Clouds are a key regulator of Earth’s surface energy balance. The presence or absence of clouds, along with their macroscale and microscale characteristics, is the primary factor modulating the amount of radiation incident on the surface. Recent observational studies in the Arctic highlight the ubiquity of supercooled liquid-containing clouds (LCCs) and their disproportionately large impact on surface melt. Global climate models (GCMs) do not simulate enough Arctic LCCs compared to observations, and thus fail to represent the surface energy balance correctly. This work utilizes spaceborne observations from NASA’s A-Train satellite constellation to explore physical processes behind LCCs and surface energy biases in the Community Earth System Model Large Ensemble (CESM-LE) project output. On average CESM-LE underestimates LCC frequency by ~18% over the Arctic, resulting in a ~20 W m−2 bias in downwelling longwave radiation (DLR) over the ~18 × 106 km2 area examined. Collocated observations of falling snow and LCCs indicate that Arctic LCCs produce precipitation ~13% of the time. Conversely, CESM-LE generates snow in ~70% of LCCs. This result indicates that the Wegener–Bergeron–Findeisen (WBF) process—the growth of ice at the expense of supercooled liquid—may be too strong in the model, causing ice to scavenge polar supercooled cloud liquid too efficiently. Ground-based observations from Summit Station, Greenland, provide further evidence of these biases on a more local scale, suggesting that CESM-LE overestimates snow frequency in LCCs by ~52% at the center of the ice sheet leading to ~21% too few LCCs and ~24 W m−2 too little DLR.
Clouds exert a strong control over the radiative energy received by Earth’s surface. Individual cloud characteristics as well as underlying surface conditions vary greatly in time and space over the globe, modulating specific radiative impacts. While clouds have been observed and studied for centuries, the Intergovernmental Panel on Climate Change (IPCC) Fifth Assessment Report (AR5) identified cloud feedbacks as one of the large puzzles remaining in accurately modeling future climate scenarios (IPCC 2013). Arctic clouds in particular have been the focus of numerous recent studies, in part due to improvements in technology that have allowed for more comprehensive in situ and remote sensing observational campaigns in the harsh region (e.g., Castellani et al. 2015; Cesana et al. 2012). Quantifying the impacts of Arctic clouds is made even more difficult given the rapid rate warming occurring at high latitudes in recent years owing to Arctic amplification (Serreze and Barry 2011). The large, ice-covered regions and their potential contribution to sea level rise makes understanding and modeling Arctic responses to future climate scenarios a high priority (e.g., Hanna et al. 2013; Clark et al. 2016). There is, however, ongoing debate over the large-scale role of Arctic clouds since compelling evidence shows that increased cloud cover can both accelerate (Francis et al. 2005; Bennartz et al. 2013) and slow down (Kay et al. 2008) cryosphere melting, depending primarily on the characteristics of the underlying surface. However, the short observational record and specific conditions of many ground-based investigations make extrapolation to a general Arctic cloud impact difficult.
Despite the low temperatures of the high latitudes, liquid-containing clouds (LCCs) are ubiquitous throughout the Arctic (Morrison et al. 2012). At subfreezing temperatures, the cloud liquid is supercooled, existing in unstable equilibrium with ice down to the homogeneous freezing point of −40°C (Wang 2013). LCCs have been observed in the Arctic for decades by flight and ship campaigns, ground-based observatories (Shupe 2011), and more recently by the spaceborne lidar carried on the CALIPSO satellite (Cesana et al. 2012). The presence of such clouds has been shown to have a large influence on downwelling longwave radiation (DLR) received at the surface, and thus also on melt over the Greenland Ice Sheet (GIS) in general (Van Tricht et al. 2016), and specifically during extreme melt events (Bennartz et al. 2013). While the occurrence of LCCs and their impact on DLR has been well documented, the physical mechanisms and microphysical properties that maintain them are not yet well understood (Morrison et al. 2012).
Since the processes that maintain supercooled LCCs in the Arctic are not well understood, modeling them correctly is difficult. Numerous studies have documented both local (Liu et al. 2011) and regional (Cesana et al. 2012; Forbes and Ahlgrimm 2014) deficiencies in the ability of global climate models (GCMs) to accurately represent cloud liquid in the Arctic. The misrepresentation of GCM LCCs has resulted in a variety of systematic temperature and radiation biases (e.g., Barton et al. 2014; Cesana et al. 2015; Forbes and Ahlgrimm 2014; Pithan et al. 2014). For example, Kay et al. (2016a) found that insufficient cloud liquid in the atmospheric component of the Community Earth System Model (CESM) contributed to a cold bias of 2°–3°C at Summit, Greenland. Parameterization of the Wegener–Bergeron–Findeisen (WBF) process, the rapid growth of ice at the expense of cloud liquid, has been identified as contributing to too few high-latitude LCCs in CESM (Tan and Storelvmo 2016) as well as the European Centre for Medium-Range Weather Forecasts (ECMWF) model (Forbes and Ahlgrimm 2014).
