After the hailstorm, two citizens of Azzano Decimo (Michele Cristofoli and Sara Santarossa) collected some of the largest hailstones from their garden and put them into their freezers. Figure 2a shows a wide distribution of hail sizes at the observer’s location. Of course, the observer collected only the largest stones (Figure 2b). Figure 2c was taken by the other observer two days later, when she decided to weigh the hailstones. ARPA FVG - OSMER (Osservatorio Meteorologico dell’Agenzia per l’Ambiente del Friuli Venezia Giulia) discovered the existence of these 9 hailstones thanks to social media and, a few months later, organized their shipping with 60 kg of dry ice (see the Supplementary Figure S1) to NCAR in Boulder, Colorado, United States, where a cold room laboratory is operated by one of the authors (Knight).

Multiple kinds of observations are used to analyze the meteorological conditions supporting the Azzano Decimo hailstorm, including data from surface meteorological stations, radiosondes, radars, satellite, and lightning sensors. The surface stations are managed by Civil Protection FVG, while the data quality control is made by ARPA FVG - OSMER. The highest-resolution sampling time is 1 min, but in this work only station measurements with 5-min resolution are used. The radiosounding data are managed by Aeronautica Militare. The Rivolto station (45.978° N, 13.058° E, 51 m AMSL; WMO ID 16045) usually performs only 00 and 12 UTC soundings, but, upon request of ARPA FVG - OSMER during alerts or severe weather outlook, they can also launch extra soundings at 06 and/or 18 UTC, as was done in the Azzano Decimo case. The sondes used are Vaisala RS41 and the full-vertical resolution data (1-Hz sampling) are analyzed. Two Doppler C-band radars cover the area of interest: the first is the Fossalon di Grado GPM-500C dual-polarization radar (45.73° N, 13.48° E, 0 m AMSL, see Bechini et al., 2002), which is managed by Civil Protection FVG and operates at 5.630 GHz (wavelenghth ∼ 5.3 cm) with an unambiguous range of 125 km, and the Pasja Ravan single-polarization radar (46.10° N, 14.23° E, 1,019 m AMSL), managed by the Slovenian Environmental Agency (ARSO) and operating at 5.633 GHz with an unambiguous range of 150 km. While the Pasja Ravan radar is 120 km far from Azzano Decimo, the Fossalon di Grado is much closer and, when scanning the volumes every 5 min, was recently ranked as one of the best radars in Europe for mesocyclone detection (van ’t Veen and Groenemeijer, 2025). Lightning data were purchased from the Meteorage company, which uses Vaisala LS7002 lightning detection sensors that are able to report both cloud-to-ground and cloud-to-cloud flashes with good accuracy (Erdmann et al., 2020). More technical details on lightning detection in this area can be found in Manzato et al. (2022b) and reference therein. Satellite data in different channels are from the EUMETSAT MSG dataset (see https://www.eumetsat.int/msg-services).

Prior to the event, the ECMWF IFS ERA5 reanalysis for 06 UTC 1 August 2021 showed a 500-hPa trough upstream, centered on France (Figure 3a). 500-hPa temperatures between − 12 and − 14 °C have overspread northeastern Italy. At 700 hPa, strong southwesterly flow associated with the trough helped transport moist Mediterranean air towards the Italian peninsula, resulting in large relative humidity near the Alps and Appenines mountains (Figure 3b). In northeastern Italy, winds showed consistent uplift associated with forcing for ascent from the approaching trough. The FVG region was within and on the cool side of a diffuse baroclinic zone associated with almost stationary cold front (not shown).

Figures 4a–c show Thetaplot diagrams (Morgan, 1992) for three consecutive soundings taken on 1 August 2021 from Aeronautica Militare in Udine–Rivolto, located approximately 86 km to the northeast of Azzano Decimo. Throughout the entire 18 − h period shown, the atmosphere over northeastern Italy was potentially unstable. CAPE increased from 821 J kg − 1 at 00 UTC to 3444 J kg − 1 by 06 UTC. This latter measurement ranks among the top 10 highest values ever observed at Udine (with a record dating back to 1992). The extremely large potential instability is supported by a low-level (0–2-km AGL) moist layer featuring Θ e up to 352 K, overlaid with low Θ e s in the mid levels. Indeed, the high Θ e values at low levels are also confirmed by surface observations at this time, which show significant advection of high- Θ e air from the Adriatic Sea (Supplementary Figure S3). Note that the exact Θ e values and computed CAPE must be considered with care, because part of the 06 UTC sounding was already contaminated by clouds; the profile is almost saturated with respect to liquid water above 750 hPa, meaning that convection in the vicinity was already ongoing at the time of sounding launch, at ∼ 05  UTC. The very low values of Θ e s in the saturated part of the sounding would increase the CAPE computed relative to a sounding not in cloud. For example, compare the large part of the Θ e s profile below 330 K at 06 UTC with the much smaller part having Θ e s < 330 K in the 00 UTC sounding. The decrease in 500 hPa temperature from − 11.6 °C to − 14.1 °C from 00 UTC to 06 UTC and consequent increase of instability can be partly due to this contamination effect, but there was also cold-air advection (from the south-southwest) in the mid troposphere (cf. Figure 3b).

