The global aerosol–climate model ECHAM6.3–HAM2.3 (E63H23) as well as the previous model versions ECHAM5.5–HAM2.0 (E55H20) and ECHAM6.1–HAM2.2 (E61H22) are evaluated using global observational datasets for clouds and precipitation. In E63H23, the amount of low clouds, the liquid and ice water path, and cloud radiative effects are more realistic than in previous model versions. E63H23 has a more physically based aerosol activation scheme, improvements in the cloud cover scheme, changes in the detrainment of convective clouds, changes in the sticking efficiency for the accretion of ice crystals by snow, consistent ice crystal shapes throughout the model, and changes in mixed-phase freezing; an inconsistency in ice crystal number concentration (ICNC) in cirrus clouds was also removed. Common biases in ECHAM and in E63H23 (and in previous ECHAM–HAM versions) are a cloud amount in stratocumulus regions that is too low and deep convective clouds over the Atlantic and Pacific oceans that form too close to the continents (while tropical land precipitation is underestimated). There are indications that ICNCs are overestimated in E63H23.
Since clouds are important for effective radiative forcing due to
aerosol–radiation and aerosol–cloud interactions (ERF
The decrease in ECS in E63H23 (2.5 K) compared to E61H22 (2.8 K) is due to changes in the entrainment rate for shallow convection (affecting the cloud amount feedback) and a stronger cloud phase feedback.
Experiments with minimum cloud droplet number concentrations (CDNCmin) of
40 cm
Clouds are the largest modulators of radiation in Earth's atmosphere. Cloud hydrometeors are generally shorter lived than other modulators of radiation in the atmosphere like aerosol particles, greenhouse gases, or changes in surface albedo through changes in land use. Also, the spatial structure of multiple clouds shows a large variability on different scales as it depends not only on large-scale motions of the air but also on convective and turbulent motions at different scales. These convective and turbulent motions in turn are driven in large part by diabatic heating (and cooling) and radiative cooling (and heating) involving cloud and precipitation hydrometeors, leading to a tight coupling between clouds and circulation (e.g., Wood, 2012; Voigt et al., 2014; Vial et al., 2016). The range of microphysical properties of cloud droplets and ice crystals adds to the complexity of clouds in Earth's atmosphere. This complexity makes clouds difficult to observe and to simulate using models, substantially contributing to the current large uncertainties in future climate projections. Therefore, it is necessary to have an increasingly realistic representation of clouds in global climate models to be able to study past and present climate forcings and to strengthen the reliability of climate projections. It is crucial to evaluate clouds in these models with reliable observations and account for the complexity in clouds in the process.
In this study we use current satellite products to evaluate the aerosol–climate model ECHAM6.3–HAM2.3 and the two precursor model versions ECHAM6.1–HAM2.2 and ECHAM5.5–HAM2.0. One problem in using satellite products is that they are produced with retrieval algorithms that have to make assumptions about the nature of the clouds (e.g., assumptions about the vertical cloud profile; Miller et al., 2016) (and other modulators of radiation) that will not always fit optimally for every cloud in the observed satellite pixels. Accordingly, current satellite products include measures of uncertainty in the retrieved cloud properties. We use these uncertainty measures to limit the evaluation only to regions where the observations are reliable. Furthermore, we apply the CFMIP (Cloud Feedback Model Intercomparison Project) Observation Simulator Package (COSP) where appropriate to account for limitations in the satellite observations (e.g., clouds cannot be observed below the level of full lidar signal attenuation by spaceborne lidar; Chepfer et al., 2010) and the different scales of the model grid compared to the satellite data as well as to ensure a comparison of exactly the same variables in the model output as in the satellite products.
To further limit the impact of observational uncertainties we use several products from independent instruments and aim to identify model biases in several of them. We also perform some of the analyses for different regions to study biases for different cloud types.
For studying past and present climate forcings it is indispensable to
constrain the effective anthropogenic aerosol forcing due to
aerosol–radiation and aerosol–cloud interactions (ERF
Section 2 gives a short description of the representation of clouds in
ECHAM6.3–HAM2.3 and of the observational products applied in the model
evaluation. In Sect. 3 the results of the cloud evaluation and the comparison
of ERF
The global aerosol–climate model ECHAM–HAM is a combination of the global climate model ECHAM with the aerosol microphysics module HAM (Stier et al., 2005). The ECHAM5 and ECHAM6 model versions used in this study are described in Roeckner et al. (2003) and Stevens et al. (2013), respectively. The ECHAM–HAM model versions and configurations used in this study are described in separate studies: ECHAM5.5–HAM2.0 in Zhang et al. (2012), ECHAM6.1–HAM2.2 in Neubauer et al. (2014), and ECHAM6.3–HAM2.3 in Tegen et al. (2019). For the sake of brevity, in the following ECHAM5.5–HAM2.0 will be referred to as E55H20, ECHAM6.1–HAM2.2 as E61H22, and ECHAM6.3–HAM2.3 as E63H23. In contrast to the one-moment cloud microphysics scheme in the ECHAM base model (Lohmann and Roeckner, 1996), ECHAM–HAM uses a two-moment cloud microphysics scheme. The two-moment cloud microphysics scheme is described in Lohmann et al. (2007) and Lohmann and Hoose (2009), with recent changes and improvements applied in E63H23 in Lohmann and Neubauer (2018). A two-moment cloud microphysics scheme is required to study aerosol–cloud interactions. In ECHAM–HAM cloud droplet activation and ice crystal nucleation from cloud condensation nuclei and ice-nucleating particles are computed along with in-cloud and below-cloud scavenging. Therefore, ECHAM–HAM simulates aerosol–cloud interactions in liquid, mixed-phase, and ice clouds. However, a two-moment cloud microphysics scheme is not only a prerequisite for simulating aerosol–cloud interactions, but the additional information from the prognostic cloud droplet and ice crystal number concentrations can also improve the simulation of clouds compared to a one-moment cloud microphysics scheme. The general representation of clouds in ECHAM–HAM is described in the literature given in this section but is briefly repeated in the subsections below for the convenience of the reader.
