A Sensitivity Study of Arctic Air‐Mass Transformation Using Large Eddy Simulation

Arctic air mass transformation is linked to the evolution of low‐level mixed‐phase clouds. These clouds can alter the structure of the boundary layer and modify the surface energy budget. In this study, we use three‐dimensional large eddy simulation and a bulk sea ice model to examine the lifecycle of clouds formed during wintertime advection of moist and warm air over sea ice, following a Lagrangian perspective. We investigate the stages of cloud formation, evolution, and decay. The results show that radiative cooling at the surface gives rise to fog formation which subsequently rises and transforms into a mixed‐phase cloud. In our baseline simulation, the cloud persists for about 5 days and increases the surface temperature by on average 17 °C. Sensitivity tests show that the lifetime of the cloud is sensitive to changes in the vapor supply at cloud top. This flux is mainly impacted by changes in the divergence rate; an imposed convergence decreases the lifetime to 2 days while an imposed large‐scale divergence increases the lifetime to more than 6 days. The largest difference in cloud radiative properties is found in the experiment with increased ice crystal number concentrations. In this case, the lifetime of the cloud is similar compared to baseline but the amount of liquid water is clearly depleted throughout the whole cloud sequence and the surface temperature is on average 6 °C cooler. The cloud condensation nuclei concentration has a weaker effect on the radiative properties and lifetime of the cloud.


Introduction
Arctic air mass transformation is the process whereby relatively warm, moist, mid-latitude air is advected over the Arctic pack ice, cools radiatively, and eventually transforms into a polar air mass characterized by low humidity and a strong surface-based inversion (Cronin et al., 2017;Cronin & Tziperman, 2015;Curry, 1983;Pithan et al., 2014Pithan et al., , 2016Pithan et al., , 2018Wexler, 1936). Rapid and strong intrusions of warm and moist air to the Arctic occurs about 16 times every winter season (October to January) and the frequency has increased in December and January during the period 1990 to 2012 (Woods & Caballero, 2016). In the initial stages of the transformation process, near-surface air becomes saturated and clouds form, frequently consisting of both liquid and ice (so called mixed-phase clouds). Due to the high longwave opacity of the cloud, the bulk of the radiative cooling is shifted from the ice surface to the cloud top (the radiative cooling is reduced at the ice surface while it is intensified at the cloud top), yielding a substantial warming of the surface and subsequent impacts on sea ice thickness (Intrieri et al., 2002;Johansson et al., 2017;Shupe & Intrieri, 2004;Woods & Caballero, 2016).
Mixed-phase clouds are ubiquitous in the Arctic , and their aggregate radiative forcing exerts an important influence on Arctic climate. Several studies have shown that the amount of surface warming induced by mixed-phase clouds strongly depends on the total cloud water amount and phase (Cronin & Tziperman, 2015;Pithan et al., 2016;Shupe & Intrieri, 2004). Thus, it is important to understand the properties and lifecycle of these clouds, as well as their sensitivity to different environmental and meteorological conditions. Cloud parameterization schemes in general circulation models often struggle to simulate Arctic mixed-phase clouds, including the clouds formed during Arctic air mass transformation (Pithan et al., 2014(Pithan et al., , 2016Wesslén et al., 2014). Despite their thermodynamically unstable state, Arctic mixed-phase clouds often persist for periods of days, and the main cause of their persistence and rapid dissipation remains unclear (Morrison et al., 2011). Recent studies have shown that Arctic mixed-phase clouds in the spring-and summertime Arctic are sensitive to the number concentration of cloud condensation nuclei (CCN) and ice-nucleating particles (INPs) as well as the large-scale divergence rate (Loewe et al., 2017;Solomon et al., 2018;Stevens et al., 2018;Young et al., 2018). Pithan et al. (2018) concluded that more observations and model experiments following air masses on their paths to the Arctic are needed to improve our understanding of Arctic air mass transformation. Previous modeling studies of this problem have been conducted with 1-dimensional column models or 2-dimensional cloud-resolving models (Cronin et al., 2017;Cronin & Tziperman, 2015;Pithan et al., 2014Pithan et al., , 2016, but never to our knowledge with a full 3-dimensional large eddy simulation (LES). Other studies using 3-dimensional LES to examine Arctic mixed-phase clouds have focused on relatively short time periods and specific stages of cloud development, (Loewe et al., 2017;Possner et al., 2017;Solomon et al., 2018;Stevens et al., 2018;Young et al., 2018). In the present study, we couple a 3-dimensional LES model to an interactive sea ice surface model to study in detail all stages of cloud development during an idealized wintertime Arctic air mass transformation, as well as the impact of the clouds on the surface energy budget. We conduct a suite of simulations to examine how different parameters affect the cloud persistence and cloud lifetime. Furthermore, we examine the sensitivity of the liquid water content and ice water content to different meteorological and microphysical parameters during the cloud formation, cloud lifetime and cloud decay. Our aim is to provide insight on critical processes governing the cloud formation and dissipation, and help guiding upcoming field campaigns dedicated to Arctic air mass transformation.
In section 2 we describe the model and the setup of the simulations. Section 3 shows the results from each simulation and depicts comparisons between them. Section 4 discusses the results and we present the conclusions drawn from the experiments in section 5.

