Application of the Pseudo-Global Warming Approach in a Kilometer-Resolution Climate Simulation of the Tropics

Clouds over tropical oceans are an important factor in the Earth's response to increased greenhouse gas concentrations, but their representation in climate models is challenging due to the small-scale nature of the involved convective processes. We perform two 4-year-long simulations at kilometer-resolution (3.3 km horizontal grid spacing) with the limited-area model COSMO over the tropical Atlantic on a 9,000 × 7,000 km 2 domain: A control simulation under current climate conditions driven by the ERA5 reanalysis, and a climate change scenario simulation using the Pseudo-Global Warming approach. We compare these

With respect to tropical low cloud changes, GCMs overall project a reduction of the low-cloud albedo, but the inter-model spread is much larger than in projections of deep convection (e.g., Vial et al., 2017;Zelinka et al., 2017). Also, there is a notorious negative cloud bias in subtropical low-cloud regions in GCMs (e.g., Kawai & Shige, 2020;. Large-eddy simulations (LES) show a more consistent climate change response of low clouds (e.g., Blossey et al., 2013) indicating a low-cloud reduction with warming, but given their small domain sizes and idealized setups, generalization of LES results to the entire planet introduces new uncertainties.
The fundamental problem behind the representation of convective clouds in GCMs is that a high horizontal and vertical resolution is required to resolve the small-scale convective circulations that drive clouds. Convective circulations represent the primary mode of vertical transport in the tropical atmosphere. If unresolved, these circulations, the clouds, as well as the vertical transport of heat and moisture associated with them have to be represented by convective parameterization schemes (e.g., Kawai & Shige, 2020). These schemes introduce substantial uncertainty in the simulation of deep-convective clouds (Suhas & Zhang, 2015), low-level clouds (Vial et al., 2016), and in how these clouds respond to climate change (Sherwood et al., 2014;Vial et al., 2017). With higher model resolution, convective parameterizations become less important and can eventually be switched off, which reduces the degree of parameterization and allows for a model formulation closer to physical first principles. For deep convective clouds, this threshold is reached at kilometer-resolution (Prein et al., 2015) which is why kilometer-resolution climate simulations are increasingly considered a major milestone toward more confident climate projections (e.g., Satoh et al., 2019;Schär et al., 2020;Schneider et al., 2017;Stevens et al., 2020). Precipitation statistics in the deep tropics have been found to be largely improved at kilometer-resolution compared to coarser models (Hart et al., 2018;Hohenegger et al., 2020;Klocke et al., 2017;Stevens et al., 2020).
Global kilometer-resolution multi-year climate simulations are not yet feasible due to computational cost , although rapid progress is evident (e.g., Satoh et al., 2012Satoh et al., , 2019Stevens et al., 2019). Instead, multi-year kilometer-resolution simulations are typically run on limited-area domains using boundary conditions from reanalysis data sets for evaluation runs (e.g., Ban et al., 2021), and from GCMs for climate change scenario simulations (e.g., Pichelli et al., 2021). Usually, a historical control simulation and a future scenario simulation are compared to extract the climate change signal. An alternative to this dynamical downscaling approach is the pseudo-global warming (PGW) approach (Adachi & Tomita, 2020;Brogli et al., 2023) in which reanalysis 10.1029/2022JD037958 3 of 24 boundary conditions are used for both the control and the scenario simulation. The climate change signal is obtained by imposing large-scale changes in the climate system on the reanalysis boundary fields of the scenario simulation. Doing so has the advantage that the biases from the GCM run do not enter the limited-area simulation, and that relatively short simulation periods can be used (Brogli et al., 2023). The PGW approach has extensively been applied in the mid-latitudes (Brogli et al., 2019;Kröner et al., 2017;Rasmussen et al., 2011;Sato et al., 2007;Schär et al., 1996;Wu & Lynch, 2000). We are aware of applications in the subtropics (Chen et al., 2020;Nakamura & Mäll, 2021), but to our knowledge, this study represents the first application of a PGW simulation at kilometer-resolution covering the full extent of the HC including the deep tropics.
We run two 4-year-long limited-area atmospheric simulations with the COSMO model at 3.3 km resolution over the tropical Atlantic with the goal to (a) evaluate how well the tropical climate and the associated distribution of clouds are represented in a kilometer-resolution atmospheric model and (b) compare the climate change response of the HC in terms of its structure, dynamics and clouds to the projections from the CMIP6 models. In a subsequent paper, a systematic analysis of the ensuing radiative feedbacks in this simulation will be presented. The following Section describes the modeling framework. Section 3 presents the results which are discussed in Section 4 and concluded in Section 5.