The scheme for stratiform clouds comprises prognostic variables for water vapor, cloud liquid and cloud ice, a cloud microphysics scheme, and a diagnostic cloud cover scheme (based on Sundqvist et al., 1989). Cloud microphysics is represented by a two-moment scheme described in Lohmann et al. (2007), Lohmann and Hoose (2009), and Lohmann and Neubauer (2018). Optionally available but not used in this study is the one-moment scheme by Lohmann and Roeckner (1996). In ECHAM6.3 changes were made in the diagnostic cloud cover scheme to enhance the cloud cover for marine stratocumulus clouds (Mauritsen et al., 2019). Condensation of cloud liquid water is based on moisture convergence (from transport by advection, turbulence, and convection) and subsequent saturation adjustment. Evaporation of cloud liquid water (or sublimation of cloud ice) occurs when the cloud cover decreases or by the transport of cloud liquid (or ice) mass into the cloud-free part of a grid box. For aerosol activation in liquid clouds the Köhler-theory-based Abdul-Razzak and Ghan (2000) scheme is used. Its implementation is described in Stier (2016). Optionally available is the Lin and Leaitch (1997) aerosol activation scheme. Precipitation is computed diagnostically. The autoconversion of cloud droplets to rain follows Khairoutdinov and Kogan (2000). The accretion of cloud droplets by rain (Khairoutdinov and Kogan, 2000) and evaporation of rain below clouds (based on Rotstayn, 1997) are also computed. Size-dependent wet scavenging of aerosol particles in-cloud and below-cloud follows Croft et al. (2009, 2010). The below-cloud collection efficiencies as a function of aerosol and raindrop or snow crystal size are read from a lookup table. The in-cloud scavenging scheme takes the nucleation scavenging and impaction scavenging of aerosol particles with cloud droplets and ice crystals into account. For nucleation scavenging the number of scavenged aerosol particles is computed for liquid cloud droplets from the cloud droplet number concentration (CDNC) (after the computation of cloud droplet evaporation and precipitation formation), and the fraction of activated aerosol particles (computed by the activation scheme). For ice crystals the aerosol particles are scavenged progressively from the largest to the smallest modes until the number concentration of scavenged aerosol particles is equal to the ice crystal number concentration (ICNC) (after the computation of ice crystal sublimation and precipitation formation) of the grid box.
A downward scavenging tracer flux is computed for each model column each model time step. In-cloud and below-cloud scavenging are sources for the downward scavenging tracer flux, while the evaporation and sublimation of precipitation are sinks for the downward scavenging tracer flux. When the sink term is larger than the source term of the downward scavenging tracer flux in a model level, aerosol mass and number concentrations will be transferred to the respective atmospheric tracers; i.e., aerosol is released from evaporating–sublimating precipitation at this model level back to the atmosphere.
The cirrus scheme follows Kärcher and Lohmann (2002) and details are given in Lohmann et al. (2008). The ice crystals in cirrus clouds form by homogenous nucleation of supercooled liquid droplets. The scheme by Joos et al. (2010) for orographic cirrus clouds can optionally be applied to account for the higher updraft velocities in orographic cirrus clouds but was not used in this study. Supersaturation with respect to ice is allowed for cirrus clouds, and therefore the depositional growth equation is solved for cirrus ice crystals (Kärcher and Lohmann, 2002). For mixed-phase clouds the heterogeneous nucleation of supercooled cloud droplets is computed via immersion and contact freezing following Lohmann and Diehl (2006). Depositional growth of cloud ice in mixed-phase clouds is computed analogous to liquid clouds based on moisture convergence and subsequent saturation adjustment. In addition, the growth of ice crystals at the expense of cloud droplets via the Wegener–Bergeron–Findeisen process (Wegener, 1911; Bergeron, 1935; Findeisen, 1938) is implemented following Korolev (2007). Snow forms by the aggregation of ice crystals, the riming of cloud droplets by snow, and the accretion of ice crystals by snow. Sedimentation of ice crystals follows Rotstayn (1997). The sublimation and melting of ice crystals and snow below clouds is also computed. Ice multiplication via rime splintering (Hallett–Mossop process) following Levkov et al. (1992) is optional (not used in this study).
The convective parameterization from Tiedtke (1989) with modifications for deep convection from Nordeng (1994) and for the triggering of convection from Stevens et al. (2013) is used. The convective parameterization uses only a one-moment cloud microphysics scheme. Detrained condensate of convective clouds is added to stratiform clouds if they exist at the level of detrainment. Whether the condensate is detrained as liquid or ice is based on the same criteria as in the two-moment stratiform cloud microphysics scheme in ECHAM–HAM. To obtain CDNC for the detrained condensate, several simplifications are applied. It is assumed that cloud droplets of convective clouds will form at cloud base. The number of activated cloud condensation nuclei (CCN) at the convective cloud base is computed using the vertical velocity from large-scale and turbulent fluxes as described in Sect. 2.1.4. It is further assumed that CDNC will be constant throughout the vertical extension of the convective clouds. At the level of detrainment these CDNCs from the convective clouds will either evaporate or be added to stratiform clouds if these exist at the level of detrainment. In the latter case a weighted average of the stratiform CDNC and detrained CDNC is computed by weighting stratiform CDNC with the stratiform liquid water content and detrained CDNC with detrained liquid water content. CDNC of the stratiform cloud is not allowed to decrease by this procedure, since cloud droplets will not evaporate in a supersaturated environment. The detrained ICNC is computed from the temperature-dependent empirical relationship of Boudala et al. (2002). An alternative convection scheme based on the Convective Cloud Field Model (CCFM) (Wagner and Graf, 2010) with representation of aerosol–convection interactions is available (Kipling et al., 2017; Labbouz et al., 2018) but not used in this study.
