The study of climate and climate change using general circulation models (GCMs) continues to advance rapidly, with impetus from widespread societal concern about anthropogenic and natural climate change, unprecedented global and field observational programs, and advances in theoretical and process-level understanding of atmospheric, oceanic, cryospheric, and terrestrial processes. The purpose of this paper is to describe recent development in the atmospheric component (AM3) of the Geophysical Fluid Dynamics Laboratory (GFDL) coupled model (CM3). AM3 is built upon the scientific and software framework of GFDL AM2 (Anderson et al. 2004). Its major developmental thrusts were chosen to enable AM3 to explore several key, emerging questions in climate and climate change that could not be addressed with AM2, such as: 1) What are the roles of aerosol–cloud interactions, specifically indirect effects of aerosols? 2) What are the dominant chemistry–climate interactions? AM3 development also aimed at enhanced capabilities for addressing emerging questions when coupled with biogeochemical and ocean models: 3) What is the interplay between climate and key biogeochemical cycles? 4) To what extent is decadal prediction possible? The model also includes advances in the dynamical core, radiation, and other components.
Addressing these scientific questions implies particular approaches to AM3 development. To model aerosol–cloud interactions using a physically based treatment of aerosol activation, parameterizations for subgrid variability of vertical velocity are important. This is because aerosol activation depends strongly on local vertical velocity, which, for both stratiform and convective clouds, can depart strongly from the large-scale (in AM3, grid-scale) average. AM3 parameterizes subgrid vertical velocities for all clouds. To study chemistry–climate interactions, AM3 specifies chemical emissions and includes large-scale and convective transport, wet and dry removal, and key tropospheric and stratospheric reactions. The AM3 stratospheric resolution has been increased, and its upper boundary has been raised, to treat stratospheric processes more comprehensively. AM3 itself does not include carbon, nitrogen, or other biogeochemical cycles, but particular attention has been given to improving its simulation of tropical precipitation in order to enhance its usefulness as a component of earth system models. AM3’s improved stratospheric resolution is also necessary for future research on phenomena such as the Southern Hemisphere annular mode, which likely plays a role in interannual variability important for decadal prediction (Thompson and Solomon 2006).
Section 2 describes the AM3 dynamical core. Section 3 presents its physical parameterizations, while appendix A presents brief summaries of the land, ocean, and sea ice models used with AM3 in CM3. Section 4 illustrates basic simulation characteristics of AM3 with prescribed sea surface temperatures and, as CM3, in coupled mode. The inclusion of aerosol–cloud interactions in CM3 links cloud radiative properties to aerosols, whose optical properties and direct effects on shortwave radiation agree better with observations than in CM2. CM3’s simulation of the increase in global-mean surface temperature from 1880–1920 to 1980–2000 is smaller than observed. The corresponding CM2 simulated increase is larger than observed. The magnitude of CM3’s underestimate is about 0.1°C larger than the CM2 overestimate.
2. Dynamical core
As in CM2.1 (Delworth et al. 2006), the dynamical core used in AM3/CM3 follows the finite-volume algorithms described in Lin and Rood (1996, 1997) and Lin (1997, 2004), with the following major modifications.
In an effort to enhance the model’s parallel computing efficiency and to improve simulation quality in polar regions, the dynamical core formulated on, and optimized specifically for, the latitude–longitude grid has been significantly modified to use a general curvilinear coordinate system. The nonorthogonal gnomonic projection in the general cubed-sphere geometry described by Putman and Lin (2007) is chosen for its excellent grid uniformity and better overall accuracy. The spatially more uniform gnomonic projection was chosen over the conformal projection for its computational efficiency. The largest dynamical time step allowed with the gnomonic grid is several times larger than that of the conformal grid. In idealized tests (Putman and Lin 2007), the solution with the gnomonic grid is also slightly more accurate. The use of the nonorthogonal coordinate system necessitated major changes to the transport operators (Putman and Lin 2007) and the need to compute both covariant and contravariant wind components (e.g., Sadourny 1972).
Compared to the original latitude–longitude grid formulation, the use of the cubed-sphere grid in the new finite-volume core greatly improved the computational efficiency owing to two major algorithmic modifications. First, the flux-form semi-Lagrangian extension (Lin and Rood 1996) needed to stabilize the (large time step) transport processes near the poles is no longer needed with the use of the cubed-sphere grid. Second, and related to the first, the polar Fourier filtering required for the stabilization of fast waves is also no longer needed. Both modifications led to greatly improved computation and communication load balancing, enabling the efficient use of 2D domain decomposition on each of the six faces of the cube.
The model’s horizontal resolution is denoted as Cn, where n is an integer number indicating the total number of cells (finite volumes) along each edge of the cube. In AM3, the model’s resolution is C48. The total number of cells on the sphere is therefore 6 × 48 × 48 = 13 824, and the size of the grid cell varies from 163 km (at the six corners of the cubed sphere) to 231 km (near the center of each face). The C48 resolution model scales roughly an order of magnitude better (can use 864, versus 30, central processing units) than its latitude–longitude counterpart (2° × 2.5° resolution) used in CM2.1, enabling nearly the full use of the GFDL 1024-core SGI Altix-3000 system.
The vertical coordinate in AM3 follows Simmons and Burridge (1981), but the number of layers has been increased to 48 (from 24 layers in AM2). The uppermost level in AM3 has a pressure of 1 Pa, a height of about 86 km for a surface pressure of 1013.25 hPa, and scale height of 7.5 km (equivalently, isothermal with a temperature of approximately 256.2 K), compared to around 35 km in AM2. The augmentation in vertical levels is aimed at resolving the stratosphere sufficiently well that its basic chemical and dynamical processes can be reasonably simulated. Between the earth’s surface and about 430 hPa, the vertical levels are positioned within 10 hPa of those in AM2. Additional layers are added at lower pressures. Table 1 shows the positions of the intermediate levels, which bound the AM3 layers.
Coefficients ak and bk for calculation of interface pressures using p = ak + bk × ps, where p is pressure and ps is surface pressure (Simmons and Burridge 1981). Pressures and heights of interface levels corresponding to a scale height of 7.5 km and ps = 1013.25 hPa are also shown. See appendix B for description and units.
