The HALO–(AC)3 field campaign took place in the Norwegian/Greenland Seas and the Fram Strait from March to April 2022. The main objective of this airborne campaign was to understand the transformation of air masses between the Arctic and the mid-latitudes and to investigate the influence of these interactions on ABL processes and cloud formation. To this end, a quasi-Lagrangian sampling strategy was applied during the campaign, in which the transformation of individual air parcels of an air mass was tracked along their trajectories using three research aircraft flying at different altitudes (HALO, Polar 5, and Polar 6) . An overview of the synoptic situation during the HALO–(AC)3 campaign is provided by .

This study focuses on a case study analysing two HALO Research Flights 10 and 11 (RF10 and RF11) conducted on 29 and 30 March 2022 (Fig. ). On these 2 d, a weak MCAO with an MCAO index of four was investigated (Appendix ). Both days were characterised by a persistent northerly wind of about 3 to 10 ms⁻¹ measured at 10 m altitude by dropsonde observations. These characteristics indicate a weak MCAO event compared with climatological values from the Fram Strait . The slowly moving air mass was probed by HALO during these 2 d. The observed MCAO case embedded within a prolonged period of extensive MCAO activity from 21 March to 12 April 2022, between two intervals of particularly strong conditions on 21–26 March and 1–2 April 2022 .

The synoptic situation in the Fram Strait was characterised by a high-pressure system centred at the east coast of Greenland and a low-pressure area in the Barents Sea (Fig. ). The cloud field was dominated by shallow convective mixed-phase clouds, with obvious street features near the sea ice edge. These cloud streets broke up further downstream in the area of C03 (Fig. ). A notable cloud-free area was located on the west coast of Svalbard (Fig. ), which can be attributed to the Foehn effect during easterly flow over Svalbard, similarly observed during other field campaigns e.g.,.

The HALO flight plans on these 2 d aimed at tracking the location of an air mass sampled on the first day at a location in the high north above sea ice using trajectory estimates based on forecast data. At four locations along the trajectory, the air mass was sampled using a circular mesoscale dropsonde pattern, with a diameter of 150 km. This technique is inspired by and has been applied in the North Atlantic and subtropical region by to measure mesoscale divergence and subsidence profiles. The centre of the initial circle (IC) (Fig. ) on the first day at 14:00 UTC was located at (84.4° N, 13.6° E) and featured only five dropsondes, a number limited by airspace restrictions. On the second day, three further circular patterns were flown along the predicted southern trajectory of the air mass. The circles sampled during the two subsequent days featured eight evenly distributed dropsondes, with two additional dropsondes launched in their inner areas. Only one of the day-two circles, (C02) at (78.6° N, 4.0° E, at 11:30 UTC), was located at the exact predicted location of the air mass as sampled on the previous day at the IC location. As shown in Fig. , this circle was situated over open ocean. Two further circles were flown on day two, one (C01) further upstream over sea ice (80.8° N, 10.7° E, at 10:30 UTC), and another (C03) was flown at a downstream location at a similar distance (76.3° N, 3.3° E, at 12:30 UTC). The Polar 5 (P5) aircraft took additional measurements in the C02 area, shortly after the departure of the HALO aircraft from this area (at 12:00 UTC), using a rectangular flight pattern with three cross-flow legs. Additional measurements were conducted by a controlled meteorological balloon (CMET) launched from Ny-Ålesund, as part of the ISLAS2022 field experiment , which took place at the same time as the HALO–(AC)3 campaign.

A combination of in-situ and remote sensing observations from the HALO–(AC)3 campaign were used to constrain the LES experiments and evaluate the model performance. A comprehensive overview of the measurements collected during HALO–(AC)3 is described by , and a summary of the datasets used is presented in Table , including the variable names, associated instruments, and references to more detailed technical and scientific descriptions.

The deployed Vaisala RD-41 dropsondes provide vertical profile measurements of the thermodynamic state, including air temperature at a resolution of 0.01 K, relative humidity at a resolution of 0.1 % and air pressure at a resolution of 0.01 h Pa at frequencies between 2 and 4 Hz and descend at speeds between 10 and 20 m s⁻¹ . The measured profiles are interpolated to a uniform vertical grid of 5 m spacing. The horizontal wind speed profiles were determined from GPS data with an estimated accuracy of 0.1 m s⁻¹ . These observations were used to derive vertical profiles of subsidence (Fig. ), geostrophic wind, and advective tendencies of temperature, humidity, and wind through a regression algorithm applied to the dropsondes deployed in each circle. A comprehensive description of the calculation technique is provided by .

To constrain the surface skin temperature Tskin, airborne measurements of the thermal-infrared imager on board HALO VELOX Video airbornE Longwave Observations within siX channels; were used, which records 2-D fields (640×512 px) of the thermal infrared brightness temperature. At typical HALO flight altitudes, this corresponds to a horizontal resolution of 10 m with a temporal resolution of 1 Hz for a field-of-view of 6.4×5.1 km. From these brightness temperature images, the skin temperature Tskin and sea-ice concentration were derived along the circular flight legs of HALO with estimated measurement uncertainties of 1.2 K and 5 %, respectively . For partly cloudy open-ocean segments in C02 and C03, we apply a cloud mask and classify pixels whose retrieved values of Tskin exceed -1.8 °C above the surrounding (cloudy) background as open ocean. The final surface skin temperature is the horizontal mean of each masked 2D-field at a 1 Hz time resolution.