The sulfur cycle model of Feichter et al. (1996) forms the base of the
sulfur chemistry module. Three sulfur species are treated prognostically:
sulfur dioxide, dimethyl sulfide (DMS), and sulfate (the latter not only in
the gas phase but also as an aerosol). Three-dimensional climatological
fields for oxidants are used, i.e., ozone (
Radiative transfer is computed with the two-stream model PSrad (Pincus and
Stevens, 2013). Turbulent fluxes in the atmosphere are computed with the
turbulent kinetic energy (TKE) scheme of Brinkop and Roeckner (1995). The
subgrid-scale vertical velocity that is needed for many cloud microphysical
processes (e.g., cloud droplet activation, ice crystal nucleation,
Wegener–Bergeron–Findeisen process) is computed from the TKE (Lohmann et
al., 2007). Next to a single characteristic updraft velocity for a grid box
that is based on TKE, there is also the option to represent the
subgrid-scale variability of updraft velocities by a Gaussian probability
density function (PDF) of updraft velocities (West et al., 2014). The
subgrid-scale variability is again assumed to be due to turbulence, and the
width of the Gaussian PDF is therefore a function of TKE. The impact of the
width of the Gaussian PDF on ERF
Changes and improvements in E63H23 are also described in Lohmann and
Neubauer (2018) and Tegen et al. (2019), and they are repeated here shortly for
the convenience of the reader. From ECHAM6.1 to ECHAM6.3 the following
improvements were made (Mauritsen et al., 2019):
new PSrad radiation scheme (Pincus and Stevens, 2013), which uses the Monte
Carlo independent column approximation for fractional cloudiness and has the
option for spectrally sparse but temporally dense calculations; update of the fractional cloud cover scheme, which improves the low bias of
marine stratocumulus clouds (this is motivated by the difficulty of
representing the strong inversions of stratocumulus-topped marine boundary
layers in global climate models; Mauritsen et al., 2019); update of the land model JSBACH (Reick et al., 2013), which uses a new five-layer soil hydrology scheme; and removal of inconsistencies in the convection scheme, convective detrainment,
and the vertical diffusion scheme to conserve the atmospheric energy budget. update of mineral dust emission parameterization, which makes use of a
satellite-based source mask for Saharan dust sources (Heinold et al., 2016); new sea salt emission parameterization based on Long et al. (2011), which
uses a temperature dependence following Sofiev et al. (2011); the latest version of the sectional aerosol module SALSA2.0 is implemented
(described in Kokkola et al., 2018) (not used in this study); and new emission datasets have been made available in an input file distribution
for E63H23 for anthropogenic aerosol emissions (Global Fire Assimilation
System (GFAS): Kaiser et al., 2012; Community Emissions Data System (CEDS):
Hoesly et al., 2018; the latter is not used in this study). The Köhler-theory-based Abdul-Razzak and Ghan (2000) cloud droplet
activation scheme (described in Stier, 2016) replaces the empirical Lin and
Leaitch (1997) activation scheme. The in-cloud scavenging scheme by Croft et al. (2010) combines a diagnostic
nucleation scavenging scheme with a size-dependent impaction scavenging
parameterization and replaces prescribed (size-dependent) aerosol scavenging
fractions. There is a changed treatment of detrained cloud water mass and number concentrations
from convective clouds: CDNC from detrained cloud water added (weighted
average) to CDNC of a stratiform cloud cannot decrease the CDNC of the
stratiform cloud; the split between liquid water and ice of detrained condensate
is made consistent between mass and number concentrations. In mixed-phase clouds the heterogeneous freezing by immersion freezing of
black carbon particles is limited to particles in the accumulation mode and
coarse mode. Ice crystals are assumed to have a shape of hexagonal plates, which covers
the whole size range of ice crystals, and the shape is consistent in all
modules. Sticking efficiency used in the accretion of ice crystals by snow has been
changed to the one used in Seifert and Beheng (2006). Two settings for minimum CDNC are available: 40 cm removal of an inconsistency in the fractional cloud cover and cloud
microphysics schemes in ECHAM6.3, which had led cloud cover to be either 0
or 1 in ECHAM6.1; removal of inconsistencies in the kappa-Köhler water uptake in HAM2.3; modularization of the two-moment stratiform cloud microphysics scheme; removal of an inconsistency for convective detrainment in the two-moment
stratiform cloud microphysics scheme to conserve the atmospheric energy
budget; removal of an inconsistency in the two-moment stratiform cloud microphysics
scheme, which led to homogeneous freezing of dry aerosol particles,
independent of availability of water vapor below CDNC–ICNC can no longer grow and in the same time step evaporate or sublimate; no more CDNC at temperatures colder than 238.15 K and no more ICNC at
temperatures warmer than 273.15 K; and and update of default settings, run templates, and run organization (the vertical
resolution is by default 47 vertical model layers; the reference year and
reference period for present-day simulations are 2008 and 2003–2012,
respectively).
The aerosol microphysics scheme HAM2.3 received the following improvements
compared to HAM2.0:
Aerosol–cloud interactions were improved from HAM2.0 to HAM2.3 by the
following changes.
The two-moment stratiform cloud microphysics scheme in ECHAM–HAM received the
following improvements from Lohmann and Hoose (2009) to E63H23.
Further technical improvements, bug fixes, and minor corrections in E63H23
include the following:
For each of the three model configurations, E55H20, E61H22, and E63H23, three
types of experiments were conducted to evaluate the clouds in the present-day
climate, ERF
Setup of the simulations for E55H20, E61H22, and E63H23.
The 10-year simulations for PD conditions were done for all model versions.