The goal of this study is to build on this previous work and to better understand the causes and impacts of the LCC bias. We use state-of-the-art observations from NASA’s A-Train satellite constellation to identify Arctic LCCs, document their radiative impacts, and determine the associated precipitation conditions, verifying the results against ancillary surface measurements at Summit Station, Greenland. Using the observed snowfall frequency from Arctic LCCs as a proxy for the WBF process, we demonstrate that it occurs too rapidly in CESM; an increased snowfall frequency scavenges too much supercooled liquid, unrealistically reducing the optical depth of the model clouds. Here we provide large-scale observations of snowfall frequency in LCCs, documenting a fundamental constraint for LCCs across the Arctic region, which can be used as a benchmark for testing future model parameterizations.
2. Datasets and methods
To explore the physical processes surrounding LCCs in the Arctic, NASA A-Train observations of LCCs, their covariability with snow, and their impact on DLR received over the Arctic surface are compared to CESM outputs. Datasets from a suite of ground-based instrumentation at Summit Station, Greenland, are used to independently corroborate conclusions drawn from the satellite data.
a. Cloud and precipitation observations from the A-Train
The A-Train constellation flies in a sun-synchronous, 98° inclination orbit, providing a detailed, multisensor perspective of the atmosphere from 82°S to 82°N (L’Ecuyer and Jiang 2010) (top left, Fig. 1). The CloudSat and CALIPSO satellites joined the A-Train in 2006 with the specific purpose of increasing knowledge and understanding of global cloud systems and processes by documenting their vertical structure. CloudSat’s 94-GHz Cloud Profiling Radar (CPR) provides sensitivity to large cloud and precipitation particles while the Cloud–Aerosol Lidar with Orthogonal Polarization (CALIOP) (532- and 1064-nm wavelengths) aboard CALIPSO is able to determine the phase of cloud layers. The synergy between these instruments lies at the intersection between cloud and precipitation regimes. When CALIOP identifies an LCC, the lidar can attenuate and provide no information below the supercooled liquid layer; however, while the longer wavelength of CloudSat’s CPR may not detect the liquid layer, it can generally penetrate the cloud to detect underlying precipitation, determining whether or not the LCC is snowing (Battaglia and Delanoë 2013).
In 2011, CloudSat experienced a battery malfunction and has since only been able to collect daytime data. As night represents nearly six months of the year in the Arctic we restrict our analysis to data from the period when both CloudSat and CALIPSO were fully operational, between January 2007 and December 2010. September 2008 and December 2009 were excluded due to insufficient CALIOP data (Z. Wang 2016, personal communication) and a battery failure on CloudSat (Keys et al. 2010), respectively. All A-Train data were binned to a ~0.94° latitude by 1.25° longitude grid to match the resolution of the CESM-LE output described below. For the purposes of this study, the Arctic is defined as the area between ~66.91° and ~81.99°N. This area represents all grid boxes within the A-Train orbit that are above the Arctic Circle, the region subject to both polar night and midnight sun. This study utilizes five data products that leverage information from both CloudSat and CALIPSO. Unless otherwise noted, the versions used are R04, released in 2013.
The 2B-CLDCLASS-lidar (2BCC hereafter) data product uses the distinct sensitivities of the CPR and CALIOP to liquid and frozen hydrometeors to determine cloud phase (Sassen et al. 2008). The CPR is particularly sensitive to the ice particles in mixed-phase clouds, while CALIOP is sensitive to the liquid droplets. For clouds with top or base temperatures below freezing but above −40°C (determined by collocated ECMWF temperature profiles) the cloud phase is determined by the relative strengths of the radar and lidar backscatters: all water clouds have strong lidar and weak radar backscatter; all ice clouds have weak-moderate lidar and strong radar backscatter; and in mixed phase, the backscatter for both instruments is strong. An example of the cloud phase partitioning is shown in Fig. 1, bottom. The lidar backscatter in Fig. 1e shows full attenuation at the top of the mixed-phase cloud in the boxed region, while Fig. 1a shows reflectivities below that layer, highlighting the utility of combining the instrument measurements. In our analysis, the bottom contiguous cloud layer phase is used. Both mixed- and liquid-phase flags are considered LCCs in this analysis.
The 2C-PRECIP-COLUMN algorithm (2CPC hereafter) uses CloudSat’s CPR reflectivities, path-integrated attenuation, and ECMWF temperature information to identify precipitation and determine its type—rain, snow, or mixed (Haynes et al. 2009). This product has been validated, showing good agreement with previous climatologies (Ellis et al. 2009; Smalley et al. 2014). While this study focuses primarily on snow, 2CPC is used as a check to identify LCCs that are producing precipitation over Arctic areas with warm ocean surfaces, such as the North Atlantic. The following 2CPC flags are counted as precipitation occurrences for this study: Snow Certain, Mixed Certain, Rain Certain, and Rain Probable.
