temperature within 500 km from the cyclone centre. With the latter, we aim to select the warmest and moistest air parcels in the warm sector of the developing cyclone. The backward trajectories are started at 1200 UTC 27 Sep 2018, which is at cyclogenesis in the operational analysis and at or shortly after cyclogenesis in most ensemble members and, hence, before cyclone dynamics starts to dominate air 330 mass advection. It is also before a warm core forms in the operational analysis and some ensemble members. Therefore, this timing allows identifying the large-scale airstreams prior to cyclogenesis and discussing their potential role as "predictors" for the development of a Medicane-like cyclone. The evolution of equivalent potential temperature (Fig. 11a-c), potential temperature ( Fig. 11d-f) and specific humidity (Fig. 11g-340 i) along the backward trajectories reveals additional differences between cluster 3 and clusters 1 and 2. In cluster 3, the air parcels in the cyclone centre (blue shading) have an average equivalent potential temperature more than 7 K lower than in clusters 1 and 2, even though 48 h before the equivalent potential temperature was very similar in all clusters. To a large part, this can be attributed to a stronger increase in specific humidity (Fig. 11g-i) especially in the period between -36 h and -12 h.
The air parcels are, in contrast to cluster 3, transported over the Mediterranean (Fig. 10a-c) where they most likely moisten due 345 to ocean evaporation. In the same period, potential temperature also increases more strongly in clusters 1 and 2, likely due to surface sensible heat fluxes. The rapid increase in potential temperature and drop in specific humidity in clusters 1 and 2 during the last 6 h clearly indicates strong latent heating due to cloud formation. This signal is almost entirely missing in cluster 3.
The air parcels selected in the warm sector of the cyclone have very different properties in clusters 1 and 2 compared to cluster 3 already at their origin, when they are much moister and warmer (green shading in Fig. 11). However, they experience 350 only a slight warming and moistening as they are transported into the warm sector, whereas in cluster 3 they are significantly moistened and heated. Nevertheless, the air parcels in cluster 3 are eventually still cooler and drier. In cluster 3, the temporal evolution of all three variables (Fig. 11c,f,i) is very similar for the air parcels selected in the cyclone centre (blue) and the ones in the warm sector (green), with the main difference that the latter are warmer and moister. In clusters 1 and 2, however, the air parcels are part of airstreams with clearly distinct properties. 355 We conclude that in the ensemble members of clusters 1 and 2, the air which originates from the Black Sea / Eastern Europe region and later constitutes the cyclone centre substantially moistens as it is transported rapidly over the Central Mediterranean (wind speeds over the Bosporus reach above 20 m s −1 in the operational analysis, see Fig. 3). Previous studies have already pointed out the importance of strong surface fluxes for Medicane formation due to dry and cold winds reaching the the upper-level PV streamer/cut-off, this favours the strong latent heating of the air masses eventually constituting the cyclone centre. In cluster 3, this process is lacking as the cyclone formation occurs in the northern part of the Mediterranean. Hence, in clusters 1 and 2, moister and warmer air is present in the cyclone centre that could favour the transition into a Medicane-like cyclone. Moreover, the PV streamer is located such that very warm and moist air from the Central Mediterranean is advected cyclonically into the warm sector, which might support the formation of a warm seclusion and transition into a Medicane-like 365 cyclone. Again, in cluster 3 this process is lacking. Hence, from the perspective of low-level processes, clusters 1 and 2 favour the formation of a low-level warm core, whereas this is not the case for cluster 3. As caveat, we mention that this analysis neglects that the cyclone formation does not occur exactly at the same time in all members, and the stage of the cyclone may be slightly different in different members. Nevertheless, it provides a basic understanding of the fundamental differences in the low-level processes between the clusters.

Explaining Medicane predictability
So far, the uncertain forecast of the PV streamer was shown to be directly linked to uncertainties in the position and magnitude of the PV cut-off, and the location of cyclogenesis and precipitation. Further, the eastward displacement of the PV streamer in cluster 3 leads to substantial differences in the low-level flow and, likely, air-sea interaction, compared to clusters 1 and 2.
We now argue that this helps to explain major differences in the vertical thermal structure of the cyclone in the three clusters 375 and the extent to which it acquires tropical-like characteristics. To this aim, we consider the cyclone phase space (CPS; Hart, 2003), which is a useful tool to diagnose the thermal structure of cyclones.
Cyclone tracks at 6-h temporal resolution are obtained for each of the 50 ECMWF ensemble members and the operational analysis using the cyclone detection and tracking method described by Picornell et al. (2001). This method was specifically de- Based on the results obtained in Sections 6.1 and 6.2, we expect cluster 1 to produce most Medicane-like systems as a strong upper-level PV cut-off is present even after cyclogenesis and there is supply of very moist and warm air in the lower levels.
Cluster 3, on the other hand, is expected to produce less Medicane-like systems, as the conditions for the formation of a low-395 level warm core are much less favorable. Finally, we expect cluster 2 to be placed between clusters 1 and 3, because conditions are favourable for the formation of a low-level warm core but likely the PV cut-off is too weak to maintain cyclonic circulation, destabilization, and forcing for ascent after cyclogenesis.