Description of the LES
We used the three-dimensional Large Eddy Simulation (LES) model MIMICA (Savre et al., 2014) to examine Arctic low-level mixed-phase clouds. In MIMICA, a 1.5 order subgrid scale closure is applied with turbulent kinetic energy (TKE) as a prognostic variable. For the surface turbulent fluxes, the model employs the Garrat formulations (Garratt, 1994) which are based on Monin-Obukhov similarity theory. The default version of the model considers five types of hydrometeors (cloud droplets, rain drops, pristine ice crystals, graupel, snow), where the mass and number concentration of each hydrometeor is explicitly calculated. For simplicity, the simulations performed in this study only include pristine ice, i.e. the categories graupel and snow are excluded, and no riming or accretion of ice crystals is allowed. Any growth of ice particles is only due to vapor deposition. A sensitivity test performed with the full ice microphysics switched on showed no substantial impact on the results. The size distribution of the hydrometeors is obtained using prescribed gamma functions (cf. Savre et al., 2014) and simple power laws are employed to describe the terminal fall speeds. The habit of the ice crystals was selected to be plates for our experiments due to the air temperature range existing within the cloudy area. Warm cloud microphysics is parameterized according to Seifert and Beheng (2001). The supersaturation is explicitly calculated at each time step, i.e. no saturation adjustment is assumed (Morrison & Grabowski, 2008).
Droplet activation and ice nucleation are described by simple parameterizations: the number of cloud droplets formed is equal to the number of activated CCN which is calculated following Khvorostyanov and Curry (2006). All particles in the simulations were assumed to consist of ammonium sulfate. The ice crystal number concentration is relaxed to a fixed background concentration following Savre et al. (2014) and Ovchinnikov et al. (2011Ovchinnikov et al. ( , 2014. Radiative transfer is modeled using a multiband 2-4 stream approximation (Fu & Liou, 1993), which takes into account not only the mixing ratio of water vapor but also that of cloud droplets and ice particles.
Above the top of the model domain (at 4 km) a prescribed vertical profile extending to the top of the atmosphere is used in the radiative transfer calculations. Periodic lateral boundary conditions are used and in the vertical direction, a sponge layer was applied at 3600 m to damp gravity waves. In the present study, the surface was treated as a simple slab represented by a 2 m thick layer of sea ice, covered with a thin layer of snow (cf. next sub-section).

Sea Ice Model
For the purposes of this study, MIMICA was extended to include a simple horizontally uniform slab consisting of a 2 m thick sea ice layer, covered with a thin layer of snow (default value of 20 cm). Figure 1 shows a schematic of the model. The surface roughness height was set to 0.0004 m to represent a snow-covered flat surface following Stull (1988) and the combined sea ice-snow slab thermal conductance (γ) was calculated using the parameterization of Maykut (1978): where k i (2 W m −1 K −1 ) is the thermal conductivity of ice, k s (0.31 W m −1 K −1 ) is the thermal conductivity of snow, H is the ice thickness and h s is the snow thickness. The surface albedo and longwave emissivity were set to 0.8 and 0.92, respectively. The one-dimensional slab surface model was coupled to the atmospheric part of the LES taking as input the domain-averaged downward longwave and shortwave radiative flux and providing the domain-average surface skin temperature as output. The temperature at the bottom of the slab was held constant at −2°C which is the sea water freezing temperature corresponding to 35 psu sea water salinity. The temperature gradient within the combined sea ice-snow slab was assumed to be linear and its thickness was held constant. Hence, the following energy conservation equation was solved for the upper boundary of the ice slab (Untersteiner et al., 1986): with the following definitions: where F swd is the solar downward radiative flux (included here for completeness, but set to zero in our simulations), I o is the transmitted solar radiative flux, F lwd is the infrared downward radiative flux, F lwu is the infrared upward radiative flux, F SH is the sensible turbulent heat flux, F LH is the latent turbulent heat flux, F c is the conductive heat flux through the sea ice. T f and T o are the temperatures at the bottom and at the surface of the sea ice layer, respectively. L is the heat loss due to melting of ice or snow, i o (0.35, Untersteiner et al., 1986) is the fraction of net shortwave radiation transmitted below 10 cm of sea ice, ρ ice is the density of sea ice, L ice (330 kJ kg −1 ) is the heat of fusion of ice. The (2) equation was simplified by setting L = 0 due to the constant thickness of the slab and took the following form: where ε (0.92) is the snow emissivity, ρ air is the density of the air, c P is the specific heat of the air, C s and C e are the bulk transfer coefficients for the sensible and latent turbulent heat fluxes (Deardorff, 1968), u is the wind speed at the reference height (height of the first vertical level of the model), T a is the air temperature at the reference height, L v is the latent heat of vaporization, q a and q o are the values of the specific humidity of the air at the reference height and at the surface.

Experiments
We used the LES to simulate an idealized case where maritime air is advected over Arctic sea ice for five days, following the Lagrangian column approach used in previous work (Cronin & Tziperman, 2015;Pithan et al., 2014Pithan et al., , 2016. This type of approach enables us to interactively calculate the surface turbulent fluxes during the whole model simulation. If a shorter time period was considered, as in previous LES studies by e.g. Loewe et al., 2017;Possner et al., 2017;Solomon et al., 2018;Stevens et al., 2018 andYoung et al., 2018, prescribed surface fluxes could be feasible option. However, as shown in Section 3, the surface turbulent fluxes do vary significantly throughout the simulation time, affecting the surface temperature and the properties of the simulated clouds. Also, this approach allows us to be consistent with previous studies and to examine the whole cloud lifecycle without using a large numerical domain to encompass the horizontal extent of the intrusion. It should be noted that the Lagrangian approach does have some limitations. The most important drawback is that the same piece of ice is continuously heated by the simulated cloud, which could lead to an overestimate in the surface temperature increase. We do not take into account the fact that in reality, the cloud would move over areas of unaffected sea ice. The date of the simulation was set to January 1 meaning that the shortwave radiation was zero. In the reference case, labeled "BASE", the initial temperature and moisture profiles follow Cronin and Tziperman (2015) and Cronin et al. (2017): the surface was set to 0°C decreasing with altitude according to the moist adiabatic lapse rate (8°C km −1 ). The initial surface relative humidity was set to 80% decreasing linearly with altitude to 20% at the top of the model domain (4000 m). These profiles are broadly realistic and representative for maritime air located near the Arctic sea ice edge. The horizontal winds were initially set to 5 m s −1 at all heights and they were nudged to the same value with a time scale of 12 h. The large-scale subsidence was set to zero. The grid spacing was 62.5 m in the horizontal direction and 15 m from the surface to 2500 m and 7.5 m from that level up to the top of the domain for the vertical. The finer resolution was applied in the upper part of the domain to better resolve gravity waves which were formed at the cloud top during cloud dissipation. The length and width of the domain was 6 km. The time step was set to 2 s. The default chosen maximum number concentrations of cloud droplets and ice crystals were 30 cm −3 and 1 L −1 respectively. These values roughly match observed concentrations of CCN and ice crystal number concentration (ICNC) in maritime air entering the Arctic region (Jackson et al., 2012;Young et al., 2018).
beneath the sea ice to the atmosphere is a main factor of the surface energy budget of the Arctic. To examine the role of this flux we increased the value of the surface thermal conductance from 0.6 to 1 W m −2 K −1 assuming that there is no snow layer above the sea ice (IBASE). The potential influence of the large-scale circulation was investigated in CONV and DIVER where a specified horizontal convergence and divergence rate, respectively, was imposed on the model domain. The rates were defined to be constant with height in the entire domain. No compensating tendencies were applied to prevent any excessive increase in temperature or drop in humidity due to the continuous effect of subsidence. The top boundary condition was treated as no-flux condition, i.e. the vertical velocity including the divergence term was assumed to be zero at the domain top.