Wherever 2CPC identifies a snowing scene, 2C-SNOW-PROFILE (2CSP hereafter) then retrieves snowfall intensity. It is not possible to directly observe surface snowfall rates using the spaceborne CPR since reflectivities at or near the surface are unreliable due to ground clutter (see Fig. 1a). The 2CSP product uses the lowest clutter-free bin of CPR reflectivity to determine if there is snowfall reaching the ground (Wood et al. 2014). If snowfall is indeed found in the lowest bin, then every other bin is examined to determine if it contains snowfall and instantaneous snowfall rates are retrieved in each bin that does, completing the vertical picture of precipitating ice content. If snowfall is not found in the lowest clutter-free bin, then the algorithm is terminated, meaning that for situations such as snowing virga (snow that falls from the cloud but sublimates before reaching the lowest clutter-free bin) no snowfall is reported. Maahn et al. (2014) found that such over/underestimations due to the surface blind zone effectively cancel each other, with satellite detection of snow event frequency within ±5% of ground-based observations at sites in Svalbard and East Antarctica. The 2CSP dataset has also been compared to polar ground-based data in Antarctica (Milani et al. 2015; Palerme et al. 2017) and Sweden (Norin et al. 2015) with encouraging agreement. A nonzero value in the clutter-free height bin nearest the surface is considered a surface snowfall event in this study (Fig. 1b).
To establish the radiative effects of LCCs, we utilize the multisensor 2B-FLXHR-lidar product (hereafter 2BFLX). The algorithm of Henderson et al. (2013) combines cloud and precipitation property measurements from A-Train satellites with ancillary information including ECMWF temperature and humidity profiles, using them collectively to initialize a radiative transfer model that then provides the desired observationally constrained fluxes and heating rates at every elevation (L’Ecuyer et al. 2008). Key to this study is that supercooled liquid detected by CALIOP is incorporated into the algorithm (Matus and L’Ecuyer 2017). Protat et al. (2014) found good agreement between 2BFLX and detailed ground-based observations from the U.S. Department of Energy Atmospheric Radiation Measurement (ARM) site in Darwin, Australia, particularly for DLR, the flux component used in this work. Van Tricht et al. (2016) evaluated downwelling fluxes from this dataset against several ground stations around Greenland and found very good agreement, with a mean bias for DLW of only −4 W m−2. An example of the DLR component estimated from 2BFLX is shown in Fig. 1c. The regions of increased radiation near the surface each correspond to cloudy regions identified in Fig. 1d, with particularly high values under liquid-containing clouds.
To better understand how LCCs modulate DLR in the Arctic, it is important to consider the radiative impact of clouds overlying an ice surface. Compared to a clear-sky scenario, clouds simultaneously reflect some incoming solar shortwave radiation (SW) that would have otherwise reached the surface and reduce the outgoing terrestrial longwave radiation (LW) radiated to space by absorbing and emitting it back toward Earth, thus increasing the DLR received by the surface. The relative contributions of these competing effects to the net surface radiation budget depend strongly on surface conditions as well as the microphysical and macrophysical properties of the cloud. The difference between the radiation balance in all-sky and clear-sky conditions—the cloud radiative effect (CRE)—quantifies the impact of clouds on the radiation received at the surface:
where is the downwelling radiative flux in all-sky conditions, meaning cloud effects are taken into account, and is the downwelling radiative flux that would occur in the same scene if clouds were removed.
To make sense of the factors that influence net CRE, it is helpful to consider its SW and LW components individually. The shortwave is largest when a bright cloud overlies a dark surface. Time of day and season determine the incident on the top of atmosphere, which is effectively zero at night and in local winter and is largest in summer at midday. A dark surface has a small albedo (α) and will absorb the majority of the SW incident on it. Since a bright cloud will block incoming solar radiation from reaching the surface, significantly more SW radiation will be absorbed in a clear-sky scenario than a cloudy one. However, when the surface below a cloud also has a high α, such as a bright ice sheet, the is small. The longwave , on the other hand, does not depend on the α of the surface, but rather on the difference in atmospheric emissivity caused by the presence of a cloud.
Cloud properties influence in different, and often competing, ways. In general, thicker clouds reflect more incoming SW from the sun and emit LW more efficiently, resulting in smaller and larger , respectively. In this study, the role of thermodynamic phase is central. When compared to a fully glaciated (or all ice) cloud, an LCC with the same total water path is brighter and more emissive (Rogers and Yau 1989), meaning it will both reflect additional SW and trap additional LW relative to an all-ice cloud.
In contrast to lower latitudes, the Arctic surface is frequently bright ice or snow, so clouds in the Arctic have a predominantly warming effect on the surface. This is illustrated in Fig. 2, showing a global map of the observed cloud impact on surface radiation ratio (CISRR), which quantifies the ratio of SWCRE to LWCRE at the surface (McIlhattan 2015):
While the full annual cycle of CISRR at any given location is dependent on many factors including incoming solar radiation, surface conditions, and cloud characteristics, the dominant radiative forcing at the Arctic surface is that of the LW. The deepest red in Fig. 2 indicates where the LWCRE is a factor of 10 larger than the SWCRE, as observed over the GIS and Antarctica. LW dominance can also be seen over the desert regions of North Africa, the Himalayas, and the stratocumulus regions off the western coasts of Africa and South America. Conversely, the deepest blues are found over the tropics, where the SWCRE is 10 times as strong as the LWCRE. It is clear that surface type plays a large role in CISRR, with ice-covered regions dominated by LW effects.