The datasets collected during the HALO–(AC)3 campaign provide all atmospheric state variables required to run Lagrangian LES experiments following the air mass as it moves south. A small number of surface parameters, such as the surface albedo, the aerodynamic roughness length for momentum, and the thermal roughness length for heat, could not be measured, but are needed to prescribe the (latitude-dependent) surface boundary condition in the LES experiments. Therefore, these data were retrieved from the European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis product version 5 ERA5;. By combining the dropsonde and VELOX observations with the surface values from ERA5, a complete set of forcing fields for the Lagrangian LES has been compiled.

For model validation, independent datasets obtained from the HALO–(AC)3 campaign are used. The Airborne Mobile Aerosol Lidar (AMALi) on the P5 aircraft provides measurements of cloud top height, which are compared to the simulated cloud top heights . The liquid water and ice water paths are derived from the Microwave Radar/Radiometer for Arctic Clouds (MiRAC) on board P5 . The ice water path is derived from the 94 GHz-radar reflectivity using the Ze–IWC relationship of . We additionally include integrated water vapour (IWV) and liquid water path (LWP) data retrieved from the HALO Microwave Package (HAMP) over open ocean (not available over sea ice) . These data allow a direct comparison with the quantities computed by the LES.

Further thermodynamic observations were performed using Lagrangian CMET soundings . The balloon was equipped with an air temperature sensor that operates over a range of -40 to 105 °C with an interchangeability of ±0.5 °C and measured relative humidity with respect to water with an accuracy of 1.5 %. Due to the cold, dry conditions, and the absence of active ventilation of the sensor package on the CMET, actual measurement errors would be substantially larger than the sensor interchangability, in particular during daytime when the balloon was drifting horizontally, creating an overall warm and dry bias. This particular CMET-04 was launched in Ny-Ålesund on 30 March 2022 at 14:00 UTC following a trajectory parallel to the simulated trajectory (Fig. ), complementing the dropsonde observations and providing additional in-situ validation data within the lower ABL.

The simulations were performed using the Dutch Atmospheric Large-Eddy Simulation (DALES) , which is an open-source model available at https://github.com/dalesteam/dales (last access: 2 October 2026). DALES has been tested and applied extensively in previous studies of ABL turbulence and clouds in various regions of the Earth , including the Arctic .

To represent Arctic mixed-phase clouds, the DALES code was equipped with the double-moment microphysics scheme of , which is predictive of both the mass and number concentrations of five hydrometeor classes (cloud water, cloud ice, rain, snow, and graupel) . The number concentration of cloud condensation nuclei (CCN) is prognostic, initialised with a constant vertical profile of 100 cm⁻³ within the range observed during the HALO–(AC)3 campaign in the Fram Strait and in the high Arctic during MOSAiC . The available INP number concentration is prescribed with an upper limit of 1×10-3 cm⁻³, following the setup of and , while the activated fraction depends on temperature and ice supersaturation through the parametrisation. Heterogeneous freezing processes are limited to temperatures below -15 °C, and the maximum number of ice particles resulting from deposition nucleation is restricted to 200 L⁻¹, while homogeneous freezing and secondary ice formation via the Hallett-Mossop process remain unchanged as represented in the model. This microphysical model setup was thoroughly tested by against a year of in-situ measurements of Arctic clouds in the central Arctic during the MOSAiC expedition.

Further model settings follow the setup described by . They include a radiative transfer scheme interactive with liquid and ice hydrometeors applying a four-stream radiative transfer solver in combination with Monte Carlo spectral integration . The optical properties of ice particles are based on the physical properties of the ice crystals derived from the microphysics scheme . Resolved advection is calculated using a fifth-order central difference numerical scheme for momentum and a κ-limiter scheme for the other prognostic scalar variables. Subgrid-scale transport uses a prognostic turbulent kinetic energy (SFS-TKE) scheme . A sponge layer is applied in the upper 25 % of the range to dampen disruptive gravity waves. The surface fluxes of heat, moisture, and momentum are calculated based on Monin-Obukhov theory using Arctic stability functions derived by . Horizontally periodic boundary conditions are applied, while large-scale forcings are prescribed as time-dependent tendency profiles to account for advection, subsidence, and horizontal pressure gradients (geostrophic wind vector vg).

To set up a Lagrangian LES experiment, the trajectories (pathways) of the air parcels forming the air mass need to be estimated. Trajectories for the low-level air parcels were calculated during flight planning using the Lagrangian Analysis Tool (LAGRANTO) and based on IFS forecast winds. For this study, this was repeated using 3-D wind fields of the ERA5 reanalysis. Air parcels were identified, and their trajectories were calculated starting every 3 km lat/lon within the IC on 29 March 2022 and passed through the circle locations on 30 March 2022 . The best match of the trajectories was found at 870 hPa in the initiation region. For the simulations, a mean trajectory was calculated by averaging all trajectories initiated within the first circle at 870 hPa. Each trajectory is included for every circle region it intersects, but is excluded from further analysis once it exits a circle region without entering the subsequent one. This procedure is illustrated in Fig. .