Experimental Setup
The limited-area model COSMO (see Section 2.3) is used in two 4-year-long simulations. The first one (CTRL) serves as a control simulation and represents current climate conditions. It is initialized and driven at the boundaries by the European Center for Medium Range Weather Forecast (ECMWF) ERA5 Re-Analysis (Hersbach et al., 2020). CTRL is used to evaluate the COSMO model against observations, and serves as a baseline for comparison with the second simulation: a climate change scenario simulation (PGW) obtained with the PGW approach (see Section 2.2).
Both, CTRL and PGW simulations are initialized on 1 August 2006 (for details on the initialization see below) and the analysis is done for the years 2007-2010 and focused on five geographic regions (Figure 1), representing The three subdomains intertropical convergence zone (ITCZ), TRD and STC comprise the regions of the three major tropical marine cloud regimes (deep convection at the ITCZ, trade-wind cumulus and stratocumulus). The HC-CS is used to compute altitude-latitude cross-sections to visualize the structure of the Hadley cell.
10.1029/2022JD037958 4 of 24 conditions over the entire tropical Atlantic (ATL), of the deep-tropical convection (ITCZ), of the shallow cumulus convection in the Trades (TRD), of the subtropical stratocumulus decks (STC), as well as a cross-section along the Hadley circulation (HC-CS). The 4-year-long analysis period is too short to fully average out inter-annual variability. The effect of this short analysis period in comparison to the typical 30-year-long climatological period is quantified in Section 3. The simulation domain covers 37.5° S-24.5° N and 54.5° W-28.0° E and consists of 2,750 × 2,065 × 60 grid points at 0.03°/3.3 km resolution, integrated with a time step of 25 s. The vertical grid stretches to an altitude of 30 km with a resolution of about 20 m near the surface, 500 m at 5 km altitude, and 1.5 km at the model top. The domain covers the deep-tropical and parts of the subtropical Atlantic (Figure 1) encompassing the full southern hemispheric branch of the HC. Although the focus lies on the Atlantic, the simulation domain includes parts of Africa and South America to enable interaction between marine and continental areas for instance through Monsoon circulations or the African easterly waves.

Pseudo-Global Warming Approach
The initial and boundary conditions of the PGW simulation are obtained following Brogli et al. (2023) using the software PGW4ERA5 v1.1 by adding the mean climate change signal (so-called climate deltas) for temperature, relative humidity, horizontal wind and sea and land surface temperature to the ERA5 boundary conditions of the CTRL simulation period. The climate deltas are a function of latitude, longitude, pressure and month, and represent the mean annual cycle of the spatial change pattern between two climate states, that is, here between a historical and a future scenario climatology. Note that, apart from the model initialization, the climate deltas are only applied at the lateral boundary conditions of the limited-area model, and at the surface for sea surface temperature (SST). The change signal PGW−CTRL in the interior of the domain is thus a model-internal response to the forcing applied at the boundaries.
The climate deltas are computed from the CMIP6 output of the MPI-ESM1-2-HR model (von Storch et al., 2017) as the difference between the Intergovernmental Panel on Climate Change SSP5-8.5 scenario (Kriegler et al., 2017) simulation during 2070-2099 and the CMIP6 historical simulation during 1985-2014. The MPI-ESM has a climate sensitivity at the lower end of the CMIP6 models. The output of the MPI-ESM is obtained as daily mean values from the CMIP6 output group CFday (on a regular latitude-longitude grid but on the native vertical coordinate) and aggregated into monthly means. Since this output group was intended for the Cloud Feedback Model Intercomparison Project  it is provided at fine vertical resolution which is desirable to accurately represent the difference in warming across the trade-wind inversion (see Brogli et al., 2023, Figure 5 and corresponding discussion). The obtained changes are displayed in Figures S1-S5 of Supporting Information S1.
The monthly mean climate deltas are then linearly interpolated to the grid and time of the ERA5 boundary files of the CTRL simulation, where the deltas are added to obtain the boundary and SST files of the PGW simulation. After modifying temperature and relative humidity, the pressure field is adjusted to restore the hydrostatic balance. The corresponding changes in microphysical species quickly adjust to the new thermodynamic environment within the model domain, and are thus not otherwise accounted for in the PGW methodology. The change of the soil temperature is computed based on the surface skin temperature climate delta assuming an exponential decay of the annual cycle signal with depth. Initial soil moisture is not modified and taken from the CTRL simulation (5 months before the analysis period begins). Greenhouse gas concentrations are held fixed during the simulation and set to 530 ppm CO 2 -eq during CTRL and 1,100 ppm CO 2 -eq during PGW consistent with the SSP5-8.5 scenario. Aerosols are identical in CTRL and PGW following the Tegen et al. (1997) climatology. Even though biomass burning over Africa is a significant source of aerosol over the Atlantic (Zuidema et al., 2016), the change of aerosol loading between CTRL and PGW is neglected here for simplicity. The same is the case for ozone.