Cloud impacts on both the SW and LW radiative balance are necessary constraints in models; however, we focus on the LW component in this work, due to its clear strong influence on the Arctic surface.
b. Ground-based observations from Summit Station
To provide an independent assessment of the physical relationships between LCCs, snowfall, and DLR at a point on the GIS, continuous ground-based observations from Summit Station, Greenland, are used. The upper left panel in Fig. 1 shows the location of the station as well as the region used for both A-Train and CESM-LE analysis in subsequent comparisons (70.68°–74.45°N, 35.63°–40.63°W).
The ground-based observations of LCC frequency and DLR used in this study are consistent with those presented in Miller et al. (2015, hereafter M15). The M15 data and corresponding precipitation occurrence measurements have been collected as part of the ongoing Integrated Characterization of Energy, Clouds, Atmospheric State and Precipitation at Summit (ICECAPS) project. The high temporal resolution data from Summit Station are averaged to monthly periods to allow comparison to the satellite measurements. Comparing the monthly statistics of the observations, rather than individual overpasses, avoids the complications of differing sensor volumes and locations (Maahn et al. 2014). A threshold of 15 000 measurements per month is used to remove data points that are biased to a particular part of the month due to instrument malfunction.
As shown in Table 1, the Summit analysis considers a different time period than the A-Train datasets used here. However, mismatches in statistics owing to interannual variability are expected to be small since the time periods are adjacent and both the ground-based and spaceborne measurements span multiple years. The annual cycles shown include the interannual variability of each dataset to indicate the range of uncertainties that might be incurred owing to the slight offset in time periods analyzed.
To identify the presence of cloud liquid, we adopt a physical retrieval similar to Turner et al. (2007) and M15. Column liquid water path (LWP) is retrieved from a total of three microwave frequencies: two channels from the Humidity and Temperature Profiler (HATPRO) microwave radiometer (23.84 and 31.40 GHz) and one channel from a high-frequency microwave radiometer (90.0 GHz). Based upon the approximate retrieval uncertainty of LWP (2–5 g m−2) a threshold of 5 g m−2 is considered a liquid containing cloud.
DLR measurements at Summit were obtained from Kipp and Zonen CG4 pyrgeometers (4.5–40 μm) located approximately 2 m above the surface. The longwave broadband radiometers are maintained by the Swiss Federal Institute of Technology, Zurich, and daily checks are performed by an on-site field technician. These direct measurements of fluxes reaching the surface should be contrasted with the indirect flux estimates from 2BFLX that are derived from the application of a broadband radiative transfer model using satellite observed atmospheric properties.
The Precipitation Occurrence Sensor System (POSS) is a bistatic, continuous-wave X-band Doppler radar that retrieves precipitation rate and occurrence within an ~1 m3 volume directly above the radar (Sheppard 2007; Sheppard and Joe 2008). POSS data processing and quality control are performed similar to Castellani et al. (2015). A precipitating LCC is defined by an observation with a POSS snowfall rate above zero and a LWP greater than 5 g m−2. These criteria provide an alternative, independent measure of precipitation frequency of LCCs to satellite-based estimates. Neither approach can be considered absolute truth, since all instruments are subject to sensitivity limitations and uncertainties including blowing snow can contaminate POSS data. Instead, these independent analyses provide a means for assessing physical consistency between two distinct perspectives.
c. Temperature data from GC-Net
Greenland surface air temperature observations are obtained from the Automatic Weather Stations (AWS) of the Greenland Climate Network (GC-Net) (Steffen and Box 2001) for the span of January 2007–December 2010. GC-Net hourly temperature measurements from two Type-E Thermocouple instruments on each AWS are averaged to monthly means for this study. Air temperature measurements collected by HMP45 Vaisala instruments on the AWS are excluded because they are subject to excess solar heating and have lower accuracy (K. Steffen 2016, personal communication). Individual months for which more than 240 h (10 days) of data are missing are excluded from this analysis. Months in which the difference between the maximum and minimum temperature measurements do not exceed 10°C are also excluded, as this indicates the probes may have become buried in snow, effectively insulating them from variations in the air temperature. For the summer analysis, stations with more than one of the four years of data missing from May, June, or July are excluded.
d. CESM and the Large Ensemble Project
Developed and housed at the National Center for Atmospheric Research (NCAR), CESM is open for use by and input from the entire Earth science community. Its fully coupled components and complex cloud microphysics offer the ideal test bed for examining the impacts of physical process specifications and parameterizations. We use output from the CESM Large Ensemble (CESM-LE) project to examine the spread of possible realizations owing to intramodel variability and compare that to interannual variations in the observational datasets.
The CESM-LE was run with historical forcings and fully coupled atmosphere [Community Atmosphere Model, version 5 (CAM5)], ocean (Parallel Ocean Program, version 2), land (Community Land Model, version 4), and sea ice (Los Alamos Sea Ice Model) components. The unique strength of the CESM-LE project is that it gives a range of climate scenarios that are possible with the given model physics and parameterization (the internal model variability) (Kay et al. 2015). CESM-LE utilizes very small (of the magnitude of roundoff error) variations in sea surface temperature of a multicentury control run to create 30 core members spanning 1920–2100. The resulting realizations give an envelope of possible climates for the model. If the CESM physics and parameterizations are representative of the real-world climate, the observations should fit within that envelope.