The lower boundary conditions are prescribed as time-varying fields along the Lagrangian simulation trajectory with hourly resolution. Surface skin temperature Tskin and sea-ice concentration “sic” are derived from VELOX observations (Sect. ), longitudinally averaged along the circular flight legs, and interpolated along the simulation trajectory. As the simulated air mass moves from sea ice through the marginal ice zone to the open ocean, this interpolation provides a temporally evolving surface boundary condition. The lower-boundary specific humidity is calculated following from the saturation specific humidity at the prescribed surface skin temperature and weighted by the local sea-ice fraction.

The prescribed large-scale forcing profiles are time-dependent and they are obtained by interpolation along the trajectory between the sonde data at the four circle locations. A regression algorithm is used to estimate horizontal gradients within each circle area, following the method of . The horizontal wind divergence and the associated subsidence ω are calculated as described by . At high latitudes, Cartesian latitude–longitude coordinates distort horizontal distances. To address this issue, introduced a regression method in spherical coordinates r,φ,θ. Assuming linear fluctuations of any variable ξ∈u,v,q,T,p within the dropsonde area, ξ at the centre of the pattern can then be approximated at every vertical level by a first-order Taylor expansion: 1ξ=ξ0+∂ξ∂θΔθ+∂ξ∂φΔφ. Here, ξ0 denotes the mean of all dropsonde measurements, and θ and φ are the polar and azimuth angles, respectively. This approach is applied in this study to estimate divergence: 2D=∇⋅vh=1r⋅sin⁡θ∂u∂φ+∂sin⁡θv∂θ. This allows for the estimation of the associated subsidence at height z by vertical integration from the surface: 3ω=-g⋅ϱ∫0zDdz.

Figure shows the resulting subsidence profiles in the four circles, provided in pressure and height units for reference, with key characteristics summarised in Table . In IC and C02, subsidence dominates the lowest 3 km, with mean downward motions of -0.74 and -0.60 cm s⁻¹, respectively, while C03 exhibits weaker subsidence throughout the profile. In contrast, C01 shows a mean upward motion over the 8 km profile. Uncertainties associated with the regression-based divergence estimate were calculated following and reach up to 0.025 cm s⁻¹ (0.002 Pa s⁻¹) below 3 km across all circles, which cannot be resolved in the figure and are therefore not shown. These estimates reflect uncertainties propagated from the regression analysis and do not represent a complete uncertainty budget, as they exclude potential representativeness errors associated with temporal evolution during the sequential dropsonde releases and deformation of the sampling pattern due to horizontal sonde drift. The method assumes approximate stationarity over the sampling period; sensitivity tests by indicate that accounting for pattern deformation changes the derived divergence and pressure velocity by less than 1 % and 10 %, respectively. The subsidence magnitudes are consistent with reanalysis-based estimates of Arctic MCAOs , but are approximately an order of magnitude lower than those reported for the tropical Atlantic, possibly due to the specific weather conditions of the sampled case .

The same regression approach using spherical coordinates can be applied to calculate the horizontal advective tendencies in the circular areas, 4-vh⋅∇ξ=-u1r∂ξ∂θ+1r⋅sin⁡θ∂ξ∂φ, and similarly, the horizontal pressure gradient force, here expressed as the geostrophic wind vector vg, 5vg=1ϱ⋅f⋅∂p∂φ1cos⁡φ∂p∂θ. These profiles, which extend up to HALO's flight altitude at about 8 km, are linearly interpolated between the four circles and kept constant south of the last circle along the simulation trajectory (Fig. ). What stands out in the interpolated mesoscale vertical pressure velocity field ω(t,z) is the distinct maximum at C02 below 4 km, which indicates a relatively strong signal in the prescribed subsidence that will play a crucial role in this study. Concerning the advective tendencies, the temperature advection -vh⋅∇T is dominated by cooling at low altitudes, below 1 km. The specific humidity exhibits low-level drying before the sea ice edge, transitioning to moistening south of C02. The horizontal wind components show low signals, except for a slight positive tendency of the meridional wind in the deepening ABL south of C01. These likely reflect Ekman-layer effects of Reynolds stresses, inducing a deviation of the wind from the geostrophic wind direction. The geostrophic wind (|vg|) expresses an eastward zonal pressure gradient force throughout the lower levels, consistent with the general synoptic situation during the 2 d. It should be noted that at a trajectory height of 870 h Pa corresponding to heights between 1.34 and 1.38 km, advective tendencies are, by definition, zero, as the air mass speed is subtracted from the wind in the Lagrangian forcing method.

Above 8 km altitude, which is well above the simulated ABL height (max 2 km), the sonde data are extended by ERA5 fields. The main objective of this procedure is to allow radiative flux density (irradiance) calculations above the model domain in which turbulence is resolved, which has a ceiling far below HALO's flight altitude (described in Sect. ). As a result, it is concluded that the use of ERA5 fields above this height has no discernible impact on the model results.

