Radiative energy budget and cloud radiative forcing in the daytime marginal sea ice zone during Arctic spring and summer

Airborne measurements of the surface radiative energy budget (REB) collected in the area of the marginal sea ice zone (MIZ) close to Svalbard (Norway) during two campaigns conducted in early spring and and early summer are presented. From the data, the cloud radiative forcing (CRF) was derived. The analysis is focussed on the impact of changing atmospheric thermodynamic conditions on the REB and on the linkage of sea ice properties and CRF. The observed two-mode longwave net irradiance frequency distributions above sea ice are compared with measurements from previous studies. The transition of 5 both states (cloudy and cloud-free) from winter towards summer and the associated broadening of the modes is discussed as a function of the seasonal thermodynamic profiles and the surface type. The influence of cold air outbreaks (CAO) and warm air intrusions on the REB is illustrated for several case studies, whereby the source and sink terms of REB in the evolving CAO boundary layer are quantified. Furthermore, the role of thermodynamic profiles and the vertical location of clouds during on-ice flow is illustrated. The sea ice concentration was identified as the main driver of the shortwave cooling by the clouds. 10 The longwave warming of clouds, estimated to about 75 W m−2, seems to be representative for this region, as compared to other studies. Simplified radiative transfer simulations of the frequently observed low-level boundary layer clouds and average thermodynamic profiles represent the observed radiative quantities fairly well. The simulations illustrate the delicate interplay of surface and cloud properties that modify the REB and CRF, and the challenges in quantifying trends in the Arctic REB induced by potential changes of the cloud optical thickness. 15


Introduction
The Arctic climate has experienced drastic transformations in the last decades (Meier et al., 2014;Koenigk et al., 2020), caused by global climate change and related uncertainty of local and remote feedback mechanisms (Goosse et al., 2018). One of the obvious Arctic climate changes is the amplified warming relative to the global warming , which 20 is commonly referred to as Arctic amplification. Also, the diminishing sea ice extent (Stroeve and Notz, 2018) represents a dramatic climate change in the Arctic, playing a decisive role (Screens and Simmonds, 2010) in Arctic amplification via overview of the campaign is given by (Knudsen et al., 2018). The instrumentation on board of both research aircraft Polar 5 and Polar 6 from the Alfred Wegener Institute (AWI) as well as the data processing is presented by Ehrlich et al. (2019b). In March/April 2019, the AFLUX campaign was performed in the same region as ACLOUD, using only the Polar 5 aircraft, but with identical remote sensing instrumentation. Between 21 March and 11 April 2019, 13 flights in the MIZ were accomplished with six hours of data in an altitude below 100 m (average 73 m), extending the 16 hours of low-level flights during ACLOUD 5 . Thus, both campaigns covered different seasons with late winter/early spring conditions during AFLUX (Section 3), and late spring/early summer conditions during ACLOUD, representing the transition from a cold season into a beginning warm melting season .
The basic instrumentation used in this study and the one by Stapf et al. (2020) has been operated during both campaigns.
Broadband radiometer for upward and downward shortwave (0.2 -4 µm) and longwave (4 -100 µm) irradiance are used to 10 quantify the REB with a frequency of 20 Hz. A 180°fish-eye camera is applied to derive the cosine-weighted sea ice fraction (I f ) and a Kelvin infrared-thermometer  provides information on the surface brightness temperature. From dropsondes and aircraft in situ observations during ascents and descents in the vicinity of the low-level flight sections, the local atmospheric thermodynamic state and basic meteorological parameters are obtained. The cloud liquid water path (LWP) was derived during the below-cloud, low-level flights by a shortwave, transmissivity-based retrieval technique . The retrieved 15 values represent an equivalent LWP (assumed cloud droplet effective radius of 8 µm), as cloud ice has not been considered in the simulations. In case of mixed-phase clouds, the retrieved values of equivalent LWP are lower compared to the real total water path (liquid plus ice).

Limitations of airborne radiation observations in the marginal sea ice zone
Due to the generally colder in-cloud temperatures during AFLUX, icing on the instruments was more likely compared to 20 ACLOUD. Especially during longer horizontal flight sections in super-cooled cloud tops, occasionally sudden and persisting ice caps occurred on the glas domes of the pyranometers. The flat glas dome of the pyrgeometers were found to be iced less frequently. In cloudy and diffuse illuminated conditions, icing is difficult to detect and a possible influence on shortwave and longwave irradiances as discussed by Cox et al. (2021), can not fully be excluded. Nevertheless, during AFLUX about 21 % of the low-level shortwave irradiance measurements were discarded after quality checks using collocated radiative transfer 25 simulations as well as manual tests of plausibility. During 5 % of the sections severe icing was observed, which also affected the longwave irradiances.
Due to a combination of slower true air speed in low-level sections (average of 56 m s −1 instead of 67 m s −1 ), more turbulent conditions, and a higher payload of the aircraft, the pitch angles have been slightly higher during AFLUX. Especially for high solar zenith angles (SZA) and optically thin clouds the alignment of the pyranometers fixed to the aircraft fuselage represents 30 a major cause of uncertainty of airborne shortwave irradiance observations (Wendisch et al., 2001). The attitude threshold of the roll and pitch angles was increased from 4 • to 5 • compared to ACLOUD, but still, 28 % of the dataset had to be discarded.
The combination of high SZA and thin clouds (sun disk visible) poses, without the ability to correct for the aircraft attitude like during cloud-free conditions (Ehrlich et al., 2019b), significantly higher uncertainties in shortwave irradiances, surface albedo, and retrieved LWP. Also, geometric issues arise for low sun conditions as the major information content for the LWP retrieval in these conditions (direct dominated irradiance), will be retrieved from larger horizontal distances to the actual point of observations. Furthermore, 3D effects of heterogeneous mid layer cloud fields occasionally introduced unreasonably high LWP values (> 200 g m −2 ), which were consequently excluded from the analysis.
The airborne observations quantify the REB in flight altitude close to the ground, however, these values may differ from 5 the surface REB depending on the atmospheric conditions. Especially over the open ocean and leads, sometimes differences between the air temperature in 60 m flight altitude and the surface temperature of up to 20 K occurred. In such cases, common during CAOs above water, radiative transfer simulations indicate a difference between cloud-free longwave net irradiances observed in flight altitude (60 m) and the ground of up to 10 W m −2 . In case of strong surface temperature inversions over sea ice, up to 2 W m −2 difference is simulated. During AFLUX, the visibility during low-level flight sections was often reduced. 10 Over leads and open water sea smoke developed and in the low boundary layer over sea ice, sometimes surface-based clouds or precipitation and also fog were observed, in contrast to the conditions observed during ACLOUD. These conditions can bias the longwave and shortwave irradiances and parameters such as surface albedo observed in flight altitude, but were not excluded because otherwise the typical springtime features of the MIZ as strong CAOs could not have been included in our analysis. 15 Another important aspect of this dataset is the temporal and spacial sampling. As the flights took place mostly during noon, the shortwave irradiance represents the maximum level during those days and does not hold for typical seasonal averaged shortwave irradiance conditions. The solar zenith angles (SZA) for AFLUX ranged between 72 • and 82 • degrees and would be representative for the daily range of SZA observed from end of April to the beginning of May. In addition, the flight operations were limited to suitable conditions for the Polar 5 and Polar 6 aircraft, which reduced, the number of warm air intrusion events 20 during AFLUX significantly and, therefore, will bias the obtained distributions of radiative parameters.
Although these airborne radiation measurements are tainted with a higher uncertainty and limitations in temporal sampling compared to ground-based observations, they have the advantage to capture the continuously and quickly changing conditions in the MIZ of the Fram Strait with up to 3 m horizontal resolution on large horizontal scales with a flight speed of 50 to 70 m s −1 . 25