Previous studies using E55H20 or E61H22 often used the year 2000 as the
reference year or 2000–2009 as the reference period for present-day
simulations (Zhang et al., 2012; Neubauer et al., 2014); therefore, we also
use the period 2000–2009 for the PD simulations of E55H20 and E61H22. For
E63H23 the default model setup has changed and the reference year and
reference period for present-day simulations are now 2008 and 2003–2012,
respectively. This has become necessary because of the relatively large
changes in aerosol emissions in recent years in several regions (Hoesly et
al., 2018) and was aided by the availability of new boundary condition
datasets. Time-varying (RCP8.5) ACCMIP MACCity (AeroCom II ACCMIP) aerosol
emissions were used for E63H23 and E61H22. The biomass burning emissions are
based on observations until 2008 in ACCMIP MACCity, and afterwards the biomass
burning emissions are from the RCP8.5 emission scenario. For E55H20 the
AeroCom I emissions for the year 2000 are applied for all years. The
greenhouse gas concentrations are set to the year 2008 (RCP8.5) concentrations
in all model versions. All model versions also use a climatology for monthly
values of sea surface temperature (SST) and sea ice cover (SIC) derived from
AMIP data (Taylor et al., 2000) of the years 2000–2015. The spectral
horizontal resolution is T63 (
To compute ERF
The radiative forcing due to aerosol–radiation interactions (RF
To compute ECS, ECHAM–HAM was coupled to a mixed-layer ocean to compute two
50-year simulations, one with preindustrial
Following Hourdin et al. (2017), who argue that estimating uncertain parameters in model development is an important process that should be made transparent, we document our tuning strategies and targets. Tuning is needed mainly to ensure that the TOA radiative fluxes are balanced, and model tuning is limited to adjusting global mean properties. We start from the ECHAM6.3 parameter settings and adapt mainly parameters related to the cloud and convection scheme for tuning E63H23. The tuning strategy and parameters for ECHAM6, as well as the impact of these parameters on the model climate, are described in Mauritsen et al. (2012, 2019). The tuning parameters for ECHAM6–HAM2 and their impact on climate are described in Lohmann and Ferrachat (2010). The parameters that were used in the tuning of the ECHAM–HAM versions and that have different values in E55H20, E61H22, and E63H23 are shown in Table 2.
Parameter settings for E55H20, E61H22 ,and E63H23. The parameters
used to tune the ECHAM–HAM versions are a scaling factor for stratiform rain
formation rate by autoconversion (
The primary tuning target for E63H23 is to match the global mean observed
shortwave (SW) and longwave (LW) TOA fluxes within the range of uncertainty
of the observations along with a TOA net radiative imbalance close to the
observed present-day value. The secondary tuning target is that the SW, LW,
and TOA net cloud radiative effect (CRE) are within the range of uncertainty
of the observations. Cloud cover (CC), liquid water path (LWP), ice
water path (IWP), total precipitation (
Global mean values of the PD simulations. Radiative fluxes are at
the top of the atmosphere. Values from observations (OBS) and multi-model means
(MMMs) for aerosol burdens are shown next to those of the three model
versions. ERF
The tuning is done with short 1-year simulations with a climatology for SST and sea ice. When a set of parameters has been found, one or more 10-year simulations are done to minimize the uncertainty in TOA net radiative imbalance. For E61H22 the default parameter values are used (Neubauer et al., 2014). For E55H20 it was necessary to retune the model with the tuning strategy described above, as the tuning in Zhang et al. (2012) was undertaken for nudged simulations, and we performed free simulations to compare the three ECHAM–HAM model versions. The largest differences in tuning between the three model versions are in the tuning parameters for the autoconversion of cloud droplets to rain and entrainment for shallow convection. The latter was adopted from the base model ECHAM6.3 (see Mauritsen et al., 2019, for a discussion of the impact of the change in this tuning parameter on climate sensitivity). In E63H23 stratiform rain formation by autoconversion will be faster than in the other two model versions. This is due to the larger value of the respective tuning parameter leading to reduced LWP, CC, and SW CRE and a more positive TOA net radiative imbalance in E63H23 (Lohmann and Ferrachat, 2010). The larger value of the tuning parameter for entrainment for shallow convection in E61H22 and the even larger value in E63H23 have the opposite effect: increased LWP, CC, and SW CRE and a more negative TOA net radiative imbalance (Mauritsen et al., 2012). For E63H23 there is a compensation by changing both tuning parameters; the most pronounced net effect is a reduced LWP compared to the other two model versions (we hypothesize that LWP is reduced since entrainment for shallow convection mainly affects low, thin clouds, whereas the autoconversion rate affects all liquid clouds; since the reflectivity of clouds depends nonlinearly on their thickness, an increase in thin low clouds can compensate for the SW CRE change with a decrease in thicker clouds but lead to a lower global mean LWP).
We list the products and the respective references for the observational products used in the model evaluation. From Moderate-resolution Imaging Spectroradiometer (MODIS Aqua) collection 6.1 (Platnick et al., 2015, 2017) and from the ESA Cloud Climate Change Initiative (CCI) Advanced Very-High-Resolution Radiometer (AVHRR-PM) v3.0 (prototype; Stengel et al., 2017a, b), histograms of cloud-top pressure vs. cloud optical depth and CC are taken. Histograms of cloud-top pressure vs. cloud optical depth were also taken from the International Satellite Cloud Climatology Project (ISCCP; Rossow and Schiffer, 1999; Pincus et al., 2012; Zhang et al., 2012) D1 data. Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) data for CC are from the GCM-Oriented CALIPSO Cloud Product (GOCCP) dataset (Chepfer et al., 2010). Cloud radiative effect data are from the Clouds and the Earth's Radiant Energy System (CERES) Energy Balanced and Filled (EBAF) TOA edition 4.0 data product (Loeb et al., 2018). Precipitation data are from the Global Precipitation Climatology Product (GPCP) 2.3 (Adler et al., 2018). Cloud-top CDNCs are from the climatology of Bennartz and Rausch (2017). LWP data are from the Multi-Sensor Advanced Climatology of LWP (MAC-LWP; Elsaesser et al., 2017), which is an updated version of the University of Wisconsin LWP climatology, and from MODIS. IWP is from satellite observations compiled by Li et al. (2012).
Table 3 includes global mean values of radiation, cloud, and aerosol-related
variables of the PD simulations of E55H20, E61H22, and E63H23 compared to
observations (OBS) or multi-model mean (MMM) values when observations are
not available. The global mean values of the radiative fluxes shown in Table 3
are tuning targets (see Sect. 2.6) and therefore cannot be used directly
for model evaluation. For E63H23 the SW and LW TOA fluxes, as well as the SW
and LW TOA CRE, are within the range of the observations. The net TOA flux of
E63H23 is also close to the observations (additional tuning could bring it
closer to the observed value but was not attempted given the large
uncertainty in, e.g., SW and LW TOA fluxes). The SW, LW, and net TOA fluxes of
E61H22 and E55H20 are outside the range of observations. This reflects the
change in the tuning targets and strategy in E63H23 and the availability of
better observations. The net CRE of E63H23 (and also E55H20 and E61H22) is
outside the observed range. It was not possible to find parameter settings
that bring the net CRE within the range of observations without bringing
one or more of the other radiative fluxes outside the range of
observations. This is a first indication of a structural problem in
ECHAM–HAM, which could be related to how ice crystals nucleate in (warming)
cirrus clouds or an underestimation of (cooling) stratocumulus. This will be
further discussed in the evaluation. The CC,
Although the global mean values are tuning targets (see Sect. 2.6), biases
in net CRE and IWP in the ECHAM–HAM versions, which could not be brought in
agreement with observations via tuning, were identified in the previous
section. Zonal mean values of observable variables can nevertheless be used
for model evaluation because tuning targets the global mean quantities.