Apart from the control setup, four additional LES experiments are conducted using DALES to investigate the sensitivity of ABL evolution to the mesoscale subsidence forcing. The experimental setup includes (i) the control simulation (Ctrl) that employs the observed subsidence (Sect. ), (ii) a simulation without subsidence (NoSub) to isolate the effects of entrainment on ABL development, and (iii) further experiments in which observed subsidence was scaled by 10 % (Sub10), 50 % (Sub50) and 200 % (Sub200). These scaling factors correspond to a range of subsidence rates covering weaker values (-0.12 cm s⁻¹) comparable to those used in LES studies of mixed-phase Arctic clouds , up to stronger rates of approximately -2.3 cm s⁻¹, and consistent with observations during the EUREC⁴. A campaign in the tropics . In all sensitivity experiments, the temporal and vertical structure of the observed subsidence was retained, while only its magnitude was modified.

The model domain spans 12.8 km × 12.8 km with 50 m horizontal resolution and 191 vertical levels, starting at 10 m with 20 m spacing up to 2.1 km and increasing resolution aloft to a ceiling height of 6.8 km with maximum grid spacing of 177 m. Periodic horizontal boundaries are applied, and large-scale forcings are horizontally homogeneous but variable in time and height along the trajectory. The prescribed subsidence acts on the evolving model profile. Continuous Newtonian nudging towards the observed profiles of {T,q,u,v} is applied above the ABL, determined by the strongest inversion in liquid water potential temperature, with a buffer layer of 100 m where the nudging time scale increases linearly up to 3 h. Nudging increases in the upper quarter of the domain to suppress gravity waves, while below the ABL, the flow evolved freely. Each experiment is 45 h long, reaching the marginal sea ice zone approximately 15 h after initialisation.

First, we examine the evolution of the ABL and clouds, focusing on the temporal development of cloud liquid and ice water contents as well as the deepening of the ABL along the simulation trajectory of the MCAO, expressed as a function of latitude in Fig. . The initial atmospheric column at circle IC includes a shallow ABL of 65 m, which remains low over the sea ice in the simulation. The ABL begins to deepen once it reaches the marginal sea ice zone (MIZ) where warmer surface conditions trigger enhanced vertical latent and sensible heat fluxes. Convective updrafts form a thin mixed-phase cloud between C01 and C02, with ice reaching to the surface and a shallow liquid layer existing near the cloud top during the initial phase of the growing ABL. As the air mass moves over the open ocean, the ABL continues to deepen, reaching values of liquid water content ql up to 0.35 g kg⁻¹. At the end of the 45 h simulation, the ABL has grown to approximately 1.5 km.

To evaluate the Ctrl simulation with respect to the ABL structure, several independent measurements of ABL height are compared with the simulation output. The ABL height was determined from dropsonde profiles using the Bulk Richardson number Rib, which quantifies the balance between buoyancy and shear . A critical threshold Ribcrit=0.25 was used to define the height of the ABL zi, consistent with the diagnostics of the IFS CY43R1 model and the ERA5 reanalysis . Lidar measurements from the P5 aircraft (AMALi) provide cloud top height (zct), which is comparable to the ABL height in MCAO cases due to the strong coupling of ABL growth to cloud development. Finally, the simulation diagnoses the height of the ABL from the minimum flux of total liquid water potential temperature , w′ϑl′‾, within the lowest 3 km of the domain. AMALi measured a mean cloud top height of 0.63±0.11 km at location C02, yielding an RMSE of 0.12 km compared to the simulation ABL height. The ABL height derived from the dropsondes was 0.07 km at C01, 0.43 km at C02, and 1.04 km at C03, resulting in an overall root mean square error (RMSE) of 0.14 km compared to the simulation. The conclusion from this evaluation against the three datasets is that the model reproduces the observed ABL height within 150 m throughout the simulation. The temporarily reduced deepening rate between C02 and C03 in the simulation is supported by the measured data, which is an interesting feature that could be related to the low-level maximum in subsidence during that period (Fig. a).

The vertical structure of virtual potential temperature (ϑv) and specific humidity (q) in the Ctrl experiment is evaluated against in-situ observations from HALO dropsondes, closest to the simulation trajectory; the variability of all soundings in each circle is indicated as blue shading in Fig. . This comparison reveals possible systematic biases in the thermodynamic state. Figure focuses on the lowest 3 km of the atmosphere to highlight the ABL properties. Above 3 km altitude, the simulation closely follows the observed vertical structure of ϑv at all three locations. Larger differences occur for q, primarily due to observed humidity layers not being captured in the smoother simulated humidity structure.

In the region of C01 (Fig. a and d), located above sea ice near its edge, an inversion in ϑv is detected near the surface at 30 m in the simulation, and at 20 m in the dropsonde profiles, with a thermodynamic wind profile above it. The simulated near surface ϑv of -25.6 °C differs by 1.4 K from the dropsonde observations of -24.2 °C and lies within the observed variability. The humidity profiles show greater variability among the dropsondes, with several observed humid layers not captured by the smoother simulated profile; nevertheless, the simulation closely follows the vertical structure of the observed profiles. The near-surface value of q at C01 was 0.5 gkm⁻¹ in the simulation and 0.54 gkg⁻¹ in the observations. The close correspondence at C01 reflects the continuous nudging that is active above the ABL top, which is still very low at this stage.