Baseline Simulation (BASE)
Figure 2(a) shows the evolution of the horizontal-mean temperature profile over the simulation, while Figure 3(a) shows the evolution of liquid and ice water mixing ratios. Individual profiles of cloud droplet Table 1 List of Experiments. Please note that unless otherwise mentioned, parameters and processes are changed relative to BASE and ice crystal mixing ratios are illustrated in Figure S1. Radiative cooling of the relatively warm maritime air mass initially peaks near the surface and a surface-based temperature inversion rapidly develops; liquid condensation near the surface begins almost immediately and a fog consisting mostly of supercooled water droplets sets in within the first hour of the simulation ( Figure 3). Surface radiative cooling is reduced by the fog, which increases the downward longwave radiative flux. Consequently, the surface net longwave radiative flux (net LW sfc , defined positive downward) increases from an initial value of around −35 W m −2 to −6 W m −2 (Figure 4(a)) within the first few hours. Surface temperature (T sfc ) at this time is around −10°C ( Figure 5). Peak radiative cooling shifts to the top of the fog, destabilizing the layer and leading to entrainment and thickening of the fog layer; the temperature inversion shifts to the top of the layer, with a well-mixed profile below. After~2 hours, the fog droplets become large enough to allow drizzle. The consequent dehydration of the well-mixed layer shifts the condensation level off the ground, transforming the fog in to a stratocumulus layer after around 4 hours.
During the second stage of the cloud evolution (until 36 h), substantial drizzle production continues with a mean surface rate of about 0.5 mm day −1 (Figure 6(a)), and the cloud layer remains relatively shallow with a depth of~100-150 m (Figure 3(a)) and a liquid water path (LWP) which increases slowly from~30 to 40 g m −2 (Figure 7(a)). At the same time, ice crystals begin to form; the ice water path (IWP) increases from zero to around 8 g m −2 while ice precipitation at the surface increases to around 0.6 mm day −1 (Figure 6(a)). By the end of the first day, the cloud top is located within the temperature and moisture inversions which are both at around 500 m (Figure 2(a), 8(a)), while the cloud base is at around 350 m; the temperature profile is moist-adiabatic through most of the layer, but a shallow surface-based inversion forms near the surface (Figure 2(a)) so that the cloud layer becomes decoupled from the surface-turbulent kinetic energy (TKE) produced by cloud-top negative buoyancy production is not strong enough to penetrate the stable surface layer (Figure 9(a)). Evaporation and sublimation of cloud drops and ice crystals, respectively, leads to an increase in the vapor mixing ratio below the moisture inversion. Still, the aforementioned decoupling results in the formation of a shallow surface-based moisture inversion at about the same height as that of the temperature (Figure 8(a)), leaving the air close to the surface quite dry. As the cloud top rises and cools, the efficiency of the Wegener-Bergeron-Findeisen (WBF) process increases and the ice crystals grow at the expense of drizzle and water droplets. When the temperature reaches −12°C, at about 870 m and after 36 h, the difference between the equilibrium vapor pressure of liquid and ice is at its maximum, cloud drop size is reduced and drizzle formation is suppressed. The evaporation of drizzle below cloud base cools the air, stabilizes the cloud layer and reduces TKE. Once the drizzle stops at around 36 h ( Figure 6(a)), TKE increases abruptly (Figure 9(a)) and the neutral layer originating from the cloud base expands towards the surface, weakening the surface-based temperature and moisture inversions (Figure 2(a), 8(a)). In addition, the loss of cloud water to drizzle ceases and so the cloud's liquid water content increases, which in turn enhances cloud-top radiative cooling and TKE generation. The positive

