Figure  depicts a first comparison between observations and the reanalyses. Figure a shows the Terra/MODIS corrected reflectance from NASA Worldview for 1 April 2022 . From the satellite imagery, it becomes clear that the Arctic sea ice northwest of Svalbard features many leads on different length scales. However, the MIZ is rather sharp, and the transition from closed sea ice to ice-free ocean water typically occurs within less than 1 km distance. Over the ice-free ocean, cloud streets due to roll convection are evident. The cloud streets form along the prevailing wind direction. Furthermore, a clear lee effect due to Svalbard's mountain ranges is seen to the west of the archipelago.

Figure b shows the corresponding fields as represented by ERA5 at 12:00 UTC noon. Due to its coarse spatial resolution, no leads are modeled in the SIC data fields, and the MIZ width is on a length scale of approximately 80 km. This is a typical MIZ width for ERA5 . Instead of cloud streets, a stratiform liquid and ice containing cloud deck is simulated, which thickens in off-ice direction. Clouds are partly also already formed over closed sea ice. In contrast, clouds in CARRA are exclusively formed over the ice-free ocean; see Fig. c. The high spatial resolution allows convection to be modeled. As a result, several distinct cloud streets are reproduced. In addition, CARRA better reproduces the sharp MIZ, which here is on the scale of around 10 km. The sharper MIZ of CARRA in comparison to ERA5 is not only a matter of spatial resolution (2.5 km for CARRA versus 30 km for ERA5). The sea-ice concentrations in ERA5 are derived from the Operational Sea Surface Temperature and Ice Analysis dataset, produced by the UK Met Office OSTIA; . OSTIA outputs daily sea-surface temperature and sea-ice concentration fields based on satellite observations, with a native resolution of 0.05°×0.05° (roughly 6 km). Yet the sea-ice data are based on the EUMETSAT OSI-SAF 401 dataset utilizing 19 GHz and 37 GHz microwave channels at along-track resolutions of coarse 69 and 37 km . On the contrary, CARRA strongly relies on the European Space Agency's Sea Ice Climate Change Initiative product SICCI; with a native resolution of 15–25 km. These data are additionally filtered based on the high-resolution sea-surface temperature fields and then regridded to the CARRA grid . Several authors noted that improved sea-ice and MIZ representation crucially improve the performance of models in the lower-tropospheric layers . As will be shown later, the magnitude of turbulent heat fluxes is directly correlated to the distribution of sea-ice versus ice-free ocean, which is the primary driver of CAO transformations. Errors in MIZ width can have significant downstream effects over several hundreds of kilometers .

Figure a shows the profiles of air temperature from observations. Over sea ice, clear temperature inversions are found. The coldest near-surface temperatures reach -27 °C, and the thickness of the inversions is around 0.6–0.9 km. As air masses spend more time over ice-free waters, they become warmer near the surface, leading to stronger coupled ABLs and the development of a typical marine stratification. This is accompanied by a steady, linear increase in the calculated BLHs and closely correlated CTHs.

By linearly interpolating all data to 100 m vertical resolution, the temperature differences ΔT between ERA5/CARRA and the observations are computed. The results are shown in Fig. b and c. Despite the reanalysis assimilating all the employed dropsondes, the ERA5 profiles show a distinct warm bias in near-surface air temperatures of mean 2 K over Arctic sea ice. Many authors reported on similar warm biases of skin and near-surface air temperatures in ERA5 . The skin temperatures are generally considered too warm as an insulating layer of snow is missing atop the floating ice, which can introduce surplus heat into the lower atmosphere . The surface warm bias turns to a mean cold bias of -1 K at altitudes of 0.25–0.50 km. Over the ocean, the mean temperature bias is much lower and reaches -0.5 K at around 1 km altitude. In the CARRA data, the near-surface temperature bias is reduced to an average of 1 K. Similar improvements over ERA5 have been reported by others . However, CARRA also faces challenges in accurately representing temperature inversions. This is reflected in the cold bias of around -1.5 K at altitudes of 0.20–0.40 km.