Derivation of cloud radiative forcing and radiative transfer simulations
The gain or loss of energy of the surface due to radiation, the REB, can be derived from the individual components of net irradiances (F net = F ↓ − F ↑ ) in the longwave and shortwave wavelength range: (1) To separately quantify the influence of clouds on the REB, the cloud radiative forcing (CRF, ∆F ) after Ramanathan et al. 30 (1989) is defined as the difference between shortwave and/or longwave net irradiances in cloudy (F net,all ) and cloud-free conditions (F net,cf ): 5 https://doi.org /10.5194/acp-2021-279 Preprint. Discussion started: 26 April 2021 c Author(s) 2021. CC BY 4.0 License.
The derivation of the CRF from airborne observations over homogeneous sea ice during the ACLOUD campaign using radiative transfer simulations for the required reference irradiances in cloud-free conditions is detailed in Stapf et al. (2020). This study extents the investigation by Stapf et al. (2020) to the entire MIZ including the open ocean. For the longwave CRF, two definitions of a radiative transfer-based estimate and an observation-based (climatological) CRF, discussed in Stapf et al. (2021), are used in this study. The challenge of calculating the shortwave CRF is to derive a surface albedo representative for 5 cloud-free conditions, although the measurements are obtained in cloudy situations, whereby clouds influence the illumination conditions of the surface and consequently alter the surface albedo depending on the surface type.
The cloud-free albedo of homogeneous snow and ice surfaces α cf,snow/ice for I f larger than 95 % is derived from observations in cloudy conditions by a retrieval detailed in Stapf et al. (2020). This approach is based on the combination of a snow albedo parameterization by Gardner and Sharp (2010) 20 In the MIZ, the surface types of water and sea ice are mixed depending on the sea ice fraction. They strongly vary on small scales, consequently influencing the surface albedo and the CRF. By assuming a Lambertian surface albedo, the cosineweighted sea ice fraction (I f ), obtained from fish-eye cameras, and its approximately linear relation to the albedo is used to weight the cloud-free, retrieved surface albedo of open ocean and sea ice: 25 for I f between 5 % and 95 %. Below and above these thresholds, a homogeneous surface was assumed to reduce uncertainties induced by misclassified surface types by the fish-eye camera images, like sea smoke or nilas. During AFLUX, the I f might have been underestimated sometimes because dark nilas (ice) were incorrectly detected as water, which is relevant for the broadband surface albedo because dark nilas are characterized by different spectral reflection properties compared to open water (Zatko and Warren, 2015). Albedo values that exceed the ice fraction range between the albedo values of snow, ice and 30 ocean, have been excluded from the analysis; they amount to 20 % and 12 % during AFLUX and ACLOUD. It needs to be considered that the full complexity of 3D radiative transfer in combination with mixtures of surface types in the MIZ can only imperfectly be described by the simplified assumption of a Lambertian surface. The radiative transfer simulations were performed using the LibRadtran package (Emde et al., 2016) and are set up as described by Stapf et al. (2020). The local thermodynamic profiles obtained from dropsondes or in situ profiles measured by the aircraft were merged with upper air radiosoundings. During ACLOUD, radiosoundings from the Polarstern research vessel of the concurrent PASCAL campaign  in the MIZ northwest of Svalbard and the Ny-Ålesund research station (Svalbard) were used. During AFLUX, the information above approximately 3 km altitude were only available from 5 Ny-Ålesund. This restriction results in slightly higher uncertainties of the simulated downward irradiances, especially with increasing distance from Ny-Ålesund.
3 Conditions in the marginal sea ice zone during the observations

Surface and cloud properties
Compared to ACLOUD, during AFLUX the sea ice edge (15 % isoline of satellite derived I f ) was situated, although earlier in 10 the season, slightly further in the North, but still in reach of the aircraft. In Fig. 1a the frequency distributions of the sea ice concentration observed during all low-level flight legs of both campaigns are compared. Both data sets indicate that the majority of the low-level flights have been performed above homogeneous sea ice, while during ACLOUD a slightly higher fraction of flights were performed over fractional sea ice with concentrations between 50 % and 95 %. The corresponding distribution of the surface albedo is presented in Fig. 1b. The surface albedo distributions are determined by the sea ice distribution and show, for sea ice (maxima at 0.6 and 0.76) due to the beginning melt season in the end of the campaign with decreasing surface albedo values. Also, the surface albedo of the open ocean is influenced by the different SZA of ACLOUD and AFLUX, which explains the broader open ocean albedo mode for AFLUX from 0.06 (cloudy) to 0.2 (cloud-free) compared to almost constant values below 0.1 for ACLOUD.
In Fig. 1c the frequency distribution of the retrieved equivalent LWP observed during both campaigns illustrates that during 5 AFLUX cloud fields with an LWP below 30 g m −2 are more frequently observed compared to ACLOUD, where clouds with LWP above 30 g m −2 are slightly more often. The median LWP in cloudy conditions above sea ice (LWP > 5 g m −2 , I f > 95 %) amounts to 34 g m −2 for AFLUX and 50 g m −2 ACLOUD. During AFLUX, often thin and low cloud fields in a shallow boundary layer over sea ice were observed, as compared to ACLOUD with more homogeneous cloud fields in a boundary layer with larger vertical extent (Section 3.2). Nevertheless, in 80 % of the conditions, the LWP was lower than 58 g m −2 10 during AFLUX and 68 g m −2 during ACLOUD underlining the character of the frequent often thin low-level clouds in this region.
Conclusions of differences between clouds above open ocean and sea ice are not feasible because of the unreliable statistical evidence due to the daily cloud field variability and insufficient sampling of the flight pattern. The AFLUX LWP distribution shows a mode at over 150 g m −2 , which can be partly addressed to one flight during a strong cold air outbreak with extraordi-15 nary optically thick clouds already above closed sea ice. Another characteristic difference between both campaigns represents the frequent presence of sea smoke and surface-based clouds during AFLUX, observed over open water leads embedded in the sea ice or during CAOs above the open ocean. Over sea ice, shallow fog-like conditions were observed more frequent during AFLUX as compared to ACLOUD, where mostly a clearly separated low cloud base was present with occasional precipitation. Due to lower cloud temperatures, also the cloud ice fraction should have been higher during AFLUX. Already the visual 20 impression of collocated satellite images indicate that every flight/day represents different cloud properties and distributions of cloud fields driven by the large scale processes and have to be considered as highly variable during both campaigns.