Journal of Geophysical Research: Atmospheres
As the cloud continues to rise beyond 48 h, the supply of water vapor to the cloud (due entirely to cloud-top entrainment, since the cloud is decoupled from the surface) becomes increasingly limited (Figure 8(a)). Ice crystals continue to grow at the expense of the cloud droplets and the LWP decreases (Figure 7(a)). Meanwhile, IWP achieves a quasi-steady value of~20 g m −2 as a result of the balance between vapor supply (from entrainment at cloud top as well as evaporation of cloud droplets) and the continuous loss of ice crystals by ice precipitation. The net LW sfc decreases slowly with time (Figure 4(a)) because the cloud base continues to rise, weakening the downward longwave flux at the surface. T sfc also gradually cools, reaching −20°C by the end of the fourth day ( Figure 5). At the same time, the air close to the surface dries by vapor depositional growth of the falling ice crystals due to the low temperature (around −15°C). As a result, the air is sub-saturated with respect to water.
This slow evolution of the cloud comes to an end when, after~100 h, the LWP begins to decline rapidly leading to complete dissipation of the cloud by 116 h (Figure 7(a)). The transition occurs because in the late stages of the cloud's lifecycle, the positive feedback that sustains the cloud liquid layer breaks down. More specifically, the liquid water content becomes too low to support significant cloud top cooling and TKE production. This leads to reduced entrainment of moisture from above and reduced regeneration of cloud droplets by in-cloud updrafts; meanwhile, ice production and removal by precipitation continues apace, resulting in enhanced depletion of cloud liquid water and further decreasing TKE production. TKE in the cloud layer rapidly drops to small values and the entire cloud glaciates and dissipates. With no cloud-top cooling to support new cloud droplet formation, the cloud-top temperature inversion weakens while the inversion at the surface becomes more pronounced (Figure 2(a)). The net LW sfc decreases strongly, dropping to −32.7 W m −2 (Figure 4(a)). This drop causes the T sfc to decrease to a minimum value of −30.4°C ( Figure 5). The subsequent increase in the turbulent sensible heat flux contributes to cooling the near-surface air. The turbulent latent heat flux also increases, but its overall effect is rather small compared to the sensible heat flux as it is less than 1 W m −2 throughout the simulation. The cooling results in supersaturation with respect to water of the near-surface air, even though the vapor concentration is low (6 times less than the initial value). A new cloud sequence begins at 116 h, with the difference that this cloud has lower liquid water content than before (Figure 3(a)) as the available moisture is lower compared to the start of the previous sequence (Figure 8(a)).

Clear-Sky State (CLEAR)
In the CLEAR experiment, a surface-based temperature inversion forms and deepens rapidly (Figure 2(b)) as no cloud is present (Figure 3(b), 7(b)) to prevent the surface from cooling through longwave emission. Frost formation at the surface is the only sink of moisture in this case which leads to the development of a surface-based moisture inversion (Figure 8(b)). There is no turbulent cloud layer, and thus the only source of TKE is the wind shear near the surface (Figure 9(b)). Figure 4(b) shows that net LW sfc is lower when the sky is clear compared to BASE (e.g~14 W m −2 difference around 48 h). Owing to this difference the T sfc is much lower than in BASE throughout the 5-day period ( Figure 5). In addition, the conductive heat flux through the sea ice is stronger than in BASE (Figure 4(b)) due to the lower T sfc . The near-surface TKE in the CLEAR case is considerable weaker than in BASE because of the much higher stability of the surface layer in CLEAR, but the surface sensible heat fluxes are roughly comparable in both cases.

Sensitivity to CCN Concentration (ICCN and DCCN)
In the experiment with increased CCN (ICCN), drizzle production stops earlier than in BASE (around 24 h). The higher CCN concentration increases the liquid water content, the TKE production and hence the elevation of the cloud so that the −12°C level (i.e. the point of maximum difference between the equilibrium vapor pressure with respect to ice and water) is reached earlier (Figure 3(c)). The higher LWP also results in a slightly faster IWP increase compared to BASE (Figure 7(c)) due to the faster cloud top elevation. In contrast, in the decreased CCN experiment (DCCN), the drizzle ceases later than in BASE and ICCN due to the lower liquid water content and slower elevation rate of the cloud (Figure 3(d), 6(d)). The LWP is also lower and the IWP increases slightly more slowly compared to ICCN and BASE (Figure 7(d)). During this stage of cloud development, the TKE among the experiments does not differ greatly; it is slightly stronger in ICCN and slightly weaker in DCCN compared to BASE (Figure 9(c), 9(d)). Consequently, the cloud turbulent layer in ICCN is thicker, and the surface-based stable layer is shallower than in BASE (Figure 2(c), 8(c)). The

10.1029/2019JD031738
Journal of Geophysical Research: Atmospheres opposite is true for the DCCN case where the cloud-capped layer is thinner and the stable layer at the surface is thicker than in BASE (Figure 2(d), 8(d)). Furthermore, the net LW sfc and T sfc of the ICCN experiment are higher compared to BASE (Figures 4(c), 5) due to the higher liquid water content, while the DCCN experiment produces less liquid water content and in return lower net LW sfc and T sfc compared to BASE (Figures 4(d), 5). The T sfc in ICCN is 1.0°C warmer than in BASE and 2.0°C warmer than in DCCN during the first day of simulation.
After the drizzle stops and LWP reaches its maximum, the net LW sfc and T sfc decrease at a similar pace in all experiments (Figures 4, 5). Eventually, the cloud dissipates in all cases. This happens 10 h earlier in ICCN and 14 h hours later in DCCN compared to BASE (note that Figures 1-8 only show a 5-day time period, but the DCCN experiment was extended for an extra day to capture the cloud dissipation). The cloud lifetime difference between the cases is most likely due to the different elevation rates of the clouds. The cloud with the higher liquid water content produces stronger turbulent updrafts and is therefore lifted at a faster pace. This yields a faster decrease of the vapor supply (Figure 8(c)) through entrainment at cloud top compared to BASE.

Sensitivity to Drizzle (NODRIZZLE)
The absence of drizzle allows cloud water to accumulate during the first 24 h of simulation (Figure 3(e)), peaking at a value an order of magnitude greater than in BASE. The initial fog layer thickens but does not lift from the surface through the first day, and a peak LWP of over 400 g m −2 is reached around 24 h (Figure 7(e)). At this time, enough ice has been formed (~20 mg kg −1 ) to generate efficient ice precipitation (~0.7 mm day −1 ) ( Figure 6(e)), and LWP begins to decrease. After~30 h of simulation, ice precipitation has depleted sufficient water for the cloud base to lift off the ground. The TKE generation within the cloud and surface turbulent layers is more efficient in NODRIZZLE compared to BASE leading to a coupling between these two layers through the first 36 h (Figure 9(e)). This feature is evident in the mean temperature profile where the surface-based inversion is absent and the cloud-capped neutral layer extends to the surface (Figure 2(e)). Until approximately 48 h, the net LW sfc is~4 W m −2 higher in NODRIZZLE than in BASE due to the higher cloud liquid water content and lower cloud base (Figure 4(e)). This leads to an enhanced cloud warming effect at the surface and the T sfc is around 3.5°C higher than in BASE ( Figure 5).
After 48 h, the IWP attains the same quasi-steady value (~21.5 g m −2 ) as in BASE. The ice precipitation is reduced to about 0.9 mm day −1 (Figure 6(e)) and after that point LWP evolves similarly to that of BASE. There is no longer a coupling between the two turbulent layers and a temperature inversion develops at the surface (Figure 2(e)). The net LW sfc and T sfc decrease in a similar manner as in BASE (Figures 4(e), 5), but the cloud glaciates about 16 h earlier due to the greater cloud-top height and thus faster decrease in the vapor supply at cloud top (Figure 8(e)).