Figure 1 shows zonal mean distributions of several quantities for the three
model versions and observations. The zonal distribution of SW CRE and LWP-LP
of E63H23 agrees relatively well with observations, whereas in E61H22 and
E55H20 the magnitude of both quantities is overestimated in midlatitudes.
The cloud cover distribution of E63H23 also agrees well with observations,
whereas E61H22 and E55H20 show an underestimation by up to 10 percentage
points in the subtropics. Biases in cloud-top CDNC are more complex, and
retrievals of cloud-top CDNC are only possible for specific clouds (e.g.,
horizontally homogeneous, unobscured, optically thick clouds) and rely on
assumptions, such as liquid water content increasing with altitude like
in an adiabatically rising cloud parcel (or at least like a constant
fraction of this liquid water content), CDNCs being constant throughout
the cloud, and further assumptions that together make cloud-top CDNC
retrievals uncertain (Grosvenor et al., 2018). We therefore expect larger
differences between observations and models for cloud-top CDNC than for
other variables. E55H20 agrees well with MODIS observations in the tropics
but overestimates cloud-top CDNC in the subtropics on both hemispheres and
midlatitudes in the Southern Hemisphere. E61H22 overestimates cloud-top
CDNC in the tropics and subtropics but underestimates it at midlatitudes in
the Northern Hemisphere. E63H23 also overestimates cloud-top CDNC in the
subtropics, but less than E61H22, and also in the tropics. The liquid
phase of clouds is therefore better represented in E63H23 than in the
previous model versions. IWP is underestimated in all three model versions.
E63H23 has the smallest bias, followed by E61H22, and E55H20 shows the
largest deviation from observed zonal mean IWP. The underestimation is
particularly large in the tropics, which is likely connected to the
parameterization of convection in ECHAM (and ECHAM–HAM). ECHAM has a low precipitation
bias over land in the tropics (Mauritsen et al., 2012; Stevens et al.,
2013). Gasparini et al. (2018) found indications that the level of
detrainment from deep convection is too low in altitude in ECHAM–HAM. They
lowered the tuning parameter for deep convective entrainment
Comparison of zonal annual mean values of E55H20, E61H22 and
E63H23 to observations,
The comparison of CRE of the different model versions with CERES CRE reveals
several biases in the representation of clouds. We therefore start by
identifying biases in CRE and then use observations for other quantities to
identify the causes of the model biases. In Fig. 2 the differences in SW, LW, and
net TOA CRE of all model versions to CERES observations are shown. In all
three model versions the (negative) SW CRE is too weak in the marine
stratocumulus regions west of the continents (the average bias in the wider
stratocumulus regions is 1.1, 8.1, and 7.0 W m
Comparison of annual mean SW, LW, and net CRE of E55H20, E61H22, and E63H23 to CERES 4.0 (Loeb et al., 2018) observations. CERES data are for 2005–2015, model data are from the PD simulations. In the top left panel the regions used for cloud-top pressure vs. cloud optical depth histograms are shown by green lines.
In Fig. 3 the cloud cover of the CALIPSO GOCCP product and all three model
versions is shown. The hatched areas in Fig. 3 are the regions where the
cloud cover of CALIPSO GOCCP, MODIS collection 6.1, and ESA Cloud CCI
(AVHRR-PM) differs by more than five percentage points. We therefore use only
the areas not marked by hatching in Fig. 3 for the model evaluation. Since
the COSP CALIPSO simulator is not implemented in E55H20, the direct model
output is shown for all model versions (see Fig. S8 for COSP CALIPSO
simulator output of cloud cover for E61H22 and E63H23). The cloud cover of
all three model versions agrees fairly well with the observations. The
largest biases are in stratocumulus regions west of the continents (
Comparison of annual mean cloud cover of E55H20, E61H22, and E63H23 to CALIPSO GOCCP observations. Areas where the cloud cover of CALIPSO GOCCP, MODIS collection 6.1, and AVHRR-PM differ by more than five percentage points are hatched. CALIPSO GOCCP data are for 2006–2010, model data are from the PD simulations (direct model output is used without a simulator).
Figure 4 shows LWP from the MAC-LWP climatology (Elsaesser et al., 2017) and
the three model versions. The retrieval of LWP has biases from both
visible and near-infrared sensors as well as microwave sensors (Seethala and
Horváth, 2010; Lebsock and Su, 2014). Visible and near-infrared sensors such
as MODIS have problems when the solar zenith angle is large and detecting
pixels of clouds at low altitudes (Lebsock and Su, 2014). Microwave sensors such as AMSR-E may retrieve LWP in cloud-free scenes, and
the split between LWP and rainwater path is difficult (Lebsock and Su,
2014). Elsaesser et al. (2017) corrected the retrieval bias of LWP of
microwave-sensor-based products in cloud-free scenes. And they recommend
using regions with low precipitation (LWP
Comparison of annual mean LWP of E55H20, E61H22, and E63H23 to
MAC-LWP observations. Areas where precipitation could influence the LWP
retrieval (LWP
To further characterize the simulation of liquid clouds in the ECHAM–HAM
model versions we also compare cloud-top CDNC of warm (cloud top warmer than
0
Comparison of annual mean cloud-top CDNC of E55H20, E61H22, and E63H23 to MODIS observations from Bennartz and Rausch (2017). Areas where the relative uncertainty in the observations is larger than 75 % are hatched. The MODIS data are for 2003–2015, model data are from the PD simulations.