3. Physical parameterizations
The basic shortwave and longwave radiation algorithms are described in Freidenreich and Ramaswamy (1999) and Schwarzkopf and Ramaswamy (1999), respectively, modified as in Anderson et al. (2004). Total and spectral solar irradiances are from the Total Irradiance Monitor (TIM) (Kopp et al. 2005), as recommended for Climate Model Intercomparison Project 5 (CMIP5) (http://www.geo.fu-berlin.de/en/met/ag/strat/forschung/SOLARIS/Input_data/CMIP5_solar_irradiance.html).
1) Subgrid variability and overlap
All-sky radiative transfer calculations account for the effect of clouds using the Monte Carlo independent column approximation (Pincus et al. 2003), which treats variability by creating a set of subcolumns consistent with cloud properties (including variability) and vertical structure (i.e., overlap). The in-cloud distribution of ice and water content in stratiform clouds is diagnosed from the cloud fraction and condensate amount (Pincus et al. 2006), and vertical structure assumes that the rank correlation of total water falls off exponentially with the distance between layers using a scale height of 1 km (Pincus et al. 2005). These formulations differ from those in AM2 and allow cloud optical properties to be used as predicted, rather than being arbitrarily multiplied by 0.85 as in AM2. The radiative properties of shallow and deep convective clouds (section 3e) are also included. Convective clouds are assumed to be internally homogeneous and to obey maximum overlap. When convective clouds occur in a subcolumn they replace any stratiform clouds in layers where both clouds occur, which slightly decreases the overall stratiform cloud amount. In AM2, convective clouds were excluded from radiative transfer.
The effective radius in each subcolumn is computed assuming that the predicted cloud drop number is uniform for each cloud type within each grid-scale column. As will become evident in sections 3d and 3f, information on the subgrid distribution of cloud drop number is available but, as a simplification, has not been incorporated into the subcolumns. In stratiform clouds and shallow cumulus, drop size depends on aerosol activation, as described in section 3f.
2) Cloud optics
The sizes of cloud droplets in stratiform and shallow cumulus clouds depend on aerosol activation and are determined using the procedures described in section 3f. In deep cumulus updraft cells, the sizes of liquid droplets follow Bower et al. (1994). Size-dependent shortwave optical properties for cloud liquid follow Slingo (1989). Longwave liquid optical properties follow Held et al. (1993) and depend on water path but not particle size. AM3 does not treat ice nucleation or link ice nucleation to crystal sizes. In shallow cumulus and stratiform ice clouds, ice particle sizes are diagnosed as a function of temperature, based on aircraft observations (Donner et al. 1997) with radiative properties following Fu and Liou (1993). In mesoscale updrafts associated with deep convection, ice crystals increase in size with distance from the top of the updraft, as in McFarquhar et al. (1999), except that McFarquhar et al.’s heights are replaced with equivalent normalized fractional distances between the top and base of the mesoscale updraft. Ice crystals in cumulus cell updrafts are assigned a generalized effective size of 18.6 μm, based on observations of the early temporal evolution (most likely dominated by deep cells) of a convective system in the Central Equatorial Pacific Experiment (Fu 1996). Solar and infrared radiative properties of ice crystals in cell updrafts and mesoscale anvils are obtained from Fu (1996) and Fu et al. (1998), respectively.
3) Gas concentrations
Historical concentrations of carbon dioxide, nitrous oxide, methane, and halocarbons (CFC-11, CFC-12, CFC-113, and HCFC-22) were obtained online (from www.iiasa.ac.at/web-apps/tnt/RcpDb/, where the Representative Concentration Pathways may also be found). Note that the methane specification for radiation differs from the methane obtained from the chemistry calculations described in section 3g. Tropospheric and stratospheric ozone are modeled as described in section 3g.
4) Aerosol optics
The effects of volcanoes are included in the AM3 and CM3 simulations described in section 4. Sulfur dioxide emissions from volcanoes are described in section 3f. Direct injection of sulfur into the stratosphere from volcanic eruptions is not included, nor is carbonyl-sulfide chemistry, a major source of background stratospheric aerosol. To compensate, in the stratosphere, a time series of volcanic optical properties is specified as in Stenchikov et al. (2006).
Aerosol optical properties (i.e., extinction efficiency, single-scattering albedo, and asymmetry factor) are based on Mie theory, assuming that all particles spherical. Lognormal size distribution is assumed for sulfate and carbonaceous aerosols. The geometric mean radius and standard deviation of the lognormal distribution for sulfate and black carbon are from Haywood and Ramaswamy (1998), and for organics from Hess et al. (1998). The mass size distribution of dust and sea salt is assumed constant within five bins from 0.1 to 10 μm. Hygroscopic growth is considered for sulfate, sea salt, and aged (hydrophilic) organic carbon. We model the hygroscopic growth of sulfate after that of pure ammonium sulfate (Tang and Munkelwitz 1994), of sea salt as pure sodium chloride (Tang et al. 1997), and of hydrophilic organics as a mixture of acids and insoluble organics (Ming et al. 2005). The refractive indices of sulfate and black carbon are from Haywood and Ramaswamy (1998), organics from Hess et al. (1998), sea salt from Tang et al. (1997), and dust from Balkanski et al. (2007), assuming a 2.7% content of hematite. Internal mixture of sulfate and aged (hydrophilic) black carbon is calculated by volume-weighted average of their refractive index. All other aerosols are assumed externally mixed.
