The GFDL Global Atmospheric Chemistry‐Climate Model AM4.1: Model Description and Simulation Characteristics

Abstract

We describe the baseline model configuration and simulation characteristics of the Geophysical Fluid Dynamics Laboratory (GFDL)'s Atmosphere Model version 4.1 (AM4.1), which builds on developments at GFDL over 2013–2018 for coupled carbon‐chemistry‐climate simulation as part of the sixth phase of the Coupled Model Intercomparison Project. In contrast with GFDL's AM4.0 development effort, which focused on physical and aerosol interactions and which is used as the atmospheric component of CM4.0, AM4.1 focuses on comprehensiveness of Earth system interactions. Key features of this model include doubled horizontal resolution of the atmosphere (~200 to ~100 km) with revised dynamics and physics from GFDL's previous‐generation AM3 atmospheric chemistry‐climate model. AM4.1 features improved representation of atmospheric chemical composition, including aerosol and aerosol precursor emissions, key land‐atmosphere interactions, comprehensive land‐atmosphere‐ocean cycling of dust and iron, and interactive ocean‐atmosphere cycling of reactive nitrogen. AM4.1 provides vast improvements in fidelity over AM3, captures most of AM4.0's baseline simulations characteristics, and notably improves on AM4.0 in the representation of aerosols over the Southern Ocean, India, and China—even with its interactive chemistry representation—and in its manifestation of sudden stratospheric warmings in the coldest months. Distributions of reactive nitrogen and sulfur species, carbon monoxide, and ozone are all substantially improved over AM3. Fidelity concerns include degradation of upper atmosphere equatorial winds and of aerosols in some regions.

1 Introduction

Atmospheric chemistry and composition are intrinsically coupled to the Earth's climate system. The pivotal role of chemistry-climate interactions in regulating climate has been recognized for over 30 years (Ramanathan et al., 1987). Advances in fundamental knowledge of atmospheric chemistry and availability of long-term measurements, combined with advances in computing, have provided an opportunity to enhance the comprehensiveness of our representation of atmospheric chemistry-climate interactions. For the fifth phase of the Coupled Model Intercomparison Project (CMIP5), Geophysical Fluid Dynamics Laboratory (GFDL) contributed its first coupled chemistry-climate model, CM3 (Donner et al., 2011), which allowed us to represent interactive aerosols and ozone, rather than prescribing concentrations from offline models (e.g., Horowitz et al., 2003). Much has been learned in the preceding years as to the strengths and weaknesses of CM3 chemistry, including the critical role of prognostic aerosol interactions (e.g., Levy et al., 2013; Naik, Horowitz, et al., 2013). As such, interactive aerosols were included in all of GFDL's fourth-generation atmospheric model development efforts, targeted for the sixth phase of the Coupled Model Intercomparison Project (CMIP6). The high computational cost of interactive atmospheric chemistry, however, was avoided in GFDL's recent CM4.0 development (Held et al., 2019) by prescribing ozone and other oxidants. A full, interactive chemistry representation, along with a comprehensive carbon cycle, was reserved in this generation of GFDL models for Earth System Model development in ESM4.1 (Dunne et al., 2020). Thus, CM4 included a focus on ocean resolution, while ESM4.1 focused on a comprehensive representation of the Earth system.

The overall goal of AM4.1 development was to merge a suite of mostly parallel sets of updates and innovations into GFDL's fourth-generation atmospheric model. These updates include a revised chemical mechanism from AM3 to AM4.1 to take advantage of new laboratory kinetic data (e.g., Mao, Fan, et al., 2013; Sander et al., 2011; as implemented by Li et al., 2016), in particular for photooxidation of biogenic volatile organic compounds (BVOCs). Analysis of deficiencies in AM3 chemistry has pointed out improper treatment of nitrate aerosols and gas-aerosol interactions and biases in wet and dry deposition. We also wanted to leverage successful follow-on development efforts from AM3 targeted to implement reduced nitrogen cycling (Paulot et al., 2016; Paulot, Paynter, et al., 2017) and improved representation of the seasonal cycle in sulfate (Paulot, Fan, & Horowitz, 2017). Finally, we wanted to provide the ability for the atmospheric model to handle a more diverse suite of land-atmosphere and ocean-atmosphere linkages for comprehensive Earth system representation of not only heat and hydrology but also CO₂, dust, reactive nitrogen, and organic carbon.

The focus of the present study is to document the atmosphere physics and chemistry developed for AM4.1, as distinguished from the physical climate–focused AM4.0 (Zhao et al., 2018a, 2018b), for standalone atmospheric applications. A more comprehensive discussion of coupled atmosphere-ocean-land Earth system interactions in ESM4.1 is provided by Dunne et al. (2020). We focus our analysis on evaluating the AMIP configuration of AM4.1 used for CMIP6, and document the differences in results between AM4.1 and other GFDL CMIP models, including AM4.0 (CMIP6) and AM3 (CMIP5). In the case of comparisons with AM3, the differences in results reflect updates to both the model configuration and the emissions (as discussed in section 2.4).

2 Model Description

A general schematic of AM4.1 forcing, dynamics, physics, aerosol, and chemistry interactions is provided in Figure 1. The following sections discuss the formulations for these components in reference to their AM4.0 (Zhao et al., 2018a, 2018b) counterparts.

Figure 1. Schematic description of forcing, dynamics, physics, aerosol, and chemistry interactions in AM4.1. Terms depicted in gray (left) are prescribed as inputs to the model, while chemical processes included in the orange box are calculated interactively within the atmospheric model. The light blue box (top) includes physical processes calculated in AM4.1. The green box (bottom left) represents the land component (LM4.1), which is coupled to AM4.1. The dark blue box (bottom right) includes specified ocean-surface boundary conditions.

2.1 Physical Model Formulation

The physical formulation of AM4.1 is similar to that of AM4.0, but the model top has been raised from 100 Pa (~45 km) to 1 Pa (~80 km), and the number of vertical levels has been increased from 33 to 49, similar to the 48-level structure of AM3. This enhanced vertical extent and resolution allows AM4.1 to represent stratospheric dynamics and chemistry and stratosphere-troposphere coupling. The time step used in the dynamical core for gravity wave and the Lagrangian dynamics is reduced from 150 s in AM4.0 to 130 s in AM4.1 for increased numerical stability.