Structure of the lower atmosphere
The different seasons are evident in the surface brightness temperature distributions shown in Fig. 1d. During AFLUX low surface temperatures of often below -20 • C and surface temperature gradients in the MIZ of up to 25 K were found in the MIZ. 25 The conditions were characterized by a strong daily variability of the surface temperatures depending on the synoptical situation (Section 3.3) and distribution of clouds in the area, but also more complex surface types like nilas (covered or uncovered by thin snow), which broadened the surface temperature distribution and underline the complexity of the surface in this region.
Instead, during ACLOUD after a still cold period end of May, a thermodynamic rather uniform surface was observed in the beginning melt season in the MIZ, with surface temperatures distributed around the melting point of snow (see also Section Besides colder surface temperatures, the thermodynamic profiles show seasonal characteristics. In Fig. 2 the average temperature profiles over areas with a sea ice concentration above 90 % (derived from daily sea ice concentrations maps from Spreen et al., 2008) are shown along with long-term references from the Surface Heat Budget of the Arctic Ocean (SHEBA)  campaign (Uttal et al., 2002) (inner Arctic) and the Norwegian young sea ICE (N-ICE2015) campaign  (sea ice and MIZ north of Svalbard) for different seasons. Thermodynamic profiles and longwave net irradiances are used from Hudson et al. (2016) and Hudson et al. (2017) for N-ICE2015, and from Persson (2011) and Moritz (2017) for SHEBA. During ACLOUD, also Polarstern radiosoundings from the concurrent PASCAL campaign  in the sea ice area have been used. The aircraft observations are only shown above the minimum flight altitude due to the strong near-surface 5 temperature variability discussed in Fig. 2d. A separation between cloudy and cloud-free profiles of AFLUX/ACLOUD appear unreliable due to insufficient statistics, and thus, is shown only for the SHEBA and N-ICE2015 campaigns.
During ACLOUD (spring/early summer), the average temperature profile is shaped by a frequently observed cloudy boundary layer with low-level clouds and a cloud top temperature inversion at altitudes between 300 and 400 m. This profile is in agreement with those observed in cloudy conditions during N-ICE2015 and SHEBA (slightly warmer) but differs in temperature inversion strength.
Compared to ACLOUD, a stronger temperature inversion above the even lower and approximately 15 K colder boundary layer was observed during AFLUX. Also, the near-surface lapse rate of the individual profiles depended more on the prevailing conditions over sea ice. They quickly adapted to the cloud conditions and transitioned between strong surface-based inversions 5 in cloud-free conditions and more neutral profiles in cloudy situations (Stapf et al., 2021).
The profiles measured during N-ICE2015 are discussed by Cohen et al. (2017) and Kayser et al. (2017). For the time period of ACLOUD (Fig. 2c), they appear to have weaker cloud-top temperature inversions compared to ACLOUD. In cloud-free conditions they appear to be less stable compared to SHEBA (Fig. 2c). Kayser et al. (2017) illustrated (their Fig. 9) that in comparison to SHEBA, the near-surface layer tends to be more unstable during N-ICE2015. Cohen et al. (2017) (their Fig. 12) 10 illustrates a less uniform transition of thermodynamic profiles from spring to summer, and a period of higher lifted inversion bases in cloudy and less stable cloud-free conditions during the April/May period in Fig. 2b. In late winter and early spring, N-ICE2015 profiles are (especially in the cloudy state) significantly warmer and less stable compared to SHEBA data due to strong synoptic storm events Cohen et al., 2017).
In addition to homogeneous sea ice, the average atmospheric profiles above open ocean are shown for AFLUX and ACLOUD. 15 During both seasons the boundary layer height and temperature increases significantly towards the open ocean. The broad distribution of near-surface stability between surface and atmosphere as well as the slightly unstable near-surface layer represents an interesting preconditioning relevant for the longwave net irradiances in this region.

30
To identify the periods dominated by cold air outbreaks during both campaigns, in Fig. 3a the difference between potential skin temperature and potential temperature in 850 hPa (marine cold air outbreak index, MCAO), as commonly applied (e.g., Papritz and Spengler, 2017;Knudsen et al., 2018), were calculated from ERA5 reanalysis data (Hersbach et al., 2020). While during ACLOUD, Knudsen et al. (2018) identified three synoptic periods starting with a cold period with CAOs in May followed by  a warm and a neutral period, during AFLUX, a more frequent shift between on-ice and off-ice flow pattern (positive MCAO index, bluish) occurred due to several low-pressures systems moving from the south of Greenland towards Svalbard. These events began with the advection of warmer air masses in the course of the low-pressure systems moving towards Svalbard, followed by strong northerly flows for several days induced by the low-pressure system located in the east of Svalbard.
With changing atmospheric temperatures and humidity, the longwave REB in the MIZ in the Fram Strait will be affected by 5 these synoptic scale processes, which can be seen in the simulated cloud-free longwave downward irradiances (F ↓ lw,cf ) ( Fig.  3b). In the beginning of AFLUX a strong warm air advection caused a 40 W m −2 higher F ↓ lw,cf compared to the ERA5 10 year climatological mean. With the beginning and intensification of a CAO period from 23 to 26 March 2019 the atmospheric temperatures as well as surface temperatures steadily decreased. The simulated in situ profiles in the MIZ (black scattered points) vary by up to 50 W m −2 distributed around the climatological mean above sea ice and increase towards the open ocean, illus-10 trating the impact of the presence/absence of sea ice on the thermodynamics. Thereby, values frequently exceed the simulated F ↓ lw,cf using radiosoundings from Ny-Ålesund, illustrating a warmer marine atmosphere compared to the local thermodynamic profiles embedded in the protected fjords of Svalbard. The sea ice surface brightness temperatures are distributed around the climatological mean and are clearly linked to the synoptic situation only in the beginning of the campaign. During the last two flights of the campaign, characterized by less dynamic conditions on large scales, a broad distribution of surface temperatures and F ↓ lw,cf due to local thermodynamics (cloud-free and cloudy areas) were observed. In the beginning of ACLOUD, the ERA5 climatology indicates an approximately 20 W m −2 higher F ↓ lw,cf compared to 5 the observed period of CAOs with colder atmospheric temperatures compared to the climatology. Due to weaker temperature gradients in the MIZ, the gradient of F ↓ lw,cf decreased significantly compared to AFLUX. In the transition from the cold towards the warm period, the F ↓ lw,cf increased by up to 30 W m −2 (mainly due to temperature advection in the free troposphere, Knudsen et al., 2018), while a second influx of warm and moist air (Knudsen et al., 2018) around the 10 June 2017 increased the F ↓ lw,cf even more, underlining the importance of increased moisture in the advected air mass. The simulated in situ profiles 10 transitioned between the F ↓ lw,cf simulated using Polarstern radiosoundings (in the sea ice) and Ny-Ålesund emphasizing the areal variability in the region northwest of Svalbard.
In general, the conditions during AFLUX appear more synoptically driven compared to ACLOUD. Nevertheless, during both campaigns, the daily variability in cloud distributions and properties in the area clearly mask trends in the REB or CRF linked to synoptic processes (for example thermodynamic atmospheric background influence on CRF, Stapf et al., 2021) and 15 underline that both aspects (synoptic scale and local effects) have to be analysed separately.
4 Radiative energy budget in the marginal sea ice zone 4.1 Seasonal two-mode structure of longwave net irradiances above homogeneous sea ice As was illustrated in Section 3.2, the atmospheric thermodynamic structure over sea ice underlies a transition from a surfacebased inversion-dominated, often also cloud-free, winter atmosphere towards a cloud-dominated boundary layer in summer. 20 As the longwave net irradiances are basically a temperature difference between the effective radiative temperature of the atmosphere and the surface, seasonal characteristics and influences on the distribution of longwave net irradiances can be expected.
In Fig. 4, the longwave net irradiance distributions observed during SHEBA, N-ICE2015, ACLOUD, and AFLUX, are shown for certain periods to illustrate the transition from winter to summer. Similar to winter conditions observed during 25 SHEBA and N-ICE2015 (Stramler et al., 2011;Graham et al., 2017), also during the ACLOUD/AFLUX campaign over closed sea ice (Ice Fraction > 90 %) mode structures appear in the frequency distribution of longwave net irradiances. However, the location of the individual modes during each season is shifted, and the difference between both increases towards the summer.
The representative cloud-free mode is found, compared to winter (SHEBA, -40 W m −2 ), at around -60 W m −2 during late winter/early spring (AFLUX), and shifts even further to the negative during summer (ACLOUD, around -75 W m −2 ). The 30 cloud-free mode during AFLUX indicates higher negative longwave net irradiances with low frequency of occurrence. This is less related to atmospheric fluctuations and rather the result of the heterogeneous surface temperatures (Fig. 2b), where for example snow-covered thin nilas or thinner ice floes appear to have significantly warmer surface temperatures compared to  For the cloudy modes (distributed around zero), also a slight trend towards the summer to more negative longwave net irradiances in those opaque cloudy conditions is indicated. The median values decrease from SHEBA winter (-2 W m −2 ) to AFLUX (-6 W m −2 ) and ACLOUD (-11 W m −2 ). Interestingly, the SHEBA longwave net irradiance distributions appear 5 during all seasons slightly less negative, representing a weaker loss of energy in the longwave wavelength range at the surface in cloud-free as well as cloudy conditions. Especially during early summer, more neutral (distributed around 0 W m −2 ) longwave net irradiances were observed along with a still frequent occurrence of positive values during SHEBA induced by a warmer cloud base compared to the surface. This state was not observed neither during N-ICE2015 nor ACLOUD and might represent a specific feature, potentially linked to the special near-surface thermodynamics or cloud properties of this region north of The parameters that control the individual modes are the surface temperatures in combination with the atmospheric thermodynamic profile that modifies, via stability/lapse rate and vertical profiles of absorber gases, the emissivity of the atmosphere or the effective cloud base temperature as a function of cloud base height and optical thickness.
With increasing temperatures from winter to summer and a large difference between atmospheric brightness and surface temperatures in cloud-free conditions, the non-linearity of the Planck radiation law is an important driver for the widening 5 between both modes and the stronger loss of energy in the longwave wavelength range of the surface in summer. By assuming a theoretical combination of a surface temperature of -20 • C and a single layer 20 K colder cloud-free atmosphere a net longwave irradiance of -65 W m −2 would result, representative for early spring conditions. Scaling the same temperature difference (similar lapse rate) to a surface temperature of 0 • C in summer, results in a stronger loss of energy in the longwave range with -83 W m −2 . In addition, the stability in cloud-free conditions degrades from winter to summer, which contributes to more 10 negative longwave net irradiances in cloud-free conditions, where also the amount and vertical distribution of humidity plays an important role.
In cloudy conditions, a small difference between effective cloud-base and surface temperatures of, for example 2 K, results in -7 W m −2 and -9 W m −2 for the above mentioned simplified spring and summer surface temperatures, respectively. Thus, for the cloudy mode only a slight negative trend due to the Planck non-linearity towards summer can be explained. The major 15 driver hereby might be the lapse rate between surface and cloud base controlling also the sign of longwave net irradiances, while for optically thinner clouds (below a saturated longwave CRF) also the strength of cloud top inversion still seems relevant.
During winter and early spring, in cloudy conditions ( systems in early spring reported by Cohen et al. (2017) as well as the thermodynamic diversity observed during AFLUX might explain the broadened cloudy and cloud-free modes, which were also observed during SHEBA, and might be typical for early spring.
With the shift to a cloud-dominated boundary layer in mid to late spring, the boundary layer becomes thicker and less stable, which supports colder cloud base temperatures relative to the surface depending on the cloud base height. For the conditions 25 over sea ice during ACLOUD, clearly negative longwave net irradiances were observed due to a clearly separated cloud base from the surface and due to the absence of dense fog as during AFLUX. Especially, the thermodynamic profiles during N- May Fig. 2b) illustrate the impact of the preconditioned less stable cloudy and cloud-free states that result in a more negative longwave net irradiance distribution (Fig. 4b, dotted blue).
The clear difference between the cloudy modes observed during ACLOUD and N-ICE2015 in early summer relative to 30 SHEBA observations, seems to be less related to the average thermodynamic profiles shown in Fig. 2c ence was distributed around 0 K (distribution not shown), illustrating slightly more stable conditions, which might additionally contribute to the broader and more neutral cloudy longwave net irradiance mode.
Considering the modes as frequently preferred states of the longwave REB in the Arctic, the seasonal cycle of cloud fraction plays a crucial role. While during SHEBA winter, the cloud-free conditions outweighed the cloudy state (N-ICE2015 mostly 5 cloud-free due to shorter period), early spring conditions seem to be more diverse, trending already slightly to a more frequent cloudy state. In early summer, N-ICE2015 and ACLOUD data indicate a dominant and less variable cloudy mode with only occasional and shorter cloud-free periods, while SHEBA data show a more divers character. With less frequent and probably shorter cloud-free periods in early summer, the thermodynamic adjustments of the cloud-free atmosphere might be less pronounced, resulting in weaker surface based inversions (N-ICE2015 April/May or May/June cloud-free profiles, Fig. 2), which 10 would contribute to more negative longwave net irradiances in the cloud-free states.
The difference between cloudy and cloud-free modes can be understood as the climatological or measurement-based CRF (Stapf et al., 2021), likely also synoptic scale circulation pattern, or regional characteristics, might be of relevance for those mode structures, or more general, the impact of the presence of clouds in the Arctic.