b. Gravity wave drag
Orographic gravity wave drag is parameterized using Stern and Pierrehumbert (1988), as described in Anderson et al. (2004). Nonorographic gravity wave drag is parameterized using Alexander and Dunkerton (1999), which treats vertical propagation of wave components of a spectrum of gravity waves with a range of phase speeds and horizontal wavelengths, assuming that the momentum associated with each wave component is deposited locally at the level of linear wave breaking. There are uncertainties in the seasonal, latitudinal, and height dependencies of gravity wave sources and sinks. Alexander and Rosenlof (2003) found that parameters related to the sources and sinks varied from the tropics to the extratropics. In the AM3 application of Alexander and Dunkerton (1999), the momentum source is represented by a broad spectrum of wave speeds (half-width of 40 m s−1) with a resolution of 2 m s−1 and a single horizontal wavelength of 300 km. The amplitude of the momentum source is 0.005 Pa in the northern middle and high latitudes, 0.004 Pa in the tropics, and 0.003 Pa in the southern middle and high latitudes, with smooth transitions around 30°N and S. The sources themselves are not observationally based but have been chosen based on the circulation they yield. For example, the asymmetry in the northern and southern sources improves the simulation of stratospheric zonal winds and polar temperatures. The wave launch height decreases smoothly from 350 hPa at the equator to near the surface at the poles. Optimizing the input parameters was eased by limiting the influence of the orographic wave drag parameterization to below 30 hPa. The scheme yields a reasonable semiannual oscillation. However, the vertical resolution employed here is not sufficiently fine to enable simulation of the quasi-biennial oscillation (Giorgetta et al. 2006).
c. Turbulence and planetary boundary layer
Turbulence and planetary boundary layers (PBLs) in AM3 are treated as in AM2. Lock et al. (2000) is used for convective PBLs and stratocumulus layers. Louis (1979) is employed for other unstable layers. Stability functions with thresholds dependent on Richardson number are adopted for stable layers. Variations in vertical diffusion coefficients are damped. Full details can be found in Anderson et al. (2004).
d. Stratiform clouds
Microphysical processes except for activation of liquid cloud drops (described in section 3f) follow Rotstayn (1997) and Rotstayn et al. (2000), as described in Anderson et al. (2004). The number of activated aerosols depends on aerosol mass, composition, and vertical velocity. To account for the effect of subgrid variability, the vertical velocity is assumed to be normally distributed within each model grid box and the activation computed by integration over this distribution following Ghan et al. (1997). The mean of the distribution is the velocity driving the stratiform condensation in the Tiedtke (1993) parameterization, and the standard deviation is related to the turbulence mixing coefficients. A minimum standard deviation of 0.7 m s−1 is imposed. The integration is performed numerically using a 64-point Gauss–Hermite quadrature. Through its control on aerosol activation, subgrid variability in vertical velocity is a major factor in the magnitude of aerosol indirect effects (Golaz et al. 2011).
Several parameters in the Tiedtke (1993) parameterization have been altered from their Anderson et al. (2004) values. The critical droplet radius for autoconversion is 8.2 μm. Autoconversion thresholds as low as 4.5 μm have been used in GCMs. [See Rotstayn (2000) for a detailed discussion]. The threshold value was 10.6 μm in AM2 (Anderson et al. 2004) and 8.0 μm in AM2.1 (Delworth et al. 2006). The erosion constants when vertical diffusion is active, when convection (shallow, deep, or both) is active without vertical diffusion, and when neither convection nor diffusion is active are 7 × 10−5 s−1, 7 × 10−5 s−1, and 1.3 × 10−6 s−1, respectively. Erosion scales are larger in AM3 than in AM2/AM2.1, but we note that Anderson et al. (2004) pointed out that the erosion scale under convective conditions in AM2 might be 40 times too small compared to large-eddy simulation (LES) results from Siebesma et al. (2003).
The ice fall speeds follow Heymsfield and Donner (1990), multiplied by a factor of 1.5. The Heymsfield and Donner (1990) fall speeds are not parameterized to include rimed particles, which have higher fall speeds than ice-only particles. The applied multiplicative factor produces slightly better agreement between the AM3 fall speeds and fall speeds observed at the Southern Great Plains site of the Atmospheric Radiation Program. Riming occurs in these observed particles and has been included in a new parameterization (Lin et al. 2011), which is likely to be used in future model development.
As the foregoing discussion indicates, the changes to autoconversion threshold, erosion constants, and fall speeds are regarded as within observational or conceptual uncertainties, given the design of the parameterizations. The changes were chosen to increase realism of the simulations, particularly with regard to radiation balance, precipitation, and implied ocean heat transports in AM3 integrations with prescribed sea surface temperatures (SSTs).
e. Cumulus convection
Deep cumulus systems consist of deep updraft cells, mesoscale updrafts, and mesoscale downdrafts (Donner 1993; Donner et al. 2001; Wilcox and Donner 2007). Several modifications have been made in AM3 for computational efficiency or simulation improvement. The plumes in the deep updraft cells are discretized on the AM3 vertical grid instead of a higher-resolution cloud grid. With the coarser plume resolution, entrainment coefficients have been increased relative to those in Donner (1993) by a factor of 1.65. Liquid/frozen-water static energy (conservative without precipitation) is used instead of temperature for plume thermodynamics. Aspects of the water budget in deep convective systems related to Rm, precipitation from mesoscale updrafts; Eme, condensate transfer from mesoscale updrafts to large-scale stratiform clouds (cf. section 3d); Cmu, condensation and deposition in mesoscale updrafts; and CA, lateral transfer of condensate from deep updraft cells to mesoscale updrafts, have been modified. In particular,
The closure for deep cumulus results in heating by cumulus convection relaxing convective available potential energy (CAPE) toward a threshold over a relaxation time scale [cf. Eq. (2) in Wilcox and Donner (2007)]. The CAPE threshold is 1000 J kg−1, and the relaxation time scale is 8 h. These choices were motivated primarily by their effect on the precipitation simulation.
Shallow cumulus follows Bretherton et al. (2004), modified as in Zhao et al. (2009). The shallow scheme of Bretherton et al. assumes a single bulk entraining and detraining plume with entrainment/detrainment rate determined by a parcel buoyancy sorting algorithm. The scheme includes a plume vertical momentum equation and a parameterization of cloud-top penetrative mixing of air between the level of neutral buoyancy and the maximum vertical extent of the plume. The buoyancy sorting representation of the inhomogeneous shallow cumulus mixing as well as the cloud-top penetrative mixing is supported by both observations (e.g., Paluch 1979; Grinnell et al. 1996) and large-eddy simulations of shallow cumulus clouds (e.g., Zhao and Austin 2005a,b). The closure for cloud-base mass flux in the plume is determined by estimates of the boundary layer turbulent kinetic energy and of the convective inhibition. The important nondimensional parameter controlling the strength of the lateral mixing [c0 in Eq. (18) in Bretherton et al. (2004)] is set to be 13.5 in AM3.