Like AM4.0, AM4.1 includes five tracers for water (specific humidity, liquid water, ice water, cloud amount, and liquid droplet number concentration) and uses the same large-scale and convective cloud parameterizations as in AM4.0. Cloud parameterizations in AM4.1 were retuned slightly compared to AM4.0 in order to improve agreement with observed top-of-atmosphere shortwave and longwave radiative fluxes, in response to initially excessive reflection from convective clouds over sub-Saharan Africa, North Indian Ocean, and the western tropical Pacific. In particular, the scale factor applied to the fall speed of ice clouds (c1 in Zhao et al., 2018b) was reduced from 0.90 in AM4.0 to 0.85 in AM4.1 to increase ice water path and decrease outgoing longwave radiation. The cloud erosion timescale (τ_eros) in convectively active regions is decreased slightly from 6.9 to 5.6 h to increase the absorbed shortwave radiation. The cloud erosion timescale under other conditions is unchanged from AM4.0.

As described by Zhao et al. (2018b), nonorographic gravity wave drag in AM4.0/AM4.1 is parameterized following Alexander and Dunkerton (1999), but the parameters used in AM4.1 are modified from those in AM4.0 to improve stratospheric circulation. In AM4.1, the magnitudes of the nonorographic gravity wave flux above 350 hPa for the tropics, northern extratropics, and southern extratropics (St, Sn, and Ss) are set to 0.004, 0.005, and 0.0035 m² s⁻², respectively.

Land hydrology and ecosystem dynamics are represented in AM4.1 by the GFDL Land Model version 4.1 (LM4.1; Elena Shevliakova, personal communication), replacing the LM4.0 model used in AM4.0 (Zhao et al., 2018b). LM4.1 includes advanced vegetation and canopy competition, fire, land-use representation, and dynamic atmospheric dust coupling.

2.2 Atmospheric Chemistry and Aerosol Scheme

AM4.1 includes interactive tropospheric and stratospheric gas-phase and aerosol chemistry. The bulk aerosol scheme, including 18 transported aerosol tracers (see Table S1 in the supporting information), is similar to that in AM4.0 (Zhao et al., 2018b), with the following updates: (1) ammonium and nitrate aerosols are treated explicitly, with the sulfate–nitrate–ammonia thermodynamic equilibrium simulated using ISORROPIA (Fountoukis & Nenes, 2007), as described by Paulot et al. (2016); (2) oxidation of sulfur dioxide and dimethyl sulfide to produce sulfate aerosol is driven by the gas-phase oxidant concentrations (OH, H₂O₂, and O₃) and cloud pH simulated by the online chemistry scheme (Paulot et al., 2016); and (3) the rate of aging of black and organic carbon aerosols from hydrophobic to hydrophilic forms varies with calculated concentrations of hydroxyl radical (OH), as described by Liu et al. (2011).

Unlike AM4.0, the AM4.1 model has an online representation of gas-phase tropospheric and stratospheric chemistry. The combined tropospheric and stratospheric chemistry scheme includes 18 prognostic (transported) aerosol tracers, 58 prognostic gas-phase tracers, five prognostic ideal tracers, and 40 diagnostic (nontransported) chemical tracers (Table S2), with 43 photolysis reactions, 190 gas-phase kinetic reactions, and 15 heterogeneous reactions (Table S3). The tropospheric chemistry includes reactions of the NO_x–HO_x–O_x–CO–CH₄ system and oxidation schemes for other nonmethane volatile organic compounds. The stratospheric chemistry accounts for the major ozone loss cycles (O_x, HO_x, NO_x, ClO_x, and BrO_x) and heterogeneous reactions on liquid and solid stratospheric aerosols as in Austin et al. (2013). The base chemical mechanism is updated from that in AM3 (Naik, Horowitz, et al., 2013), using gas-phase and heterogeneous chemistry updates from Mao, Horowitz, et al. (2013) and Mao, Paulot, et al. (2013), similar to the configuration described by Schnell et al. (2018). We include heterogeneous reactions of HO₂, NO₂, N₂O₅, and NO₃ on the surfaces of all simulated aerosol types, with specified gamma values (given in Table S3). Note in particular that γ (HO₂) is reduced from the value of 1 recommended by Mao, Horowitz, et al. (2013) to 0.2. We also include the heterogeneous oxidation of SO₂ on aerosols following Zheng et al. (2015). The heterogeneous chemistry included in AM4.1 has a much stronger effect on oxidant levels than that in AM3, which used γ(N₂O₅) = 0.1, γ(NO₃) = 0.1, γ(NH₃) = 0.05, γ(HO₂) = 0, and γ(NO₂) = 0, applied only to sulfate aerosols. The chemical system is solved using an implicit Euler backward method with Newton-Raphson iteration, as in Horowitz et al. (2003). Photolysis rates are calculated interactively using the FAST-JX version 7.1 code, as described by Li et al. (2016), accounting for the radiative effects of simulated aerosols and clouds.

Dry deposition velocities for all aerosols are calculated interactively using a wind-driven resistance method (Gallagher et al., 2002), accounting for the effect of whitecaps over the ocean (Williams, 1982). The treatment of wet deposition accounts for slower removal by frozen precipitation due to the Bergeron process in mixed-phase clouds (Liu et al., 2011). Dry and wet deposition for gases are as described by Paulot et al. (2016).

Optical properties of aerosols are precalculated, as in AM4.0, using Mie theory assuming sphericity. The extinction efficiency, single scattering albedo, and asymmetry parameter are tabulated as a function of wavelength, aerosol type, aerosol size (for dust and sea salt), and relative humidity. Sulfate is assumed to be internally mixed with black carbon for the calculation of optical properties. Unlike AM4.0, radiative effects of nitrate aerosol are included in AM4.1 (as in Paulot, Paynter, et al., 2017).

2.3 AMIP (1980–2014) Simulation Configuration

We conduct AMIP simulations with AM4.1 over the period 1979–2014 using observed gridded sea surface temperature (SST) and sea-ice concentration boundary conditions from the reconstructions of Taylor et al. (2000). Historical reconstructions of monthly solar spectral irradiances are from Matthes et al. (2017). For radiation calculations, global monthly mean concentrations of greenhouse gases (GHGs), including nitrous oxide (N₂O), and ozone-depleting substances (ODSs, including CFC-11, CFC-12, CFC-113, and HCFC-22) are specified from Meinshausen et al. (2017). Global mean mixing ratios of methane (CH₄) and N₂O are specified at the surface as lower boundary conditions for chemistry. Carbon dioxide (CO₂) mixing ratio is restored to observed global-mean values with a one-year timescale. The simulated global-mean CO₂ and CH₄ concentrations are used for radiation calculations.