As the air mass moved over the marginal sea ice zone (MIZ) and the open ocean to the location of C02 (Fig. b and e), surface warming caused an increase in the near surface air temperature ϑv,0 in the simulation of +14.5 K to -11.1 °C, while observed warming was even stronger with a measured increase in ϑv,0 of +15.6 K to -8.6 °C. The inversion height was approximately 590 m in both profiles, but the simulated inversion was less steep. Below the inversion, the model shows a cold deviation of 3.3 K, while reproducing the observed mixed-layer structure and inversion height. The increase in evaporation over the open ocean led to an increase in humidity, with surface moisture q0 of 1.5 g kg⁻¹ in the simulation and 1.6 g kg⁻¹ in the observations. Both show a pronounced drop in humidity near 590 m, with the observed jump being stronger and sharper than the simulation, induced by the prescribed advection of qt (shown in Fig. c). Within the mixed layer, where temperature and humidity are uniform with height, the model underestimates the humidity by an average of 0.2 g g⁻¹. Above the temperature inversion, additional differences result from observed moisture layers that are not captured in the simulation.

Further south, over the open ocean at C03 (Fig. c and f), the ABL has deepened further and ϑv has continued to increase, with a simulated surface value of ϑv,0 of -7.5 °C and an observed value of -4.6 °C representing an underestimation of near-surface warming by 2.9 K. The modelled inversion base is at 830 m, compared to 1000 m in the observations, with the simulated inversion being less steep. Below the temperature inversion, the model underestimates ϑv by approximately 2 K. Therefore, there is a consistent cold bias of -2 to -3 K across all circles, a tendency also found by in the limited-area configuration of the ICON model for all observed MCAO cases of HALO–(AC)3 and may reflect uncertainties in the prescribed surface skin temperature and limitations of the bulk surface-flux parametrisation. The humidity in the boundary layer at C03 has also increased further, reaching 1.7 g kg⁻¹ in the simulation and 2 g kg⁻¹ in the observations, with the model underestimating the surface layer of high humidity directly above the ocean. The simulated humidity jump is also weaker than observed. Below the temperature inversion, the model underestimates q by 0.2 g kg⁻¹.

Additional measurements taken with the CMET sonde drifting south along a similar trajectory further support these results. At the same latitude as C01, the CMET flew at an altitude of 340 m at a distance of 225 km to the simulated trajectory. The CMET reached the latitude of interest about 2.5 h later than the simulated air parcel. Here, ϑv=-8.8 °C and q=0.52 g kg⁻¹ were measured, which largely corresponds to the simulation in terms of humidity (difference of only 0.02 g kg⁻¹), but with ϑv being 5.7 K warmer, which may be due to the proximity to Svalbard and the associated local warming effects. As the trajectory of CMET was closer to Svalbard and travelled through the cloud-free region associated with lee-effects, likely causing a warming of the lower layers. The CMET reached the latitude of C02 7.5 h later at a distance of 132 km from the simulated trajectory. The CMET measured at 902 m above sea level a temperature of ϑv=-1.8 °C and q=0.25 g kg⁻¹, which is 3.7 K warmer and 0.2 g kg⁻¹ drier than the simulation. At C03, the CMET was 2 h later and 149 km away from the simulation at an altitude of 1780 m. A virtual potential temperature ϑv of 2.7 °C was measured, 1.8 K warmer than in the model and q=0.78 g kg⁻¹, only 0.06 g kg⁻¹ more humid. As the CMET moved south and the influence of Svalbard diminished, the measured air mass increasingly resembled that observed by the dropsondes, with the humidity in all three circles corresponding very closely to the simulated q values measured by the dropsondes.

The overall conclusion from these results is that the LES model reproduces the observed evolution in the vertical thermodynamic structure after C01 to a reasonable degree. This is not trivial, as the nudging takes place above the temperature inversion only, and below it, the simulated ABL is free to evolve. A slight cold and dry bias is evident within the ABL, and the inversions at its top height appear less pronounced than observed. These differences may arise from (i) uncertainties in the surface boundary condition Tskin, and (ii) shortcomings in the bulk flux parametrisation at the surface. Nevertheless, the overall correspondence between simulation and observation indicates that the model adequately reproduces the thermodynamic structure of the observed MCAO, providing a robust basis for further analysis.

The model performance in representing various column-integrated quantities is assessed next. In Fig. a, the simulated integrated water vapour (IWV) is evaluated against observations from two sources. The first source are the data from the HAMP radiometer on board HALO which measured along the flight trajectory, and the second is the IWV determined from the dropsonde humidity profiles, averaged over the ten sondes in each circular pattern. The simulation accurately replicated the observed IWV at C01 of 1.9 kg m⁻² with a small difference of 0.1 kg m⁻². At C02, the dropsondes measured 2.7 kg m⁻² IWV, reproduced with +0.04 kg m⁻¹ in the simulation. At C03, the observed IWV was 4.4 kg m⁻², which the simulation slightly overestimated by +0.3 kg m⁻². This temporal evolution of IWV closely follows the HAMP observations, with an RMSE of 0.2 kg m⁻² compared to the simulation. These results suggest that the net input of humidity into the deepening ABL is well captured by the model.

