1. Introduction
Marine boundary layer (BL) clouds are found over large areas of the world's subtropical oceans and have a strong impact on the global energy balance. The representation of marine BL clouds in global models remains a major cause of uncertainty in climate change predictions (Bony and Dufresne 2005), and within the context of numerical weather prediction (NWP), the correct description of boundary layer clouds and transport on shorter time scales is necessary to accurately predict the large-scale circulation and tropical winds. It is a challenge common to many general circulation models (GCMs) to represent the correct contrast and transition between areas with predominantly overcast (stratocumulus) and broken (trade cumulus) low-cloud conditions. Williams and Tselioudis (2007) find that much of the variance between GCMs in their response to climate change stems from the stratocumulus and transition cloud regimes. In a more recent study, Williams and Webb (2009) voice concern over the fact that none of the GCMs investigated simulate the trade cumulus regime well; a conclusion similar to that found by Medeiros and Stevens (2011) when exploring the representation of low clouds in several GCMs. Trenberth and Fasullo (2010), while focusing primarily on the southern oceans, also illustrate typical errors in top-of-the-atmosphere (TOA) radiation found in reanalysis and climate models: the trades are an area of net loss energy at TOA, while the stratocumulus regions are areas of net gain.
The Integrated Forecast System (IFS) used for global operational numerical weather prediction at the European Centre for Medium-Range Weather Forecasts (ECMWF) is subject to similar model errors. Jakob (1999) identified an overestimate of trade cumulus cloud cover by 10%–15% and an underestimate of stratocumulus cloud cover by 15% as two of the main deficiencies in the cloud fields of the ECMWF reanalysis when compared to International Satellite Cloud Climatology Project (ISCCP) observations. These results were supported by a study examining the TOA radiative impact of low clouds in the operational model (Chevallier et al. 2001). Although some progress has been made to reduce these biases, systematic errors in low-cloud occurrence and TOA radiation are still evident in today's operational model. Satellite observations are crucial to identifying and quantifying these model errors, but are not always sufficient to establish a direct link to the model parameterizations that cause the systematic errors. For example, it is not always clear whether errors in TOA radiation are due to cloud frequency of occurrence, cloud amount when present, or cloud optical properties. Yet, this distinction is valuable information when looking for ways to improve model parameterizations. It is possible to disentangle some of these contributions using satellite observations (e.g., Ahlgrimm et al. 2009; Ahlgrimm and Köhler 2010), but ground-based observations from heavily instrumented sites such as those operated by the Atmospheric Radiation Measurement (ARM) program are arguably even better suited to the task. Exploiting the synergy of vertically resolved cloud properties and coincident radiation measurements, Ahlgrimm and Forbes (2012) link errors in surface irradiance at the ARM Southern Great Plains (SGP) site to specific aspects of the IFS, such as the frequency of low-cloud occurrence, low-cloud fraction, liquid water path, and effective radius. While these results were confined to a single location only, opposite and partially compensating errors were identified for overcast and broken low-cloud regimes consistent with the global TOA error in shortwave radiation, helping to identify possible causes of the systematic model errors. Extending this approach from the continental SGP site to a second site dominated by maritime conditions will test whether the conclusions drawn at the SGP site are indeed applicable to marine low clouds as well.
A second challenge facing global models is the representation of drizzle. In the NWP environment, accurate prediction of precipitation is of obvious value. Not only is the integrated precipitation amount of interest, but so are the frequency and intensity. Drizzle also forms an integral part of the water and energy cycle in the cloudy marine boundary layer and is observed frequently (Leon et al. 2008). Without a faithful representation of drizzle, other aspects of the marine BL cannot be modeled accurately. An example is the contrast between pockets of open cells, which are generally associated with more intense drizzle, and the surrounding overcast and weakly drizzling stratocumulus areas (Stevens et al. 2005). Modeling cloud albedo and radiative properties in this example is linked to the representation of drizzle. Typically, the occurrence of light precipitation is overestimated in global models (Sun et al. 2006). Stephens et al. (2010) show that the ECMWF model, in common with a number of other GCMs, overestimates precipitation frequency over ocean by up to a factor of 2 compared to CloudSat observations, with a concurrent underestimate in precipitation intensity.
During 2009 and 2010, the ARM mobile facility located on Graciosa Island in the Azores collected detailed observations on cloud micro- and macrophysics, radiation, and precipitation over a 19-month-long period. This dataset provides the opportunity to comprehensively evaluate the ECMWF model's performance in a location dominated by marine boundary layer clouds and provide guidance for parameterization changes. Section 2 describes the model and observational data used in this study. In section 3, the model performance is discussed, and areas for improvement are identified. Results from a single-column version of the ECMWF model with improved parameterizations are described in section 4, with an initial assessment of impacts in the global model in section 5. Conclusions are drawn in section 6.
2. Data and methods
a. Observations
The Department of Energy's Atmospheric Radiation Measurements (ARM) Mobile Facility (AMF) was located on Graciosa Island in the North Atlantic from June 2009 through December 2010 as part of the Clouds, Aerosol, and Precipitation in the Marine Boundary Layer (CAP-MBL) field campaign, providing a 19-month-long record of observations. The instrument site (39.09°N, 28.03°W, 14-m elevation) was chosen for its exposed situation at low altitude in close proximity to the coast. The prevailing circulation carries maritime air toward the island, limiting island effects. The instrumentation includes the W-band ARM cloud radar, micropulse lidar, Vaisala ceilometer, microwave radiometer, longwave, and shortwave radiometers, among others. The radar and lidar/ceilometer provide complementary vertically resolved information on hydrometeors in the atmospheric column. A comprehensive overview of the meteorological state and observed clouds is provided by Rémillard et al. (2012, hereafter REM12). The observations show that low clouds are the most frequently observed cloud type (present 40%–50% of the time) and precipitation below cloud base is frequently present (30%–40% of the entire time), but half of the time it evaporates before reaching the surface. The boundary layer is also observed to be decoupled to some degree most of the time. The monthly mean statistics on cloud and precipitation occurrence in REM12 are used here to evaluate the ECMWF model. In addition to the monthly statistics, hourly mean cloud properties and surface radiative fluxes are used for direct comparison with the modeled clouds and radiation. The cloud properties are derived using the Cloudnet algorithm (Illingworth et al. 2007). The algorithm combines radar reflectivity, lidar backscatter, and the microwave radiometer retrievals to determine the presence, type, and properties of hydrometeors in the column. The target classification product has a temporal resolution of 30 s with approximately 43-m vertical spacing. Targets classified as containing cloud (including liquid and ice) or cloud plus drizzle are used to calculate a vertically resolved cloud fraction over an hour-long period. These profiles are used in conjunction with hourly mean downwelling shortwave and longwave radiative fluxes calculated from the 60-s ARM Sky Radiometers on Stand for Downwelling Radiation (SKYRAD) data stream (Stoffel et al. 1996). Frequency distributions of observed in-cloud liquid water path (LWP) are calculated from the 30-s Cloudnet retrieval of the two-channel microwave radiometer observations. Hourly samples with inferior quality observations of either surface radiation or hydrometeor profile are excluded from the evaluation. The target classification is only performed when both lidar and radar are available, and surface radiation measurements must be available for at least 50 out of 60 min within an hourly period.