In Fig. 6 the IWP of all three model versions and IWP satellite observations compiled by Li et al. (2012) are shown. Li et al. (2012) used three different CALIPSO plus CloudSat ice water products and two different methods to remove the contribution of convective clouds and precipitation from the products. Figure 6 displays the compiled mean IWP of the datasets of Li et al. (2012), and areas where the relative standard deviation of the different datasets is larger than 75 % are hatched. The regional distribution of the occurrence of IWP of all three model versions agrees in general quite well with the observations, although it is biased low in all ECHAM–HAM model versions. This could already be seen in the respective global mean and zonal mean values (see Sect. 3.1 and 3.2). Similar to what was found in the analysis of zonal mean IWP the underestimation is largest in the tropics (see Sect. 3.2).
Comparison of annual mean IWP of E55H20, E61H22, and E63H23 to CALIPSO–CloudSat observations from Li et al. (2012). Areas where the relative standard deviation of the different datasets compiled in Li et al. (2012) is larger than 75 % are hatched. The CALIPSO–CloudSat data cover the years 2006–2010, model data are from the PD simulations.
Cloud ice mass vertical profiles can be obtained from CALIPSO plus CloudSat
observations. The global mean vertical profile of ice water content (IWC) is shown in Fig. 7 for
all three model versions and the compiled mean IWC from Li et al. (2012).
IWC is underestimated above 700 hPa in all model versions. In E63H23 the
maximum of IWC is at the same pressure level as in the observations, at
about 350 hPa, whereas in E61H22 and E55H20 the maximum of IWC is at higher
altitudes at about 300 to 250 hPa. This can be explained by changes in ICNC
and subsequent changes in precipitation formation and ice crystal
sedimentation. ICNC changed between the model versions because the way
detrained ice crystals are added to existing stratiform clouds has changed
since E61H22. The shape of the ice crystals has been made consistent in all
modules since E61H22, and a bug in E61H22 was removed in E63H23, which led to
homogeneous freezing of dry aerosol particles independent of the availability
of water vapor below
Comparison of global annual mean IWC as a function of pressure of E55H20, E61H22, and E63H23 to CALIPSO–CloudSat observations from Li et al. (2012). Gray shading indicates the uncertainty in the CALIPSO–CloudSat observations. The CALIPSO–CloudSat data cover the years 2006–2010, model data are from the PD simulations.
The regions where IWP is underestimated in the three model versions
correspond to the regions where the three model versions underestimated LW
CRE in Fig. 2 (in particular in the tropics). There are also regions where
LW CRE is overestimated (see Fig. 2) in the three model versions, although
IWP is underestimated (see Fig. 6). This is an indication that ICNC is too
large in the three model versions (the vertical profile of IWC agrees fairly
well with observations, although the IWC magnitude is underestimated in all
three model versions). As IWP is larger in E63H23 than in E61H22 and E55H20
but the overestimation in LW CRE is smaller in E63H23 than in E61H22 and
E55H20, this is an indication that ICNC and the size of the ice crystals are
closer to reality in E63H23 than in E61H22 and E55H20. The overestimation of
LW CRE in E61H22 around 60
Next to E61H22 there is also a bias of net CRE south of 60
In Fig. 8 the total precipitation of all model versions and GPCP2.3 (Adler et al., 2018) is shown. Areas where the relative uncertainty of the GPCP2.3 data is larger than 75 % are hatched. Despite the biases in the representation of clouds in the three model versions identified above, the geographical distribution and magnitude of the annual mean precipitation of all model versions agree well with the observations. Only in the intertropical convergence zone (ITCZ) and South Pacific convergence zone (SPCZ) do the areas and magnitude of precipitation differ from the observations, corresponding to differences in cloud cover and IWP (Figs. 3 and 6, respectively). Cloud cover, IWP, and precipitation are low in the central Pacific and central Atlantic ITCZ but relatively large in the ITCZ west of Central America, east of South America, over the Philippines, and west of Southeast Asia. In the SPCZ cloud cover, IWP, and precipitation are relatively large compared to the respective observations. ECHAM underestimates tropical precipitation over land and overestimates tropical precipitation over ocean (Mauritsen et al., 2012; Stevens et al., 2013). This bias can also be seen in Fig. 8 for all ECHAM–HAM versions. Since ECHAM and ECHAM–HAM use the same parameterizations for convective clouds, this bias is very likely inherited from the base model ECHAM.
Comparison of annual mean precipitation (stratiform
In Fig. 9 histograms of cloud-top pressure vs. cloud optical depth of
ECHAM–HAM are compared to ISCCP, AVHRR-PM, and MODIS observations. The COSP
simulator (Bodas-Salcedo et al., 2011) was not implemented in E55H20 so we
only compare E61H22 and E63H23 to the observations. We applied the
COSP–ISCCP simulator for E61H22 and E63H23 for comparison to ISCCP and
AVHRR-PM. The COSP–MODIS simulator is only implemented in E63H23 so we
compare only E63H23 to MODIS. The histograms were produced for four regions
(shown in Fig. 2): wider stratocumulus regions, midlatitudes, tropics, and
60
Histograms of cloud-top pressure vs. cloud optical depth of E61H22 and E63H23 compared to ISCCP, MODIS, and AVHRR-PM observations for different regions. The definition of the four regions shown is described in the text and the regions are shown in Fig. 2. The ISCCP data are for 2000–2008, MODIS data are for 2003–2012, AVHRR-PM data are for 2003–2012, and the model data are from the PD simulations.
Figure 10 shows a Taylor diagram (Taylor, 2001) comparing SW and LW CRE, cloud
cover, LWP-LP, cloud-top CDNC, IWP, and precipitation of the three model
versions to the respective observations. The standardized deviations of
LWP-LP had to be scaled by a factor of
Taylor diagram for comparison of SW and LW CRE, cloud cover,
LWP-LP, cloud-top CDNC, IWP, and precipitation of E55H20,
E61H22, and E63H23 to observations as REF. The standardized deviations of
LWP-LP are scaled by a factor of
Several biases in the representation of clouds in the three ECHAM–HAM model
versions could be identified. The common problem of GCMs in their
representation of convective and boundary layer clouds is also present in
the three ECHAM–HAM model versions. Stratocumulus clouds are underestimated
in all three model versions. Shallow convective clouds are underestimated in
E61H22 and E55H20. In E63H23 the cloud cover and LWP in regions where
shallow convective clouds are common agree well with observations, but the
cloud-top CDNCs are overestimated, leading to an overestimation of SW CRE in
these regions. Deep convective clouds over the Atlantic and Pacific oceans
form too close to the continents (see Figs. 3, 6 and 8) in E63H23 and ECHAM
(Stevens et al., 2013). For the tropical Atlantic this is a common bias in
GCMs with coarse horizontal resolution (Siongco et al., 2014).