2.4 Emissions

Annually varying time series of monthly anthropogenic and biomass burning emissions of ozone precursors and aerosols (and their precursors) are from the Community Emissions Data System (CEDS; Hoesly et al., 2018) and the data set of van Marle et al. (2017), respectively, developed in support of CMIP6. Wildfire emissions are distributed vertically between the surface and 6 km, with location- and biome-dependent vertical profiles, as recommended by Dentener et al. (2006), similar to the treatment in AM3 (Donner et al., 2011; Naik, Horowitz, et al., 2013). Natural emissions of NO_x, CO, non-methane volatile organic compounds (NMVOC), hydrogen (H₂), and ammonia (NH₃) are generally the same as those considered by Naik, Horowitz, et al. (2013), namely from the Precursors of Ozone and their Effects in the Troposphere (POET) inventory for present day (corresponding to year 2000) (Olivier et al., 2003). Emissions of NH₃ from sea bird colonies, not accounted for in AM3, are included in AM4.1 following Riddick et al. (2012). The treatment of marine ammonia emissions is also revised as described below.

Biogenic emissions of isoprene and monoterpenes are calculated online using the Parameterized Canopy Environment Emission Activity (PCEEA algorithm; Guenther et al., 2006) in the Model of Emissions of Gases and Aerosols from Nature (MEGAN v2.1; Guenther et al., 2012) as a function of simulated air temperature and shortwave radiative fluxes, implemented as described by Rasmussen et al (2012). Leaf area indices for 17 plant functional types are based on AVHRR and MODIS data and are mapped to five vegetation types (Emmons et al., 2010). These vegetation types and leaf area indices are independent of those simulated by the LM4.1 dynamic vegetation model, due to a lack of coupling between the dynamic vegetation properties simulated by LM4.1 and the atmospheric emissions module. We do not apply the soil moisture or CO₂ responses from Guenther et al. (2012). Future model development plans include coupling biogenic emissions to LM4.1. Sea salt emissions are based on the parameterization of Monahan et al. (1986) as in CM3 (Donner et al., 2011), but are modulated by sea surface temperature following Jaeglé et al. (2011). Ocean ammonia emissions are calculated following Paulot et al. (2015), using the simulated seawater concentration of NH₄⁺ in ESM4.1. Other marine emissions, including primary organic aerosols (POA) and dimethyl sulfide (DMS), are calculated similarly to in CM3. DMS emissions are calculated using an empirical formula as a function of a prescribed monthly climatology of DMS concentration in sea water (Lana et al., 2011) and calculated wind speed at 10 m, as described by Chin et al. (2002). Thus, oceanic emissions of POA, DMS, ammonia, and sea salt are dependent on the simulated meteorology in the model.

Emission totals for year 2014 are shown in Table 1. Time series of annual global emissions in AM4.1 (using CMIP6 inventories) are shown for select species in Figure 2 and compared with corresponding totals in AM3 (using CMIP5 inventories).

Table 1 Annual Total Emissions for Year 2014 in AM4.1

Species	Units	Anthro	Biomass burning	Biogenic/natural	Ocean	Animals	Soil	Ship	Aircraft	Total
ACET	Tg C a−1	1.47	0.98	15.09	0	0	0	0	0	17.53
BC	Tg C a−1	7.83	1.77	0	0	0	0	0.17	0	9.76
C2H4	Tg C a−1	4.88	3.82	0	0	0	0	0.14	0	8.83
C2H5OH	Tg C a−1	2.40	0.07	4.82	0	0	0	0	0	7.29
C2H6	Tg C a−1	5.22	2.71	0.80	0.78	0	0	0.17	0	9.67
C3H6	Tg C a−1	9.50	5.85	0.85	1.29	0	0	0.16	0	17.66
C3H8	Tg C a−1	5.05	0.53	1.63	1.05	0	0	0.49	0	8.76
C4H10	Tg C a−1	52.93	2.34	0	0	0	0	1.10	0	56.38
C10H16	Tg C a−1	0	1.24	57.37	0	0	0	0	0	58.61
CH2O	Tg C a−1	1.00	1.94	0	0	0	0	0	0	2.94
CH3OH	Tg C a−1	0.30	3.24	85.61	0	0	0	0	0	89.14
CO	Tg a−1	612.40	356.68	159.24	19.80	0	0	0.69	0.57	1,149.37
DMS	Tg a−1	0	0	0	42.72	0	0	0	0	42.72
DUST	Tg a−1	0	0	0	0	0	2,507.67	0	0	2,507.67
H2	Tg a−1	24.50	9.01	0	2.98	0	2.98	0.03	0	39.48
ISOP	Tg a−1	0.00	0.57	499.78	0	0	0	0.00	0	500.36
NH3	Tg a−1	60.82	4.30	0	3.89	0.15	2.95	0.02	0	72.13
NO	Tg N a−1	35.52	6.23	3.29	0	0	3.59	6.89	0.93	56.45
OM	Tg a−1	31.26	26.81	0	16.21	0	0	0.20	0	74.48
SSALT	Tg a−1	0	0	0	6,254.24	0	0	0	0	6,254.24
SO2	Tg S a−1	51.26	1.14	3.59	0	0	0	4.44	0.14	60.56

Figure 2. Global annual totals (in Tg a−1, using mass as indicated on y-axis label) for anthropogenic (fossil fuel + biomass burning + ship + aircraft) emissions of NO, CO, SO2, NH3, BC, and primary OM in AM3 (blue, CMIP5 emissions) and AM4.1 (red, CMIP6 emissions) AMIP simulations.

Sources of secondary organic aerosols (SOA) include an anthropogenic source from oxidation of the simulated C₄H₁₀ hydrocarbon tracer by hydroxyl radical (with a 10% per-carbon yield) and a biogenic pseudo-emission assuming a 10% per-carbon yield from emissions of BVOCs, including isoprene and monoterpenes, from vegetation. This yield is in the range of values suggested by recent studies using more detailed schemes for SOA production (e.g., Bates & Jacob, 2019; Pai et al., 2020). In year 2014, the sources of SOA are 83.84 Tg a⁻¹ from BVOCs and 3.49 Tg a⁻¹ from anthropogenic hydrocarbon oxidation.