For this work we use the atmospheric data from the initial 30 members of the CESM-LE, which is provided in ~0.94° latitude by 1.25° longitude resolution. While our preference would be to use the same years of model output as the A-Train observations, that is not possible for all variables. All datasets from the CESM-LE output are available as monthly averages, whereas only a subset of variables is output as daily averages or 6-hourly instantaneous values and only for select years. Monthly averages are appropriate for radiation comparisons, but to determine covariability between precipitation and LCCs, instantaneous values of both parameters are required. Thus, 6-hourly output from January 2002–December 2005 are used here. These years were the closest available years to the A-Train observations. This mismatch of years is, however, likely to have little impact on the results of this study, since the model physics do not vary in time. The fraction of LCCs that snow across the Arctic domain is, therefore, uniform in time and between ensemble members, an assertion that will be demonstrated below. A summary of the time periods used for all CESM variables can be found in Table 1.
When comparing model output to observations, it is important to be aware of scale and sensitivity. This is particularly important when looking at clouds, which occur on scales much smaller than a GCM grid box can resolve. Clouds and their microphysics are necessarily parameterized. Ideally satellite simulators should be used to emulate satellite sampling, attenuation, and retrieval assumptions in the model in order to provide accurate comparison values (Kay et al. 2012). Unfortunately, the CESM-LE members were not run coupled to the GCM Oriented Cloud CALIPSO Product (GOCCP) simulator and no simulator of the CloudSat 2CSP algorithm currently exists. To mitigate mismatches between observational and model definitions of clouds and precipitation, we apply thresholds that are consistent with A-Train and Summit Station instrument sensitivity. Model LCCs are defined as grid boxes containing instantaneous values greater than 5 g m−2 of vertically integrated cloud liquid, and model LCCs are defined as precipitating if they have instantaneous precipitation values higher than 0.01 mm h−1. The sensitivity of model results to the choice of various LCC thresholds is assessed in the following section.
a. Satellite and model comparisons
1) LCC biases throughout the Arctic
The frequencies of occurrence of LCCs in A-Train observations and CESM-LE outputs are presented in Fig. 3. Because of the warm waters brought by the Gulf Stream, the North Atlantic region has year-round open water and warmer surface temperatures than the rest of the Arctic. This leads to increased cloud liquid in all seasons, with LCCs observed 60%–80% of the time. LCCs are observed least frequently over the GIS, owing to its high elevation and frozen surface. However, LCCs are observed over Summit Station at the top of the GIS ~24% of the time (Table 2), consistent with the findings of Van Tricht et al. (2016) (~28% over the total GIS). The remainder of the Arctic goes through large seasonal swings in surface conditions and weather patterns, leading to changes in moisture availability and larger variability in LCC frequency. Throughout the annual cycle, CESM-LE produces a spatial distribution that mimics observations, but the magnitude is too low. During the sea ice maximum season (February–April), the model generates some LCCs over the North Atlantic (~40%); however, virtually none occur over the remainder of the Arctic. When sea ice is at a minimum (August–October), CESM-LE again yields an accurate spatial distribution, with the highest LCC fraction over the North Atlantic and the smallest fraction over the GIS; however, overall the magnitude is ~20% too low.
Thus, despite exhibiting a generally correct spatial distribution throughout the seasons, the magnitude of LCC fraction in the model is consistently biased low relative to observations. This bias is shown quantitatively in the area-weighted averages for the 66.91°–81.99°N region shown in Fig. 4a, where CESM-LE consistently generates fewer LCCs than the A-Train observes. Each blue circle represents a single year’s monthly value as observed by the A-Train and each red × represents a 4-yr average from one ensemble member for that month, allowing the observed interannual variability to be compared to the intramodel spread. Note that the small model spread shown in the red shaded region of Fig. 4a clearly illustrates that the previously discussed slight difference in CESM-LE time period considered (Table 1) is not a large source of error. As noted in the previous section, if the model physics and parameterizations are correct, the observed values should fit within the ensemble spread. It is clear that observed LCC frequency does not lie within the envelope of model outputs. Previous studies utilizing simulator packages (Kay et al. 2012) and with instrument-specific sensitivity thresholds (Cesana et al. 2015) confirm these results over the Arctic region.
To assess the impact of the LWP threshold used to define model LCCs, Fig. 5 compares the annual average difference between observed and modeled LCCs with a cutoff of 5 g m−2 (as in Fig. 3) and with a considerably smaller CESM-LE LCC threshold of 0.1 g m−2, far below the satellite instrument sensitivity. The difference plot in the right panel demonstrates that this lower threshold yields too many LCCs in the warmer, ice-free surface regions of the model, compared to the observations. Yet over the remainder of the Arctic, and in particular over the GIS, too few LCCs are produced in CESM-LE even at this substantially reduced LWP threshold.
2) Impact on downwelling LW at the surface
Since LCCs enhance the DLR at the surface, it is reasonable to anticipate that this low bias in LCC frequency ought to result in too little DLR in the CESM-LE (Forbes and Ahlgrimm 2014). Indeed, comparing model DLR to that from 2BFLX indicates that the model significantly underestimates DLR at the surface. Figure 6 shows that observed DLR exhibits a similar spatial pattern as observed LCC frequency. CESM-LE represents this spatial pattern fairly well, but the magnitude is consistently too low, with a difference of up to 50 W m−2 in some regions. The largest differences occur in winter (November–January) around Greenland and the North Atlantic. The annual cycle shown in Fig. 4b depicts the consistent systematic underestimation of LW energy intercepted by the Arctic surface.