Both deep and shallow cumulus diffuse large-scale horizontal momentum in proportion to their mass fluxes, as in Anderson et al. (2004). The nondimensional constant γ in Eq. (1) of Anderson et al. (2004), which is a factor with the cumulus mass flux in the term added to the vertical diffusion coefficient, takes the value 0.26 in AM3. The Anderson et al. (2004) value is 0.20.
Finally, moist adiabatic adjustment (MAA) (Manabe et al. 1965) has been retained since a saturated atmosphere at grid scale should not be unstable or moist beyond saturation. The parameterizations for deep and shallow cumulus do not preclude these conditions, which produce small amounts of precipitation relative to other sources.
The changes in entrainment coefficients for deep convective plumes, water-budget partitioning for deep convective systems, maximum heights for mesoscale circulations in deep convective systems, and lateral entrainment for shallow cumulus relative to Donner (1993) and Bretherton et al. (2004) account for implementation issues in AM3 and simulation deficiencies using the referenced values. The Donner entrainment coefficients for deep cumulus updrafts were selected to yield cumulus vertical velocities in agreement with observations on a finer vertical grid for the cumulus updrafts than is used in AM3. Among other simulation characteristics, shortwave cloud forcing is sensitive to the water budget partitioning for deep convective systems. Restricting the maximum height for mesoscale circulations in deep convective systems is necessary to prevent excessive water vapor in the stratosphere. Increasing the lateral entrainment in shallow cumulus limits excessive low cloudiness. The nondimensional constant γ related to diffusion of horizontal momentum in convection has been changed from its value in Anderson et al. (2004), where its impacts on tropical precipitation and surface wind stresses are noted. As with the stratiform parameter settings discussed in section 3d, these changes resulted in improved realism in key aspects of the atmospheric circulation—for example, precipitation and implied ocean heat transports (important for coupling and, in turn, depending strongly on shortwave cloud forcing and surface wind stress).
These changes are within observational or conceptual constraints. As noted above, the entrainment coefficients in Donner (1993) were chosen based on observed vertical velocities within the framework of the one-dimensional plume model used to represent updrafts in the cumulus parameterization. The coarser grid, used for computational efficiency in AM3, changes the relationship between entrainment coefficients and vertical velocities in the plumes and can be compensated for by changing the entrainment coefficients. Leary and Houze (1980), the basis for the moisture-partitioning parameters in Donner (1993), show these parameters to depend strongly on assumptions regarding the extent of mesoscale activity in deep convection. The extensive range in observed sizes of tropical cloud clusters suggests a wide range in the extent of mesoscale activity associated with deep convection and that the altered moisture-partitioning parameters remain within observational constraints. [Mapes and Houze (1993) report a range from under 2000 to 100 000 km2.] The strong temperature inversions generally observed at the tropopause support restricting the maximum height of the convective mesoscale circulations not to exceed the local tropopause as a reasonable approximation to observed behavior. Lateral entrainment in shallow cumulus departs by only 10% from Bretherton et al. (2004), whose value was obtained from a large-eddy simulation based on a single observed case. The value of γ in AM3 is 30% higher than the upper range reported to be consistent with cloud models in Anderson et al. (2004). That range required assumptions regarding the vertical structure of the mean flow and the relationship between convective mass fluxes and rain rate, which can be relaxed to allow γ to vary 30%.
In the AM3 integration described in section 4a, deep convective cells dominate in the middle and upper troposphere in the tropics, but at pressures of 100 to 200 hPa the mass fluxes in mesoscale updrafts are comparable to those in the cells (Fig. 1). Mesoscale downdrafts have the smallest mass fluxes among the convective components but can extend to the PBL where changes by these downdrafts in thermodynamic and moisture structure can impact surface fluxes. Shallow cumulus can coexist with deep convection and, although its vertical extent is not imposed, generally is confined below about 500 hPa. Deep convection can only occur when the level of zero buoyancy is at a pressure less than 500 hPa. Both are called from the same atmospheric state. In AM3, deep convective precipitation dominates in the tropics, while stratiform precipitation prevails in the middle latitudes (Fig. 2a). The small values of precipitation associated with MAA indicate that the other precipitation parameterizations generally preclude the development of oversaturated, unstable conditions. The midlatitude maxima in precipitation from the MAA coincide with the edges of the faces of the cubed-sphere in the dynamical core. Relative to precipitation reported by version 2 of the Global Precipitation Climatology Project (GPCP v.2) (Adler et al. 2003), AM3 produces 16% excessive precipitation. In CM3, described in section 4, sea surface temperatures depart from the observed values specified in the AM3 integrations when AM3 is coupled to ocean and sea ice models, with appreciable effects on precipitation patterns (Fig. 2b). Most notably, a double intertropical convergence zone (ITCZ), not evident in GPCP v.2, is apparent. This double maxima occurs in all of the parameterized sources of precipitation, despite wide variations in the ways in which the occurrence of precipitation in these parameterizations is related to large-scale flows. The departure of CM3 precipitation patterns from AM3 patterns is typical when coupling atmospheric and oceanic GCMs and is evidently a consequence of a chain of interactions between the ocean and atmosphere components (e.g., Zhang et al. 2007).
AM3 calculates the mass distribution and optical properties of aerosols based on their emission, chemical production, transport, and dry and wet removal. The transport processes include advection, convection, and eddy diffusion by turbulence. The chemical production of sulfate includes gas and aqueous-phase oxidation of sulfur dioxide by radicals, ozone, and hydrogen peroxide, which are calculated explicitly by the chemical mechanism described in section 3g. Dry deposition includes gravitational settling and impaction at the surface by turbulence. Wet deposition takes into account in- and below-cloud scavenging by large-scale and convective clouds.
Anthropogenic and biomass burning emissions of sulfur dioxide, black carbon, and organic carbon are from Lamarque et al. (2010). Dimethyl sulfide (DMS) emission is calculated using an empirical formula as a function of seawater DMS concentration and wind speed at 10 m, as described by Chin et al. (2002).