Figure b compares the simulated LWP against two measurements. The LWP retrieved from HAMP onboard HALO exhibits a much higher short-term variability than in the simulation, resulting in an RMSE of 30 g m⁻². Despite this, the overall temporal evolution aligns with the simulated LWP, with liquid water strongly increasing after the MIZ. The MiRAC instrument on P5 showed even higher variability due to its high spatial resolution and low flight altitude above the clouds. When averaged, MiRAC observations yield a mean LWP of 0.01 g m⁻², which appears much lower than the simulated LWP of 12 g m⁻², but the simulation value lies within the spread of the LWP observed by MiRAC. These results suggest that, while the observed LWP fluctuates more strongly due to higher resolution, capturing individual cloud streets, than the smoother, domain-average model output, the simulation captures the general trend and magnitude of the evolution of liquid water throughout the event.

The simulated ice water path of all frozen hydrometeors IWP shown in Fig. c is only compared to MiRAC measurements aboard P5. The IWP retrievals by HALO were excluded due to large uncertainties for this case. The average observed IWP from MiRAC was 0.6 g m⁻², but similar to the LWP observations, the variability due to the high resolution of individual clouds is large and values range up to 6.2 g m⁻². This high variability may also result from the substantial uncertainties associated with IWP retrievals. The Ze–IWC relationship of exhibits uncertainties around a factor of 2, caused by the strong temperature dependence of ice particle size distributions and densities. At warmer temperatures, the ice water content IWC tends to be overestimated, whereas at colder temperatures it is generally underestimated. This temperature sensitivity is particularly relevant for 94 GHz radar, where Mie-scattering further increases the retrieval uncertainty. Compared to the simulated IWP in the location of the P5 measurements of 2.58 g m⁻², an RMSE of 73.34 g m⁻² is found. It should also be noted that the P5 observations do not coincide with the air mass trajectory but were collected at a later time, limiting the comparability of the data (Fig. ). A notable feature of the latitudinal evolution of the simulated IWP is the local maximum reached in the simulation just before C02; the realism cannot be verified due to the absence of reliable IWP measurements along the whole trajectory.

The surface heat fluxes are key drivers of ABL evolution in MCAOs. To evaluate the simulated sensible (H) and latent (E) heat fluxes, we estimate these data from dropsonde observations following the thermodynamic method suggested by . This method assumes steady-state conditions, negligible radiative divergence and cross-flow advection, and minimal entrainment above the mixed layer. Here, following , the steady-state assumption allows thermodynamic differences between successive downstream profiles to be interpreted as along-flow advective changes and linked to the vertically integrated surface heat and moisture input. Then, the sensible and latent heat energy fluxes can be calculated as 6HDS=ϱ⋅cp⋅uML∫zi0Δϑ-Δydz-HC,

7EDS=ϱ⋅Lv⋅uML∫zi0Δq-Δydz+HC, where ϱ represents the air density, cp the specific heat capacity, Lv the latent heat of evaporation, uML the mean horizontal wind speed within the mixed layer and zi the height of the ABL. Δy is the horizontal distance between two profiles. The contribution of condensational heating within the cloud layer is represented by the following formulas:

8HC=ϱ⋅Lv⋅uML∫zizbΔl-Δydz, where zb represents the height of the cloud base, defined as the lowest altitude at which the relative humidity reaches saturation with respect to the ice or liquid. Meanwhile, Δl refers to the difference in ql between the profiles, which is approximated as the upper limit of ql within the cloud layer. This is estimated as the difference of qt within the cloud and the average qt,BC‾ below the cloud layer, following . In the model, the surface heat fluxes can be sampled continuously as

9HCtrl=ϱ⋅cp⋅w′⋅ϑ′‾,10ECtrl=ϱ⋅Lv⋅w′⋅q′‾, with w′ϑ′‾ and w′q′‾ representing the turbulent kinematic fluxes of temperature and moisture at the surface.

The simulation produced HCtrl values up to 830 W m⁻² and ECtrl values up to 102 W m⁻² over the open ocean. In particular, the sensible heat flux is substantially higher than the climatological mean sensible flux of about 200 W m⁻² during MCAOs, while the latent heat flux is comparable to the climatological mean latent heat flux of 100 W m⁻² reported by for MCAOs in the region west of Svalbard. The difference in potential temperature between the near-surface atmosphere and the ocean surface causes a direct response in the sensible heat flux, while the latent heat flux is limited by the saturation specific humidity at the sea surface temperature . In the observed region, H was generally more efficient than E in extracting energy from the ocean surface, but the relative importance of E increased as the air is mass advected away from the sea-ice edge over warmer waters. South of 75° N, the mean values of HCtrl and ECtrl were 69 and 17 W m⁻², respectively, consistent with the findings of .

Dropsonde-derived HDS values considerably underestimate simulated surface sensible heat fluxes by 564 W m⁻² between C01 and C02 and 306 W m⁻² between C02 and C03. We speculate that this offset is mostly due to numerous assumptions in the method of deriving fluxes from observed profiles. To improve comparability between model and measurements and gain further insight, the same diagnostic technique is applied to the simulated profiles (Fig. ), yielding much better agreement. The associated RMSE between HDS and HCtrl, Prof is 95 W m⁻². Between C02 and C03, the thermodynamic method yields a simulated value of only HCtrl, Prof=21 W m⁻², compared to HDS=170 W m⁻², largely underestimating the surface sensible heat flux. Note that the dropsonde-derived EDS in the marginal ice zone (C01–C02) is very small (2.6 W m⁻²) compared to the simulated fluxes that exceeded 100 W m⁻², but increased to 16 W m⁻² between C02 and C03, closer to the simulated values. The RMSE of 8.9 W m⁻² between EDS and ECtrl, Prof indicates that although the observational profile-based method struggles to fully capture the true flux magnitudes in the early MCAO phase, it reproduces their relative variability well when applied consistently.