Focusing on the wind profiles in Figure 4, we see that the 700-500-hPa layer has a mean wind of 21 m   s − 1 from about 210° (SSW) at both 00 and 06 UTC. On the other hand, the lowest 500 m AGL reveals a wind shift from 67° to 190° from 00 to 06 UTC. The significant veering of the low-level winds from east-northeasterly to south-southwesterly during this period marks the retreat of the surface frontal boundary and helps advect moist maritime air from the Adriatic sea inland, contributing to the increase in instability over this period. More generally, the lowest winds shown in Figure 4b suggest that the storm inflow was fed by maritime air. Additionally, by 06 UTC there is weaker shear, both in low levels (e.g., sfc − 3 -km bulk shear decreases from 14.7 to only 6.0 m s − 1 ) and in the 1 − 6 -km layer (bulk shear decreasing from 27 m s − 1 to 21 m s − 1 ), while the hodograph featured less turning at low levels. Both weaker low-level shear and “straighter” hodographs are thought to be favorable environmental conditions for hail production (e.g., Kumjian and Lombardo, 2020; Kumjian et al., 2021; Nixon et al., 2023; Lin and Kumjian, 2026).

Because of the partial convective contamination of the 06 UTC Udine sounding, we extracted an uncontaminated profile from the ERA5 reanalysis at 04 UTC, at the grid point closest to Azzano Decimo. (Surrounding gridpoints were also examined, and the profiles do not change considerably.) Using the radar-estimated storm motion (7.8 m s − 1 from 260°), we constructed a storm-relative hodograph (Figure 4d). The storm-relative hodograph exhibits numerous features thought to be conducive to hail production in supercell storms (e.g., Kumjian and Lombardo, 2020; Nixon et al., 2023; Lin and Kumjian, 2026), including being mostly straight, having weak sfc − 1 -km shear (4.3 m s − 1 ) paired with strong 1 − 6 -km shear (25.9 m s − 1 ). Further, additional shear in the 6–9 km layer could contribute to increasing updraft area (Warren et al., 2017), analogous to the increase in updraft area with increasing sfc-6-km shear (e.g., Dennis and Kumjian, 2017). Increased updraft area is known to be conducive to hail production by providing longer possible residence times for growing hailstones (e.g., Nelson, 1983; Kumjian and Lombardo, 2020; Kumjian et al., 2021; Lin and Kumjian, 2022; 2026). The subtle weakening of the midlevel storm-relative flow component perpendicular to the ∼ 2 − 6 -km shear vector has been recently shown to favor hail production by promoting weaker horizontal winds along a favorable hail trajectory within the updraft (Lombardo and Kumjian, 2025).

De Martin et al. (2025) recently found that the Udine soundings associated with the record-breaking hailstone on 24 July 2023 had exceptional profiles of water-Vapor Transport ( V T = V × q v , where V is horizontal wind speed and q v is vapor mixing ratio), well exceeding the summertime (June, July, August) climatological values from 2018-2023. Their statistical study showed that V T is particularly skilled in discriminating large hail. Hence, we investigate if the first August 2021 V T values were also particularly high. Figure 4e shows that the 00 and 12 UTC soundings had particularly high V T values between 0.1 and 4.5 km AGL, including > 0.1 m s − 1 between 2 and 4.5 km at 00 UTC and between 0.1 and 2.5 km at 12 UTC. Even the partially contaminated sounding of 06 UTC has V T values above the 90-percentile ( ∼ 0.08 m s − 1 ) of the updated JJA 2018–2024 distribution between 0.1 and 0.6 km (S-SW low-level jet feeding the storm inflow) and between 3.5 and 5 km AGL. However, the exceptional V T values in the mid-troposphere reached during 24 July 2023 were not equaled on 1 August 2021.