Secondary organic aerosols are produced by terrestrial and oceanic sources. Terrestrial production includes natural and anthropogenic sources. The natural source includes oxidation of terpenes emitted from plants, which yields particulate organics (Dentener et al. 2006). The yield factor varies from 0.11 per molecule at latitudes lower than 20° to 0.55 per molecule at the poles. The anthropogenic source follows Tie et al. (2005), where 10% of the butane oxidized by hydroxyl radicals becomes particulate organics. The oceanic source is the O’Dowd et al. (2008) organic sea-spray source function. Anthropogenic and natural secondary organic aerosol production is 11.3 and 31.5 Tg yr−1, respectively.
Dust emission follows the parameterization by Ginoux et al. (2001) and is based on the preferential location of sources in topographic depressions. Sea salt particles are emitted from the ocean according to Monahan et al. (1986).
For volcanoes, time-invariant sulfur dioxide emissions are specified to be the total sulfur emissions recommended by AeroCom (Dentener et al. 2006) for continuous degassing and (time-averaged) explosive emissions, multiplied by a factor of 0.25. These emissions are injected 500 to 1500 m above volcano tops for explosive emissions and over the upper third of volcanoes for continuously degassing volcanoes and are thus confined to the troposphere. The factor applied is justified by the need to scale the total sulfur emissions to include only sulfur dioxide emissions and to simulate realistic sulfur dioxide and sulfate abundances in otherwise clean regions with volcano sources, noting that considerable uncertainty exists in volcanic emissions. Owing to the absence of some chemical processes important for the formation of stratospheric volcanic aerosols, for example, related to carbonyl sulfide, and the absence of direct injection of volcanic aerosols into the stratosphere, a stratospheric signature for volcanoes is imposed through the specification of a time series of spatial distributions of optical properties, as noted in section 3a.
Following Cooke et al. (1999), we assume that 80% of black carbon and 50% of organics emitted are hydrophobic, the rest being hydrophilic. Hydrophobic black carbon and organic aerosols undergo aging processes to become hydrophilic with e-folding times of 1.44 and 2.88 days, respectively. Secondary organic aerosols are treated as hydrophilic.
Chemical processes related to aerosol formation are discussed in section 3g. Aerosols are removed by dry deposition at the surface and scavenging in stratiform and convective clouds. Dry deposition velocities for aerosols are calculated interactively using a wind-driven resistance method in which the surface resistance is calculated as an empirical parameter (reflecting surface collection efficiency) divided by the friction velocity (Gallagher et al. 2002).
Cloud scavenging of aerosol species is calculated following Giorgi and Chameides (1985). The fractional removal rate is equal to its in-condensate fraction multiplied by the fractional removal rate of condensate by precipitation. For hydrophilic aerosols, an in-condensate fraction (ranging from 0.07 for dust to 0.3 for sulfate in large-scale clouds, and from 0.12 for dust to 0.4 for sulfate in convective clouds) is prescribed. These fractions qualitatively correspond to the relative solubilities and cloud drop nucleation properties of the aerosols, but the quantitative values are selected (globally) to provide a reasonable simulation of the global mean and regional patterns of aerosol optical depth (AOD). Below-cloud aerosol washout, for large-scale precipitation only, is parameterized as described by Li et al. (2008).
Interactive simulation of aerosols from emissions in CM3 is a major change in approach from CM2.1 (Delworth et al. 2006) in which aerosol concentrations were specified. AM3 uses different emissions inventories and optical properties than AM2. AM3 also includes internal mixing and couples wet deposition to cloud microphysics. A detailed evaluation of aerosol properties is beyond the scope of this paper. Here, two fundamental CM3 aerosol properties, aerosol optical depth and coalbedo (ratio of absorption optical depth to total optical depth), are compared with AERONET observations to show improved correlation relative to CM2.1. As analyzed in detail by Ginoux et al. (2006), the CM2.1 aerosol distribution tended to overestimate AOD in polluted regions, while underestimating biomass-burning AOD by a factor of 2 or more, relative to annual-mean AOD measured by AERONET sun photometers (Holben et al. 1998) (Figs. 3a and 3b). Ginoux et al. also indicate that sea-salt mass was largely underestimated but compensated in marine environments by excessive sulfate scattering. The best represented environment was in dusty regions. Figures 3c and 3d show a reduction in these biases, particularly in biomass burning regions, but also in polluted regions. Note that the model results are averaged from 1981 to 2000, while most AERONET sun photometers began to operate in the mid-1990s or early twenty-first century. Since sulfur emission has decreased since the mid-1990s, simulated AOD values are likely higher than observed. The improved AOD simulation in AM3 is primarily due to changes in emissions and the treatment of optical properties. The treatments of chemistry, transport, and deposition are similar in AM2 and AM3, but differing large-scale and subgrid transports produce some AOD changes due to these also. Coalbedo measures aerosol absorption, and the model absorption has largely decreased from CM2.1 to CM3, agreeing much better with AERONET to generally within a factor of 2 at most stations (Fig. 4). This major change, which is particularly evident over regions of biomass burning, is due to several factors but primarily a decrease of black carbon emission. The decrease in black carbon emission, from 11 Tg yr−1 in AM2 (Horowitz 2006) to 8.2 Tg yr−1 in AM3, is partly compensated by increased absorption due to internal mixing of sulfate and black carbon. Unlike the direct measurement of AOD by sun photometers, coalbedo is retrieved by an inversion of Almucantar data (Dubovik and King 2000), and, to limit error of the retrieved values, only data with AOD greater than 0.45 are inverted. Thus, AERONET coalbedo is representative of heavily polluted, but not pristine, environments. Another bias to consider is that AERONET values are at 440 nm (blue), while the simulated aerosol properties are only archived at 550 nm (green). The subsequent bias will depend on the spectral variation of aerosol absorption. In biomass burning, smoke absorbs more in the green than the blue part of the solar spectrum, so the model coalbedo at 550 nm should be higher than at 440 nm. In dusty environments, the opposite should be true. These biases may partially explain the persisting discrepancies in Figs. 4c and 4d for CM3.