Longwave four-mode structures in the marginal sea ice zone
In contrast to the long-term ice floe camps as operated during SHEBA or N-ICE2015, the airborne observations during the 20 ACLOUD/AFLUX campaign have been conducted in the heterogeneous MIZ. Therefore, the longwave net irradiances can be analyzed with respect to the surface type. In Fig. 5, the frequency distribution of longwave net irradiance is shown in a twodimensional space spanned by the all-sky surface albedo and longwave net irradiance. This type of plot provides combined information on surface (sea ice concentration, albedo, and shortwave illumination conditions, cloudiness) and cloud conditions (cloud base height, optical thickness). 25 For both seasons, four maxima (modes) are obvious. For each surface type, ocean (depending on the SZA and illumination/cloudiness α < 0.1 for ACLOUD, α < 0.2 for AFLUX ) and closed sea ice (depending on seasonal snow/ice property, α > 0.58-0.75 for ACLOUD, α > 0.7 for AFLUX ), the cloudy (less negative) and cloud-free (more negative) modes are displaced (longwave net irradiance space). Therefore, instead of two modes over homogeneous sea ice, the MIZ is characterized by a four-mode structure. Only on small scales, e.g. over sea ice with leads or nilas, the longwave net irradiances are shifted by the surface temperatures alone, since no rapid thermodynamic adjustment occurs there.
Especially the frequently observed off-ice flows (CAOs) during AFLUX strongly modify the REB in the MIZ and shift the longwave net irradiance modes by 50 W m −2 towards the ocean. During the ACLOUD campaign, with less dynamic Similar to the more variable mode structures in spring compared to early summer shown in Fig. 4, the appearance of the modes is more clearly distinguished and apparently more homogeneous during ACLOUD. Less variable surface temperature, 15 thermodynamic profiles and a mostly cloudy boundary layer with radiative opaque clouds and similar cloud base heights compared to more variable and dynamic conditions during AFLUX represents clearly a difference between both seasons.
Due to the special characters of CAOs and warm air intrusions and their major impact on the REB, these synoptic scale processes are discussed in more detail in the following. (e.g., Brümmer, 1997;Pithan et al., 2018). Besides strong turbulent fluxes, also the longwave REB is shaped by these extreme conditions.
During AFLUX, a series of CAOs (23 to 25 March 2019, Fig. 3) was observed driven by a low-pressure system located east of Svalbard. Airmass trajectories from these three days (not shown) indicate that the air mass originated from further south, namely from the Norwegian and Barents Seas, advected around a low-pressure system located in the east of Svalbard. These The presence of leads can influence the REB of the sea ice in the MIZ during CAOs by two processes. The first one is the 20 production of thin, low clouds in the boundary layer evolving downstream of leads ( Fig. 6a above sea ice) and the second one is a general warming and increase of the moisture content in the boundary layer.
Above sea ice, the observed longwave net irradiances fluctuate strongly between cloud-free scenes (-60 to -70 W m −2 ) and values up to -20 W m −2 below optically thick clouds. The thin and partly fog-like low clouds downstream of the leads produced a radiative transfer based longwave CRF of 10 to 30 W m −2 . The simulated cloud-free downward irradiances (indicator for 25 atmospheric temperature) shown in Fig. 6b increase for that day between the northern and southern east-west low-level section by 7 W m −2 , while the southern section exhibits already a 5 K warmer surface (-20 • C). Also during 31 March (no low-level clouds above sea ice), a steady increase of the simulated cloud-free downward longwave irradiance (Fig. 6b)   the absence of clouds (Fig. 6b, 31 March near the ice edge). Even in the presence of evolving roll clouds, strongly negative longwave net irradiances between -40 and -70 W m −2 were observed during AFLUX (Fig. 6b). and illustrate this compensation. The thermodynamic profile adapts relatively slowly to the warmer ocean surface. Thereby, the longwave net irradiances increase continuously (Fig. 6b, e.g. 24 March 2019) with increasing distance from the ice edge, which is primarily caused by the temperature adaption of the boundary layer (simulated longwave downward irrradiance) and effective cloud base temperature to the new surface temperature. The potential of CRF increases downstream, in addition to the increasing cloud optical thickness and cloud fraction, due to the transformation of the thermodynamic profile toward a negative lapse rate and weaker cloud top inversions. It is important to note that the temperature of the free troposphere often remains unchanged and mainly the evolving moist boundary layer is warming rapidly along the path southward (exemplary illustrated in Fig