b. Model
Data from the operational global ECMWF IFS model at the grid point nearest the ARM site is available hourly from June 2009 through December 2010. Over this period, model cycles 35R2, 35R3, 36R1, 36R2, and 36R4 were operational, all of which use 91 vertical levels. In cycle 36R1 (from 26 January 2010), the horizontal grid spacing of the model was decreased from approximately 25 to 16 km (the triangular truncation of the spherical harmonics was changed from wavenumber 799 to wavenumber 1279). As a result, the grid point nearest the ARM site changes location slightly from 39.01°N, 28.12°W to 39.15°N, 28.09°W, corresponding to distances of 12 and 8 km, respectively, from the ARM site location at 39.09°N, 28.03°W. Both model points are ocean points, and the slight shift in location has no noticeable impact on the time series. The radiosonde observations from the Graciosa site were of good quality and included in the IFS operational analysis, from which the operational forecasts were initialized.
The IFS parameterizations include the convection scheme described by Bechtold et al. (2008) and a boundary layer parameterization following the eddy diffusivity/mass flux (EDMF) approach (Köhler et al. 2011). The radiation parameterization is described in Morcrette et al. (2008) and uses the Monte Carlo independent column approximation (MCICA) approach with a cloud generator (Räisänen et al. 2004) and a generalized cloud overlap assumption (Hogan and Illingworth 2000) to represent subgrid-scale cloud heterogeneity and overlap in the vertical. The parameterization of cloud and precipitation is based on the work of Tiedtke (1993) with numerous modifications, including improvements to the formulation of subgrid-scale precipitation overlap (Jakob and Klein 2000), the numerics of the cloud scheme, and the representation of ice supersaturation in cloud-free air (Tompkins et al. 2007). A major upgrade to the cloud scheme was implemented in cycle 36R4 (operational from 9 November 2010), introducing separate prognostic variables for cloud ice and liquid, and changing from a diagnostic to a prognostic representation of precipitating rain and snow hydrometeor categories. For marine warm-phase boundary layer clouds, the switch from diagnostic to prognostic rain is the most relevant aspect of the model upgrade, and has an impact on the precipitation occurrence, as will be discussed in section 3. This newer version of the cloud parameterization is also used for the single-column model experiments.
Model profiles from forecast hours 12 to 35 are extracted from the daily operational 1200 UTC model forecasts (i.e., verification time from 0000 to 2300 UTC) to form a continuous time series of hourly data to compare with the observations at Graciosa. The use of 12–35-h forecast steps avoids any spinup effects in the cloud fields but keeps the lead time short and thus model drift minimal. Hourly samples that have been discarded from the observations due to poor quality are also removed from the model data.
A single-column version of the IFS with 91 levels using a recent model cycle (CY38R2) is used for sensitivity tests. The single-column model (SCM) is forced with 3-hourly operational model data and run with a time step of 900 s. Temperature, humidity, and horizontal winds are relaxed toward the operational model profiles over a time scale of 4500 s. The relaxation to the background aims to keep the atmospheric state close to that of the full model, while allowing the model physics to evolve unhindered. The SCM is not particularly sensitive to the time scale itself; doubling the relaxation time scale does not significantly change the results or conclusions.
An offline subgrid-scale processing routine is run on the operational model and single-column model output to calculate the vertical overlap of cloud, liquid water paths, and the precipitation fraction for the model evaluation against observations in the following sections. This routine is an offline version of the subgrid-scale calculations used in the model's radiative transfer scheme. Cloud overlap is calculated across 112 subcolumns following the generalized overlap assumption based on Räisänen et al. (2004). The number of subcolumns is chosen to match the number used in the shortwave radiative calculations. The in-cloud water content at each level (i.e., the grid-box mean water content divided by cloud fraction) is then assigned to each cloudy subcolumn, so that the overall fraction of cloudy subcolumns equals the grid-box mean cloud fraction at each level. To account for the inhomogeneity of the condensate, an invariant function is applied to the subcolumn condensate [see Ahlgrimm and Forbes (2012) for a more detailed discussion]. The liquid water path of the individual subcolumns provides the model equivalent to the observed in-cloud water path and will be used in the LWP distributions discussed in section 3. The cloud amount on subcolumns is also used to derive a precipitation fraction, since this is not an archived model variable. After cloud overlap has been determined, precipitation on each model level is assumed to be maximally overlapped with cloud above, so long as the precipitation flux exceeds a minimum threshold. For example, a cloud with 50% cloud fraction and a precipitation flux above the threshold will produce a 50% precipitation fraction in the layer below, and any subsequent model layers until the precipitation flux (reduced by evaporation) drops under the threshold. This may not be a realistic assumption, but reflects the current state of the model, which assumes homogeneous in-cloud condensate for the microphysics processes. Only the liquid precipitation flux is considered and the precipitation threshold is chosen as 4.3 × 10−7 kg m−2 s−1 (the equivalent of 0.037 mm day−1); the amount of liquid condensate necessary to produce a radar reflectivity factor of −56 dBZ based on the ECMWF radar forward model (Di Michele et al. 2012). The detectability threshold of the W-Band (95 GHz) ARM Cloud Radar (WACR) is about −56 dBZ at 1-km height, and increases to about −40 dBZ at 3-km height. Most of the model's cloud bases fall below 1 km; thus, a precipitation flux threshold equivalent to −56 dBZ seems appropriate and also matches the assumptions in REM12.