Lightning NO_x emissions are calculated interactively as a function of subgrid convection in AM4.1, as diagnosed by the double-plume convection scheme described by Zhao et al. (2018b). The lightning NO_x source is calculated as a function of convective cloud-top height, following the parameterization of Price et al. (1997), and is injected with the vertical distribution of Pickering et al. (1998), as in AM3 (Naik, Horowitz, et al., 2013). The global total production of NO_x by lightning is 3.59 Tg N for year 2014.

Dust emissions are calculated dynamically online in the land component, LM4.1, as a function of wind speed, topography, vegetation cover, snow cover, soil moisture, and land type, as described by Evans et al. (2016).

As in AM3, direct stratospheric injection of SO₂ from volcanic eruptions and emissions of carbonyl sulfide (COS) are not considered in AM4.1. Instead, we specify time series of stratospheric aerosol optical properties, accounting for not only the volcanic contribution to stratospheric aerosol abundance but also other natural and anthropogenic contributions. Tropospheric emissions of SO₂ from continuously degassing and explosive volcanoes are treated in the same way as in AM3 (Donner et al., 2011), with a climatological total of 3.59 Tg S a⁻¹.

3 Results: Physical Climate Simulation (AMIP, 1980–2014) 3.1 Surface Air Temperature

Comparison of surface air temperature over land with observations from CRU TS (Figure 3) illustrates the substantial decrease in overall root mean square error (RMSE) achieved in AM4.1 (RMSE = 1.92°C) from the previous generation full-chemistry AM3 (RMSE = 2.18°C) and similar, if slightly degraded, pattern to AM4.0 (RMSE = 1.85°C). The most notable difference from AM3 to AM4.0 and AM4.1 is an improvement in boreal warm biases and South American cold biases.

Figure 3. Annual mean surface air temperatures (°C) in AM4.1 AMIP simulation (1980–2014) and CRU-TS-3.22 observations (1979–2013). Differences between simulated and observed surface air temperatures in AM4.1, AM4.0, and AM3 AMIP simulations.

3.2 Precipitation

Comparison of precipitation with observations from GPCP v2.3 (Figure 4) also illustrates the substantial decrease in overall RMSE achieved in AM4.1 (RMSE = 0.83 mm d⁻¹) from the previous generation full-chemistry AM3 (RMSE = 1.02 mm d⁻¹) and a similar pattern to AM4.0 (RMSE = 0.85 mm d⁻¹). The most notable difference from AM3 to AM4.0 and AM4.1 is an improvement in Amazon dry biases and in wet biases over Australia and the Indian Ocean.

Figure 4. Annual mean precipitation (mm day−1) for 1980–2014 in AM4.1 AMIP simulation and GPCP v2.3 observations. Differences between simulated and observed precipitation in AM4.1, AM4.0, and AM3 AMIP simulations.

3.3 Circulation

Comparison of zonal mean zonal winds with the ERA40 reanalysis (Figure 5) illustrates a substantial decrease in overall RMSE in AM4.1 (RMSE = 1.32 m s⁻¹) from the previous generation high-top full-chemistry AM3 (RMSE = 1.75 m s⁻¹). The AM4.1 RMSE is greater than that in the low-top AM4.0 (RMSE = 1.00 m s⁻¹), owing to a westerly wind bias in the equatorial stratosphere, and a weak, equatorward-shifted Arctic stratospheric jet in AM4.1. The representation of the stratospheric wintertime westerly polar jet associated with the Antarctic vortex is significantly improved in AM4.1 (not shown) compared with AM3 (Donner et al., 2011), in which the westerlies were excessively strong (leading to a too-cold Antarctic vortex). We plan to work towards further improving the stratospheric circulation in future versions of AM4.1 through improvement in our representation of parameterized gravity wave drag. Tropospheric circulation patterns in AM4.1 are very similar to those in AM4.0.

Figure 5. Annual mean zonal mean zonal wind (m s−1) in AM4.1 AMIP simulation (1980–2014) and ERA40 reanalysis (1981–2000). Differences between simulated and observed zonal winds in AM4.1, AM4.0, and AM3 AMIP simulations.

3.4 Stratospheric Variability

Comparison of statistics for sudden stratospheric warmings with the ERA40 reanalysis (Figure 6) illustrates an improvement in AM4.1 with respect to capturing events in the coldest months (December–January), which were largely missed in AM4.0 (Zhao et al., 2018a), even though AM4.0 already performs quite well among low-top atmospheric models (Charlton-Perez et al., 2013). In the surrounding months (November, February), AM4.1 overestimates warming events, whereas AM4.0 matches the reanalysis data fairly well.

Figure 6. Monthly and annual (ANN) stratospheric sudden warming (SSW) frequency for 1870–2014 from AM4.0 and AM4.1, and 1957–2002 from ERA40. SSW is defined as in Charlton and Polvani (2007). Error bars indicate the 95% confidence interval (the statistical test of the SSW frequency is calculated as in Charlton et al., 2007).

3.5 Radiation Fluxes

Comparison of top-of-atmosphere (TOA) net radiation with CERES EBAF observations (Figure 7) illustrates the substantial decrease in overall root mean square error (RMSE) achieved in AM4.1 (RMSE = 7.2 W m⁻²) from the previous generation full-chemistry AM3 (RMSE = 8.6 W m⁻²) and similar, if slightly degraded, pattern to AM4.0 (RMSE = 6.8 W m⁻²). The most notable difference from AM3 to AM4.0 and AM4.1 is associated with an improvement in areas of tropical convection along the intertropical convergence zone (ITCZ) that had previously been too absorbing and increased absorption in northern boreal regions that had been previously too reflective, as discussed by Zhao et al. (2018a). The most notable differences between AM4.0 and AM4.1 are associated with a decrease in the global TOA from a near-zero bias in AM3 (0.02 W m⁻²) to a slight negative bias in AM4.0 (−0.14 W m⁻²) and substantial low bias in AM4.1 (−0.80 W m⁻²). This increase in bias is due in part to the increased albedo of northern boreal regions associated with snow masking depth in LM4.1 (Elena Shevliakova, personal communication and also in part to differences over Antarctica associated with the prescribed albedo of snow on glaciers that was modified late in the development cycle of ESM4.1 to address Southern Ocean dynamics, as discussed by Dunne et al. (2020).

Figure 7. Annual mean net radiation flux at top of atmosphere (W m−2) in AM4.1 AMIP simulation (1980–2014) and CERES EBAF v2.8 observations (2000–2015). Differences between simulated and observed net radiation flux in AM4.1, AM4.0, and AM3 AMIP simulations.