Evolving boundary layer radiative energy budget
The unique feature of strongly negative longwave net irradiances during CAOs might become relevant also for the REB of 10 the entire boundary layer, which partly converts the upward irradiance by absorption and re-emission. The components of boundary layer energy budget were quantified by Brümmer (1996Brümmer ( , 1997 during multiple flights in CAO with in situ box pattern close to the surface and at the cloud top. In contrast to their approach, we quantify the contribution of radiative fluxes to the boundary layer REB using idealized radiative transfer simulations of thermodynamic profiles observed by the dropsondes released during AFLUX. We do so because the highly variable in situ radiation observations in the heterogeneous conditions 15 during CAOs might be particularly error prone for this application. A further advantage of the usage of radiative transfer simulations is that it enables the analysis of continuous vertical profiles of longwave net irradiances and source and sink terms of the REB in the boundary layer column. This approach also allows the computation of potential radiative temperature change rates, which cannot be derived reliably from in situ observations due to the heterogeneous cloud structures in CAOs. The Over sea ice the presence of thin clouds induces a longwave net irradiance divergence between 28 and 43 W m −2 throughout the entire boundary layer, representing likely the major energy sink with radiative temperature changes of up to 0.4 K h −1 (entire boundary layer). As soon as those clouds reach the open water at the ice edge, the absorption of longwave irradiance in 5 the cloud base will reach, or even exceed, the divergence at the cloud top depending on the individual cloud optical thickness and boundary layer height. The stronger the temperature gradient between air and sea surface temperature and the thicker the boundary layer (relatively cold cloud top), the stronger is the contribution to the boundary layer warming within the first 1 to 2 hours. In the frequently observed cloud-free area close to the ice edge (Gryschka et al., 2014) (before the development of roll clouds), the longwave irradiances will always contribute to a warming of the boundary layer as illustrated by the cloud-free at the cloud base weakens due to the decreasing spread between surface and cloud base temperature. At the same time, the divergence at the radiative cloud top increases steadily up to 100 W m −2 , so that this layer represents a strong sink of energy for the entire boundary layer. Consequently, with increasing distance from the ice edge a cooling of the entire boundary layer by the longwave irradiances will dominate. In case of an increasing cloud base height (spread between effective cloud base and surface temperature) the base warming and ocean surface cooling might increase again further downstream, which is not 5 considered in the simulations as it was not observed during AFLUX.
The calculated radiative temperature change rates (Fig. 7b) 8a) indicate that the cloud base temperature was consistently colder than the surface resulting in purely negative longwave net irradiance. Also, due to the absence of surface based fog, the surface still lost energy in the longwave wavelength range despite the advection of warm air. The observed longwave net irradiances over the ocean were slightly more negative (-10 W m −2 to -15 W m −2 ) due to a cloud base around 300-400 m. The first section above sea ice (optically thick and low clouds) shows For the first two cases, a direct warming effect on the surface due to positive longwave net irradiances was not given. This was observed, however, during the 21 March 2019, where optically thick clouds in levels between 900 m and 1400 m were 30 embedded in a warm air mass from a strong warm air intrusion event the day before (Fig. 8i, Fig. 3). During the low-level flight from east to west, the surface temperatures decreased from -8 • C to -15 • C and indicated along with the temperature profiles that the inversion was not necessarily surface-based, likely due to the heterogeneous sea ice observed during that flight or the radiative impact of the clouds on the surface. The potential cloud base temperature was found in the east and west at around -10 • C. Consequently, as the surface temperature changed, the surface longwave net irradiance increased from east to west by almost 25 W m −2 , peaking in positive 17 W m −2 with a strong direct warming effect on the surface (surface gains energy). In the time series of surface brightness temperatures (Fig. 3), this event exhibits the warmest sea ice temperatures observed during AFLUX. After about an hour, due to an extensive stationary stair case pattern in the west, the light blue profile (Fig. 8i) was observed, illustrating the dynamic conditions during that flight. The atmospheric temperature decreased above the near-surface inversion by up to 3 K. Consequently, for the last low-level section towards southeast already a neutral lonwave net irradiance day, strongly positive total net irradiances (Fig. 8f) were observed towards the west, due to a decreasing cloud optical thickness enhancing the shortwave transmissivity.
The three cases underline the importance of thermodynamic profile transformation in combination with the cloud macro- The total CRF obtained for both campaigns is presented together with the averaged satellite-derived sea ice concentration during ACLOUD (and AFLUX) in Fig. 9. It becomes clear that the variability of involved parameters makes the MIZ a truly complex region. 20 On larger scales the tendency of warming effects of clouds above sea ice and cooling effects above the ocean can be seen in To illustrate the role of ice fraction in the transition from a warming to a cooling effect of clouds in the MIZ, the CRF, accounting for the surface-albedo-cloud interaction, is shown in Fig. 10 as a function of the retrieved cloud-free albedo. In addition, mean values of observed CRF for certain ranges of ice fraction and SZA are shown, as well as radiative transfer 30 simulated values using the averaged atmospheric profile and LWP during ACLOUD above sea ice ( Fig. 1 and 2) as a reference.
In the simulations, a snow grain size representative for fresh snow as observed during AFLUX and the beginning of ACLOUD is used with a specific surface area of 28 m 2 kg −1 (snow grain size of 117 µm) and an impurity load of 0.1 ppmw. The simulated . Distribution of total CRF obtained for the ACLOUD/AFLUX campaigns during cloudy scenes (LWP > 1 g m −2 ). Campaign averaged satellite derived sea ice concentration (Spreen et al., 2008) of ACLOUD in the grey color scale. 80 % (red) and 15 % (light blue) isolines of I f during ACLOUD (solid) and AFLUX (dashed). Values of total CRF can exceed the color scale. cloud is located in a frequently observed vertical extent between 100 and 300 m and an average cloud droplet effective radius of 8 µm typical for this region (Mioche et al., 2017).
During early spring with weak solar insolation (SZA < 80 • ), clouds tend to warm the Arctic surface, regardless of the surface type. With decreasing SZA and constantly high snow albedo values in early spring, the sea ice concentration becomes the major surface parameter that controls the shortwave cooling potential of clouds. In comparison to the simulations, the 5 observed mean CRF in the MIZ is mostly found in the lower range of ice fraction values of the simulations. This might be an indication that for fractional ice cover, the remaining ice floes might be less reflective than pure snow scaled as a function of sea ice concentration. A possible reason could be the presence of more new and thin ice in the vicinity of broken ice floes, which have a lower albedo relative to the ice-fraction-scaled snow of early spring conditions. For ACLOUD, also a stronger melt stage of the sea ice towards the open ocean might have reduce the albedo. In mid-June, even for homogeneous sea ice, the 10 daily mean total CRF shifts toward a cooling effect due to the beginning melt season and the related decrease in surface albedo .  Figure 10. 2D histogram of total CRF during ACLOUD/AFLUX as a function of retrieved cloud-free surface albedo during cloudy scenes (LWP > 1 g m −2 ). Overlayed scatter plot shows the mean total CRF/cloud-free albedo for certain SZA (scatter) and I f (isolines) ranges. Thin dashed lines represent idealized simulations accounting for the surface-albedo-cloud interaction as a function of SZA (horizontal isolines) and ice fraction (vertical isolines) assuming a fixed cloud with a LWP of 50 g m −2 and snow specifications given in the text. optical thick clouds above sea ice, like during ACLOUD, mainly due to the shortwave CRF (depending on SZA and LWP), as the longwave CRF is close to saturation.
Optically thinner clouds during AFLUX influenced also the difference between surface albedo in cloudy and cloud-free conditions (in addition to other parameters). On average, the albedo was 0.014 higher in cloudy conditions during AFLUX. As expected, this indicates a much weaker impact of the surface-albedo-cloud interaction compared to late spring early summer 5 conditions during ACLOUD , due to fresher and colder snow, optically thinner clouds, and higher SZA. For clouds with a equivalent LWP of below 10 to 15 g m −2 often also a higher cloud-free surface albedo was retrieved compared to cloudy conditions. This weakened the shortwave cooling effect of clouds, due to the strong shift between direct-dominated and diffuse surface albedo for high SZA. Nevertheless, on average, a 25 % stronger shortwave cooling effect (-16.8 instead of -13.4 W m −2 for a LWP > 1 g m −2 ) was induced by the surface-albedo-cloud interaction during AFLUX, compared to a 10 doubling of the shortwave CRF during ACLOUD .
From spring to early summer, the ice fraction has to be considered together with the cloud optical thickness as the main parameter controlling the CRF as the snow surface albedo during this period shows only minor fluctuations. The radiative transfer simulated grid in Fig. 10 using fixed snow properties and LWP underlines these findings. However, the diversity of cloud conditions in this region prevents further conclusions or estimates of specific ice fraction values that represent the turning 15 point from cooling to warming effects of clouds.
Radiative transfer simulations further indicate that optically thick clouds above the open ocean in the presence of solar radiation are capable of compensating the radiative effect of a vanished sea ice only for an already low sea ice albedo in the beginning melt season. The increase of LWP which would be needed to maintain net irradiances during the transition from sea ice to the open ocean under these conditions depends on the SZA and the initial LWP above sea ice. Even the highest SZA of 55 • would require an increase of more than 70 g m −2 of LWP. Insufficient statistics regarding cloud properties above the open 5 ocean and sea ice limit further conclusions on this topic, but this compensation appears unlikely in the observed season and conditions (prior to the onset of melting sea ice) and might be more of relevance in the melt pond dominated season.