3) Frequency of precipitation in LCCs
To help explain the LCC bias, we now explore one potential microphysical pathway for the removal of supercooled cloud liquid. Since supercooled water has a higher equilibrium vapor pressure than ice crystals, mixed-phase clouds are inherently unstable. Unless there is sufficient cooling or moistening such that the liquid saturation can be maintained, the constituent ice crystals will grow at the expense of the evaporating liquid droplets, glaciating the cloud and initiating precipitation if the ice crystals grow large enough. This mechanism is known as the Wegener–Bergeron–Findeisen (WBF) process (e.g., Wang 2013; Bergeron 1935; Findeisen 1938) and can quickly convert LCCs to precipitating ice clouds.
While the rate of the WBF process cannot be directly measured via satellite, coincident measurements of LCCs and falling snow provide an indication if that process is occurring (Storelvmo and Tan 2015). Here we leverage the unique capabilities of CloudSat and CALIPSO to quantify this process across the entire Arctic region. For every individual A-Train observation of a LCC (as determined by the 2BCC), 2CSP and 2CPC were used to determine if precipitation was present. The seasonal maps in Fig. 7 highlight a stark difference in physical process between the A-Train observations and the CESM-LE output. The observations show that precipitation occurs under Arctic LCCs ~13% of the time evenly across all seasons, with minor regional variations around that value. Since satellite observations represent a composite over the full life cycle of clouds, the nearly uniform precipitation frequency hints that the WBF process likely exerts a fundamental constraint on the lifetime of Arctic LCCs once snowfall develops. Thus these statistical composites indicate that an average Arctic LCC spends less than 15% of its lifetime producing precipitation. This merely represents the mean behavior of a large ensemble of Arctic clouds, the life cycles of individual clouds likely span the complete spectrum of durations and phase, but this statistical mean has relevance for the overall energy balance in the Arctic (Kay et al. 2008; Kay and L’Ecuyer 2013; Van Tricht et al. 2016).
The model ensemble average not only exhibits larger spatial and temporal variations than observations (Fig. 7), but also a substantial high bias in LCC snowfall frequency, with the seasonal Arctic average of model LCCs precipitating ~59% of the time in MJJ and ~78% of the time in NDJ. Arctic LCCs occur less frequently in the model than in observations; however, when they do occur, the model LCCs are much more likely to produce snow than observed LCCs. This result is further demonstrated in the annual cycle of precipitating liquid-containing clouds (Fig. 4c). The dark blue line depicting the average observed frequency of LCCs is nearly flat at ~13%, and shows the smallest interannual variability of all the three variables in Fig. 4. The model ensemble average indicated by the dark red line shows both a distinct annual cycle as well as the largest intramodel variability of the three variables examined. The dashed blue line in Fig. 4c demonstrates that even if all 2CPC “possible” precipitation events are included, the observed precipitation frequency only rises to ~24%, still considerably less than that in the model. The large difference in precipitation rates of LCCs between observations and CESM-LE shown here indicates that the WBF process could be too strong in the model (Morrison et al. 2012).
While CESM-LE is able to initiate some LCCs in the Arctic region, they produce falling snow too often and scavenge supercooled cloud liquid too effectively from the atmosphere. The resulting model LCC deficit does not produce the correct surface radiative balance, with potential implications for downstream surface processes such as ice melt, albedo changes, and sea level rise.
b. Ground-based support—Summit Station
Similar to the overall Arctic region, the annual cycle for Summit Station in Fig. 8a shows that CESM-LE exhibits a low bias in LCC frequency compared to A-Train observations over the ~70 000 km2 region shown in in top left of Fig. 1. In fact biases in this region are even more pronounced; the model produces almost no LCCs at any time of year. The continuous observations from Summit Station, overlaid in black, show that the ground-based measurements are very consistent with the A-Train in both the frequency of LCCs and the annual cycle. The month of June shows a slight disagreement, which could be due to the sampling of offset time periods. Despite the differences in instrumentation, sampling, and observation period, both sets of observations demonstrate a very consistent annual cycle, increasing confidence in the A-Train’s ability to capture Arctic LCC occurrence accurately.
DLR at Summit Station is shown in Fig. 8b. This comparison is more direct than the LCC frequency, since it is not subject to attenuation or differing detection thresholds. The relationship between the CESM-LE output and the A-Train observations again follows the overall Arctic relationship, with the surface receiving significantly more DLR in the observations than in the model, particularly in parts of the year when the LCC bias is largest. The ground-based measurements agree remarkably well with the A-Train observations, lending strong support for the ability of the spaceborne instruments to constrain the factors that govern DLR. The offset values in January and February could again be due to the sampling periods, although the magnitude and shape of the annual cycle between the two observational datasets are very consistent overall. The difference between the A-Train and the model values peaks in July, at ~45 W m−2, with an annual average difference of ~20 W m−2.