Figure shows the evolution of ql and qi along the MCAO trajectory. In the no-subsidence (NoSub) setup, the cloudy ABL deepens the fastest, approaching the parabolic deepening rate (∼t1/2) after the air mass moves over the open ocean . In this setup, the cloud mass is the largest, both in liquid and frozen form. The higher the prescribed subsidence forcing, the more the ABL deepening driven by turbulent/convective entrainment is suppressed by subsidence. This in particular applies during the time period between C02 and C03, where the observed low-level subsidence has its maximum (Fig. a). This leads to the development of thinner clouds and less ql and qi. In the Sub200 case, only a very thin, shallow mixed-phase cloud is observed.

To gain more insight into the spatial structure of the cloud field, Fig. presents horizontal cross-sections of ql and qi through the centre of the simulated domain for all five simulations. These locations are sampled at regular intervals along the trajectory, beginning at C01 near the sea-ice edge and extending southward to 73.25° N at the end of the simulation period. The grey shading highlights regions in which the cloud layer is decoupled from the surface, following the criterion of . This diagnostic is defined by the thresholds Δqt=q‾t,bot-q‾t,top>0.5 g kg⁻¹, based on the layer-mean total-water mixing ratio in the upper and lower 25 % of the boundary layer, and Δϑl=ϑ‾l,top-ϑ‾l,bot>0.5 K, calculated analogously for the layer-mean liquid-water potential temperature. These metrics have been associated with the onset of boundary-layer decoupling and the subsequent transition from roll convection to open-cell structures in MCAOs . What stands out during the decoupled phase is the presence of liquid-phase shallow cumulus clouds, rising into the capping mixed-phase cloud layer.

Across the simulations, the onset of decoupling exhibits a systematic dependence on the prescribed subsidence: in cases with reduced subsidence, the decoupling occurs closer to the sea-ice edge and is accompanied by overall larger maxima in both liquid and ice water content. While all simulations show the initial formation of a thin ice cloud at C01, substantial differences arise beginning at C02, most notably in cloud-top height and the vertical distribution of liquid and ice. The Sub200 simulation produces only a shallow ice cloud capped by a thin liquid layer, with little variation in cloud-top altitude throughout the trajectory. In contrast, the simulations with weaker subsidence develop considerably deeper clouds, characterised by more complex internal structure and enhanced condensate mass.

The impact of subsidence on the developing ABL is characterised by various bulk quantities, as shown in Fig. . The integrated water vapour IWV (Fig. a) is similar in all simulations, but the boundary layer (Fig. d) in the NoSub case reaches substantially greater depths (2 km), following a parabolic deepening rate due to unimpeded entrainment at the ABL top. All simulations exhibit a similar strong initial boundary layer growth within the MIZ and immediately downstream of the sea ice edge (up to C02), driven primarily by intense surface heat fluxes in this region (Fig. ). However, after this initial phase, the Sub200 simulation boundary layer actually becomes shallower, with the strong subsidence keeping zi below 1 km. In contrast, the Ctrl, Sub50, and Sub10 simulations all maintain continued ABL deepening, though zi remains below the parabolic limit visible in NoSub.

The liquid water path LWP (Fig. c) exhibits an initial peak near the sea ice in all simulations, coinciding with the region of strong boundary layer growth at approximately 80° N on average. The RMSE in the location of the peak between the simulations is 0.01° N, and the strength of the peak decreases with increasing subsidence by 18 g m⁻², with an overall RMSE of 4.9 g m⁻². A second, more pronounced LWP maximum occurs later at 76.3° N in NoSub, with a maximum value of 173 g m⁻². For Sub50 and Ctrl, this secondary peak shifts to lower latitudes and reduced magnitudes, occurring at 76° N (167 g m⁻²) and 74.4° N (133 g m⁻²), respectively. In Sub200, the liquid water path does not recover after the initial maximum and shows only a gradual increase later in the simulation, beyond 74° N.

The ice water path IWP (Fig. e), including all frozen hydrometeors, displays a similar spatial evolution, with an initial peak forming in the MIZ around 79.4° N, consistent in all simulations. Subsequently, IWP increases over the open ocean following the sea ice edge. Compared to LWP the ice growth begins slightly earlier (also visible in Fig. ), which is consistent with previous studies of Arctic MCAOs . The peak IWP in NoSub occurs at 75.7° N with 48 g m⁻², whereas for Sub50 and Ctrl it shifts to 75.7 and 73.8° N respectively with reduced magnitudes of 41 and 31 g m⁻² respectively. The Sub200 simulation exhibits no such shift or comparable peak values, indicating a systematic suppression of ice growth under stronger subsidence.

The ratio of ice to liquid water, IWPLWP, calculated only for regions with a significant liquid content (LWP> 1 g m⁻²), highlights these differences more clearly (Fig. b). The initial shallow cloud developing above the MIZ shows a similar structure among NoSub, Sub10, Sub50, and Ctrl, while Sub200 displays a slightly larger ice fraction in the initial cloud. However, because of the overall thinness of both liquid and ice layers, this difference cannot be regarded statistically significant. In later stages, LWP in Sub200 falls below 1 g m⁻², producing a discontinuity in the IWPLWP ratio, while in the remaining cases a clear shift towards later growth of IWP is evident.