Of importance for hailstone growth is an understanding of the mixed-phase region within the cloud, which can differ substantially from the ambient environmental air in cases with extremely large instability such as this one. We can apply parcel theory to the 06 UTC sounding and obtain a rough estimate of the temperature profile experienced by the 30 hPa-high most unstable parcel, which is centered at 964 hPa ( ∼ 360  m AGL). Figure 5 shows the environmental profile (continuous line) and the lifted parcel profiles computed assuming 1) pseudoadiabatic ascent using the “Tv” method described in Manzato and Morgan (2003) (dashed), and 2) reversible adiabatic ascent (i.e., rising parcel retains all condensed water), using the “Tvc” method described in the same work (dotted). In both cases, the lifted parcel experiences temperatures < 0 °C above 5.2 km, whereas the homogeneous freezing point ( ∼ − 38 °C) is reached at 10.7 km and 11.4 km, respectively, for the “Tv” and “Tvc” methods. These altitudes are approximately 1-2 standard deviations above the mean most-unstable parcel glaciation level heights reported in Spychalla and Kumjian. (2025b) for a large dataset of significantly severe supercell storms in the U.S (Warren et al., 2021). Given that severe storm environments in Europe tend to be less extreme than those in the U.S. (e.g., Taszarek et al., 2020), these results suggest the potential for a particularly intense and deep storm with an in-cloud mixed-phase region likely is between about 5 and 11 km.

In this section we present pictures of the nine hailstones that were thin-sectioned. Each of the hailstones was sectioned from one of the remaining halves of the originals, and thinned to between about 0.5 and 1 mm and photographed with transmitted light (photos on the left in Figure 14; Supplementary Figures S4–S7) and between crossed polarizing filters (photos on the right in Figure 14; Supplementary Figures S4–S7). All of the thin-section prints are at the same scale. In the transmitted light photographs, the dark areas are air pockets filled with ground ice from the sawing; light areas represent transparent ice, and grayish areas represent opaque ice. In S-5 (Supplementary Material Figure S5), the band saw blade came off its wheels while completing the thin-section cut, breaking the hailstone, so the photo is incomplete. A similar mishap happened during the sectioning of S-2 (Figure 14, top) and C-2 (Supplementary Figure S6, bottom) but otherwise the outer edges of the thin-sections are largely faithful to the shapes of the stones as collected. The large hailstones are highly nonspherical with prominent lobes, which is consistent with past studies of hailstone shapes (e.g., Knight and Knight, 1970b; Knight, 1986; Knight and Knight, 2005; Shedd et al., 2021). In particular, notice the exaggerated growth of single lobes along the long axis of hailstone C3 (Supplementary Figure S7, bottom row). Although such structures have been observed before, it is large enough to be puzzled by its origin. However, there is a hint of other lateral lobes beginning to form in hailstones S1 (Figure 14) and Supplementary Figure S3 (Supplementary Figure S4 bottom), suggesting that it is not a fluke. Notably, other giant and gargantuan hailstones have exhibited prominent elongated lateral lobes (e.g., the 18-cm Villa Carlos Paz stone reported in Kumjian and Co-authors, 2020).

All of the hailstones exhibit substantial wet growth outer layers, which seems to be a common characteristic of large hail (e.g., Knight and Knight, 2005; Kumjian and Co-authors, 2020). For example, Soderholm and Kumjian (2023) found that, across a sample of thin sections of hailstones with varying sizes and shapes, 71 % had a final wet growth layer that contributed > 30 % to the total wet growth area. This suggests that the final growth stage of large hailstones occurs at higher temperatures, consistent with the descending part of their trajectories (e.g., Kumjian and Lombardo, 2020).

Nearly all of the hailstones analyzed exhibit some growth layers with approximately symmetric ring structures suggestive of “symmetric tumbling,” as opposed to random tumbling that would give rise to spherical symmetry (e.g., Knight and Knight, 1970a; Lin et al., 2024); C-4 (Supplementary Figure S7, top) is a particularly good example of such symmetric tumbling. The mirror-like symmetry of the internal structure along the major axis of the hailstone is suggestive of rapid spinning about its minor axis, which wobbles about the horizontal.

Also of note are the incomplete layering in hailstones S2 (Figure 14, top row, towards the hailstone’s left side), S4 (Supplementary Figure S4, top row, towards the hailstone’s top side), S5 (Supplementary Figure S5, in the middle-right and outer lower-left side), C2 (Supplementary Figure S6, bottom row, hailstone’s right side), and C4 (Supplementary Figure S7, top row, left and right sides of hailstone). The latter could arise from the “symmetric tumbling” described above; However, other asymmetries suggest changes to the tumbling behavior that lead to preferential growth only on certain sides of the hailstones. In other words, hailstone falling behaviors can be highly complex (see the discussion in Lin et al., 2024).