3. Model performance
a. Monthly mean statistics
The IFS model's monthly mean occurrence of cloud cover and precipitation occurrence over the 19-month-long period is shown in Fig. 1 together with the corresponding observations reproduced from REM12. The model's total cloud cover (i.e., the projected cloud fraction at the surface accounting for overlap) is averaged for all samples within a month to estimate the monthly mean cloud occurrence. The model's cloud fraction is an instantaneous area fraction representative of the grid-box area, while the cloud fraction derived from the observations is a fractional occurrence over an hour-long period. We assume here that the temporal cloud fraction observed is representative of the area fraction in the vicinity of the site, a common and necessary approximation when using ground-based observations for model validation (Hogan et al. 2001; Brooks et al. 2005; Bouniol et al. 2010). Observed cloud occurrence (Fig. 1a, red curve) varies between 60% and 80% with relative minima in August–September. The seasonality of the model's cloud occurrence is in good agreement with the observations, though the magnitude is slightly too low in most months, underestimating the total cloud occurrence by about 5% on average. Shown as symbols are the cloud occurrences from a more recent cycle of the IFS (CY38R2) for 5 months (July 2009, June–August 2010, December 2010). This cycle includes the new cloud scheme with prognostic rain and is the same version of the model used in single-column mode. It is shown here to illustrate how this more recent version of the model compares to the various older cycles that were operational during the mobile facility deployment. Cloud occurrence is comparable for the operational IFS and CY38R2.
(a) Monthly mean total cloud occurrence from the operational IFS at the grid point nearest the AMF. Observations are shown in red and operational model results in blue. Values from the full forecast of model cycle CY38R2 are marked by symbols for 5 months during the period. Monthly mean occurrences of (b) low clouds (solid) and deep boundary layer clouds (dashed), (c) midlevel cloud (solid) and high cloud (dashed), and (d) precipitation at cloud base (solid), at the surface (dashed), and for virga (dotted).
Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00153.1
REM12 distinguish between low, mid-, high, and deep boundary layer clouds based on detected cloud base and top. For the model, cloud base and top are determined within each hourly cloud fraction profile at levels where the cloud fraction exceeds 2%. As found in previous studies (Hogan et al. 2001; Bouniol et al. 2010; Ahlgrimm and Forbes 2012), the cloud classification is not very sensitive to the choice of this threshold. Following REM12, clouds with their bases above 7 and 3 km are labeled as high and middle, respectively. Clouds with tops below 3 km are considered low, while clouds with bases not exceeding 3 km and tops greater than 3 km fall into the deep boundary layer (or deep BL) cloud category. These classes are not mutually exclusive; that is, low and high clouds may exist at the same time, for example. Each detected cloud type is assigned a cloud fraction based on the calculated cloud overlap, which uses the same subcolumn approach with generalized overlap assumptions as for the radiation scheme (Räisänen et al. 2004; Hogan and Illingworth 2000). Figures 1b and 1c show the monthly mean fraction of each of the four cloud types. Low clouds are most commonly observed with up to 60% occurrence in the summer months and about 40% occurrence over winter. The model frequently underestimates the occurrence of low clouds (on average by 10%), but month-to-month variability is captured very well. The other three cloud types are more common in winter and less common in the summer seasons. Occurrence of midlevel and deep BL cloud is well modeled, but high-cloud occurrence is overestimated by up to 20%. In part, this difference in high cloud may be due to the reduced sensitivity of radar and lidar to upper-level clouds, and is not taken into account in the model profile. However, previous studies, albeit at other locations, have found that the ECMWF model overestimates high-cloud occurrence even when corrected for the decreased instrument sensitivity (Hogan et al. 2001; Bouniol et al. 2010). Thus, at least part of this error is likely to be real. Considering the focus of this study is low clouds, the overestimated high-cloud occurrence will not be pursued further in this paper, but will remain a topic for future investigation.
Shown in Fig. 1d is the monthly precipitation occurrence. As described in the previous section, a precipitation fraction is derived by using the archived precipitation flux to determine the presence of precipitation in the model columns, as well as maximally overlapping cloud and precipitation. The monthly mean precipitation fractions in the lowest model level and at cloud base are plotted in Fig. 1d as the precipitation occurrence at the surface and at cloud base, respectively. The difference between occurrence at cloud base and the surface is shown in the “virga” category.
REM12 describe precipitation occurrence as the frequency with which hydrometeors are detected anywhere in the column below cloud base. Precipitation is detected on average 36% of the time (Fig. 1d). Considering a total cloud occurrence of 75%, this indicates that the observed clouds produce precipitation about half of the time. Surface precipitation is observed 19% of the time and virga (i.e., where precipitation is observed at cloud base but evaporates before reaching the surface) are observed 17% of the time without a pronounced seasonality. Therefore, on average, the precipitation detected below cloud base fully evaporates about half of the time before reaching the ground. In the model, precipitation is found to occur 56% of the time. Given a cloud occurrence of 70%, a much higher proportion of cloud is producing precipitation in the model than is found in the observations (Fig. 1d). Precipitation occurrence at cloud base is therefore overestimated by a factor of about 1.5 in the model. At the surface, the model has precipitation 37% of the time, giving an overestimate of surface precipitation occurrence compared to observations close to a factor of 2. During the winter season, virga is less common in the model, which is associated with heavier precipitation fluxes at cloud base during the winter months. Figures 2a and 2b illustrate the differences between the observations and the full model and for the occurrence of cloud, and for precipitation at cloud base and at the surface. Precipitation occurrence between cloud base and the surface is reduced only by a third in the model compared to a reduction of a half in the observations.
Schematic of the total percentage occurrence of cloud cover, precipitation at cloud base and precipitation at the surface for (a) observations, (b) the full 3D model, (c) the reference SCM, and (d) the SCM with modified parameterizations.
Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00153.1
While agreement between the operational model cycles and CY38R2 is good for cloud occurrence, there are differences apparent in the precipitation occurrence. Precipitation occurrence at cloud base is similar (although very slightly higher) in the more recent cycle CY38R2, but there is a smaller reduction of precipitation occurrence between cloud base and the surface compared to earlier model versions. The agreement between the operational model cycle during the last month (December 2010, CY36R4) and CY38R2 is better, since both of these cycles include the new cloud scheme with prognostic rain. The differences arise from a more frequent presence of small, but above-threshold precipitation amounts, which are more readily preserved in the new prognostic precipitation variables. Thus, precipitation occurrence is dependent on the chosen precipitation flux threshold, and differences between model cycles are reduced at higher thresholds (not shown). However, as discussed in the previous section, the threshold here was chosen for good reason and to provide a comparable measure of precipitation presence as applied to the observations. The main point is that the overprediction of light rain is present in all the model versions.
The conclusion that the ECMWF model produces surface precipitation too frequently is not new. Stephens et al. (2010) use CloudSat radar data to estimate an overprediction of precipitation frequency by a factor of 1.5–2 in the IFS and other models, depending on the region. However, with the vertically resolved profiles from the ARM instruments we gain additional insight into the cause of this model error: compared to observations, the model produces precipitation at cloud base too frequently, and the proportion of precipitation that evaporates in the subcloud layer is lower in the model than the observations would suggest, particularly in the winter months. Potential sources of error include the precipitation formation processes, evaporation, the representation of the cloud and precipitation fraction, and the representation of inhomogeneity in the cloud and subcloud environment. The focus here is on warm-phase processes, given the dominance of low stratiform oceanic cloud in this region. Systematic errors in the parameterization of the rain autoconversion and accretion processes could lead to an overestimation of precipitation occurrence at cloud base. If the precipitation produced is too intense, this may then impact evaporation rates, and errors may also be reflected in the occurrence at the surface. Incorrect representation of rain evaporation itself may also contribute to errors in the occurrence at the surface. Section 4 will explore the sensitivity of precipitation occurrence to changes in the autoconversion, accretion, and evaporation parameterizations.
b. Radiative impact of low clouds
To isolate errors in surface radiation associated with low clouds, a subset of approximately 2500 matched hourly samples is selected that contains only low clouds in both the model and observations (no mid- or high-level clouds are present). Figure 3 shows the bias in downwelling shortwave and longwave radiation associated with these low-cloud samples. It is evident that surface irradiance is overestimated by about 40 W m−2 at noon, or about 24 W m−2 in the diurnal mean. Two instruments were present at the site to measure longwave radiation and produced measurements with a bias of 6–7 W m−2 against each other. Since the instruments were not cross calibrated, it is not clear which of the two measurements is more accurate. Since the model underestimates the downwelling longwave radiation compared to either measurements (by 6 and 12 W m−2, respectively, in the diurnal mean), the sign of the model error appears to be robust, though its magnitude is unclear. Note that the noisy nature of the composite is due to the subsetting procedure; within the subset, hourly samples are not necessarily consecutive. The model error in surface radiation is consistent with an overall underestimate of low-cloud cover, though errors in cloud radiative properties may also contribute.
Diurnal composite of (a) surface irradiance and (b) downwelling longwave radiation at Graciosa for a subset of samples where the model (black) and observations (gray) both have only low clouds present. Two observed estimates of the longwave radiation are available, which have an offset of 6–7 W m−2.
Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00153.1
Figure 4a shows a joint histogram of observed and modeled total cloud cover for this subset of samples. In a perfect forecast, all samples should lie along the diagonal, indicating that the same cloud cover is modeled as observed. While model and observations frequently agree on overcast conditions (top-right corner of the histogram), there are also many samples where observed conditions are overcast, yet the model produces broken cloud cover or vice versa. Overall, there are more overcast hourly samples observed than modeled; the model does not produce overcast low-cloud conditions frequently enough. Figure 4b shows the mean surface irradiance bias (model minus observation) associated with each bin of the joint histogram. A negative bias is expected above the diagonal where the model predicts more cloud cover than is observed. In the area below the diagonal, the model underestimates the cloud cover; thus, a positive bias is expected. Along the diagonal itself lie samples with correctly forecast cloud cover. If the radiative properties of these clouds were correct, there should be no bias in surface irradiance for these samples. However, there is a negative bias associated with broken cloud cases (<70% total cloud cover) and a positive bias for cases with overcast conditions (≥70% cloud cover). This indicates that the model's broken clouds are too reflective, while clouds in overcast conditions are not reflective enough. Table 1 (top row) lists the mean bias for broken/overcast samples with correctly forecast cloud fraction (i.e., those samples located along the diagonal in Fig. 4). On average, surface irradiance is underestimated by 15 W m−2 for broken low clouds with correctly forecast cloud fraction, and overestimated by 9.3 W m−2 for correctly forecast overcast samples. Figure 4c shows the corresponding mean bias in downward longwave radiation at the surface for each bin of the joint histogram (using the set of longwave observations in closer agreement with the model). Here, values along the diagonal show that the downward longwave radiation is overestimated for broken clouds (on average by 10.5 W m−2) and underestimated for overcast conditions (−4.2 W m−2); a signal consistent with the surface irradiance bias. These opposite and partially compensating biases in surface radiation for overcast and broken low-cloud conditions were also found at the continental ARM SGP site (Ahlgrimm and Forbes 2012). Since overcast samples occur much more frequently in the observational record than do broken cloud samples, the net radiation bias is dominated by the overcast regime. This is also evident from Fig. 3, which shows model radiation biases consistent in sign with those for the overcast cases (top-right quadrant in Fig. 4).
Joint histogram of (a) observed and modeled total cloud cover (0–1) for low-cloud samples. (b) Mean surface irradiance bias (W m−2) associated with each bin in the joint histogram. (c) As in (b), but for downward longwave radiation bias (using the LW observations with the offset closest to the model). Numbers in (a) refer to samples within each quadrant with above/below 70% modeled and observed cloud cover.
Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00153.1
Number of samples and average surface radiation bias (W m−2) associated with broken/overcast low-cloud samples with correctly forecast cloud fractions for three versions of the model (operational IFS, SCM control, and SCM experiment).