Clear-sky downward shortwave radiation in CM3 is generally larger in CM3 than CM2.1 and closer to observations from the Baseline Surface Radiation Network (BSRN; http://gewex-rfa.larc.nasa.gov) (Fig. 5). The increases in clear-sky downward shortwave radiation are due to reduced aerosol direct effects in CM3. Improved agreement of CM3 simulations of downward clear-sky surface shortwave radiation, optical depths, and coalbedo with BSRN and AERONET provides strong evidence that the direct effects of aerosols are more realistically simulated in CM3.
Aerosol activation into cloud droplets follows the parameterization detailed in Ming et al. (2006). Sulfate and sea salt aerosols are treated as pure ammonium sulfate and sodium chloride, respectively, in terms of cloud condensation nuclei (CCN) efficiency, while organic aerosol is partially soluble (Ming and Russell 2004). Black carbon is assumed to be insoluble and externally mixed with soluble species. Note that sulfate and black carbon are treated as an internal mixture for radiation calculation. The assumed size distribution of organic aerosol has two lognormal modes, which are characterized by the number concentrations (N), median diameters (Dp), and standard deviations (σ). The specific parameters for organic aerosol are N1:N2 = 17:3, Dp1 = 0.01 μm, σ1 = 1.6, Dp2 = 0.15 μm, and σ2 = 2. (The subscripts denote different modes.) Sea salt has one more mode to account for giant CCN (N1:N2:N3 = 340:60:0.75, Dp1 = 0.01 μm, σ1 = 1.6, Dp2 = 0.15 μm, σ2 = 2, Dp3 = 0.62 μm, and σ3 = 2.7). The size distributions of organic and sea salt aerosols remain unchanged regardless of ambient conditions. Sulfate aerosol is assumed to have the same size distribution as organic aerosol (called distribution 1) if its concentration is above 0.3 μg m−3. In this case, most of the sulfate mass is in the accumulation mode (0.1–1 μm). Otherwise, it is partitioned between distribution 1 and distribution 2, for which N1:N2 = 17:3, Dp1 = 0.01 μm, σ1 = 1.6, Dp2 = 0.03 μm, and σ2 = 2, depending on the abundance of primary aerosols (i.e., organics, sea salt, black carbon, and dust). The fraction of the sulfate mass in distribution 1 is 0 when the concentration of primary aerosols is less than 0.5 μg m−3 and increases linearly to 1 when it exceeds 1.0 μg m−3. Note that a considerable fraction of the mass in distribution 2 is in the nucleation mode (less than 0.1 μm). This choice is based upon the consideration that gas-to-particle conversion in polluted conditions occurs mainly through condensation onto preexisting particles, as opposed to nucleation.
Updraft velocities at cloud base and at the time of cloud formation are used to drive aerosol activation within shallow cumulus and stratiform clouds, respectively. Vertical velocities for shallow cumulus are provided directly by the Bretherton et al. (2004) shallow cumulus parameterization. The procedure for generating the probability distribution functions for updraft velocities in stratiform clouds is described in section 3d. Due to the absence of ice nucleation and limited treatment of microphysics generally in deep convection (in which substantial vertical accelerations can occur well above cloud base, leading to activation above cloud base), aerosol activation is not treated in deep convection. The consequences of this omission are not clear, and the matter is a high priority for future research.
A major motivation for including aerosol activation in AM3 is to enable simulation of cloud droplet sizes, which in turn partially determine the radiative and macrophysical properties of clouds, that is, aerosol indirect effects. Droplet sizes have been evaluated using a simple simulator for the Moderate Resolution Imaging Spectroradiometer (MODIS; King et al. 2003) satellite. For every subgrid column generated with the stochastic cloud scheme of Pincus et al. (2005) and Pincus et al. (2006) (cf. section 3a above), the radii for these liquid cloud layers in the top two units of cloud optical depth are averaged to produce a MODIS-like cloud-top radius. All cloudy subgrid columns are given equal weight in calculating the grid-mean radius.
Many general patterns from MODIS (collection 5) are captured in AM3, including increases in droplet sizes in the oceans off the east coasts of most continents and the January-to-July decrease in droplet sizes over subtropical South America and Africa (Fig. 6). The amplitudes of the changes are generally smaller in AM3 than in MODIS and the locations of maxima and minima differ between AM3 and MODIS, however. Satellite retrievals are themselves uncertain, and MODIS drop sizes are much larger than estimates from other retrievals (e.g., Han et al. 1994).
g. Tropospheric and stratospheric chemistry
In AM3, the chemistry models of Horowitz et al. (2003) for the troposphere and Austin and Wilson (2006) for the stratosphere are merged. The chemical system is solved using a fully implicit Euler backward method with Newton–Raphson iteration, as in Horowitz et al. (2003). Merging the two models consisted mainly of augmenting the tropospheric model with species (including halogens and atomic hydrogen) and reactions, primarily gas-phase halogen reactions, stratospheric and mesospheric photolysis reactions, and heterogeneous reactions on stratospheric aerosols. Reaction rates follow recommendations from Sander et al. (2006). The oxidation of sulfur dioxide and dimethyl sulfide to form sulfate aerosol is fully coupled with the gas-phase chemistry. Clear-sky photolysis frequencies are calculated using a multivariate interpolation table derived from the tropospheric ultraviolet–visible radiation model (Madronich and Flocke 1998) with an adjustment applied for the effects of large-scale clouds, as described by Brasseur et al. (1998).
Monthly mean dry-deposition velocities for gas-phase species [except for ozone and peroxyacetyl nitrate (PAN)] are from Horowitz et al. (2003) and were calculated offline using resistance in series (Wesely 1989; Hess et al. 2000). Deposition velocities for ozone were taken from Bey et al. (2001) and those for PAN from a Model for Ozone and Related Chemical Tracers, version 4 (MOZART-4) simulation in which it was calculated interactively to reflect the updates described by Emmons et al. (2010).
Cloud scavenging of gas-phase species is treated as for aerosols (section 3f), except that the in-condensate fraction is determined by Henry’s law equilibrium. Below-cloud washout is calculated only for large-scale precipitation, is based on Henry’s law, as in Brasseur et al. (1998), and is assumed to operate over the full extent of the below-cloud grid cells.