Longwave cloud radiative forcing
In theory, the longwave CRF of clouds is characterized by a strong non-linear dependence of CRF with increasing cloud optical 10 thickness for optically thin clouds. In the MIZ, the LWP and the atmospheric thermodynamics change on small scales and the seasonal characteristics of atmospheric profiles were discussed in section 3.2. Both aspects might control the longwave CRF potential (convergence for high LWP) of clouds. In Fig. 11a the observed relation between LWP and radiative transfer based longwave CRF during AFLUX and ACLOUD in the MIZ northwest of Svalbard is shown as a 2D histogram together with the frequency distribution separated for both campaigns (Fig. 11b). 15 A strong increase of longwave CRF for an equivalent LWP below 25 g m −2 to values between 60 and 80 W m −2 illustrates the sensitivity of downward irradiance with respect to clouds. The importance of radiative opaque low-level clouds becomes obvious in Fig. 11b, where distinct modes for values above 50 W m −2 underline that the majority of cloud conditions in this region are represented by this cloud type characterized by a powerful longwave warming effect on the surface. During AFLUX (early spring), a higher frequency of intermediate longwave CRF values are caused by more diverse conditions, including 20 optically thin, almost haze-like, low clouds or partly mid-layer clouds or even cirrus, which were less frequently observed during ACLOUD as single cloud layers. An increased ice water content of clouds during AFLUX likely contributed to a weaker longwave CRF for optically thin clouds.
For radiative-opaque conditions (LWP > 50 g m −2 , ∆F lw > 50 W m −2 ), the modes of longwave CRF average 77 W m −2 for ACLOUD. As was shown in the seasonal longwave net irradiances mode structures over sea ice in Fig. 4, the climatological 25 CRF (spread between cloudy and cloud-free modes), shows a distinct seasonal dependence due to thermodynamic specifications. Also the radiative transfer-based estimate for AFLUX (Fig. 11b blue) represents this feature with a reduced mean value for opaque low-level clouds of 71 W m −2 . The difference between these two CRF modes (Fig. 11b) is consistent with the reported difference of the climatological CRF of 56 W m −2 during AFLUX, and 63 W m −2 during ACLOUD. Nevertheless, the presence of clouds appears to be less impactful on the longwave REB as estimated from the radiative-transfer based CRF, 30 due to thermodynamic adjustments (Stapf et al., 2021).
To quantify the impact of thermodynamic profiles on the radiative transfer-based CRF, simulations of the average AFLUX and ACLOUD atmospheric thermodynamic profile (Fig. 2) are shown along with the observations in Fig. 11a. The simulated longwave CRF (above sea ice) converges to similar values of 77 W m −2 during ACLOUD and 72 W m −2 during AFLUX, which is the direct cause of the less negative cloud-free reference net irradiances during AFLUX, induced by colder temperatures and a more pronounced temperature inversion above the boundary layer (Fig. 2). The impact of the lapse rate and temperature is even more powerful towards the open ocean as indicated by the simulations (Fig. 11a), which significantly increase the warming potential of clouds especially during AFLUX. While during ACLOUD regularly higher values of CRF 5 around 80 to 90 W m −2 were observed, during AFLUX these values were surprisingly not observed. However, low-level flights with larger distance from the ice edge were limited during AFLUX, which might explain the lack of observed stronger longwave warming effects of clouds above the open ocean. During the CAO on 24 March 2019 (AFLUX), discussed in section 4.3.1, the longwave CRF steadily increased downstream of the ice edge and a further increase appears likely.
Besides the regional and more variable characteristics of CRF observed during AFLUX/ACLOUD in the MIZ, the longwave Counts Figure 12. Total CRF as a function of the derived LWP during ACLOUD observed for scenes with a retrieved cloud-free surface albedo between 0.75 and 0.85 for the specific SZA accounting for the surface-albedo-cloud interaction. Simulated range of total CRF for the average ACLOUD atmosphere (Fig. 2) with a cloud between 100 m and 300 m and the observed SZA and surface albedo range during the campaign.
Open circles show the averaged observations (8 g m −2 bins).
around 75 W m −2 appears to be a good estimate for the low boundary layer clouds frequently observed in this region from spring to autumn.