Our previous study at the ARM SGP site (Ahlgrimm and Forbes 2012) suggests that the errors in surface irradiance can be partially linked to errors in model cloud LWP. In-cloud LWP values from the model are calculated using the cloud overlap assumptions described in section 3; that is, the model column is divided into subcolumns according to the generalized overlap assumption and a fixed inhomogeneity function is applied to the in-cloud condensate. The resulting in-cloud LWP on subcolumns is what the model's radiative transfer sees. In Fig. 5, the normalized frequency distributions of LWP from observations and model subcolumns are compared. Considering the retrieval uncertainty in this product [approximately 10–20 g m−2, Gaussiat et al. (2007)], observations with less than 20 g m−2 are discarded, as are corresponding model subcolumns with LWP below this threshold. The observed distributions are markedly different for broken and overcast clouds, with lower values (<50 g m−2) much more frequent for broken clouds. The model's LWP distribution is shifted to higher values than observed for broken clouds (Fig. 5a). In overcast cases, the peak of the distribution is captured well, but the model does not produce higher LWP clouds often enough (Fig. 5b). This is consistent with the overestimated surface irradiance found in these overcast samples. The in-cloud LWP distribution for all low-cloud samples (Fig. 5c) is similar to that for the overcast samples because this cloud type dominates. Two other problems contributing to the systematic model error were identified in the previous study at ARM SGP, and have yet to be addressed in the IFS. First, the model does not make consistent assumptions for the treatment of in-cloud heterogeneity. A simple, situation-independent distribution of cloud water is assumed in the radiation scheme, but cloud properties are treated as homogeneous in the microphysics routines. Second, the optical properties of the cloud, particularly due to the cloud droplet effective radius, may not be optimal. However, these aspects are left for future investigation and are not addressed here.
Normalized frequency distribution of observed (gray) and modeled (black) samples with (a) broken low clouds (<70% cloud cover), (b) overcast low clouds (70+% cloud cover), and (c) for all low-cloud samples.
Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00153.1
4. Sensitivity experiments with the single-column model
The SCM version of the ECMWF IFS at cycle CY38R2 is used to explore parameterization choices and parameter sensitivities to address some of the systematic errors identified in section 3 including
too low occurrence of overcast conditions in the model,
over-/underestimation of LWP in broken overcast/low-cloud conditions and related surface radiation biases,
too frequent precipitation occurrence at cloud base, and
too little precipitation evaporation and related errors in precipitation occurrence at the surface.
a. Triggering of shallow cloud production
Boundary layer clouds are created via three pathways in the IFS model. First, the eddy diffusivity mass flux (EDMF) parameterization (Köhler et al. 2011) treats transport throughout the subcloud and cloudy layer when the lower-tropospheric stability is large and the surface buoyancy flux is positive. The scheme produces direct tendencies for the prognostic cloud condensate and fraction variables due to turbulent mixing and transport. This aspect of the scheme is aimed at modeling stratocumulus-type situations in relatively well-mixed boundary layer conditions and generally leads to clouds with high fractions (“overcast”). Second, if the conditions necessary for the EDMF scheme to create cloud are not met, the shallow convection scheme is called subsequently and can trigger the production boundary layer cumulus clouds under decoupled conditions. The shallow convection scheme can then produce convective precipitation and also detrains condensate and cloud fraction as source terms for the cloud scheme. The shallow convection scheme generally produces clouds with lower cloud fractions. Third and last, the cloud scheme can generate cloud through grid-scale lifting or radiative cooling and the microphysics acts upon the updated condensate and cloud fraction from all schemes, including “large scale” rain production through autoconversion–accretion and evaporation of rain in subsaturated air.
b. Autoconversion and accretion
c. Rain evaporation
d. Results from SCM simulations
Table 2 shows the mean error (model minus observations) for cloud cover occurrence, precipitation occurrence, and surface radiative fluxes for the operational IFS time series (discussed in section 3; designated full IFS in the first row), the control version of the SCM (SCM control), and the SCM with the modified boundary layer cloud triggering, rain autoconversion–accretion, and evaporation parameterizations described above (SCM expt). To illustrate the contribution of each changed model component to the results, mean errors for SCM experiments with each component switched on individually are also listed in Table 2 (SCM parcel, SCM evap, SCM auto–acc). The SCM errors are not identical to the full model, but they are of the same sign and close enough for us to be confident the SCM is representing the errors in the full model. The SCM control has a slightly higher cloud occurrence than the full IFS (slightly better agreement with observations), but also has a greater precipitation occurrence at the surface (worse agreement with the observations). In this respect, it reflects the differences between the operational model cycles and CY38R2, as discussed previously. The errors in cloud and precipitation occurrence and surface radiation are significantly reduced in the new version of the SCM combining all three changes (SCM expt) compared to either the SCM control or the full IFS.
Columns 1–4 show the differences in percentage occurrence (model − obs) of clouds and precipitation, averaged over the 19-month-long period for the full IFS, the control version of the SCM (SCM control), the SCM plus sensitivity experiments on parcel ascent (SCM parcel), evaporation (SCM evap), and autoconversion–accretion (SCM auto–acc) added separately, as well as the SCM experiment with all three changes added together (SCM expt). Columns 5–7 show low-cloud-only daily mean model bias (model − obs, in W m−2) in surface irradiance and downward longwave radiation against the two longwave observational records.
Figure 6 shows the cloud and precipitation occurrence time series from the SCM experiment with modified parameterizations (SCM expt) alongside the SCM control and observations. The total cloud occurrence has increased (Fig. 6a). This is mainly the result of an increase in low-cloud occurrence (Fig. 6b). The bias in low-cloud occurrence over the entire period is reduced from 5% to almost zero for the modified SCM. This can be attributed to the changes in the BL parcel and in the autoconversion–accretion treatment (see Table 2). The new autoconversion–accretion parameterization leads to a delay of precipitation onset with respect to the cloud water content and reduces the bias in cloud-base precipitation occurrence from 19% to 5%. Precipitation occurrence bias at the surface is also reduced from 26% to 9% (Fig. 6c), in part due to precipitation being generated less often in the clouds, and partially due to the enhanced evaporation below cloud base. Figures 2c and 2d illustrate the impact on cloud occurrence and precipitation at cloud base and at the surface, as compared with the observed values and operational IFS time series. Although the precipitation occurrence is still overestimated, the modified SCM shows significant improvement both at cloud base and at the surface.