Sensitivity to Ice Crystal Number Concentration (IICNC)
In the IICNC simulation-where the ice crystal number concentration is doubled compared to BASEcloud development is similar to BASE during the first 6 h of simulation: a fog forms, the cloud liquid water content increases and the fog eventually lifts and transforms into a cloud (Figure 3(f)). However, the higher ICNC results in a faster increase in the ice water content and a lower LWP compared to BASE (Figure 7(f)). The cloudy layer is also much shallower (Figure 3(f)) and the TKE production is lower (Figure 9(f)). Consequently, the cloud and surface turbulent layers are not coupled and a sharper surface temperature inversion forms during the first day of the simulation (Figure 2(f)). At about 35 h, the LWP reaches its maximum at a value about half that in BASE (Figure 7(f)). Beyond this point, net LW sfc drops at a faster pace than in BASE (Figure 4(f)) leading to a much faster cooling of the surface ( Figure 5) and an even sharper surface temperature inversion.
At 56 h, IWP in the IICNC case becomes lower than in BASE despite the greater ice crystal number concentration due to the lower total water content of the cloud. Furthermore, the downward longwave radiation emitted by the cloud is lower compared to BASE as the cloud is optically thinner and the cloud base higher. Thus, the surface temperature inversion continues to sharpen and deepen with time. After 101 h, the LWP is only about 0.5 g m −2 limiting the TKE production, vapor supply through cloud top entrainment, and growth of the IWP (Figure 7(f)). The cloud becomes tenuous but does not completely dissipate. Nonetheless, at 105 h the cloud water mixing ratio is less than 0.0005 g kg −1 and the evolution of the net LW sfc and T sfc resembles 10.1029/2019JD031738 Journal of Geophysical Research: Atmospheres that of CLEAR (Figures 4(f), 5). The cloud lifetime is therefore estimated to be 105 h in this case. In contrast to the BASE case, a second cloud is not formed after the end of the first cloud lifecycle. The air between the surface and about 800 m altitude is cooler compared to BASE due to the reduced longwave emission after 65 h. Thus, all the air below the cloud liquid layer becomes supersaturated with respect to ice and the ice crystals grow by vapor deposition all the way down to the surface. This effect combined with the strong stability of the surface layer, which limits the influx of vapor from above, leads to strong dehydration of the air close to the surface and the formation of a sharp surface-based moisture inversion (Figure 8(f)). Consequently, in the IICNC case there is not enough humidity at the end of the first cloud sequence for supersaturation with respect to water to be reached after the cloud dissipates.

Sensitivity to Horizontal Convergence and Divergence (CONV and DIVER)
In DIVER, large-scale subsidence limits lifting of the cloud top compared to BASE (Figures 2(g), 3(g)). The cloud-top inversion is considerably stronger than in BASE (compare Figures 2(a) and 2(g)), and the greater cloud-top stability reduces TKE production (Figure 9(g)) and cloud-top entrainment rates. Nevertheless, there is still a coupling between the cloud and surface turbulent layers between 24 h and 50 h of simulation, as the cloud base height is much lower (Figures 2(g), 9(g)). The IWP (which is the integrated ice mixing ratio from the surface to cloud top) is much lower in DIVER compared to BASE (Figure 7(g)) simply because cloud top is much lower in DIVER, but within the cloud there is almost no difference in ice water content between the two simulations. As a result, ice precipitation is also very similar in the two simulations ( Figure 6(g)). However, the lower elevation of the cloud in DIVER means that free-troposphere humidity above the capping inversion is greater than in BASE, which entails a greater supply of moisture to the cloud via cloud-top entrainment after 66 hours of simulation (Figure 8(g)). This effect leads to a slower LWP decrease in DIVER compared to BASE (Figure 7(g)) and results in greater cloud persistence. We have extended this simulation to 6 days and found that the cloud has still not dissipated at the end of this period. Regarding the surface energy budget, the net LW sfc evolves similarly to BASE during the first 48 h (Figure 4(g)). Thereafter, the net LW sfc is higher than in BASE due to the difference in LWP (Figure 7(g)); before 48 h, the LWP in DIVER is lower than in BASE which results in reduced emission of longwave radiation from the cloud towards the surface and lower T sfc ( Figure 5). Thus, the longwave radiation emitted from the surface is also decreased and the net LW sfc does not differ much from BASE. After 48 h, the LWP in DIVER increases and leads to higher T sfc and net LW sfc (Figures 4(g) and 5).
When horizontal convergence is applied, the cloud rises much faster than in BASE and the cloud development is clearly different (Figure 3(h)). During the first 37 h both LWP and IWP are higher in CONV than in BASE (Figure 7(h)) due to the more efficient water condensation. Beyond this time, IWP increases very rapidly with time, rapidly depleting the liquid water content (Figure 7(h)). The imposed convergence results in a slow large-scale ascent that allows the air at upper levels to saturate and form clouds above the main cloud deck. These secondary clouds glaciate and crystals sediment into the lower cloud deck. Consequently, ice crystal formation and growth is enhanced in the main cloud deck contributing to the rapid IWP growth. Eventually, the cloud glaciates and the cloud lifetime is about 67 h shorter compared to BASE.