The photographs of thin slices through crossed polarizing filters reveals the crystal fabric structure of hailstones (e.g., Knight and Knight, 1968). Often, there are large crystals found in the central parts of hailstones, as we have seen in other large hail from previous analyses (e.g., Knight and Knight, 2005). The thin sections show quite extensive recrystallization except in the bubbly layer–the small air bubbles in such layers inhibit grain boundary migration, which is the mechanism of grain coarsening. In particular, large single-crystal centers in C4 (Supplementary Figure S7); S1, S2 (Figure 14) are notable.

Hailstone sampling involves cutting small pieces from each hailstone growth layer, melting each piece, and hermetically sealing it in a vial. The sampling was performed from the outside in. For each sample, a short note was written to describe where it was taken from the hailstone thick slice; these notes were later used to build the “composite” images shown in Figures 15, 16. A total of 28 samples were taken for the four “C” hailstones, and a total of 37 sampling were taken for the five “S” hailstones. Then, the 65 vials were then shipped to the Ice cores, Isotopic and Environmental Geochemistry Laboratory of the Ca’ Foscari University of Venice to analyze their water isotopes with a Picarro cavity ringdown spectrometer ² .

Figures 15, 16 show the isotope ratios collocated with the approximate area along the hailstone cross section where the samples were extracted, scaled to cover the extent of the sampled fractions. Dotted lines refer to samples with possible evaporation issues (discussed below). There exists inter-hailstone variability in their radial isotopic profiles. For example, C2 (Figure 15, top right) does not show much variability in δ 2 H and δ 18 O values in its different layers, whereas some (e.g., C3; Figure 15, bottom left) reveal considerable fluctuations. Where radial variation exists, generally, local minima (i.e., more negative δ 2 H and δ 18 O values) are found in the opaque ice, dry-growth layers, with greater values in the clearer, wet-growth ice. In addition, the hailstone centers generally exhibit less negative δ values than the adjacent fractions (e.g., Supplementary Figures S3,S4 in Figure 16, middle row). However, exceptions exist: Supplementary Figure S5 (Figure 16, bottom row) has somewhat greater δ 2 H and δ 18 O values in its outer wet-growth layer compared to its center.

The meteoric water lines describe the linear relationship between δ 18 O and δ 2 H in precipitation. On a global scale, this relationship is represented by the Global Meteoric Water Line (GMWL; Craig, 1961), which exhibits an approximate 8:1 slope between δ 18 O and δ 2 H, reflecting the relative proportions expected from equilibrium fractionation processes. The Local Meteoric Water Line (LMWL) derived for precipitation in Friuli Venezia Giulia for the period 1984–2015 (1,365 single sample, R 2 = 0.98) (Masiol et al., 2021) is given by Equation 1: δ 2 H = 7.8   C I   7.7 ; 7.9 ⋅ δ 18 O + 8.9   C I   8.3 ; 9.4 (1) with 95th percentile confidence intervals (CI) of Equation 1 calculated by ordinary nonparametric bootstrap resampling over 2000 replicates. The hail samples collected in this study plot almost exactly along the LMWL (Figure 17a), indicating that the formation of the hailstones occurred under near-equilibrium conditions. This interpretation is further supported by the deuterium excess or “d-excess” ( d = δ 2 H − 8 δ 18 O ) values of the hailstone layers near the intercept value of the LMWL (Figures 15, 16). This also suggests the hailstones comprise water from a single source (almost certainly the Adriatic Sea, given the meteorological set up and the low-level inflow described in Section 3.1).

Nonetheless, several layers display markedly lower d-excess values, falling well below the LMWL line (Figure 17a) and suggesting partial evaporation effects during their formation. Some of those layers show very low d-excess values ( ≪ 0‰) corresponding to remarkably high δ 2 H and δ 18 O values. The extremely low d-excess values are never located in the superficial layers but occur in the interior of the hailstones, lying between layers with normal values. This spatial pattern argues against post-sampling evaporation, metamorphism, or contamination, which would be expected to affect the outermost layers first and produce more systematic shifts in d-excess rather than isolated anomalies confined to the interior. Instead, these unusually low values most likely reflect specific conditions or processes acting during hailstone growth, rather than artifacts introduced after sampling.