The fraction of LCCs in which precipitation is observed by the A-Train over Summit Station, shown in Fig. 8c, is consistent year round (~8%) and similar to the Arctic annual average (~13%), despite large seasonal variations in LCC frequency (~11% in January to ~57% in August). At Summit Station, LCCs are too infrequent in model winter to obtain a precipitation frequency; however, in the summer months, the precipitation frequency of the few present LCCs (~64%) is again biased high compared to the A-Train observations (~8%). Ground-based observations reinforce this conclusion, suggesting a precipitation frequency of ~12% averaged annually. Because of the differing sensitivities of the POSS and CPR and their different sample volumes and retrieval heights (the POSS is measuring snowfall at the ground, the CPR at the top of the blind zone), perfect agreement between these estimates cannot be expected. However, the similarity of these estimates strongly supports the conclusion that Arctic precipitation processes are fundamentally different in the GCM than is supported by observations.
c. Implications for modeling the Greenland Ice Sheet
These differences between the observations and model provides strong evidence of links between model cloud, snowfall, and DLR biases, but these findings alone do not fully demonstrate the extent of the impacts of these biases. Recent work has shown how important clouds, and supercooled liquid–containing clouds in particular, are to the mass balance of the GIS. In their study on the cloud enhancement of GIS melt, Van Tricht et al. (2016) showed that clouds increase meltwater runoff by about one-third when compared to clear-sky conditions, and while LCCs make up ~42% of observed cloud cover, they are responsible for half of this effect.
Figure 9 provides evidence that the missing LCCs over the Greenland surface during the summer months (May–July) may have similar impacts on modeling cryospheric processes in CESM. In the top set of images, the LCC bias is highlighted, followed by the DLR bias in the center set—both indicate the strongest bias occurs at the center of the ice sheet. This result is consistent with the underestimation Lenaerts et al. (2016) found in CESM DLR over the Antarctic ice sheet. The bottom images compare 2-m air temperatures from CESM to ECMWF reanalyses and GC-Net (filled circles). CESM exhibits a cold bias of ~2°–3°C relative to the reanalyses over large portions of the ice sheet, similar to the results of Kay et al. (2016a) over Summit Station. The average CESM-LE cold bias over only the AWS network station sites is ~1.2°C. Note that while A-Train instruments do not observe the 2-m air temperature, the reanalysis values are consistent with the atmospheric temperature inputs used in 2BFLX.
It could be argued that the DLR bias is not due to the LCC bias, but instead due to temperature or water vapor biases. To test this, annual averages of A-Train and CESM-LE DLR, clear-sky DLR, and LW cloud forcing are compared in Fig. 10. The top and center rows of Fig. 10 indicate that if clouds were not present, the bias in the DLR would be much smaller, implying that cloud effects are responsible for much of the model DLR bias. This is more conspicuous in the bottom row of Fig. 10, which shows the difference in DLR cloud forcing between the model and the observations. Clearly the clouds play a much stronger role in the observed radiative balance over the GIS than they do in the model simulations. The small clear-sky bias shown in the right panel of the center row of the figure could be indicative of a feedback by which too few LCCs in all-sky conditions cause too little DLR, leading to colder temperatures in the lower atmosphere and further reducing atmospheric emission to the surface. In this way a lack of model clouds can result in a small DLR bias that persists in clear-sky conditions.
4. Conclusions and discussion
Our work confirms the LCC bias that has been described in previous studies. The CESM-LE project has an ensemble average of ~27% for the frequency of Arctic LCCs, while the spaceborne A-Train observations of the same region show the value to be ~45%. Given the impact that cloud liquid has on DLR, it is not surprising to see that this results in a bias in the DLR seen at the surface. The magnitude of this DLR bias is quantified here and suggests that the CESM-LE ensemble average for the region (~222 W m−2) is ~20 W m−2 below the observed ~242 W m−2. Model biases inferred from satellites are further corroborated using ground-based observations from Summit Station, Greenland. The key findings are summarized in Table 2.
Further examination of the covariability between observations of falling snow and Arctic LCCs suggests that CESM significantly overestimates the frequency with which LCCs produce snowfall, hinting at an overactive WBF process through which too much supercooled liquid is scavenged by precipitating ice. The composite life cycle of Arctic LCCs provided by the satellite observations exhibit a nearly uniform precipitation frequency, indicating that less than ~15% of an average Arctic LCC’s lifetime is spent producing precipitation. The increased model snowfall frequency removes too much supercooled liquid, unrealistically reducing the optical depth of model clouds. This, in turn, causes the model to underestimate the amount of DLR absorbed at the surface leading to a cold bias and, by extension, likely too little surface melting via the mechanisms described in Van Tricht et al. (2016).
While the CESM-LE does not contain enough LCCs relative to observations, it does contain some throughout the Arctic and throughout the year. This demonstrates that the model physics allow for the presence of supercooled cloud liquid, but are not able maintain such clouds for long enough periods. The results presented here show that an overactive WBF process leading to too frequent snowfall is a likely cause for this deficiency. A-Train observations demonstrate that LCCs in the Arctic region only precipitate ~13% of the time, while CESM-LE’s LCCs precipitate ~70% of the time.