Sensitivity to Surface Heat Conductivity (IBASE)
A higher surface conductance increases the conductive heat flux and consequently the T sfc . The higher T sfc increases the upward emitted longwave radiation at the surface, leading to a lower net LW sfc and higher T sfc than in BASE throughout the experiment (Figures 4(i) and 5). The cloud in IBASE rises at almost the same pace as in BASE, but the cloud is thicker and contains more water until 48 h of simulation (Figure 7(i)). The higher T sfc decreases the turbulent sensible heat flux (Figure 4(i)) and reduces the stability of the air near the surface (Figure 2(i)). Thus, the surface and cloud turbulent layers are coupled between 24 h and 48 h (Figure 9(i)) and the cloud gains additional vapor from drizzle evaporating and ice crystals sublimating close to the surface (Figure 8(i)). After 48 h, the turbulent layers decouple in IBASE and the LWP starts to decrease (Figure 7(i)) as the vapor supply at the surface and at the cloud top is reduced (Figure 8i). A temperature inversion develops at the surface (Figure 2(i)). The IWP is slightly higher than in BASE, which leads to a slightly faster cloud dissipation (~4 h shorter cloud lifetime).

Summary of Sensitivity Experiments
It is clear from our experiments that the presence of a cloud has a very large impact on the surface temperature and energy balance. The difference in the mean net surface longwave flux between CLEAR and BASE is around −11 W m −2 , and this energy deficit results in a mean surface temperature drop of about 17°C. A comparable temperature increase has been observed during the SHEBA (Surface Heat Budget of the Arctic Ocean) (Uttal et al., 2002) campaign when the atmosphere transitioned from an opaque cloudy state to a clear one (Persson et al., , 2016Stramler et al., 2011). We designed our sensitivity experiments to examine how the formation, lifetime, and radiative properties of the simulated cloud were affected by different microphysical and dynamical conditions. Table 2 displays the cloud lifetime for each case, while Figure 11 shows the difference in LWP, IWP, net LW sfc , and T sfc between the various simulations and BASE, averaged over the time when a cloud was present. The radiative properties of the cloud were mainly influenced by the amount of cloud liquid water, while the driver of cloud sustenance was the existence of a supercooled liquid layer at cloud top, which generated longwave cooling, top-down convection, and subsequent entrainment of relatively moist free-tropospheric air into the cloud layer. In addition, sub-cloud updrafts provided some recycled moisture to the cloud from evaporating cloud droplets and sublimating ice crystals. In our model setup, entrainment of air at cloud top was the main source of moisture to the cloud. This process was therefore key for determining the cloud lifetime. We define the entrainment velocity at the cloud top as the rate of cloud top elevation minus the subsidence rate. The vapor flux at the top of the cloud can subsequently be calculated as the product of the entrainment velocity and the vapor mixing ratio at that corresponding height. The sink of moisture from the cloud system is precipitation, which in our case first occurred through drizzle (until~20-30 hours of simulation) and then through ice sedimentation. The vapor deposition and ice sedimentation rates were affected by the ratio between ice and liquid water, which to some extent also impacted the cloud lifetime.
The largest change in cloud lifetime compared to BASE was obtained in DIVER and CONV (Table 2). In DIVER, the entrainment velocity is lower than in BASE (Figure 10(a)). However, the lower altitude of the cloud also meant that the cloud top was in contact with higher water vapor concentrations in the overlying free troposphere (Figure 10(b)). As a result, the vapor supply via entrainment in DIVER was more sustained than in BASE, with a lower vapor flux before~66 h of simulation (due to lower entrainment rates) but a higher vapor flux afterwards (due to the higher vapor concentrations) (Figure 10(c)), which led to a longer cloud lifetime. It is worth noting that the turbulent latent heat flux at the surface is positive and less than 1 W m −2 in both experiments. Thus, the surface cannot provide the liquid cloud layer with substantial  moisture. The liquid layer is throughout the simulations sustained mainly by the vapor supply at cloud top, which consequently determines the lifetime of the cloud. When the vapor supply is diminished, the vapor flux decreases and the cloud dissipates. The cloud does entrain warmer air from above the temperature inversion in both simulations, but the radiative cooling at cloud top is sufficient to cool the air within the liquid layer. In conjunction with the moistening by entrainment, the cloud top remains supersaturated with respect to water until the moisture supply is depleted. In CONV, the stronger lifting gave rise to pockets of ice cloud formation at the top of the stratocumulus deck and rapid glaciation through the WBF process. In all other sensitivity experiments, the change in cloud lifetime was less than 15%, a difference that was mostly governed by different ice-to liquid water ratios within the cloud.
The impact of changes in dynamical and microphysical settings on cloud radiative properties was found to be small in most of the simulations, in particular after the peak LWP was reached between~30-40 hours of simulation ( Figures 5 and 11). After this time, the amount of supercooled liquid was mainly governed by a balance between the vapor supply, ice production, and ice precipitation, while parameters affecting the warm microphysical properties of the cloud were of less importance. The largest difference in cloud radiative properties was found in the experiment where the ice crystal number was increased (IICNC). In this case, the amount of liquid water was clearly depleted compared to BASE and the surface temperature was on average about 5.5°C cooler. Before peak LWP was reached, the liquid water content varied more between the different experiments and so did the surface temperature (up to~6°C). The most notable change in LWP occurred when drizzle was artificially suppressed; the cloud accumulated~350 g m −2 more water compared to BASE, i.e. a more than 300% increase. Interestingly, an increase of the CCN concentration from 30 cm −3 to 60 cm −3 led to a clear increase of the water content compared to BASE (~8 g m −2 change in LWP), while a lowering of the CCN concentration to 15 cm −3 only resulted in a small difference in the LWP (~2 g m −2 change). However, the change in cloud radiative properties and surface temperature was about the same, or even larger, in DCCN compared to ICCN. It is also worth noting that an increased ice crystal number concentration Figure 11. Differences between sensitivity tests and BASE of liquid water path, ice water path, surface net longwave fluxes and surface temperature, calculated as averages during the lifetime of the cloud and over the whole model domain. The corresponding averages of BASE for the liquid water path, ice water path, surface net longwave flux and surface temperature are 33.7 g m −2 , 13.1 g m −2 , −11.7 W m −2 , and − 14.8°C, respectively.