The mean absolute errors (MAEs) of ERA5 and CARRA with regards to measurements are computed. Output from both reanalyses as well as dropsondes is interpolated to a common vertical coordinate of altitude above ground in 10 m steps. To evaluate especially the crucial ABL representation, MAEs are averaged vertically from the surface up to the observation-derived BLHs, plus an additional 200 m margin to capture the dipole pattern of errors. Table  summarizes the results separately for dropsondes released over sea ice and ice-free ocean. For air temperature over ice, CARRA shows a slightly smaller MAE of 1.00 K versus 1.14 K for ERA5. Over the ice-free waters of Fram Strait, these errors are significantly reduced in both products, yielding a MAE of 0.39 K in CARRA and 0.44 K in ERA5.

Next, the vertical profiles of specific humidity are examined. Figure a depicts the observed profiles, as extracted from the dropsonde measurements. Over sea ice, a uniform and dry ABL is found, where maximum values of around 0.6 gkg-1 are measured. Near-surface layers are the driest, at around 0.4 gkg-1. The longer the air masses reside over the sea, the more water vapor is picked up by the lower air mass layers through evaporation from the ocean surface.

Over sea ice, ERA5 shows a mean near-surface moist bias of 0.05 gkg-1 (Fig. b), as well as a slight dry bias close to the BLHs. Once air masses drift over the sea, a strong dry bias is found throughout the ABL. It increases over time and reaches down to -0.5 gkg-1, which corresponds to about 30 % of the observed values. CARRA shows different patterns (Fig. c). Over the closed ice pack, the lowest 0.2 km shows a negligible humidity bias. However, in higher layers above 0.5 km, a slight moist bias is seen. During the off-ice drift, at first a slight moist and later dry bias becomes obvious; however, this is much smaller compared to the ERA5 reanalysis. The same patterns are found in the quantified MAEs within the ABLs; see again Table . Notably, over the ice-free ocean, CARRA's MAE of 0.069 gkg-1 is significantly lower than ERA5's MAE of 0.112 gkg-1.

Figure  gives an overview of the quasi-Lagrangian matches calculated with reference to the dropsondes. All matches are colored by the time air masses spent over ice-free ocean. As described in the Methods (Sect. ), these approximately 24 200 matches are a function of height above ground because not only are the zonal and meridional winds height-dependent, but also the vertical velocity is used for the three-dimensional trajectory calculations. This allows air masses to ascend or descend along their horizontal flow. The matches cover 150 km along the prevailing wind direction over the Arctic sea ice and about 200 km along the CAO evolution over ice-free ocean.

Naturally, the question arises of how reliable the trajectory calculations presented here are. In previous studies, sometimes additional criteria were applied to prove the reliability of trajectories. These are similar hydrocarbon fingerprints between matches or an inert perfluoromethylcyclopentane tracer being deployed . However, such a use of tracers is only possible in case of in situ sampling. In the CAO presented here, the assimilation of the high-density grid of dropsondes serves as crucial input for ERA5 and CARRA. This can also be seen in the comparison of wind profiles as shown in Figs. S3 and S4. With the exception of the nearest-surface layers, a close match is seen between dropsondes and both reanalyses. Also, as trajectories are only calculated over short spatiotemporal scales (on the order of 1–4 h, 50–200 km), small errors can not add up as much. For some research questions, it might be more valuable to investigate transformations over larger spatiotemporal scales, such as was done, for example, for aerosol and hydrocarbon species . However, as will be demonstrated, the highly important thermodynamic evolution occurs on timescales of a few hours and below. If too much time passes between two matching observations, the “net” rates, e.g. of (Δθ/Δt)net and (Δq/Δt)net introduced in Eqs. () and (), would smooth out short-lived effects even more; models would be increasingly needed to disentangle the net rates calculated over longer time frames into sections of more or less intense transformations. Finally, the approach presented here of initializing and then registering matches for a large number of trajectories within a radius of 20 km around HALO's location is also essential to better assess the statistical significance of matches. Notably, all deviations seen between the aforementioned observed and modeled wind profiles result in an error of less than this 20 km radius over 1–3 h of drift.

The evolution of thermodynamic properties in the form of heating and moistening rates is analyzed in a quasi-Lagrangian framework, grouped by time air masses spent over ice-free ocean.