Finally, the impact of subsidence on the total surface precipitation rate (Ptot) (Fig. f) is broadly similar to the IWP. The NoSub simulation reaches precipitation rates up to 60 mm h⁻¹ towards the end of the simulation, compared with 25 mm h⁻¹ at the peak near 74° N for Ctrl. In Sub200, precipitation remains confined to the initial shallow cloud phase, with maximum rates of only 2.1 mm h⁻¹.

The results so far indicate that the behaviour of the cloud ice phase is particularly affected by the strength of subsidence. To gain further insight, the relative contributions by the three frozen hydrometeor classes in the LES microphysical scheme are investigated. The integrated column water paths of cloud ice (CIWP), snow (SIWP), and graupel (GIWP) are shown in Fig. . The behaviour of CIWP (Fig. a) closely follows that of the total ice water path (IWP), reflecting its dominant contribution to the total ice mass. The maximum values of CIWP reach 38 g m⁻² at 73.4° N for NoSub and 32 g m⁻² at 73.7° N for Ctrl, with a systematic decrease in amplitude under stronger subsidence. The SIWP (Fig. c) distribution exhibits a similar latitudinal progression, with peak values occurring later and at reduced magnitudes as subsidence increases. However, unlike CIWP, SIWP does not sharply decline after the peak but plateaus afterwards, reaching 9.6 g m⁻² at 73.3° N for NoSub and 5.5 g m⁻² at 73.4° N for Ctrl. The graupel path GIWP (Fig. b) exhibits the most distinct and localised peak. In NoSub the peak occurs at 76° N with a magnitude of 8.4 g m⁻², generally shifting southward with increasing subsidence to 74° N for Ctrl, weakening in amplitude to 2.7 g m⁻². The Sub200 experiment does not develop graupel. In this southward progression, the width of the graupel peak remains similar.

Graupel is typically formed in convective updrafts, as it requires substantial vertical velocities. Accordingly, the southward shift and progressive weakening of the peak in GIWP with increasing subsidence probably reflects a similar shift in convective dynamics. This motivates a deeper investigation of the behaviour of boundary-layer internal decoupling. Previous MCAOs studies have established that such decoupling plays a central role in the transition from organised roll convection to open-cellular convection. Enhanced precipitation, often driven by secondary ice production and riming, has been shown to promote evaporation and sublimation below the cloud, thereby cooling and moistening the sub-cloud layer. This preconditioning in water vapour at lower levels accelerates the breakup of stratocumulus decks and favours the onset of cellular convection. As a result, the surface supply of moisture and heat to the cloud layer is reduced, weakening roll structures and promoting the development of cellular patterns . A slightly different form of decoupling that does not involve precipitation has been proposed earlier by , showing that during deepening marine subtropical stratocumulus-topped ABLs a layer of negative buoyancy flux w′ϑv′‾ can form below cloud base, as a result of top-down and bottom-up driven mixing becoming too far separated in the vertical. This negative buoyancy flux layer then disrupts the ABL-deep circulation, inducing decoupling. The latitudinal development of the buoyancy flux is shown in Fig. .

Both mechanisms of decoupling and the link to graupel formation are investigated next. The buoyancy-based decoupling mechanism is assessed by defining a decoupling point Dw′ϑv′‾ as the time point at which w′ϑv′‾ becomes negative below cloud base for the first time, and remains so until the end of the simulation (indicated as ▾ in Fig. ). The humidity-based decoupling mechanism involving precipitation evaporation is assessed by defining a second decoupling point, this time inferred from Δqt and Δϑl denoted Dqt,ϑl, as discussed above (Sect. ). Comparing these two definitions gives insight into how exactly decoupling works in MCAOs and how it affects the shifts in ice phase hydrometeors.

Figure compares the locations of the largest peaks in liquid water path (LWP_max), graupel water path (GIWP_max), and surface precipitation rate (Ptot, max) with the two decoupling points Dqt,ϑl and Dw′ϑv′‾, for all simulations except Sub200 (in which no graupel appears). A key result that stands out is that buoyancy-based decoupling (vertical axis) structurally happens first, in all simulations (also indicated in Fig. ). This is followed by the peak in LWP, then integrated graupel, then surface precipitation, and finally the precipitation-based decoupling point. This sequence of events typically covers 1° of latitude in this MCAO case, and shifts southward with increasing prescribed mesoscale subsidence. The order of appearance of the various maxima is consistent with the idea that (i) internal buoyancy-decoupling initiates events and causes stronger surface-driven convection, (ii) driving stronger precipitation formation (mainly in graupel form) that efficiently depletes cloud liquid water mass as discussed by, (iii) forming intense graupel for a while, (iv) associated with enhanced surface precipitation which then (v) further stabilises the lower boundary layer through evaporation below cloud base, suppressing vertical mixing and further enhancing decoupling . The last step is also suggested by the slight lowering of liquid cloud base height visible in Fig. . This sequence of events ultimately yields a fully decoupled structure with respect to Δqt and Δϑl at Dqt,ϑl.