10.1029/2019JD031738
Journal of Geophysical Research: Atmospheres led to more ice production during the first two days of simulation, but this eventually led to an overall reduction in the liquid water content, a lower ice production, and a decrease of IWP compared to BASE.

Discussion
We have examined how clouds form, evolve and dissipate as relatively warm and moist maritime air is advected over Arctic sea ice. One specific aim was to understand the impact of these low-level mixed-phase clouds on the surface energy budget and the temperature of the sea ice and how this impact changes during a cloud lifecycle spanning over several days. For these reasons, we have used a slab sea-ice model to interactively calculate the surface turbulent heat fluxes and surface temperature instead of using prescribed values. As a result, the evolution of each component of the surface energy budget could be analyzed and the effect of the clouds on the surface and vice versa could be estimated. Figure 11 shows that the average LWP and net LW sfc generally correlate well except for the IBASE and DIVER cases. In IBASE, the higher surface conductance leads to higher T sfc , i.e. the surface will emit more longwave radiation than in BASE, which also affects the reemission of the cloud. This result indicates that the surface characteristics, such as the surface conductance, may be as important for the surface energy budget as the optical properties of the cloud. From this perspective, the ice thickness and perhaps even more importantly the thickness of a snow pack are important variables as they affect conductive heat flux through the sea ice (F c ).
In our simulations, the cloud sequence generally expressed four different regimes: i) fog ii) drizzle (dominated by warm microphysics) iii) transition iv) decay (dominated by cold microphysics). To our knowledge, no continuous, Lagrangian observation of an Arctic air mass long enough to capture all the stages in our simulations has so far been attempted. We can therefore not directly compare our results with an equivalent observational dataset. However, individual stages of the cloud life cycle identified in our simulations do have observed counterparts. For example, fogs and surface-based mixed-phase clouds have been observed e.g. during the SHEBA campaign Stramler et al., 2011). Their occurrence was linked to an increase in net LW sfc from about −40 to around 0 W m −2 (Stramler et al., 2011). Doppler radar and lidar depolarization measurements at the SHEBA site showed the internal structure of these clouds in detail which were classified as all-liquid and mixed-phase clouds (Shupe et al., 2005(Shupe et al., , 2006. Hopefully, present and future Arctic measurement expeditions, such as MOSAiC (Multidisciplinary drifting Observatory for the Study of Arctic Climate), will shed light on the complete lifecycle and evolution of mixed-phase clouds conducting Lagrangian measurements. Also, information regarding drizzle reaching the surface is scarce. To our knowledge, no measurements of drizzle are currently available for the wintertime Arctic in wintertime. Precipitation observations from e.g. the SHEBA campaign  make no distinction between warm and cold precipitation. Thus, there are no observational constraints on the drizzle produced in our simulations.
During initial cloud stage simulated by the model, which typically only lasted a few hours, the surface air cooled, saturated and a fog was initiated. The longwave cooling shifted to the top of the fog and a mixed-phase cloud formed. The supercooled liquid layer and longwave cooling at cloud top was the main source of TKE and thus also the source of entrainment and water vapor supply to the cloud. During the second regime, there was more or less a balance between the TKE production, cloud-top vapor entrainment, liquid water content and drizzle (cf. loop illustrated by brown dashed arrows in Figure 12). As the temperature decreased, ice growth became more efficient and the ice crystals grew due to the WBF process, eventually inhibiting drizzle production. When the drizzle stopped, after about 30 hours, a transition stage started during which the liquid and ice water content of the cloud increased rapidly until there was once again a balance between the vapor flux and precipitation, the latter now in the form of ice sedimentation (cf. path given by purple arrows in Figure 12). After approximately 45 hours, the cloud entered its fourth, decaying stage. The liquid was continuously depleted due to the WBF process, and the TKE production, entrainment and vapor flux at the top of the cloud were reduced. Finally, after about four to five days (except in the case of convergence), the cloud dissipated. The longevity of the low-level Arctic clouds in our simulations was closely tied to the survival of a supercooled liquid layer. Sources and sinks that affected the liquid water content, in particular the cloud-top vapor supply, governed the cloud lifetime. Consequently, in our model setup, the large-scale subsidence rate was found to be crucial as it modified the entrainment rate and vapor concentrations at cloud top, a result that corroborates the findings of Young et al. (2018). The ratio between ice and liquid also modulated the depletion of vapor from the system (and the cloud lifetime) during the decaying stage of the cloud.
We have assumed that the CCN concentration and ICNC are constant, i.e. that there is an infinite source of cloud condensation nuclei and INPs to the system. With this setup, the cloud persisted in its mixed-phase state for many days. Several studies (e.g. Loewe et al., 2017;Mauritsen et al., 2011;Stevens et al., 2018) have indicated that the pristine Arctic may be CCN-limited, which can cause mixed-phase cloud dissipation. The LWP, and the radiative properties of the cloud during the first three stages of cloud evolution, were indeed sensitive to CCN concentration. When the CCN concentration was doubled, LWP and IWP increased which is consistent with the studies by e.g. Possner et al. (2017), Stevens et al. (2018) and Solomon et al. (2018).
Decreasing the CCN concentration resulted in a decrease in LWP and IWP, but their change was relatively small, i.e. the effect was non-linear, which also corroborates the results of e.g. Solomon et al. (2018). However, even if the LWP response was relatively small when the CCN concentration was decreased (compared to when it was increased), our study shows that the change in longwave emission and surface temperature was relatively high (Figure 11) since the cloud became more tenuous. The radiative properties of the cloud were thus more susceptible to CCN changes at low CCN concentrations.
In agreement with Possner et al. (2017), Stevens et al. (2018) and Solomon et al. (2018), we also found that a doubling of the ICNC decreases the LWP (Figure 11). The higher ICNC pushed the cloud into a tenuous regime and the surface temperature decreased rapidly. However, while the previous studies showed that an increase in ICNC also led to an increase in IWP, we only obtain a consistent result if we consider a simulation period shorter than three days. If a longer time is considered (i.e. the whole lifetime of the cloud) the IWP was actually lower in our simulation with increased ICNC than in BASE resulting in an overall decrease of the average IWP. In other words, the average change in IWP depends on which part of the cloud sequence we consider. Observations of INPs in the Arctic are rare, but a recent compilation by Wex et al. (2019) shows that the typical number of INPs for cloud-top temperatures around −15°C (i.e. approximately the temperature of the cloud top during the second and third cloud stages in our simulations) is probably not more than 0.02 L −1 . The ICNC that we have used in BASE should therefore be considered high, and an increase to 2 L −1 extreme, if no secondary ice production mechanisms are active. At temperatures below −25°C (i.e. the cloud top temperature towards the end of the cloud sequence), INP concentrations around 1 L −1 are more plausible.
The influence of divergence on the LWP was also different in our study compared to Young et al. (2018). We found a decrease in LWP with increasing divergence (Figure 11), while they found an overall increase. The reason is most likely the initial water vapor profiles and BL structure in the two cases. In our case, entrainment at cloud top was the main source of moisture, and an imposed divergence resulted in a decrease of this source. It was only during the decaying stage of the cloud sequence that the vapor supply at cloud top was Figure 12. Schematic of fluxes between the different water reservoirs of the cloud system. The symbols denote water vapor (q v ), liquid water (q c ), ice water (q i ) and rain water (q r ).
higher in DIVER compared to BASE as the top in the former case then was located at an altitude with higher vapor concentrations.