First, the evolution of diabatic heating rates is analyzed. Figure a shows the vertically resolved diabatic heating rates based on the quasi-Lagrangian dropsonde observations. No significant differences exist between using ERA5 or CARRA winds as input. The MAE between ERA5 and CARRA heating rates in the lowest 1.5 km above ground and averaged for all times over ice-free ocean amounts to only 0.13 Kh-1. Note that air parcels of the first category (tocean=0–1 h) are drifting almost exclusively over sea ice and leads (Fig. ). For these air masses, a maximum near-surface warming of around 1.8 Kh-1 is found. This possibly stems from some of the leads crossed by the trajectories. However, the heat is contained within the very shallow ABL, and all heating rates above the BLHs of around 0.30 km are around 0 Kh-1. After crossing the MIZ and reaching the ice-free ocean (tocean=1–2 h), a very intense surface warming is seen, where values larger than 6 Kh-1 are found. This heating is starting to be mixed upwards into the increasingly deep ABL, which reaches BLHs of around 0.70 km. After the initial rapid exposure of the cold and dry Arctic air masses to the much warmer ocean surface, the heating in the lowest layers declines rapidly and stays around 2 Kh-1. The vertical mixing is now dominating and leads to almost homogeneous mixing within the lowest 0.75 km of the troposphere. Interestingly, some layers above show regions with negative heating rates, i.e., a net cooling of air masses at altitudes around the BLHs. An analysis of ERA5 temperature tendencies indicates that this is not a sign of a net cloud-top radiative cooling effect but instead of the mixing of colder near-surface air with the original, overlying warmer air (Fig. S5). Previous airborne studies have shown that in the early stages of CAOs, the heat budget has several sources. Notably, the strong surface heat fluxes are reinforced by entrainment fluxes with the warmer inversion aloft. The relative contributions of the heat sources change with distance from the sea-ice edge. Over leads in the MIZ, entrainment heat fluxes exceeding 30 % of the surface heat fluxes have been observed in single cases . Shortly after passing the sea-ice edge, this ratio can initially increase to 80 %, which then again decreases to 30 % for fetches over 150 km . In the region of deep, cellular convection, condensation can even become the dominating contributor to air-mass heating . Overall, these previous studies indicate that the diabatic cooling near cloud tops corresponds to a warming below cloud tops, caused by the entrainment fluxes.

The ERA5-derived heating rates reflect the general features of the observations (Fig. b). This can be attributed to the assimilation of all 40 dropsondes into ERA5. However, some important differences are found. Over the Arctic sea ice, ERA5 shows BLHs almost twice as high as that seen in observations. While the slight surface warming of around 1.6–1.8 Kh-1 is also seen in ERA5, it shows excess heat that it mixes upwards towards the BLH. The intense surface warming at tocean=1–2 h is not represented in ERA5. This is in agreement with the sea-ice distribution shown on the overview map in Fig. b, which revealed a wide MIZ on the order of 80 km – much wider than that shown in the observations. As a result, the initial stage of the CAO is delayed in ERA5. The later stages (tocean>2 h) of the CAO, however, are represented rather well, yet again with an exaggerated vertical mixing. The essential feature of negative heating rates in higher altitudes is captured.

Figure c shows the heating rates extracted from the CARRA product. Generally, these settle in between the observations and ERA5. All BLHs are lower than in ERA5 and significantly closer to the observed values. For tocean=1–2 h, the observed intense warming rate higher than 6 Kh-1 is also not represented fully, yet it is much better than in ERA5. A maximum value for the near-surface heating of around 4 Kh-1 is found, which is homogeneously mixed upwards up to 0.40 km altitude. This might in part be caused by the much sharper MIZ, which is closer to reality (Fig. c).