Physically, d-excess is largely conserved during condensation but is highly sensitive to kinetic fractionation during evaporation, especially under dry, non-equilibrium conditions (Froehlich et al., 2002). In modern precipitation, very low or even negative d-excess values are typically associated with strong evaporation of raindrops or cloud droplets as they encounter warm, unsaturated air, where kinetic effects preferentially remove lighter isotopologues and drive the remaining vapor and condensate off the meteoric water line.

The combination of very low d-excess and relatively high δ 2 H and δ 18 O values may suggest episodes of intense kinetic fractionation, perhaps occurring when liquid water on the growing hailstone surface underwent evaporation before being incorporated into the ice structure (see discussion in Section 2.2.2). Despite growing in cloudy conditions, hailstones undergoing wet growth have ∼ 0 °C surface temperatures that can be much greater than the ambient environment, and thus the ambient vapor pressure is below what is needed to maintain that liquid in equilibrium, leading to evaporation. Further, this partially evaporated liquid water can also seep into porous rime ice and subsequently freeze, which may account for the opaque ice (dry growth) low d-excess layers. In addition, collection and subsequent freezing of raindrops (e.g., Kumjian and Lombardo, 2020) could lead to such low d-excess layers, particularly if the raindrops’ journeys included sub-saturated conditions prior to being lofted into the hail growth region (Kumjian et al., 2014). However, even without considering those points far from the GMWL, there is a large variability of the isotopic values, with deuterium ranging from about − 80 ‰ to − 40 ‰ δ 2 H of V-SMOW (Vienna Standard Mean Ocean Water, the global standard, i.e., the reference for isotopic analysis).

Application of the “Jouzel” model (Jouzel et al., 1975; Jouzel et al., 1985), which relates a growth layer’s isotope content to a specific atmospheric level (temperature), can provide at least qualitative insights into the hailstone’s growth history in the mixed-phase region of the cloud. However, application of the model with the large isotope variability, including outlier points, causes the model to fail to converge. Given the kinetic fractionation issues (which invalidate the main assumptions of the Jouzel model outlined in Section 2.2.2), we removed the data from hailstones C1 and C4 (which had the most extreme outlier points) before applying the Jouzel model. Doing so allowed the model to converge, and the resulting profiles are shown in Figure 17b.

From this isotope analysis, one can infer that most of the hail growth seems to have occurred in the upper portions of the cloud, i.e., between 8 and 10 km, at temperatures between − 16 and − 31 °C. That result is consistent with the temperature profile and the estimate of homogeneous freezing level above 11 km made applying the Lifted Parcel Theory (cf. Figure 5), and with the high radar reflectivity core ( > 55 dBZ) – a threshold indicative of hail–extending at least up to 10 km (Figure 9a), which corresponds well to the maximum growth height estimated from the isotopic analysis, as shown in Figure 9b. Thus, we infer that these particularly large hailstones have a relatively large residence time at high altitudes, which is consistent with the large vertical extent of high-reflectivity echoes discussed in Sections 3, 3.3. Further, the hailstones generally follow paths without multiple large vertical excursions, which is seemingly consistent with the simple, arching trajectories suggested by some modeling work (e.g., Dennis and Kumjian, 2017; Kumjian and Lombardo, 2020; Lin and Kumjian, 2022; Spychalla and Kumjian 2025b) and recent observations from Hailsondes (Soderholm et al., 2025). These results are also consistent with a similar analysis performed for hailstones collected in China (Lin et al., 2025).

Isotope analysis from the hailstone embryos (layer A in Figure 17b) show a wide range of formation temperatures from just below 0 °C to about − 25 °C. Interestingly, the hailstones starting below 7 km reached a final size ≤ 7.5 cm, while those starting above 8 km reached a final size ≥ 7.0 cm. Hence, there is a moderate ( R = − 0.6 ) negative correlation between final hailstone size and embryo origination height, suggesting that the largest hailstones from this sample entered the updraft at lower altitudes and began their growth on ascent. Though the sample of hailstones is small and the size range is narrow, this finding seems consistent with results from a recent idealized modeling study by Spychalla and Kumjian. (2025a), who found that embryos initialized near the bottom of the hail growth zone had the greatest probability of producing the largest hail.

Lastly, an insoluble particle analysis were performed on two of these hailstones, but since they did not contain the embryos anymore (these were cut away during the isotope analysis) the results are shown only in the Supplementary Material (Section 1.2; Supplementary Figure S8).