3.6 Lightning Flash Frequency

Figure S1 shows the lightning flash frequency retrieved from the spaceborne Optical Transient Detector (OTD) and Lightning Imaging Sensor (LIS) (Cecil et al., 2014), compared with simulated values from AM3 and AM4.1. In both AM3 and AM4.1, lightning flash frequency is parameterized as a function of convective cloud top height, following Price et al. (1997), but the two models use different parameterizations of cumulus convection (Zhao et al., 2018b). While the overall correlation between model and observations is lower in AM4.1 than AM3, there are some notable areas of improvement in the representation of flash frequency, including a reduction of the high biases present in AM3 over the Amazon and the maritime continent, improving agreement with observations.

4 Results: Simulation of Atmospheric Composition 4.1 Ozone

In this section, we evaluate model simulations of ozone, including surface ozone concentrations relevant for air quality and column ozone abundances relevant for climate.

4.1.1 Surface Ozone

We focus on the seasonal mean of the maximum daily 8-h average (MDA8) surface ozone over the period 2005–2014, when observations are available from densely clustered monitoring sites across northern mid-latitude populated regions (Figure 8 for MAM, Figure 9 for JJA). Observations were obtained from the Tropospheric Ozone Assessment Report (TOAR) Database for 2005–2014 (Schultz et al., 2017) and a monitoring network operated since 2013 by China's Ministry of Environmental Protection (CNMEP, http://106.37.208.233:20035/). Observations are averaged onto the same 1° × 1° grid as AM4.1. We compare simulated ozone from the AM4.1 AMIP simulation with that from the AM3 AMIP simulation.

Figure 8. MAM mean surface MDA8 ozone mixing ratios (ppbv) for 2000–2008 from (a) TOAR observations regridded to the same 1° × 1° grid as AM4.1, (b) AM3 AMIP simulation, (c) AM4.1 AMIP simulation. Here, mn is the mean and rmsd is the root-mean-square deviation between observations and simulations.

Figure 9. Same as Figure 8, but for JJA.

Surface MDA8 ozone in AM3 is biased high by 12 ppb on average during MAM (Figure 8b) and by up to 20 ppb over the eastern U.S. during summer (Figure 9b), as documented in previous studies (Fiore et al., 2014; Lin, Fiore, Cooper, et al., 2012; Lin, Fiore, Horowitz, et al., 2012; Lin et al., 2017; Rieder et al., 2015). AM4.1 shows substantially reduced biases in mean ozone for both spring and summer over the eastern U.S. and Europe (Figures 8c and 9c). This dramatic improvement in the simulation of surface ozone concentrations results from a combination of updates to the chemical mechanism from AM3 to AM4.1, including updates to the isoprene oxidation scheme (Mao, Paulot, et al., 2013) and the representation of heterogeneous reactions (Mao, Fan, et al., 2013), and the change from CMIP5 emissions in AM3 to CMIP6 emissions in AM4.1 (section 2.3). The shallow surface layer of the model (30 m thick) may also have an impact on the comparison with surface sites. Zhao et al. (2018a) found a significant improvement in diagnosed 2-m temperatures associated with this shallower surface layer.

To further explore the causes of the differences in surface ozone abundances between AM3 and AM4.1, we conduct two additional simulations—an AM4.1 simulation with nudged meteorology and an additional AM4.1 nudged simulation with AM3-like chemistry (AM4.1_AM3Chem; Lin et al., 2019). The two experiments use the same CMIP6 emissions and have nearly identical meteorology (as a result of the nudging), allowing us to isolate the influence of changes in chemistry alone. Seasonal-mean MDA8 ozone from these simulations are plotted in Figures S2–S5.

Similar to the results from AM3 (Figures 8b and 9b), surface MDA8 ozone in AM4.1_AM3Chem is biased high by 11 ppb on average during spring (Figure S2b) and by up to 20 ppb over the eastern U.S. during summer (Figure S3b). Switching the chemistry scheme from AM3 to AM4.1 leads to substantial reductions in mean ozone biases for both spring and summer over the eastern U.S. and Europe (Figures S2c and S3c), but the model underestimates springtime MDA8 ozone over central eastern China by 20 ppb (Figure S2c versus CNMEP observations in Figure S2a).

Observations show more severe springtime ozone pollution over central eastern China and Mexico than in the U.S. and Europe. This regional contrast is not simulated in either of our experiments. Particularly, the enhanced heterogeneous chemistry in AM4.1 (section 2.2; Mao, Horowitz, et al., 2013; Mao, Fan, et al., 2013) likely leads to excessive heterogeneous loss of HO_x and NO_x radicals over eastern China and Mexico, where aerosol loadings are high during the spring season. For summer over the southeastern U.S., where high mean-state ozone biases are found in many current-generation CTMs and CCMs (Fiore et al., 2009; Young et al., 2018), the AM4.1 experiment shows remarkable agreement with observations. However, on the basis of analysis conducted for an intensive field campaign, Travis et al. (2016) suggested that the common model biases in simulating summertime ozone over the southeastern U.S. may reflect a combination of excessive NO_x emissions (too high by 50%) and the deep model surface layer that cannot resolve near-surface ozone gradients. A balanced view is needed to interpret the reduced ozone biases in the AM4.1 experiment.

Our results suggest the complexity of various sources, sinks, transport, and chemistry in influencing the simulation of surface ozone. In the future, process-based assessments, not only for means but also for variability and extreme events, are needed to fully evaluate how the choices of different emission data sets, chemical mechanisms, and deposition schemes affect simulations of surface ozone and related tracers.

4.1.2 Tropospheric Ozone Column

We compare climatological annual mean tropospheric ozone columns simulated by AM3 (mean over 2000–2008) and AM4.1 (2005–2014) with those derived from the OMI-MLS (Ziemke et al., 2019) (Figure 10). In the analysis shown here, AM3's native ozone output on model levels is used to calculate tropospheric ozone column using the WMO tropopause definition, while for AM4.1, the tropospheric ozone column (tropoz) is diagnosed at every time step, by applying the WMO tropopause definition using model simulated temperature. The global mean tropospheric ozone columns simulated by AM3 and AM4.1 are 35 DU and 31 DU, respectively, compared to the OMI/MLS value of 30 DU. While AM3 showed consistent high biases globally except over the Antarctic, AM4.1 shows an interhemispheric pattern in the biases with high values in the Northern Hemisphere mid-latitudes and over continents and low values in the Southern Hemisphere extra-tropics. This pattern is consistent with global chemistry-climate models evaluated against the OMI/MLS climatology by Young et al. (2013) for a slightly different time period. An interesting feature in AM4.1 is the strong positive bias over Oceania, possibly related to the different biomass burning emissions applied in the two models. AM3 exhibited an average high bias of 21.7%, which has been reduced to 7.3% in AM4.1; accordingly the RMSE has been reduced considerably, from 7.1 DU in AM3 to 4.6 DU in AM4.1.