Monthly mean cloud and precipitation occurrence from the SCM with new parameterizations. (a) Total cloud occurrence. (b) Low cloud (solid) and deep BL (dashed) cloud occurrence. Precipitation occurrence at (c) cloud base and (d) at the surface and (e) for virga. Symbols mark values for select months of the full IFS forecast cycle CY38R2.
Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00153.1
The changes to the autoconversion–accretion processes also lead to a shift of the in-cloud LWP distribution toward higher values (Fig. 7a). The increase in low-cloud occurrence and LWP reduces the mean surface irradiance error associated with low clouds from 27 to 7 W m−2 (Fig. 7b and Table 2). The mean error in downward longwave radiation is also reduced from −5 to −1 W m−2 for the radiometer with observations closer to the model prediction (Fig. 7c). The improvement in the downward longwave radiation stems primarily from an increase in low-cloud cover due to the EDMF BL parcel and autoconversion–accretion changes (Table 2). The cloud-base height does not change significantly in the SCM experiments.
(a) Normalized frequency distribution of in-cloud LWP for low-cloud samples. Observations are shown in red; results from two versions of the SCM indicated by green dashed (control) and blue solid (experiment) curves. (b) Diurnal composite of surface irradiance from approximately 2500 low-cloud samples. (c) Diurnal composite of downwelling longwave radiation for low-cloud samples.
Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00153.1
5. Global impact
The model parameterization changes described in section 4 and tested in the SCM environment are assessed here in the full global model environment. For the implementation, the new parameterizations for autoconversion, accretion, and evaporation are the same as for the SCM, but the parameter sensitivity for the cloud triggering in the modified boundary layer scheme was found to be higher in the full IFS due to feedbacks via the large-scale circulation not represented in the SCM. The results described below therefore use a modified parameter setting for Eq. (2) (cϵ = 0.8), which results in a comparable pattern of behavior for the BL parcel in the full model as in the SCM experiments.
An ensemble of four 1-yr-long forecasts at T159 resolution (spectral truncation equivalent to a grid resolution of 125 km) are produced to provide an initial assessment of the global impact on the model climate. The model seasonal and annual means are compared with a range of satellite datasets for cloud [ISCCP, Moderate Resolution Imaging Spectroradiometer (MODIS)], precipitation (Global Precipitation Climatology Project, GPCP), and radiation (Clouds and the Earth's Radiant Energy System, CERES). The global annual mean and root-mean-square error of the differences for cloud cover and precipitation are similar in both the reference model (CY38R2) and the model with modifications, with a small increase in low-cloud cover in the marine stratocumulus regions (not shown). However, the impact of the cloud changes can be seen more clearly in the general improvement in the TOA net shortwave radiation (Fig. 8), consistent with the improved representation of broken and overcast marine boundary layer clouds and surface irradiance found in the SCM at Graciosa. Figure 8a shows the annual mean TOA net shortwave bias against the CERES Energy Balanced and Filled (EBAF) data product (Loeb et al. 2009) for the reference model cycle CY38R2. The spatial pattern of the error is similar to that found for other GCMs and reanalyses (Trenberth and Fasullo 2010), with negative values over the trade wind regions (broken cloud regime, too reflective) and positive biases in the stratocumulus regions and the Southern Ocean (overcast cloud regime, not reflective enough). Figure 8b shows the differences between the model with the three combined parameterization modifications and the EBAF TOA net shortwave radiation. The TOA shortwave bias is reduced over large areas of the globe, including the regions of marine stratocumulus and trade cumulus, as seen in the absolute error difference between the reference model and the modified model versions (Fig. 8c). Although there are some regions that are slightly farther from the observations, the improvement dominates with up to 10 W m−2 locally and the annual root-mean-square error reduces from 10.3 to 9.1 W m−2. Although these parameterization changes do not remove the systematic errors in the global radiation and the spatial pattern of bias is still present, there is an overall improvement in the TOA net shortwave radiation, which suggests the reduced errors in the comparison at Graciosa are more widely applicable to boundary layer cloud around the globe. A more comprehensive assessment of the parameterization changes in the global model, and in particular the impact on precipitation occurrence, is beyond the scope of this study and will be the focus of future research.
(a) Annual mean net shortwave radiation bias (W m−2) of the ECMWF model against CERES EBAF for (a) IFS CY38R2 (control) and (b) IFS CY38R2 with three parameterization changes as discussed in the text (experiment). Areas with blue shading indicate regions where clouds are too reflective, while areas with red shading underestimate cloud albedo. (c) Difference in absolute error between IFS with new parameterizations vs the control version of the IFS. Areas with blue shading indicate a reduction in error, while red areas indicate an increase in error.
Citation: Monthly Weather Review 142, 2; 10.1175/MWR-D-13-00153.1
6. Conclusions
Ground-based radar, lidar, and radiometric observations from the ARM Mobile Facility on Graciosa Island during the Clouds, Aerosol, and Precipitation in the Marine Boundary Layer (CAP-MBL) field campaign are used to investigate and identify systematic errors in the ECMWF model's representation of surface radiation, low clouds, and the occurrence of light rain (drizzle). A time series of data at the Graciosa site extracted from the operational global model for the 19-month-long period is compared to the observations. The model is found to underestimate low-cloud occurrence by 10% and overestimate high-cloud occurrence, but is able to capture the month-to-month variation of all cloud types. Consistent with previous studies, drizzle occurrence is overestimated at cloud base by a factor of 1.5 and at the surface by a factor of 2, and given the dominance of warm-phase low-level cloud in this region, systematic errors in both the representation of the warm-rain autoconversion–accretion process and the evaporation of rain in the subcloud layer are likely to contribute. Considering individual hourly periods, opposite and partially compensating biases in downward surface radiation can be identified for broken and overcast low-cloud regimes, consistent with previous results from the continental ARM Southern Great Plains (SGP) site (Ahlgrimm and Forbes 2012). The broken cloud regime, typical of trade cumulus, has liquid water paths that are too high and are too reflective. The overcast regime, typical of stratocumulus, underestimates liquid water paths and is not reflective enough.