In the early stage of CAOs, the primary source for turbulent heat fluxes is the warm ocean surface . To study the evolution of the heating profiles along the CAO, it is thus essential to investigate the surface sensible heat fluxes (SSHFs). A comparison between the observation-derived and reanalysis-based SSHFs is given in Fig. a. All data are grouped by time spent over ice-free ocean. Over the Arctic sea ice and MIZ, the observation-derived SSHFs show a mean of below 50 Wm-2, with the 95th percentile peaking above 150 Wm-2. However, as was outlined in Sect. , the computation of turbulent heat fluxes over the sea ice and MIZ based on dropsondes is prone to high uncertainty. Both reanalyses show higher values, especially ERA5. For tocean=0–1 h, this corresponds to heating rates in both reanalyses being slightly too large and a mixing that is exaggerated.

After the air masses cross the MIZ, large values of SSHF of around 520 Wm-2 are observed. Such values are typical in CAOs . ERA5 significantly underestimates the observed SSHFs by at least 130 Wm-2, while CARRA slightly exaggerates it. Later into the CAO, the observed SSHFs drop slightly yet are again best captured by CARRA. In general, the reduction of SSHFs over time is expected, as the temperature difference between the sea surface and the overlying air is reducing. This can potentially be counterbalanced, for example, by increasing underlying sea-surface temperatures, increased winds, or decreased surface roughness .

The different parameters that are required for the calculations of SSHF as shown in Eq. () are investigated in Fig. S6. Notably, over ocean both reanalyses significantly underestimate U10m, with CARRA being always closer to the observations. The horizontal thermal gradient between sea ice and the ice-free water surface causes a marked off-ice breeze, an analogue to sea–land breezes. Similar to that reported by , in our case, U10m reached its maximum near the sea-ice edge, and the off-ice acceleration due to thermal contrasts is estimated to be around 2.6 ms-1h-1 (calculations can be found in the Appendix). Therefore, the U10m in CARRA might be closer to observations than ERA5 as (i) the MIZ is narrower in CARRA and (ii) the discussed near-surface warm bias over sea ice is weaker in CARRA. Previous studies have also found ERA5 underestimates the highest near-surface winds over the ocean next to the MIZ, as well as SSHFs and SLHFs over the MIZ . Feeding coarse-resolution sea-ice data (with a MIZ of around 80 km, such as in ERA5) into higher-resolution models was also found to smear out the rapid increases in air temperature, wind speed, and surface fluxes .

In order to evaluate whether the differences between CARRA and ERA5 discovered for 1 April 2022 are of a systematic nature, a climatological comparison of SSHFs from both reanalyses for 1991–2022 can be found in Fig. S8. It shows that during CAO conditions, CARRA SSHFs are systematically larger than ERA5 SSHFs, and this is consistent over several decades. It is especially pronounced over ocean and corroborates our results for the case study of 1 April 2022. Similar systematic differences in the output surface turbulent heat fluxes have been reported also for comparisons of other reanalyses . Underestimated fluxes result in too-low uptake rates for heat and moisture, particularly close to the ice edge . However, similar studies like are required for a deeper systematic evaluation of ERA5 versus CARRA, for example, to disentangle the combined effects on U10m as caused by MIZ width, parameterized surface roughness, or synoptic patterns.

Figure a shows the vertically resolved moistening rates based on the quasi-Lagrangian dropsonde observations. The MAE on the observed moisture uptake rates based on ERA5 versus CARRA trajectories as input is very low at 0.01 gkg-1. For the air masses mostly sampled over sea ice, only minimal moisture uptake is found. The highest uptake at tocean=1–2 h reaches around 0.4 gkg-1h-1 at the surface. For longer times over the ice-free ocean, this moisture is then quickly mixed upwards. The magnitude of upward mixing partly exceeds the moisture uptake near the surface at later stages.

Figure b shows the corresponding rates as extracted from ERA5. ERA5 underestimates the near-surface moistening rates significantly; also layers further up show rates which are 2–3 times too low. CARRA performs better than ERA5 (Fig. c). Not only are the near-surface moistening rates closer to observations, but also the upward mixing is more realistic. An insufficient moistening rate within the lower troposphere during a CAO can be caused by (i) an insufficient supply of moisture from the surface, i.e., too-low SLHF, and/or (ii) an exaggerated removal of water vapor from the atmospheric column. Here, we check both factors separately.