Figure 10. Climatological mean tropospheric ozone column in AM3 (upper left; Dobson Units, DU), AM4.1 (lower left; DU), and the % bias compared to the OMI/MLS satellite estimate of the Tropospheric Ozone Column (Ziemke et al., 2019) for AM3 (upper right; %) and AM4.1 (lower right; %). RMSE is provided in DU.

4.1.3 Total Ozone Column

Figure 11 shows the evaluation of modeled time series of total column ozone against two data sets for 1980–2015, namely, Multi-Satellite Merged Total Column NASA and NOAA product from Frith (2013; SBUV; open triangles) and version 3.4 of the National Institute of Water and Atmospheric Research—Bodeker Scientific (NIWA-BS; closed circles) total column ozone database. AM3 results are plotted for 1980–2008 period, while AM4.1 results are for 1980–2014. The comparison is shown for the annual average globally, in the tropics, and in southern and northern mid-latitudes, and for March in the Arctic and October in the Antarctic. Globally (Figure 11a), absolute values of total column ozone for AM3 were biased high compared to both data sets, whereas AM4.1, on the other hand, is biased low. Both models generally capture the trend in total column ozone, although the evaluation of AM3 is truncated at 2008. As suggested by the greater correlation coefficients for AM4.1 compared with AM3, AM4.1 is better able to capture the observed interannual variability and trends of global mean total column ozone. In the tropics (Figure 11b), total ozone column values remain lower than observed in AM4.1, as opposed to higher in AM3. Consistent with observations, both models simulate negligible trends in total column ozone in the tropics; however, AM4.1 exhibits greater skill in capturing the observed evolution of total column ozone. In the northern mid-latitudes (Figure 11c), AM4.1 differs more from observations than AM3 does, although with fairly similar skill in simulating the observed time evolution of total column ozone. The comparison is opposite for the southern mid-latitudes (Figure 11d), where AM4.1 is much closer to observed values than AM3 with similar correlations. In the Arctic in March (Figure 11e), AM4.1 reproduces the observed total ozone column values slightly better than AM3, however both have fairly low skill in reproducing the observed evolution. In the Antarctic in October (Figure 11f), AM4.1 exhibits greater skill in simulating ozone depletion compared to AM3 both in terms of trends and absolute values. This improvement likely results from the improved dynamical representation of the Antarctic polar vortex in AM4.1 (section 3). Overall, AM4.1 compares slightly better against observations of total column ozone than AM3.

Figure 11. Comparison of time series of total ozone column (DU) for the annual mean (a) global mean (90°S–90°N), (b) tropics (25°S–25°N), (c) northern mid-latitudes (35°N–60°N), (d) southern mid-latitudes (35°S–35°N), and for the (e) March mean in the Arctic (60°N–90°N), and (f) October mean in the Antarctic (60°S–90°S) from AM3 (red) and AM4.1 (blue) against NASA and NOAA observations from the multisatellite merged ozone total column (Frith, 2013) (SBUV; open triangles) and version 3.4 of the NIWA-BS total column ozone database (Bodeker et al., 2005) (NIWA; closed circles). The numbers in each panel indicate linear correlation coefficient (R) for model against each of the measurement data sets (top for NIWA and bottom for SBUV).

4.2 Carbon Monoxide

The simulated tropospheric CO columns are evaluated against CO retrievals from the MOPITT (Measurements of Pollution in The Troposphere) instrument in Figure 12. We use the MOPITT V8 Joint (NIR + TIR) retrievals (Deeter et al., 2019) during 2001–2014, which are available from the NASA Earthdata archive (https://earthdata.nasa.gov). The model is interpolated to the gridded monthly MOPITT observations and the averaging kernel for each grid is applied to the simulated monthly mean CO profiles.

Figure 12. Absolute difference in tropospheric CO column between AM3 and MOPITT (left panel) and AM4.1 and MOPITT (right panel) for winter (December–January–February, DJF, top) and summer (June–July–August, JJA, bottom).

The tropospheric CO columns are in general higher in AM4.1 than AM3, in better agreement with MOPITT retrievals in terms of magnitudes (RMSE reduced from (2.6–2.7) × 10¹⁷ cm⁻² to (1.6–1.8) × 10¹⁷ cm⁻²) and spatial distribution (r² increased from 0.7–0.9 to 0.8–0.9). Compared to AM3, AM4.1 reduces the underestimations in column CO in the Northern Hemisphere, but overestimates column CO in the Southern Hemisphere, especially during summer. This is in part due to lower OH levels in AM4.1 than AM3.

To evaluate surface CO, we use measurements from a globally distributed network of air sampling sites maintained by the Global Monitoring Laboratory (GML) of the National Oceanic and Atmospheric Administration (NOAA) (Pétron et al., 2019; data available at ftp://aftp.cmdl.noaa.gov/data/trace_gases/co/flask/). Surface CO observations during 1988–2014 are used to evaluate model performance (Figure 13).

Figure 13. Comparison of surface CO mixing ratios (ppbv) from AM4.1 (red) and AM3 (blue) against NOAA Global Monitoring Division (GMD) flask observations (Pétron et al., 2019, for 1988–2014). Left panels show model bias (top) and correlation coefficient (bottom) versus observations, plotted by station latitude. Right panels show monthly time series comparisons at selection stations. The root mean square error (RMSE) and correlation coefficient (R) are indicated on plots.

AM4.1 simulates higher surface CO concentrations than AM3 over the Southern Hemisphere, and slightly overestimates surface CO concentrations by <5 ppb when compared to surface observations. Over the Northern Hemisphere, AM4.1 largely reduces the negative biases that occurred in AM3, with a mean bias of ±20 ppb over most GMD sites. This is consistent with the comparisons to the MOPITT retrievals shown above. In addition, compared to AM3, AM4.1 better captures the seasonal cycles (with correlation coefficient R > 0.5) at most sites and better captures the latitudinal gradient as well (R = 1.0 versus R = 0.9).