Three specific parameterizations are identified as contributing to the model error: 1) triggering of cloud in the boundary layer and shallow convection parameterizations, 2) the autoconversion–accretion parameterization, and 3) the parameterization of drizzle evaporation. The single-column version of the ECMWF IFS model (SCM, cycle CY38R2), forced with data from the operational model for the 19-month period, is used to assess the impact of changes to the parameterizations. Although the results from the SCM are not identical to the full 3D model, they are close enough to be representative of the full-model systematic errors. A consistent representation of the test parcel entrainment is introduced in the boundary layer and shallow convection parameterizations to determine the lifting condensation level. A more nonlinear autoconversion–accretion parameterization for rain based on Khairoutdinov and Kogan (2000) replaces the Sundqvist (1978) scheme. This reduces the generation of rain when cloud liquid water contents are low, but rapidly increases the conversion rate as precipitation increases. Finally, a new parameterization of rain evaporation based on Abel and Boutle (2012) is included that represents the increased numbers of small drops observed in drizzling stratocumulus and, hence, enhances evaporation rates for light precipitation. Together, the three model changes in the SCM improve the occurrence of overcast low clouds (reducing the difference with observations from 5% to almost zero) and increase their liquid water path (closer to the observed PDF). Precipitation occurrence is reduced at cloud base (reducing mean differences with observations from 19% to 5%) and at the surface (reducing mean differences from 26% to 9%). Surface irradiance bias is reduced from about 27 to 7 W m−2. Downwelling longwave errors are also reduced but difficult to quantify given an offset of 6–7 W m−2 for the two instruments on site.
The SCM sensitivity results show an improvement in the representation of the broken cloud and overcast regimes, which reduces compensating errors. This is particularly evident in the shortwave radiation at the surface, which is affected by cloud cover and liquid water path. Therefore, we might expect a general improvement in the full 3D model shortwave radiation (both at the surface and at the top of the atmosphere) from these parameterization changes. A global observation dataset of top-of-the-atmosphere net shortwave radiation from CERES is used as a measure of the impact of the changes in an ensemble of four 1-yr simulations of the full global model in climate mode. The annual mean absolute error is reduced over large areas of the globe and locally by up to 10 W m−2. The global evaluation of the parameterization changes performed here is limited but shows the changes are one step toward addressing some of the long-lived systematic errors in the ECMWF IFS model—systematic errors that have similarities in other GCMs (Trenberth and Fasullo 2010). A more thorough assessment against observations is required to evaluate other aspects of the global model performance, particularly the impacts on precipitation occurrence, as well as changes to forecast skill for different model resolutions and lead times.
While promising, these improvements do not yet fully resolve the model's low-cloud problems and further developments to the representation of cloud, precipitation, and radiative impacts will be explored in the future, including different approaches to the formulation of boundary layer and shallow convection parameterizations, the transition from the stratocumulus-to-cumulus regime, warm-phase cloud effective radius, and the treatment of in-cloud heterogeneity for both radiation and microphysical processes.
Acknowledgments
This study was funded by the U.S. Department of Energy's Office of Science through a fellowship of the Atmospheric System Research program under Grant DE-SC0005259. The authors are grateful to Ewan O'Connor for his help in obtaining the Cloudnet processed observations for Graciosa and to Jasmine Rémillard and colleagues for sharing the results of their observational study. The authors thank ECMWF colleagues Irina Sandu, Peter Bechtold, and Anton Beljaars for helpful and informative discussions and the three anonymous reviewers for their valuable comments and suggestions.
REFERENCES
Abel, S., and I. Boutle, 2012: An improved representation of the raindrop size distribution for single-moment microphysics schemes. Quart. J. Roy. Meteor. Soc., 138, 2151–2162, doi:10.1002/qj.1949.
Ahlgrimm, M., and M. Köhler, 2010: Evaluation of trade cumulus in the ECMWF model with observations from CALIPSO. Mon. Wea. Rev., 138, 3071–3083.
Ahlgrimm, M., and R. Forbes, 2012: The impact of low clouds on surface shortwave radiation in the ECMWF model. Mon. Wea. Rev., 140, 3783–3794.
Ahlgrimm, M., D. Randall, and M. Köhler, 2009: Evaluating cloud frequency of occurrence and cloud-top height using spaceborne lidar observations. Mon. Wea. Rev., 137, 4225–4237.
Bechtold, P., M. Köhler, T. Jung, F. Doblas-Reyes, M. Leutbecher, M. Rodwell, F. Vitart, and G. Balsamo, 2008: Advances in simulating atmospheric variability with the ECMWF model: From synoptic to decadal time-scales. Quart. J. Roy. Meteor. Soc., 134, 1337–1351.
Bony, S., and J.-L. Dufresne, 2005: Marine boundary layer clouds at the heart of tropical cloud feedback uncertainties in climate models. Geophys. Res. Lett.,32, L20806, doi:10.1029/2005GL023851.
Bouniol, D., and Coauthors, 2010: Using continuous ground-based radar and lidar measurements for evaluating the representation of clouds in four operational models. J. Appl. Meteor. Climatol., 49, 1971–1991.
Boutle, I., S. Abel, P. Hill, and C. Morcrette, 2013: Spatial variability of liquid cloud and rain: Observations and microphysical effects. Quart. J. Roy. Meteor. Soc.,doi:10.1002/qj.2140, in press.
Brooks, M., R. Hogan, and A. Illingworth, 2005: Parameterizing the difference in cloud fraction defined by area and by volume as observed with radar and lidar. J. Atmos. Sci., 62, 2248–2260.
Cheinet, S., 2004: A multiple mass flux parameterization for the surface-generated convection. Part II: Cloudy cores. J. Atmos. Sci., 61, 1093–1113.
Chevallier, F., P. Bauer, G. Kelly, C. Jakob, and T. McNally, 2001: Model clouds over oceans as seen from space: Comparison with HIRS/2 and MSU radiances. J. Climate, 14, 4216–4229.
Di Michele, S., M. Ahlgrimm, R. Forbes, M. Kulie, R. Bennartz, M. Janisková, and P. Bauer, 2012: Interpreting an evaluation of the ECMWF global model with CloudSat observations: Ambiguities due to radar reflectivity forward operator uncertainties. Quart. J. Roy. Meteor. Soc., 138, 2047–2065.