Total cloud radiative forcing
Combining the longwave and shortwave CRF in early spring (AFLUX), the total CRF above sea ice was positive throughout the campaign due to the high surface albedo and SZA. For these conditions, the total CRF is primarily controlled by the longwave 5 CRF distribution (Fig. 11b), and thus, also the total CRF distribution appears similar. The opaque cloudy total CRF mode during AFLUX is shifted slightly from the distribution shown in Fig. 11b to around 50 W m −2 by the weak shortwave cooling effect. In late spring and early summer (ACLOUD), however, the shortwave cooling influence on the total CRF becomes more relevant.
In Fig. 12 the relation of observed total CRF and LWP during ACLOUD is illustrated for a selected range of retrieved 10 cloud-free albedo values between 0.75 and 0.85. The simulated CRF of the averaged ACLOUD atmospheric profile and an assumed low-level cloud are show as a reference. The total CRF exceeds a tipping point for an increasing LWP, above which a significantly decreasing CRF is observed, potentially transitioning to a total cooling effect, depending on the SZA and surface albedo conditions.
The observations approximately peak in the LWP range between 15 g m −2 and 25 g m −2 , fluctuating between a total CRF of 15 20 g m −2 and 40 W m −2 . These values are representative for optically thin partly also broken cloud fields in low altitudes and have been frequently observed during ACLOUD. These clouds are known for a strong warming effect on the surface (Bennartz et al., 2013). More often however, clouds in the range between 30 to 55 g m −2 were observed for this surface albedo range.
These clouds are characterized by a slightly weaker warming effect over sea ice and a more homogeneous character (less variable total CRF).
In general, the observed distribution follows the range described by the simulations, expect for low LWP. In this range, the total CRF frequently exceeds the theoretical weak CRF for those optically thin clouds. It should be noted that the LWP is derived by the shortwave transmissivity. Especially in the case of broken thin cloud conditions, the direct component of 5 downward shortwave irradiance is partly not attenuated by clouds due to openings in the cloud field. In combinations with the still present diffuse shortwave irradiance by the surrounding clouds, the shortwave transmissivity suggests cloud-free conditions or even exceeds the value for cloud-free conditions due to the 3D effects. The longwave irradiance is mainly diffuse and thus even for small cloud openings in the cloud field high values of downward longwave irradiance are maintained. In combination, a stronger total warming effect of clouds can often be observed during these conditions compared to values expected by the 10 1D simulations. With increasing LWP of a cloud field, also the probability of openings decreases and these effects become less frequent. The total CRF values distributed below the simulated range can be attributed to conditions with a higher cloud base and a reduced longwave warming effect for a certain LWP value, which were rarely observed during ACLOUD.
If the 3D features are relevant for the areal-averaged surface REB with respect to simplified 1D radiative transfer and cloudfraction assumptions, for example in regional climate models, or if they can be neglected on larger areas might be estimates by 15 3D radiative transfer studies similar to Benner et al. (2001), whereby the scale of heterogeneities and structures in the cloud field might be important for the non-linear increase of longwave CRF for optically thin clouds.

Total cloud radiative forcing and Arctic amplification above sea ice
As was shown in the previous sections, the radiative transfer simulations assuming a frequently observed low layer of Arctic boundary layer clouds represent the conditions during both campaigns fairly well and might be justified for this cloud type 20 from spring to autumn in agreement with other studies in this region. Using these simulations, the tipping point of maximum total CRF (as shown in Fig. 12) was estimated for the observed atmospheric conditions in Fig. 13a depending on SZA, LWP, and cloud-free snow albedo for various grain sizes. This point also represents the pair of values of the highest total (solar plus longwave) net irradiances on sea ice (background of Fig. 13a).
Those tipping points are of great importance in order to assess whether future changes in LWP (cloud optical thickness), 25 presumably forced by the Arctic amplification, will induce an increase or decrease of CRF and net irradiances over sea ice.
These simulations hold for an increase of LWP in already cloudy conditions (typically high cloud fraction in summer), where the radiative transfer-based CRF can be used to estimate trends. In case of increasing cloud fractions, one might rather consider a observation-based CRF approach (analysis of longwave modes structures) to quantify the impact of the increasing presence of clouds based on the observed REB climatology (Stapf et al., 2021). surface REB. It should be mentioned that not only the absolute values of net irradiances and CRF depend the snow grain size, but also the expected trend in net irradiances (e.g. May). This relation clearly illustrates that potential changes in net irradiances in the future Arctic will depend on a delicate representation/estimate of cloud optical thickness and sea ice properties.

Summary and conclusion
We have presented REB and CRF observations in the heterogeneous environment of the MIZ northwest of Svalbard in spring 5 and early summer. The REB in this region is driven by the sea ice concentration and the thermodynamic transition between the relatively warm ocean and the cold sea ice. Small-scale fluctuations of surface properties and thermodynamic profiles make this region a complex terrain for radiative processes. Several characteristic features have been observed during AFLUX and ACLOUD.
The longwave net irradiance two mode structure (above sea ice) is subject to a seasonal shift toward more negative irradiances 10 in both modes and a more frequent cloudy state towards the summer. The individual distribution of a state is likely linked to seasonal thermodynamic profile characteristics, which control the distribution of cloud-free and cloudy net irradiances.
Compared to SHEBA, the distributions north of Svalbard during spring and summer obtained from N-ICE2015, AFLUX, and ACLOUD, indicate a shift toward the negative in both states, likely due to less stable near-surface atmospheric conditions in this region. 15 For both surface types in the MIZ (open ocean, sea ice), two separated modes of longwave net irradiance are observed.
On small horizontal scales in the MIZ they are induced by the obvious change of surface temperature between sea ice and leads/nilas. On larger horizontal scales the adjustment of the atmospheric thermodynamics and cloud properties relocate the longwave modes over the open ocean. Strongly negative longwave net irradiances in cloudy as well as cloud-free conditions represent a special feature in the MIZ due to the large spread between air/atmospheric and surface temperatures, especially 20 during CAOs.
In the evolving boundary layer of CAOs, the longwave radiation represents a sink of energy, counteracting the strong sensible and latent heat fluxes that warm the boundary layer. Our simulations indicate that only for an area close to the ice edge, in the presence of particularly strong temperature gradients, a short-term warming contribution due to longwave radiation to the boundary layer warming of up to 40 W m −2 can be expected. This contribution transforms into a cooling of the boundary layer 25 by up to -75 W m −2 further downstream, depending on the individual cloud and thermodynamic evolution.
The presence of leads/nilas upstream of the ice edge induces enhanced downward longwave irradiances via cloud production and boundary layer warming/moistening during CAOs and thus likely contributes to a warming of the ice floes in the MIZ.
The impact of warm air intrusions on the local sea ice REB in the MIZ is sensitive to the thermodynamic profiles along the flow and the vertical location of clouds. We observed that the strongest direct warming potential of clouds in the longwave 30 wavelength range is found for an advection of clouds aloft embedded in the warm air masses, rather than fog embedded directly in strong surface based temperature inversions, and not necessarily for the optically thickest clouds in the presence of solar irradiance.
In spring and early summer with a relatively constant and high surface albedo of sea ice, the sea ice concentration is the primary parameter controlling the total CRF in the MIZ.
The low altitude of the observed clouds (especially during ACLOUD) enabled the reproduction of longwave CRF by simplified radiative transfer simulations. The results indicated a seasonal dependence of maximum longwave CRF (increasing towards summer) driven by atmospheric thermodynamic characteristics. In comparison to other studies in the area around 5 Svalbard, values of around 75 W m −2 seem to be typical.
The maximum total CRF was found for optically thin and partly also broken low-level clouds (3D effects of combined diffuse longwave and direct shortwave irradiance), whereby a delicate function of LWP and sea ice albedo controls the tipping point, from which a further increase of cloud optical thickness will result in lower net irradiances and weaker total CRF, as often observed during ACLOUD.