Comparisons of surface CO concentrations over pristine sites show significant improvement in AM4.1 over AM3 across latitudes from South to North. In the Southern Hemisphere, such as at South Pole (SPO), Ushuaia (USH), and Easter Island (EIC) sites, the underestimation of surface CO concentrations by AM3 are reduced in AM4.1. In the Northern Hemisphere, such as at Mauna Loa (MLO), Barrow (BRW), and Alert (ALT) sites, both surface concentrations and monthly variations are improved significantly in AM4.1 compared to AM3. These improvements are mainly associated with improved chemistry in AM4.1.

4.3 Aerosols

We first evaluate concentrations of aerosols in surface air. Figure 14 (top panels) compare simulated concentrations of sulfate and nitrate aerosols from AM4.1 with observations over the United States from the IMPROVE network. The model successfully captures the wide range of observed sulfate aerosol concentrations. While nitrate concentrations are well correlated with observations (R = 0.74), simulated concentrations are generally too high (normalized mean bias [NMB] = +80%). This bias is larger than in Paulot et al. (2016), where nitrate aerosols are assumed to deposit rapidly like nitric acid. Simulated concentrations of sulfate and nitrate in precipitation are compared with observations from the NADP network in the lower panels of Figure 14. The rainwater abundances of sulfate and nitrate are well correlated with observations, but with a low bias for sulfate (NMB = −19%) and a high bias for nitrate (NMB = +35%).

Figure 14. Comparison of AM4.1 (2000–2014) against IMPROVE (a,b) and NADP (c,d) observations of concentrations in surface air (top) and in precipitation (bottom) of sulfate (left) and nitrate (right).

Figure 15 compares simulated concentrations of sulfate, dust, and sea salt aerosols from AM3, AM4.0, and AM4.1 with observations from the University of Miami network (Savoie & Prospero, 1977). The model successfully captures the wide range of observed sulfate aerosol concentrations. For sulfate, the RMS error versus observations is reduced in AM4.1 (0.20 μg m⁻³) from AM3 and AM4.0 (both 0.22 μg m⁻³), and the correlation is improved (r = 0.93 in AM4.1, r = 0.89 in AM3 and AM4.0). The agreement between simulated and observed dust improves from AM3 to AM4.0, but then degrades in AM4.1, reflecting the shift from prescribed to interactive source regions for dust in LM4.1. The RMSE for simulated sea salt is reduced significantly in AM4.1 (0.35 μg m⁻³) compared with AM3 (0.47 μg m⁻³) and AM4.0 (0.49 μg m⁻³), as a result of updates to the emissions and deposition parameterizations in AM4.1.

Figure 15. Comparison of simulated (AM3, 1979–2008; AM4.0, 1980–2014; and AM4.1, 1980–2014) and observed (University of Miami) annual mean surface concentrations (μg m−3) of (first row) sulfate, (second row) dust, and (third row) sea salt sodium at 28 locations and (bottom) their ratios (simulated/observed) at each location (for AM4.1 only). Shaded contours indicate simulated surface concentrations (top colorbar) and symbols indicate the ratio of simulated/observed concentrations (bottom colorbar, symbol points upwards if ratio greater than one, downwards if less than one).

We next evaluate the simulated AOD against measurements from the AERONET sunphotometer network (Holben et al., 1998) in Figure 16. Here we use the quality assured and cloud screened level 2 version 2 AOD data (Smirnov et al., 2000). For comparison, we also show the results from AM4.0 (middle) and AM3 (bottom). Both AM4.0 and AM4.1 exhibit higher correlation (0.89 and 0.9) and lower RMS (0.07 and 0.08) with AERONET observations than AM3 (0.81 and 0.09, respectively). In particular, the large positive biases in the tropics and equatorial regions are reduced, which reflects the more efficient removal of aerosol by convective precipitation (Paulot et al., 2016). AM4.1 exhibits a greater positive bias than AM4.0 over the Midwest United States, associated with higher dust loading and nitrate aerosol (not included in AM3 and AM4.0).

Figure 16. Comparison of simulated aerosol optical depths (550 nm) with AERONET observations over the 2000–2014 period for (top) AM4.1, (middle) AM4.0, and (bottom) AM3 AMIP simulation. Dashed lines in left panels denote slopes of 0.5 and 2. Color in right panels shows the percentage difference between model and AERONET (i.e., 100% × [model − AERONET]/AERONET).

Figures 17 and S6 and compare the regional monthly mean AOD simulated by AM3, AM4.0, and AM4.1 with observations from the MODIS (Levy et al., 2007) and MISR (Kahn et al., 2009) instruments. AM4.0 and AM4.1 have reduced the seasonal contrast between winter and summer months, in better agreement with observational constraints. The spring maximum over East Asia and the North Pacific is also better captured with AM4.0 and AM4.1. The AM3 high biases over the Caribbean Sea and maritime continent are reduced consistent with the comparison against AERONET.

Figure 17. Monthly climatology (2003–2014) of aerosol optical depth simulated by AM3 (purple line), AM4.0 (green line) and AM4.1 (orange line) and measured by MODIS (TERRA: star, AQUA: cross) and MISR (filled circles) satellite instruments. Each panel represents a spatial average over the corresponding region on the background map. The numbers in each box show the correlation coefficients (left) and normalized root mean square error (right) compared to MODIS-TERRA (purple: AM3, green: AM4.0, orange: AM4.1).

These improvements primarily reflect changes in the treatment of aerosol removal, including reduced removal by frozen precipitation formed by the Bergeron process and more efficient scavenging by convective precipitation (Paulot et al., 2016). AM4.1 exhibits greater bias over Asia than AM4.0, which primarily reflects higher optical depth from dust and ammonium nitrate. Uncertainties in Asian SO₂ and NH₃ emissions (Zhang et al., 2009) and aerosol hygroscopic growth may also contribute to the AM4.1 high bias over this region (Paulot et al., 2018).