Gaussiat, N., R. J. Hogan, and A. J. Illingworth, 2007: Accurate liquid water path retrieval from low-cost microwave radiometers using additional information from a lidar ceilometer and operational forecast models. J. Atmos. Oceanic Technol., 24, 1562–1575.
Hogan, R. J., and A. J. Illingworth, 2000: Deriving cloud overlap statistics from radar. Quart. J. Roy. Meteor. Soc., 126, 2903–2909.
Hogan, R. J., C. Jakob, and A. Illingworth, 2001: Comparison of ECMWF winter-season cloud fraction with radar-derived values. J. Appl. Meteor., 40, 513–525.
Illingworth, A., and Coauthors, 2007: Cloudnet: Continuous evaluation of cloud profiles in seven operational models using ground-based observations. Bull. Amer. Meteor. Soc., 88, 883–898.
Jakob, C., 1999: Cloud cover in the ECMWF reanalysis. J. Climate, 12, 947–959.
Jakob, C., and S. A. Klein, 2000: A parametrization of the effects of cloud and precipitation overlap for use in general-circulation models. Quart. J. Roy. Meteor. Soc., 126, 2525–2544.
Jakob, C., and A. P. Siebesma, 2003: A new subcloud model for mass-flux convection schemes: Influence on triggering, updraught properties, and model climate. Mon. Wea. Rev., 131, 2765–2778.
Kessler, E., 1969: On the Distribution and Continuity of Water Substance in Atmospheric Circulations. Meteor. Monogr., No. 32, Amer. Meteor. Soc., 84 pp.
Khairoutdinov, M., and Y. Kogan, 2000: A new cloud physics parameterization in a large-eddy simulation model of marine stratocumulus. Mon. Wea. Rev., 128, 229–243.
Köhler, M., M. Ahlgrimm, and A. Beljaars, 2011: Unified treatment of dry convective and stratocumulus-topped boundary layers in the ECMWF model. Quart. J. Roy. Meteor. Soc., 137, 43–57, doi:10.1002/qj.713.
Lebsock, M., H. Morrison, and A. Gettelman, 2013: Microphysical implications of cloud-precipitation covariance derived from satellite remote sensing. J. Geophys. Res., 118, 6521–6533, doi:10.1002/jgrd.50347.
Leon, D. C., Z. Wang, and D. Liu, 2008: Climatology of drizzle in marine boundary layer clouds based on 1 year of data from CloudSat and Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO). J. Geophys. Res., 113, D00A14, doi:10.1029/2008JD009835.
Loeb, N. G., B. A. Wielicki, D. R. Doelling, G. L. Smith, D. F. Keyes, S. Kato, N. Manalo-Smith, and T. Wong, 2009: Toward optimal closure of the earth's top-of-atmosphere radiation budget. J. Climate, 22, 748–766.
Medeiros, B., and B. Stevens, 2011: Revealing differences in GCM representations of low clouds. Climate Dyn., 36, 385–399.
Morcrette, J., H. Barker, J. Cole, M. Iacono, and R. Pincus, 2008: Impact of a new radiation package, McRad, in the ECMWF integrated forecasting system. Mon. Wea. Rev., 136, 4773–4798.
Neggers, R., A. Siebesma, and H. Jonker, 2002: A multiparcel model for shallow cumulus convection. J. Atmos. Sci., 59, 1655–1668.
Räisänen, P., H. Barker, M. Khairoutdinov, J. Li, and D. Randall, 2004: Stochastic generation of subgrid-scale cloudy columns for large-scale models. Quart. J. Roy. Meteor. Soc., 130, 2047–2067.
Rémillard, J., P. Kollias, E. Luke, and R. Wood, 2012: Marine boundary layer cloud observations in the Azores. J. Climate, 25, 7381–7398.
Siebesma, A., 1998: Shallow cumulus convection. Buoyant Convection in Geophysical Flows, E. J. Plate, Ed., Springer, 441–486.
Stephens, G. L., and Coauthors, 2010: Dreary state of precipitation in global models. J. Geophys. Res.,115, D24211, doi:10.1029/2010JD014532.
Stevens, B., G. Vali, K. Comstock, R. Woods, M. C. Van Zanten, P. H. Austin, C. S. Bretherton, and D. H. Lenschow, 2005: Pockets of open cells and drizzle in marine stratocumulus. Bull. Amer. Meteor. Soc.,86, 51–57.
Stoffel, T., and Coauthors, 1996: Atmospheric Radiation Measurement (ARM) Climate Research Facility. Updated hourly. SKYRAD60S. 2009-03-05 to 2010-12-31, 39.0911 N 28.0297 W: ARM Mobile Facility (GRW) Graciosa Island, Azores, Portugal (M1). Atmospheric Radiation Measurement (ARM) Climate Research Facility Data Archive: Oak Ridge, Tennessee. [Available online at http://dx.doi.org/10.5439/1025281.]
Sun, Y., S. Solomon, A. Dai, and R. W. Portmann, 2006: How often does it rain? J. Climate, 19, 916–934.
Sundqvist, H., 1978: A parameterization scheme for non-convective condensation including prediction of cloud water content. Quart. J. Roy. Meteor. Soc., 104, 677–690.
Tiedtke, M., 1993: Representation of clouds in large-scale models. Mon. Wea. Rev., 121, 3040–3061.
Tompkins, A. M., K. Gierens, and G. Rädel, 2007: Ice supersaturation in the ECMWF integrated forecast system. Quart. J. Roy. Meteor. Soc., 133, 53–63.
Trenberth, K. E., and J. T. Fasullo, 2010: Simulation of present-day and twenty-first-century energy budgets of the southern oceans. J. Climate, 23, 440–454.
Williams, K., and G. Tselioudis, 2007: GCM intercomparison of global cloud regimes: Present-day evaluation and climate change response. Climate Dyn., 29, 231–250.
Williams, K., and M. Webb, 2009: A quantitative performance assessment of cloud regimes in climate models. Climate Dyn., 33, 141–157.
Wilson, D., and S. Ballard, 1999: A microphysical based precipitation scheme for the UK Meteorological Office numerical weather prediction model. Quart. J. Roy. Meteor. Soc., 125, 1607–1636.