Figure b compares the SLHFs between observations and the two reanalyses. Over sea ice, very low SLHFs with a median below 25 Wm-2 are found. Over the ocean, both ERA5 and CARRA underestimate the SLHF, and they are close to each other. Surprisingly, ERA5 always predicts slightly higher SLHFs than CARRA, which at first seems not to agree with the much lower moistening rates. Also the climatological comparison under CAO conditions shows that ERA5 SLHFs exhibit a constant bias towards larger values than in CARRA (Fig. S8). Overall, these findings hint towards mechanisms in ERA5 leading to an exaggerated removal of water vapor, namely cloud processes and precipitation. This is investigated in the next section.

To help understand possible errors in the reanalyses connected to cloud physics, HAMP radar reflectivities from aboard HALO as well as in situ Nevzorov measurements of cloud liquid water and ice contents by the Polar 6 aircraft are utilized . A deeper investigation of cloud microphysical processes, such as riming, precipitation formation, or cloud street aspect ratios, is outside the scope of this article. However, details on these processes specifically including the CAO on 1 April 2022 are reported by and .

Figure a shows the measured HAMP radar reflectivity profiles averaged for 2 min around the time of each dropsonde release and up to 6 h into the CAO. The profiles are plotted as function of time air masses spent over ice-free ocean. For the locations over sea ice, either very low (radar reflectivity below -25 dBz) or no radar signals were recorded. As the air moves onto open waters, the radar reflectivities increase, and cloud tops are seen to rise linearly. Radar reflectivities for the first time cross the threshold of -5 dBz only for tocean>1 h. This indicates the presence of precipitation .

As Polar 6 has a range much lower than HALO, it was able to only sample the first 3 h of the CAO. Figure b–d shows the height-resolved measured specific cloud liquid water contents qliq. Over the closed sea ice (Fig. b), in the sampled lowest 0.6 km above ground, no cloud liquid water was found, yet with low amounts at the top of the ABLs. After the drift across the MIZ, noticeable amounts of cloud water of up to 0.07 gkg-1 are seen up to around 1  km altitude, which corresponds to the altitude of moisture uptake. Surprisingly, the liquid water content decreases in the next time step. This might be correlated with an increase of the frozen hydrometeors (i.e., cloud ice and snow) depicted in Fig. S7a. Several in situ probes confirm the occurrence of riming during the flight of Polar 6 . With the air temperatures always in the range of -25 to 0 °C (see Fig. ), mixed-phase clouds are possible, and also the typical pattern of a supercooled layer above the ice layer is reproduced.

Figure c shows the cloud structures as reproduced by ERA5. ERA5 tends to overestimate the amount of liquid water present in the clouds. A similar enhanced abundance of liquid-bearing clouds especially over sea ice has been reported for the IFS, the model behind ERA5 . In the CAO case here, this is in contrast to CARRA (Fig. d). With the exception of missing the strong increase in liquid clouds at tocean=1–2 h, CARRA matches the observations well. The overabundance of liquid-bearing clouds in ERA5 might explain the exaggerated vertical mixing of heat in ERA5. Even though ERA5 featured lower SSHFs compared to CARRA, the more pronounced cloud tops might lead to stronger entrainment fluxes. Also, the condensation into cloud droplets releases additional heat to the surrounding air masses.

Furthermore, total precipitation at the surface is much higher in ERA5 than in CARRA, which creates an additional sink for atmospheric moisture already over sea ice (Fig. ). Precipitation over sea ice is unlikely based on the presented HAMP radar reflectivities. The -5 dBz threshold is only exceeded at tocean>1 h. This threshold translates to a precipitation rate of around 0.02–0.09 mmh-1 . The statistical comparison between ERA5 and CARRA presented in Fig. S8 substantiates the finding of this case study. ERA5 has a strong bias to form liquid-bearing clouds already over sea ice and the MIZ. Over the ocean, there also is a strong bias towards higher cloud liquid hydrometeor contents. The ERA5 clouds systematically precipitate more strongly over the MIZ and ocean than clouds in CARRA. The significance of our findings is reinforced by , who showed that issues such as an overabundance of low, liquid-bearing clouds can propagate into higher-resolution models through large-scale forcings.