4.4 Hydroxyl Radical (OH) and Methane Lifetime

Here, we evaluate the climatological mean hydroxyl (OH) radical simulated by AM4.1, as OH is the primary atmospheric oxidant determining the abundance and lifetime of several short-lived climate forcers, including methane. The simulation of OH depends on the chemical mechanism, particularly the representation of isoprene photooxidation (Archibald et al., 2010; Bates & Jacob, 2019). Differences in emissions, meteorology, and photochemical mechanisms across models also lead to differences in OH (Wild et al., 2020). Climatological mean (1980–2014) global airmass-weighted tropospheric OH simulated by AM4.1 is 10.4 × 10⁵ molecules cm⁻³, about 18% lower than that simulated by AM3, but is within the range of values reported for ACCMIP models for the 2000s (Naik, Voulgarakis, et al., 2013). Consequently, the mean whole-atmosphere chemical lifetime of methane (calculated as the global methane burden divided by global total loss) in AM4.1 is 8.5 years; lifetime against loss by reaction with tropospheric OH is 9.7 years, which is 13% greater than the AM3 value of 8.6 years (1981–2000), but still lower than the observationally derived estimate of 11.2 ± 1.3 years (Prather et al., 2012). Figure 18 shows the comparison of tropospheric OH distribution for 12 regions simulated by AM4.1 with estimates from AM3, ACCMIP ensemble mean, and the climatology of Spivakovsky et al. (2000). AM4.1 simulates reduced OH levels compared to AM3 throughout the troposphere, possibly because of differences in emissions and chemical mechanisms between the two model versions. In particular, the lower lightning NO_x in AM4.1 versus AM3 acts to lower OH because of the strong sensitivity of OH to lightning NO_x emissions (Murray et al., 2013). Relative to the Spivakovsky et al. climatology, AM4.1 exhibits a reduced high bias compared with AM3, but has too low OH, particularly in the tropical upper troposphere.

Figure 18. Climatological (1980–2014) annual mean airmass-weighted tropospheric OH concentration averaged globally (top-most row) and regionally for individual atmospheric subdomains from AM4.1 (black) compared with those from AM3 (1980–2008, red), ACCMIP ensemble mean (orange), and climatological mean values from Spivakovsky et al. (2000) (purple). Values for AM4.1 and AM3 also show +/− standard deviation about the mean.

5 Sensitivities to Greenhouse Gases, Aerosols, and SST Perturbations

Table 2 shows the net radiative flux perturbations that result from historical changes in anthropogenic forcing agents and from idealized changes in CO₂ and SST. Comparison of these radiative metrics between AM3, AM4.0 and AM4.1 indicates that effective radiative forcings (ERF) from preindustrial to present-day changes in greenhouse gases and aerosols are nearly identical between AM4.0 and AM4.1. However, the ERF from quadrupling CO₂ is significantly lower in AM4.1, mostly because of the inclusion of interactive ozone (colder stratospheric temperatures reduce the rate of ozone chemical loss) but also partially resulting from increased dust emissions from LM4.1 (related to increased fires under elevated-CO₂ conditions). The Cess feedback, the change in net radiative flux resulting from an increase of SSTs by 2K, is significantly more negative in ESM4.1 (corresponding to a weaker Cess sensitivity), likely resulting from increased emissions of salt, dust, and BVOCs with increasing temperatures in ESM4.1. While comparison with previous-generation models is complicated by changes in the AMIP configuration since the AM3 model simulations were conducted (in particular, updating the “present-day” conditions from representing 1990 conditions to 2014 conditions), some assessment of these differences can be made using AM4.0 simulations conducted for 1990 conditions (as in Zhao et al., 2018a). The most important differences between AM3 and AM4.0 are a decrease in the magnitude of the negative aerosol ERF from AM3 to AM4.0, an increase in the 4xCO2 ERF consistent with an update to the treatment of CO₂ radiative bands (Zhao et al., 2018b), and a strengthening of the negative Cess feedback. The decrease in the magnitude of the aerosol ERF from AM3 to AM4.0 has been attributed by Zhao et al. (2018b) to a decrease in the strength of the aerosol indirect effect, resulting from the increase in horizontal resolution and improvements to the representations of aerosol convective wet deposition (Paulot et al., 2016) and aerosol activation.

Table 2 Effective Radiative Forcings and Feedbacks (in W m⁻²) to Greenhouse Gases, Aerosols, Anthropogenic Forcings, Land Use, Quadrupling CO₂, and SST Perturbations

	AM3 (1990)	AM4.0 (1990)	AM4.0 (2014)	AM4.1 (2014)
GHG ERF	2.63	2.61	3.14	3.22
Aerosol ERF	−1.69	−0.96	−0.73	−0.70
Anthro ERF	N/A	N/A	2.33	2.37
LU ERF	N/A	N/A	−0.33	−0.28
4xCO2 ERF	7.19	N/A	8.23	7.72
Cess feedback (SST + 2 K)	−2.86	−3.59	−3.64	−4.14

Note: The AM3 (1990) and AM4.0 (1990) experiments are identical to those described by Zhao et al. (2018a). The AM4.0 (2014) and AM4.1 (2014) experiments use 30-year climatological simulations, i.e., the piClim experiments from the Radiative Forcing Model Intercomparison Project (RFMIP; Pincus et al., 2016).

6 Summary

AM4.1 includes considerable advances in resolution and physics as in AM4.0 (Zhao et al., 2018a, 2018b) as well as a comprehensively revised suite of chemistry parameterizations to improve consistency in treatment across species and with advances in the underlying science over the last decade. AM4.1 is able to maintain the fidelity of AM4.0 while substantially increasing in comprehensiveness and associated climate-chemistry interactions and feedbacks. This development effort has also led to considerable improvement in model fidelity compared to GFDL's previous-generation coupled chemistry-climate model (AM3) with respect to observed atmospheric composition for aerosol, CO, ozone, as well as climate phenomena such as sudden stratospheric warmings.

Acknowledgments

We thank GFDL's Model Development Team, Modeling Systems Division, and Operations group for keeping ESM4.1 development and implementation going through many technical challenges. We thank the GFDL Data Portal team for their tireless work on developing and maintaining the vast infrastructure needed to enable publication of data output from our CMIP6 models to the Earth System Grid. We are also grateful to the many groups that freely share the observations used here to evaluate various aspects of chemistry and composition simulated by the models and to the input4MIPs activity (https://esgf-node.llnl.gov/projects/input4mips/) for making available the boundary condition and forcing data sets needed for performing model simulations. This manuscript benefited from internal GFDL reviews by Leo Donner and Songmiao Fan. Discussions with Songmiao Fan were helpful in improving the formulation of dry and wet deposition of aerosols.

Data Availability Statement

Data are provided at 10.22033/ESGF/CMIP6.1407. Model code is provided at https://data1.gfdl.noaa.gov/nomads/forms/esm4/. The input data are provided at ftp://data1.gfdl.noaa.gov/users/ESM4/ESM4Documentation/GFDL-ESM4/inputData/ESM4_rundir.tar.gz.