10
The MIZ represents a complex environment regarding radiative transfer and thermodynamic processes. Aside from representing only a small area of the Arctic, understanding and modelling the processes in this area is a challenge, but also an opportunity to improve the understanding of the REB in the inner Arctic. The extreme changes of boundary layer thermodynamics and its impact on the longwave REB in the MIZ, likely overdraw the importance of the thermodynamics. However, even slight shifts in the longwave net irradiance modes structures above homogeneous sea ice underline the importance of the 15 linkage between low-level atmospheric profiles and stability, in cloudy as well as cloud-free conditions.
The agreement between radiative transfer simulations of simplified low-level clouds and averaged thermodynamic profiles, is encouraging, because improvements in the representation of the boundary layer thermodynamic profile in climate models might enable a powerful improvement of the surface REB during that season.
The delicate tipping point of total CRF and surface net irradiances, however, underlines the requirements of a precise repre-20 sentation of the cloud/surface regime in the specific season, as the impact of changing cloud fraction and optical properties in a future Arctic depends on a complex interaction of microphysical snow and ice properties and, precisely, these cloud properties.
The sign of cloud feedbacks depends on the cloud/surface regime in the individual models, which will likely vary on regional, seasonal scale, or in terms of the observed Arctic amplification on the real character of the rarely observed low-level clouds in the Arctic. "Inserting" specific cloud and surface regimes, either from climate models, or observed from an increasing number 25 of datasets in the inner Arctic (e.g., Achtert et al., 2020), in individual radiative lookup tables, might give some evidence in expected trends.
Being in the right regime of cloud and surface properties is a challenge for climate models. Considering insufficient statistics on real cloud properties over the entire inner Arctic, as well as diverging/variable estimates from satellite, models, and observations (Zygmuntowska et al., 2012;Cesana et al., 2012;Achtert et al., 2020), conclusions on potential 30 future cloud impact in the Arctic appear challenging.
Data availability. The combined ACLOUD and AFLUX low-level dataset (REB, CRF) is made (currently) available on the PANGAEA database. The ACLOUD broadband irradiance dataset is available in Stapf et al. (2019) and is made (currently) available for AFLUX on PANGAEA. Broadband radiation data and radiosounding data of the N-ICE2015 campaign are used from Hudson et al. (2016) and Hudson et al. (2017), for SHEBA from Persson (2011) and Moritz (2017 Air temperature, relative humidity, and pressure in situ profiles from both aircraft are used from (Hartmann et al., 2019), dropsondes from Ehrlich et al. (2019a) and Becker et al. (2020). The radiosoundings from Ny-Ålesund are available from (Maturilli, 2020) and for Polarstern 5 from (Schmithüsen, 2017). Relevant for the derivation of CRF above the ocean is the change between diffuse (dotted line in Fig. A1a) and the albedo in theoretical cloud-free conditions (dominated by direct radiation). Similarly, the CRF estimates using the observed ocean albedo CRF by up to 40 %. Especially during CAOs in early spring with strong wind periods the cooling potential of clouds will be enhanced while for weak wind areas or leads embedded in sea ice the cooling effect will be weaker.
An additional source of uncertainties represents the background aerosol, which can not be measured in case of cloudy conditions. An increasing AOD will reduce the fraction of direct irradiance and consequently manipulate the cloud-free ocean 25 albedo required for the estimate of CRF. Looking at the potential uncertainties especially for SZA above 80 • the derivation of CRF becomes challenging and uncertainties in the radiative transfer simulations become critical. Additional uncertainties remain due to the impact of reduced fetch in sea ice leads, which might reduce the surface roughness and whitecaps for a Appendix B: Estimate of radiative flux convergence/divergence during cold air outbreaks To estimate the absolute contribution of radiative (diabatic) processes to the temperature changes of the boundary layer during CAOs, radiative transfer simulations of thermodynamic profiles obtained from dropsondes were used. The temperature profiles for a CAO on 23 March 2019 (AFLUX) are shown as an example in Fig. A2a. In the simulations, the clouds have been assumed to be surface-based (frequently observed during AFLUX CAOs) with a sub-adiabatic liquid water content (LWC) 5 profile (constant fraction f ad = 0.5 of the adiabatic value) with a cloud droplet effective radius of 8 µm. The inversion base is assumed to be the level of the cloud top. The northern most profile (coldest boundary layer) was recorded over the sea ice covered area (assumed surface albedo 0.8), while the following dropsondes were ejected above the open ocean, whereby a simplified, constant surface skin temperature of zero degrees was assumed.
For the sea ice profile, the simulated profiles of longwave net irradiance (Fig. A2b) indicate a flux divergence (radiative 10 cloud top cooling) throughout the whole boundary layer due to the optically thin cloud. As soon as these cold and thin clouds are advected above the sea ice edge (located between the two northernmost profiles) the surface temperature changes by almost 20 K and induce strongly outgoing longwave net irradiances of in this case up to -72 W m −2 even in cloudy conditions. With the increasing optical thickness of the clouds, a part of the upward irradiance is absorbed in the cloud base (radiative cloud base warming) indicated by increasing longwave net irradiances, until at a specific altitude, the highest net irradiances are observed.
From this level, the relatively warm cloud emits radiation towards the cold upper atmosphere inducing the cloud top cooling (radiative cloud top, decreasing longwave net irradiances). The diabatic temperature change rates (potentially) caused by the absorption or emission of radiation in a certain atmospheric layer: with the air density ρ and specific heat c p are shown in Fig. A2d. With increasing LWC, the net irradiance in the cloud will saturate close to zero and will reduce the vertical extent of the radiative cloud top. This enhances the potential temperature change rate but not the integrated flux divergence. The cloud top cooling potential during CAOs increases with increasing distance from the ice edge. This is caused by a sufficiently high LWC inside the cloud layer causing maximum (neutral) 10 longwave net irradiances inside the cloud (Fig. A2b) and, more importantly, by a decreasing temperature spread between the cloud top and the upper atmosphere (Fig. A2a). A relatively warm cloud top inversion results in a higher downward irradiance at the cloud top, which results, together with a colder cloud top temperature, in an weakly negative longwave net irradiance at the cloud top, commonly found over the sea ice or close to the ice edge. With the temperature adaption of the boundary layer to the new environment, also the temperature spread at the cloud top decreases, resulting in consistently increasing negative longwave net irradiances and divergence in the radiative cloud top downstream supported also by the non-linearity of the Plank radiation law.
For the radiative cloud base, the temperature adaption of the boundary layer reduces the spread between surface and air temperature and consequently the surface longwave net irradiances become less negative (observations in Fig. 6b) and the flux convergence at the cloud base downstream decreases.

5
The total boundary layer (surface to cloud top) flux divergence in this particular case is always negative (cooling the entire boundary layer). Only the first profile above the ocean shows the same amount of gain and loss of energy. In case of a strong air-surface temperature contrast in this area, also a gain of energy in the entire boundary layer is possible due to more negative net irradiances in the cloudy state which depends on the individual strength of the CAO. For cloud-free conditions close to the ice edge (dashed profiles Fig. A2b) the boundary layer will gain energy. Also in hypothetical conditions without roll 10 cloud formation the boundary layer strongly absorbs the upward irradiances, shifting towards a weak cooling downstream with increasing distance from the ice edge.
The shortwave convergence of fluxes is shown for two SZA in Fig. A2c and is in general weak and might only be relevant in late spring, where it has to be consider as a compensation of the flux divergence of the whole boundary layer in the longwave range.

15
Vertically integrated flux divergences and potential radiative temperature change rates are discussed in Fig. 7 for four different CAO observed during AFLUX with suitable flight pattern and dropsonde releases.
Author contributions. All authors contributed to the editing of the article and to the discussion of the results. JS processed the radiation data, merged the datasets, performed the radiative transfer simulations, analyzed the data, and drafted the article. MW, AE, and CL designed the