Summary and Conclusions
The 3-dimensional LES model MIMICA was coupled to a slab sea-ice model in order to simulate an idealized case of wintertime moist, marine air advection over Arctic sea ice. Radiative cooling of the air mass led to formation of a mixed-phase cloud. This cloud underwent a sequence of four different stages: an initial fog stage, a drizzle stage dominated by warm cloud microphysics, a transition stage where drizzle was replaced by ice sedimentation, and a decay stage dominated by cold cloud microphysics. Without any imposed divergence or convergence, and with CCN and ice crystal number concentrations set to 30 cm −3 and 1 L −1 , respectively, the mixed-phase cloud survived for five days. During this time, the cloud warmed the surface by approximately 11 W m −2 or 17°C. As the first cloud dissipated, a new fog and mixed-phase cloud rapidly formed again, due to the cool and relatively moist surface air.
We examined the sensitivity of the cloud lifetime and radiative properties of the cloud to a number of microphysical and dynamical parameters, including CCN and ice crystal number concentration, drizzle formation, surface conductivity, and large-scale divergence or convergence. Two cases displayed substantially different cloud lifetimes compared to the others: the case with imposed divergence (DIVER) and the case with imposed convergence (CONV). The driver of cloud sustenance was the supercooled liquid layer at cloud top. A change in either the vapor supply to this layer (as in DIVER) or in the sink of liquid due to the WBF process (as in CONV) prolonged or shortened the cloud lifetime by up to three days. The influence of the divergence rate on the lifetime of the cloud conceptually agrees with the findings by Young et al. (2018). However, in our simulations, large-scale subsidence resulted in a decrease of the LWP and IWP, which is opposite compared to Young et al. In our case, entrainment of air at cloud top was the main source of water vapor to the cloud and increased divergence reduced the entrainment rates for over half of the total simulation time.
The LWP and radiative properties of the cloud were surprisingly similar in most of the experiments, in particular after the cloud entered the decaying stage. The largest difference in LWP and cloud radiative properties was found in the experiment with increased ice crystal number concentrations. In this case, the amount of liquid water was clearly depleted throughout the whole cloud sequence compared to the baseline simulation and the surface temperature was on average 6°C cooler. Before the cloud entered the decaying stage, the liquid water content and the radiative properties of the clouds varied more between the different experiments. In agreement with previous findings (e.g. by Possner et al., 2017;Solomon et al., 2018;Stevens et al., 2018), the LWP (and IWP) of the mixed-phase cloud increased with increasing CCN concentration. However, the change was non-linear so that a CCN increase at low levels gave a smaller impact on the LWP than a change at high CCN levels. In contrast, an increase in the CCN concentration at low CCN levels increased the surface longwave warming of the cloud more than an increase at high CCN levels. In other words, mixed-phase clouds formed during Arctic air mass transformation may be close to the limit of black body emission. Accurately simulating LWP changes around that limit is therefore crucial when simulating the radiative impact of Arctic mixed-phase clouds.
Our results also confirm the aforementioned studies in that the LWP decreases strongly with an increased ICNC due to a more efficient WBF process. However, we found that the change in IWP was less straightforward. If the first three stages of the cloud lifecycle was considered, then the IWP increased in our simulations, in agreement with previous studies. If the whole lifecycle of the cloud was considered, then the average IWP decreased due to the overall lower liquid water amount available for freezing. Furthermore, the ice crystal number concentrations that we prescribed as default in our simulations are high compared to observations of typical Arctic ice nucleating particle concentrations. The experiment with a higher number of ice crystals may therefore be considered as an extreme, unless there are active secondary ice production mechanisms.
Though our simulation setup is nominally Lagrangian, in fact the same piece of sea ice was affected by the cloud during the whole simulation period, as in previous work (Cronin & Tziperman, 2015;Pithan et al., 2014Pithan et al., , 2016. The sea ice thickness was also held constant and we did not take into account its growth rate which subsequently could affect the conductive heat flux. All these aspects should be evaluated in future studies of Arctic wintertime air mass transformation and mixed-phase cloud evolution.