Siongco et al. (2017) discuss different ways this bias in the tropical Atlantic
precipitation could be reduced in ECHAM6. IWP is underestimated in all three
model versions, in particular in the tropics, whereas LW CRE and the
vertical profile of cloud ice agree rather well with observations. This
indicates that ICNC may be overestimated in all three model versions (since
LW CRE depends on the cloud temperature (
In Fig. 11 global maps of SW and LW ERF
Global maps of SW, LW, and net ERF
The treatment of surface albedo over land, ocean, and sea ice has changed
substantially from ECHAM5 to ECHAM6 (see Stevens et al., 2013), which has an
impact on SW ERF
It is interesting to note that although biases in the simulation of clouds
in stratocumulus regions are reduced in E63H23, there seems to be no
increase in ERF
Most of the differences between the model versions discussed above are
differences in SW ERF
Global maps of all-sky and clear-sky net RF
Tegen et al. (2019) found an improved aerosol representation in biomass
burning regions when GFAS biomass burning emissions, multiplied by a scaling
factor of 3.4 as recommended by Kaiser et al. (2012), replace the default
ACCMIP biomass burning emissions. Therefore, we performed an E63H23
simulation with GFAS biomass burning emissions multiplied by 3.4
(E63H23-GFAS34). E63H23-GFAS34 has a weaker ERF
We would like to point out that our simulations include interactions between
sulfate and mineral dust. On the one hand, (anthropogenic and natural)
gaseous sulfate may coat mineral dust particles; this leads to a transfer of
dust from insoluble modes to soluble modes in the models, which increases
the wet deposition of dust (and leads to decreased present-day mineral dust
burdens; see Table S2). On the other hand, mineral dust particles
provide surfaces onto which (anthropogenic and natural) gaseous sulfate may
condensate, leading to a dampening of the nucleation of new particles.
Similar interactions between sulfate and mineral dust have been found by Fan
et al. (2004) (using the Geophysical Fluid Dynamics Laboratory (GFDL) global
chemical transport model; Mahlman and Moxim, 1978), Bauer and Koch (2005),
and Bauer et al. (2007) (using the Goddard Institute for Space Studies
(GISS) climate model, modelE; Schmidt et al., 2006; Hansen et al., 2005).
The forcing from these interactions between sulfate and mineral dust is
included in our estimates for ERF
RF
The equilibrium climate sensitivity (ECS) is strongest in E55H20 (3.5 K),
weaker in E61H22 (2.8 K), and weakest in E63H23 (2.5 K) (Fig. 13). The
corresponding ECS values in the base model versions are: ECHAM5: 3.4 K
(Randall et al., 2007; their Table 8.2), ECHAM6.1: 2.8 K (Block and
Mauritsen, 2013; Meraner et al., 2013) and ECHAM6.3: 2.8 K (Mauritsen et
al., 2019); i.e., changes in ECS between the ECHAM–HAM model versions are
driven substantially by changes in the ECHAM base model versions. Note that
the ECS value for ECHAM6.3 is from a simulation with abruptly quadrupled
Global mean ERF
The largest differences between E61H22 and E63H23 in terms of
ERF
The liquid phase of clouds is better represented in E63H23 than in the
previous model versions because the low bias in cloud cover in the
subtropics is reduced and the zonal distribution of LWP agrees with
observations. This also leads to a better agreement of the SW CRE with
observations in E63H23. Important reasons for these improvements are the
change in the fractional cloud cover scheme for marine stratocumulus clouds,
the removal of an inconsistency that had led to either 0 or 1 cloud
cover in ECHAM6.3, and subsequent changes in model tuning (Mauritsen et al.,
2019). Furthermore E63H23 uses the Abdul-Razzak and Ghan (2000) activation
scheme and the Long et al. (2011) sea salt emission parameterization
(temperature dependent; Sofiev et al., 2011), which leads to higher CDNCs when LWP is large. The Abdul-Razzak and Ghan (2000)
activation scheme is more physically realistic than the empirical Lin and
Leaitch (1997) activation scheme used in the previous model versions as it
is Köhler theory based and therefore takes into account the size of the
aerosol particles and their chemical composition. Although the Abdul-Razzak
and Ghan (2000) activation scheme has limitations under certain conditions
(the assumption that the aerosol particles are in equilibrium with its
environment is not valid in all conditions; Phinney et al., 2003) and does
not account for preexisting cloud droplets during cloud droplet activation
(Barahona et al., 2010), it certainly helps to improve the representation of
cloud droplets in E63H23. The performance of the new Long et al. (2011)
(temperature dependent; Sofiev et al., 2011) and the old
Guelle et al. (2001) sea salt parameterizations in E63H23 was
analyzed by Tegen et al. (2019). The new temperature dependence leads
to increased sea salt emissions
where the sea surface temperature is warmer than 20
Components of the net global mean cloud feedback parameter of
E61H22 and E63H23 for low (cloud-top pressure (CTP)
Also the ice phase of clouds has improved in E63H23 compared to previous
model versions. The low bias in IWP is reduced in E63H23 and the global mean
vertical IWC is within the observational range (Fig. 7). This is because the
Seifert and Beheng (2006) sticking efficiency used in E63H23 leads to a less
efficient removal of ice crystals by snow. A subsequent reduction in the
tuning parameter for stratiform snow formation by aggregation further
increases IWP in E63H23. Only a few laboratory studies for sticking efficiency
have been conducted, and even fewer theories for sticking efficiency have been
developed (Phillips et al., 2015). We find that the simple formulation of
Seifert and Beheng (2006) for sticking efficiency for the accretion of ice
crystals by snow improves the simulation of cloud ice in E63H23.