Calculating the stratospheric sources of reactive chlorine and bromine directly by transporting and photolyzing source gases (CFCs and halons) is computationally expensive and sensitive to any circulation biases in the model. Thus, we parameterize the source of reactive chlorine and bromine as a function of tropospheric concentrations of source gases (lagged by the stratospheric “age of air”), as described in appendix A of Austin and Wilson (2010). The parameterization uses observed source gas distributions to estimate, essentially, the fractional rate at which source gases entering the stratosphere are photolyzed and converted into reactive halogen species along their transport path through the stratosphere. Also as described in Austin and Wilson, heterogeneous reactions are included on ice and nitric acid trihydrate polar stratospheric clouds (PSCs) and in liquid ternary solution (LTS) aerosols. The PSCs are taken to be in thermodynamic equilibrium with the local conditions and calculated as in Hanson and Mauersberger (1988). The reaction rates in LTS are treated as in Carslaw et al. (1995). Mass accommodation coefficients and reaction probabilities are taken from Sander et al. (2006).
Compared to the observed climatology developed by Atmospheric Chemistry and Climate (AC&C)/Stratospheric Processes and their Role in Climate (SPARC) for CMIP5, following Randel and Wu (2007), general features of the annual-mean, zonally averaged ozone for the period 1980–99 are well reproduced (Fig. 7). The tropical peak in ozone mixing ratios is correctly simulated to occur near 10 hPa, but is overestimated by more than 10%. Ozone at high latitudes is underestimated compared with the observations, likely resulting from a deficiency in model transport. The seasonal variation of total column ozone (Fig. 8) is very similar to values retrieved by the Total Ozone Mapping Spectrometer (TOMS) (Stolarski and Frith 2006) for the decades of the 1980s and 1990s. In the 1980s, before significant ozone destruction, the model shows low tropical ozone, consistent with observations throughout the year. In middle and high latitudes, the annual variation is well reproduced, but the column ozone amounts are biased low in high northern latitudes, reflecting the bias shown in Fig. 7. In the Southern Hemisphere, the peak column amounts in austral spring near 60°S are simulated to be larger than observed. Similar features are also present in the 1990s. Note that the excessive column ozone at 60°S (with respect to TOMS) occurs despite the underestimate of peak ozone mixing ratios near 5 hPa at this latitude (with respect to the AC&C/SPARC climatology). This apparent discrepancy results from overestimate of ozone by AM3 compared with AC&C/SPARC in the lower stratosphere (30–100 hPa), where most of the ozone column resides. The simulated ozone hole is deeper than observed and lasts longer into (austral) summer, although it is smaller in physical area. In the annual mean, the biases are generally small (Fig. 8e), under 5%, but are larger in the Southern Hemisphere and dominated by the spring period indicated above.
Chemistry was not calculated online in AM2. Instead, decadally varying monthly mean tropospheric ozone concentrations (and aerosol concentrations; see section 3f) were prescribed in AM2 on the basis of calculations using the MOZART chemical transport model driven with present-day climatological meteorology, as described by Horowitz (2006). MOZART was configured with the gas-phase chemical mechanism of Horowitz et al. (2003), the sulfate and carbonaceous aerosol mechanism of Tie et al. (2005), and the mineral dust scheme of Ginoux et al. (2001). Stratospheric concentrations of ozone in MOZART were constrained by relaxation to present-day climatological values [as described by Horowitz et al. (2003)]. Stratospheric ozone concentrations in AM2 were prescribed following Randel and Wu (1999).
4. Basic simulation characteristics
a. Boundary conditions and integrations
AM3 and the land model were integrated with prescribed sea surface temperatures, sea ice coverage, and sea ice albedo to demonstrate their behavior with realistic boundary conditions. These integrations will be contrasted in this section with observations and with simulations in which AM3 served as the atmospheric component of CM3.
Observed sea surface temperatures and sea ice for the uncoupled integrations are from Rayner et al. (2003). Except as noted below, the period of integration is 1980 to 2000, with averages taken from 1981 to 2000. Initial conditions for the atmospheric model are drawn from the AM3 developmental integrations.
For the coupled integrations, CM3 was spun up for several centuries with 1860 trace gas concentrations and emissions, as described in sections 3a and 3f. During the 1860 spinup, solar irradiance was held constant at 1860 values (no solar cycle). Following the spinup, time-varying trace gas concentrations and emissions were imposed over the period 1860–2005. During the 1860–2005 simulation, solar irradiance varied following the observational time series recommended by CMIP5, as described by Fröhlich and Lean (2004). (As mentioned in section 3a, the solar irradiance has been scaled uniformly to correspond to the TIM scale, as recommended by CMIP5.) Anthropogenic aerosols (through both direct and indirect effects) and trace gases force the climate between 1860 and 2000. The CM3 global-mean temperature (for a five-member ensemble) increases by 0.32°C from the 1881–1920 period to the 1981–2000 period. The corresponding increases in the Climate Research Unit (CRU) observations (Brohan et al. 2006), Goddard Institute for Space Studies (GISS) observations (http://data.giss.nasa.gov/gistemp/tabledata/GLB.Ts+dSST.txt), and a five-member CM2.1 ensemble (Knutson et al. 2006) are 0.56°, 0.52°, and 0.66°C, respectively. Observed warming is intermediate between the CM2.1 and CM3 warming. In the following sections, CM3 analyses are restricted to 1981–2000 averages. Considerable interensemble variability is likely at higher time resolution.