For each cluster we identify how many members form a DWC cyclone. As shown in Table 1, this is the case for 15 out of the 19 members (79%) in cluster 1. In cluster 3, only 2 out of 18 members (11%) develop a DWC cyclone, and, as expected, 400 cluster 2 shows an intermediate scenario with 6 out of 12 members (50%) producing a DWC cyclone.
Cluster 1 not only produces significantly more DWC cyclones but also shows stronger upper-level warm cores (indicated by the higher −V U T values in Fig. 12) and longer duration of the DWC stage (indicated by the number of DWC steps in Fig. 12). Note that, especially for cluster 3, the number of DWC steps has to be considered with caution, due to the small sample size.
Nonetheless, these results show that cluster 1 tends to produce not only more but also more robust DWC cyclones. Interestingly, 405 the Medicane in the operational analysis has an upper-level warm core that is on the weaker side of what members in cluster 1 forecasted and is about as strong as the upper-level warm cores produced in clusters 2 and 3. But it maintains a deep warm core about twice as long as the average Medicane in all clusters.
Overall, these findings suggest that cluster 1 provides the best synoptic environment out of the three clusters for a Medicanelike system to form. However, the strength of the upper-level warm core in cluster 1 is overestimated compared to the analysis 410 and the variability among the cluster members is large. This indicates that the synoptic setting in cluster 1 has the potential for much stronger Medicane-like systems than the one that actually occurred. On the contrary, the duration of the deep warm-core stage is strongly underestimated. We conclude that, once the PV streamer is forecasted well, sub synoptic-scale processes including the detailed interaction between the surface cyclone and upper levels become limiting factors to accurately predict the cyclone evolution including its vertical structure. It is clear that the uncertain position of the PV streamer was the dominant factor limiting the predictability of Zorbas for both its location and its vertical thermal structure. The first aspect, the direct influence on the cyclone formation is straightforward, as the PV streamer and cut-off provided the main forcing for cyclogenesis. Regarding the second aspect, the cyclone structure, 425 we identified two possible large-scale key ingredients relevant for the transition of the extratropical cyclone into a sub-tropical cyclone and, consequently, a tropical-like system, i.e. a Medicane. First, low-level advection below the eastern side of the streamer had to be such that dry and cold air masses from Eastern Europe and the Black Sea region were transported rapidly over the Mediterranean. They took up moisture by surface fluxes, experienced diabatic heating and ended up in the cyclone centre, where they helped to form a strong low-level warm core [similar to the first case in Miglietta and Rotunno, 2019) ]. The 430 location of cyclogenesis and hence the extent to which this process could be active was directly linked to the position of the PV streamer. Additionally, the PV streamer had to be far enough west for its induced circulation to reach the region with very warm and moist low-level air over the Central Mediterranean and advect it cyclonically around its tip into the warm sector of the cyclone. As a second ingredient, the upper-level PV cut-off had to be strong enough to maintain the cyclonic circulation and destabilize its immediate surrounding to favour deep convection even after cyclogenesis. This is reminiscent of the second 435 Medicane case in Miglietta and Rotunno (2019) and the case discussed in Fita and Flaounas (2018). In most members of cluster 1, both conditions were fulfilled, whereas in cluster 2 the second and in cluster 3 the first condition was mostly missing.
The uncertainties in the position of the PV streamer after 72 h forecast lead time could be clearly linked to relatively large-scale uncertainties in the initial conditions on the stratospheric side of an upper-level jet streak over the Gulf of Saint Lawrence. They propagate along the dynamical tropopause and strongly amplify in the left exit of a jet streak over the North Atlantic. At the 440 same time, the strong QG forcing for ascent in this region enables a coupling of upper and lower levels and initiates a baroclinic wave. This wave and the associated rapid growth, as expected from baroclinic instability, are crucial for the in situ amplification of the uncertainties in this case. Non-linear tropopause dynamics then leads to the rapid downstream development of the uncertainties eventually resulting in the uncertain PV streamer and the development of Zorbas. The contributions of diabatic airstreams, such as warm conveyor belts, were negligible for the uncertainty amplification in this case. The described amplifi-445 cation process could be an important element to better understand the amplification of forecast uncertainties also in other flow situations, especially in the storm track regions. Further case studies as well as more climatological analyses are needed to quantify its relevance.
Since the seminal work of Lorenz (1969), the growth of very small uncertainties on convective scales to large-scale uncertainties, so called upscale error growth, has been discussed as theoretical limit of atmospheric predictability (e.g. Zhang et al., 450 2007). The picture of the flapping wings of a butterfly influencing the development of a storm much later has become well known outside research. However, recent studies suggested that the practical limits of atmospheric predictability often come from uncertainties on much larger scales, even if they are very small compared to the average kinetic energy on that scale (e.g. Durran and Weyn, 2016). This study provides an illustrative example that large-scale, but relative to the background kinetic energy small, uncertainties in an upper-level jet streak over North America can dominate the forecast uncertainty of a storm in 455 the Mediterranean.
Finally, we note that many other factors may be also relevant for the exact evolution of the Medicane even if the PV streamer is at the right location. The details of the interaction between lower and upper levels, for example, likely influence the formation the warm core, the intensification, and the track of the cyclone. This in turn determines if the cyclone remains over the sea and is able to intensify, or, if it makes landfall and decays. These aspects are subject of a follow-up study.