Furthermore, the altitude of the global mean maximum IWC agrees well with
observations in E63H23, whereas in E61H22 and E55H20 it is at higher
altitudes than observed. This can be explained by the changes in ICNC described
in Sect. 2.1.5, such as the use of a consistent ice crystal shape
(hexagonal plates), removal of an ICNC bug, or the changed treatment of
detrained ice crystals. The subsequent changes in precipitation formation
and ice crystal sedimentation can then lead to a different vertical
distribution of cloud ice. In E61H22 the global ICNC burden is considerably
higher than in the other two model versions because of an inconsistency
between cloud droplet activation, condensation, vertical transport of CDNC,
and homogeneous freezing of cloud droplets in cirrus clouds, which led to
homogeneous freezing of aerosol particles even when the water vapor pressure
was too low for homogeneous nucleation. The higher ICNCs are also responsible
for the LW component of ERF
While the global mean values of RF
The weaker ECS in E63H23 compared to E61H22 can be linked to changes in cloud feedbacks. There are indications for a stronger cloud phase feedback in non-low clouds due to increased CDNC and changes in cloud water in E63H23. A stronger (cooling) cloud phase feedback will lead to less warming in the future. Similarly, the less positive cloud amount feedback of low clouds (related to model tuning in ECHAM6.3) in E63H23 contributes to the weaker ECS in E63H23 compared to E61H22.
The changes and improvements in E63H23 (including changes in the base model
version ECHAM6.3) have therefore not only improved the representation of
clouds in E63H23 compared to previous model versions, but they have also had an
impact on ERF
Clouds in the current (E623H23) and previous (E55H20 and E61H22)
versions of the ECHAM–HAM global aerosol–climate model were evaluated using
a suite of global observational datasets for clouds and precipitation.
Improvements in E63H23 compared to previous model versions for cloud water
include the following:
a more physically based activation scheme (Abdul-Razzak and Ghan, 2000); changes in the treatment of CDNC detrained from convective clouds (as
described in Sect. 2.1.5); and an increase in low clouds (Mauritsen et al., 2019), all of which together lead to a more realistic LWP globally. different sticking efficiency for the accretion of ice crystals by snow
(Seifert and Beheng, 2006); consistent ice crystal shapes throughout the model (Lohmann and Neubauer,
2018); changes in mixed-phase freezing (as described in Sect. 2.1.5); and the removal of an inconsistency in ICNC in cirrus clouds, all of which together lead to a more realistic IWP globally.
For cloud ice the improvements include the following:
The sum of the changes leads to improved cloud radiative effects. Although
the representation of shallow convective clouds has improved in E63H23,
stratocumulus clouds are still underrepresented. The comparison of the
different model versions showed that the misrepresentation of certain cloud
types can lead to compensating biases in other clouds via model tuning.
Therefore, if the bias in stratocumulus clouds in E63H23 can be reduced,
this could also improve the representation of other cloud types. Reasons for
the bias in stratocumulus clouds identified by Neubauer et al. (2014) in
E61H22 were, e.g., turbulent mixing that is too strong at cloud top, the shallow
convection scheme triggering too often, and/or a lack of vertical resolution.
Future work will focus on addressing these difficult issues.
Deep convective clouds over the Atlantic and Pacific oceans form too close
to the continents, which leads to biases in the geographical distribution of
precipitation in E63H23 and ECHAM (while tropical land precipitation is
underestimated). While the geographical (except for deep convective clouds)
and vertical distributions of cloud ice agree well with observations in
E63H23, IWP remains biased low. The combination of observations of IWP, LW
CRE, and the vertical distribution of cloud ice indicate that ICNC may be
overestimated in ECHAM–HAM. Previous studies with ECHAM–HAM showed that the
bias in ICNC and IWP can be reduced when heterogeneous freezing of ice-nucleating particles and/or water vapor deposition on preexisting ice
crystals are accounted for in cirrus clouds.
Estimates of ERF the more realistic simulation of cloud water, but also the new activation scheme in E63H23 and the new sea salt emission parameterization, the removal of an inconsistency in ICNC in cirrus clouds leading to a weaker
LW ERF
which lead to a weaker SW ERF
Since there are reductions in both SW and LW ERF
ECS is weaker in E63H23 (2.5 K) than in E61H22 (2.8 K) (and E55H20; 3.5 K).
The decrease compared to E61H22 is due to the following:
changes in the entrainment rate for shallow convection adopted from the base
model ECHAM6.3 (which leads to a less positive feedback of cloud amount of
low clouds in some regions) and a stronger cloud phase feedback.
Although the differences in both ERF
The ECHAM-HAMMOZ model is made freely available to the scientific community
under the HAMMOZ Software License Agreement, which defines the conditions
under which the model can be used. More information can be found at the
HAMMOZ website (
Scripts can be found at
Data can be found at
The supplement related to this article is available online at:
DN designed the evaluation methodology with comments from the coauthors. DN performed the experiments and the analysis of the data. CS and SF provided the support needed to run the ECHAM–HAM model versions. All coauthors were involved in the development of the ECHAM-HAMMOZ model. DN wrote the paper with comments from coauthors.
The authors declare that they have no conflict of interest.
The ECHAM-HAMMOZ model is developed by a consortium composed of ETH Zurich, the Max-Planck-Institut für Meteorologie, Forschungszentrum Jülich, the University of Oxford, the Finnish Meteorological Institute, and the Leibniz Institute for Tropospheric Research; it is managed by the Center for Climate Systems Modeling (C2SM) at ETH Zurich. This study has received partial funding from the Center for Climate System Modelling (C2SM) at ETH Zurich. The computing time for this work was supported by a grant from the Swiss National Supercomputing Centre (CSCS) under project ID s652 and from ETH Zurich. The CERES satellite data were obtained from the NASA Langley Research Center Atmospheric Science Data Center. We would like to thank Sebastian Rast and the anonymous reviewers for useful comments and suggestions.
This research has been supported by the Swiss National Science Foundation (grant no. 200021_160177), the European Union's Seventh Framework Programme (FP7/2007-2013) project BACCHUS (grant no. 603445), the European Research Council project RECAP (grant no. 724602), and the Academy of Finland (grant nos. 308292 and 307331).
This paper was edited by Samuel Remy and reviewed by two anonymous referees.