b. Radiation and surface fluxes
Annual-mean shortwave absorption by the earth–atmosphere system in AM3 and the Earth Radiation Budget Experiment (ERBE) (Harrison et al. 1990) (Fig. 9) agree within 5 W m−2 over most of North America, the central Pacific Ocean, and southern Europe. AM3 exhibits negative biases in the tropical Indian and western Pacific Oceans, where excessive cloudiness and precipitation occur. Positive biases characterize the oceans off the subtropical west coasts of Africa, South America, and North America where marine stratocumulus is inadequate. Problematic marine stratocumulus persists from AM2 (Anderson et al. 2004), perhaps not surprisingly, given that the parameterizations for boundary layers and cloud macrophysics have not been changed in ways expected to remedy this deficiency. The marine stratocumulus biases are slightly smaller in the CM3 integrations than the AM3 integrations, suggesting a response to a small change in SSTs. Simultaneously, negative biases in the tropical oceans, consistent with a double ITCZ, emerge in the CM3 integration. A positive bias over the Amazon, consistent with insufficient convection, is considerably more apparent in the CM3 integration than in the AM3 integration. Positive biases exceeding 12 W m−2 between about 60°S and Antarctica are evident in CM3 but not AM3. The behavior of the corresponding fields for outgoing longwave radiation (OLR) is consistent with the shortwave changes (Fig. 10). In particular, the AM3 OLR exhibits negative biases in the tropical Indian Ocean and west Pacific where excessive high cloudiness occurs in association with deep convection (Fig. 10c). The double ITCZ in CM3 is evident in the splitting of the negative tropical OLR bias in the Pacific Ocean, separated by a zone of positive bias (Fig. 10d). The positive OLR bias over the Amazon in CM3 results from insufficient high cloudiness and convection (Fig. 10d).
To present a statistical summary of the radiation balances in AM3 and CM3, Taylor diagrams (Gates et al. 1999; Taylor 2001) (Fig. 11) are constructed using ERBE observations from 1985–89 (Harrison et al. 1990) and observations from the Clouds and the Earth’s Radiant Energy System (CERES) satellites from 2000–05. The CERES observations are analyzed in several ways: CERES-ES4-ERBE-like, CERES-SRB-GEO, CERES-SRB-nonGEO (Wielicki et al. 1996), and CERES-Energy Balanced and Filled (EBAF) (Loeb et al. 2009). (Observations are available online at http://eosweb.larc.nasa.gov/PRODOCS/ceres/table_ceres.html). Shortwave and net radiation have similar root-mean-square (rms) errors and correlation relative to observations for both AM3 and CM3. ERBE and CERES observations differ by about as much as the modeled results do from the CERES results, and the various CERES analyses differ little among themselves. AM3 and CM3 OLR rms differences from ERBE are two to three times larger than those of shortwave and net radiation. Note that the rms differences in Fig. 11 are normalized by the standard deviation of the ERBE observations and that the ERBE shortwave standard deviation is also two to three times larger than that of the ERBE OLR. The spread among the CERES observations themselves is somewhat greater for shortwave and longwave cloud forcing (Figs. 11d,e) than for shortwave absorption and OLR, as are the differences between ERBE and CERES observations. AM3 and CM3 differ more between themselves than they did for OLR and shortwave absorption, consistent with the cloud differences between AM3 and CM3 evident in Figs. 9c,d and 10c,d, for example, in the ITCZ and regions of marine stratocumulus. Pincus et al. (2008) note that cloud forcing is a more difficult field for models to simulate than total fluxes, which are to an appreciable extent controlled by the geometry of solar insolation. In that light, it is noteworthy that shortwave cloud forcing in AM3 compares more favorably with ERBE and CERES than AM2 (Fig. 11d). Correlations and rms differences between both atmospheric models and observations are comparable for longwave cloud forcing, but AM3 has more spatial variability than observed, while AM2 has less.
AM3 and CM3 include the Cloud Feedback Model Intercomparison Project’s Observation Simulator Package (COSP; http://cfmip.metoffice.com/). Among its components, the package includes simulators for the Cloud–Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) satellite lidar (Chepfer et al. 2008) and CloudSat radar (Bodas-Salcedo et al. 2008), which permit comparison of model cloud fields to the vertical structure of clouds provided by these new instruments. As an example, CALIPSO observations of cloud fraction for January 2007 (Chepfer et al. 2010) and the simulated cloud fractions from AM3 show broad qualitative agreement, while showing biases consistent with other fields sensitive to cloudiness (Fig. 12). For example, AM3 simulates smaller cloud fractions than CALIPSO observes off the west subtropical coasts of North America, South America, and Africa, consistent with positive ERBE shortwave biases in these regions (Figs. 9c,d). CALIPSO reveals excessive cloudiness in the AM3 Arctic, a region especially important for climate change. These errors are not evident in ERBE shortwave radiation (Fig. 9c). Unlike CALIPSO, ERBE shortwave radiation cannot separate cloud from clear sky with highly reflective surfaces in the Arctic.
For coupling AM3 with ocean models, the surface energy balance (including latent and sensible heat fluxes, in addition to radiative fluxes) is crucial and not related trivially to the top-of-atmosphere radiation balance. The implied ocean heat transport (OHT) is the heat transport implied in the ocean to balance surface fluxes. Although considerable uncertainty exists in diagnosing implied ocean heat transports from observations (e.g., Large and Yeager 2009; Griffies et al. 2009), agreement between these transports in uncoupled atmospheric models and observational estimates has been found to favor successful coupling with ocean models. The AM3 implied that OHT generally falls within or close to the observational estimates of Ganachaud and Wunsch (2003) and Trenberth and Caron (2001), except for the Indo-Pacific Ocean south of 30°S (Fig. 13).
The AM3 midlatitude westerly jets in the troposphere are about 10% stronger than in the ERA-40 reanalysis (Uppala et al. 2005) (Fig. 14). A small area of weak, spurious westerlies appears in the equatorial stratosphere around 10 hPa, and stratospheric westerlies at polar latitudes can be over 50% stronger than in the 40-yr European Centre for Medium-Range Weather Forecasts Re-Analysis (ERA-40). In the troposphere, westerly biases are smaller in CM3 than AM3 in the Southern Hemisphere but larger in the Northern Hemisphere.
Wind stresses in uncoupled models, along with implied OHT, are important to successful coupling. Wind stresses over the Atlantic and Pacific Oceans for AM3 and CM3 are generally within or close to the observational estimates from the Comprehensive Ocean–Atmosphere Datasets (COADS) (da Silva et al. 1994; Woodruff et al. 1987), ECMWF reanalysis (Gibson et al. 1997), and the European Remote Sensing Satellite (ERS) scatterometer (CERSAT-IFREMER 2002) (Fig. 15). The largest AM3 Pacific departures from observations are in the Southern Hemisphere, where CM3 stresses agree better with observations. The largest Atlantic departures for CM3 are in the Northern Hemisphere, where AM3 agrees better with observations.