Identifying that snow falls too frequently in CESM LCCs is the first step toward improving model representation of Arctic clouds. More work is required to deduce which of the numerous individual processes and parameterizations involved in producing supercooled cloud liquid and snowfall require modification to resolve this discrepancy. The Community Atmosphere Model version 5 (CAM5) employed in CESM-LE (Park et al. 2014) utilizes a two-moment bulk-stratiform microphysics scheme (Morrison and Gettelman 2008) that includes parameterized subgrid variability; prognostic number concentrations (N) and mixing ratios (q) for both cloud ice and cloud liquid and their responses to ice (nucleation, melting, evaporation, sublimation, and sedimentation) and liquid (nucleation, autoconversion, accretion, and heterogeneous/homogeneous freezing) phase processes as well as convective detrainment. The WBF process in CAM5 is assumed to proceed if both cloud ice and cloud liquid are present; however, it is only one of three possible options for a real-world mixed-phase cloud: both ice and liquid can grow, both can deplete, or ice can grow at the expense of the liquid (Korolev 2007). Fan et al. (2011) used a cloud-resolving model (CRM) and aircraft observations to show that in mixed-phase clouds the WBF process strongly depends on vertical velocity, and only occurs in ~50% of cloud volume. Updrafts in the remaining volume are strong enough to support simultaneous growth of both ice and liquid particles.
CESM CAM5 has the ability to output terms that describe the fate of water in all of its phases. Conversion rates between phases of hydrometeor types are in general referred to as “tendency terms.” The tendency terms that quantify the creation and depletion of the cloud liquid provide a more direct assessment of the regional cloud process (Kay et al. 2016b). These terms are not available as standard output within the CESM-LE data but a separate 4-yr branch simulation (January 2011–December 2014) based on CESM-LE member 1 was executed, outputting all monthly average tendency terms. Once again, the specific years simulated are not important, as the parameterizations and model physics should remain constant; however, we chose a span of the same length and similar time period as our previous analysis for consistency.
Figure 11a illustrates that the microphysical processes dominate the removal of cloud liquid within the Arctic region. Dynamic processes (the “deep convection” term) are responsible for a small fraction of the cloud liquid removal near the surface, but this is likely due to activity only over the North Atlantic where warm waters are brought north by the Gulf Stream. While not the focus of this work, the processes that create cloud liquid in this region are primarily gridscale condensation (“macrophysics”) and detrainment from the shallow convection parameterization (“shallow conv. detrainment”). The “vertical diffusion” and “shallow conv. transport/mixing” are not dominant in the Arctic. A discussion of how modified cloud liquid processes impact CESM LCCs over the Southern Ocean can be found in Kay et al. (2016b). The breakdown into individual microphysical processes is shown in Fig. 11b. Vapor deposition is shown as a minor source of cloud liquid, while homogenous freezing (cloud liquid to ice) is a minor sink. Sedimentation provides a minor sink as well above the 950-hPa level, and a minor source below. By far the largest microphysical sink of cloud liquid is the transition to precipitation, supporting our hypothesis that the supercooled cloud liquid in the model is being removed by an overactive WBF process.
Isolating the tendency terms only over the Summit region shown in Fig. 12, we find a consistent picture; microphysical processes are still the dominant removal mechanism. Of those processes, transition to precipitation is the largest sink of cloud liquid. For this region, dynamic processes are not a sink at low atmospheric levels, as they were in the overall Arctic. Thus a closer examination of the CAM5 tendency terms provides more direct evidence that by far the dominant removal process for cloud liquid within the Arctic region is conversion to precipitation. This supports both our use of the covariability between snow and LCCs as a proxy for the WBF process and our hypothesis that precipitation caused by an overactive WBF process in the model likely removes too much supercooled liquid.
It is beyond the scope of this study to discuss specific strategies for modifying CESM microphysics to better represent LCCs. Rather the primary contribution of this work is to provide evidence that this important model LCC bias is linked to excessive snowfall occurrence throughout the Arctic and provide and observationally based benchmark of the true precipitation frequency in Arctic LCCs to guide future model development. Satellite observations measure the state of our atmosphere and can only be used to imply processes, not directly to observe them. To constrain the fundamental microphysical processes in the Arctic atmosphere, we must examine and modify the parameterizations and physics that produce the atmospheric states in coupled GCMs like CESM. Only when models can be made to produce what is seen in real-world observations can we be confident that the observationally implied processes are well understood.
This research was supported by the NASA Grant NNX14AB35G. A-Train data were provided by the CloudSat Data Processing Center (http://www.cloudsat.cira.colostate.edu/). Summit Station, Greenland, data were collected by the ongoing Integrated Characterization of Energy, Clouds, Atmospheric state and Precipitation at Summit (ICECAPS) project. Model data came from the CESM Large Ensemble Community Project using supercomputing resources provided by NSF/CISL/Yellowstone. GC-Net collected and provided the Greenland surface temperature observations. The Swiss Federal Institute of Technology (ETH) provided the broadband radiometer measurements at Summit. We thank Jennifer Kay for helpful discussions along the way. We thank Matthew Shupe for processing the POSS data and David Turner for developing the code for physical retrieval for the LWP.