COSMO Model
The COSMO model is a fully compressible non-hydrostatic atmospheric model originally developed as a numerical weather prediction model (Baldauf et al., 2011) and later evolved into a regional climate model (Rockel et al., 2008). Here a COSMO version capable of exploiting Graphics Processing Units is employed Leutwyler et al., 2016). This version of COSMO has been extensively validated in kilometer-scale configurations including a 10-year-long reanalysis-driven simulation over Europe (Leutwyler et al., 2017), validation of clouds (Hentgen et al., 2019), and surface winds (Belušić et al., 2018). The model discretizes the horizontal and vertical dimensions on a rotated latitude-longitude grid and a generalized Gal-Chen coordinate, respectively. The model equations are integrated in time with a split-explicit third-order Runge-Kutta scheme (Baldauf et al., 2011;Klemp & Wilhelmson, 1978;Wicker & Skamarock, 2002). Horizontal advection is treated with a fifth-order advection scheme except for moist quantities which are integrated using a positive-definite second-order scheme (Bott, 1989). The upper boundary is treated following (Klemp & Durran, 1983) and no relaxation of the model top toward the boundary files is performed.
Radiative transfer is computed following the δ-two-stream approach after Ritter and Geleyn (1992). The subgrid-scale vertical turbulent fluxes are parameterized with a TKE-based model (Raschendorfer, 2001). Cloud microphysics is parameterized using the single-moment bulk scheme after Reinhardt and Seifert (2006). The parameterizations for deep and shallow convection are switched off as this was previously found to give a reasonable representation of clouds in the COSMO model at kilometer-resolution (Heim et al., 2021;Vergara-Temprado et al., 2020). At the surface, the second-generation land-surface model TERRA_ML (Heise et al., 2003) with the groundwater-runoff scheme after Schlemmer et al. (2018) is used on land grid points.
Soil moisture profiles are initialized based on a 12-year-long soil spin up COSMO simulation at 24 km grid spacing. The resulting soil moisture conditions serve as initial condition for a 5-month-long spin up at full (3.3 km) resolution, initialized on 1 August 2006 (for CTRL and PGW). Over ocean grid points, sea-surface temperature is prescribed from the ERA5 reanalysis (or the PGW shifted ERA5, respectively) with a temporal resolution of 3 hr. A number of empirical model parameters are adjusted to improve the representation of low clouds in comparison to previous simulations over the extratropics: The vertical turbulent length scale is set to 200 m. The minimum threshold for eddy-diffusivity for heat and momentum under stable conditions are set to 0.25 m 2 s −1 (see Possner et al. (2014) for more details about these parameters).

Observational Data Sets
The following observational data sets are used to evaluate the simulations: • The Clouds and the Earth's Radiant Energy System (CERES) Energy Balanced and Filled (EBAF) Top-of-Atmosphere (TOA) Edition-4.0 Data Product (Loeb et al., 2018) provides monthly values of TOA radiation at 1° horizontal resolution. • The Satellite Application Facility on Climate Monitoring (CM SAF) TOA radiation (Clerbaux et al., 2013), based on the Geostationary Earth Radiation Budget instrument, provides monthly values of TOA radiation at 45 km horizontal resolution. • The Global Precipitation Measurement (GPM) Integrated Multi-satellitE Retrievals for GPM (IMERG) data set (Huffman et al., 2019): provides precipitation observations at daily frequency and 0.1° horizontal resolution. • The ERA5 reanalysis (Hersbach et al., 2020) is a gridded reanalysis data set. It is obtained from the CDS data store (Copernicus Climate Change Service (C3S), 2017) and used at 3-hourly frequency and 0.25° horizontal resolution.

Results
We start by looking at a cloud visualization to provide an overview of the cloud phenomena occurring within the domain. . The visualization shows atmospheric cloud liquid and ice water content in white and light-blue-to-white colors, respectively, as well as surface precipitation in yellow-to-blue colors. Areas of high atmospheric water vapor content over oceans are visualized using purple shading. The land surface is rendered based on the model surface albedo and vegetation types, with a desert-to-green color gradient that is modulated by the soil moisture content to imitate the seasonal cycle of vegetation density. The panels on the right-hand side show close-up views of (a) a mid-latitude frontal system, (b) a tropical cyclone, (c) a mesoscale convective system, (d) deep convection at the marine intertropical convergence zone, (e) marine stratocumulus clouds, (f) marine shallow cumulus (or trade-wind cumulus) clouds. An animation of this visualization can be obtained via https://doi.org/10.3929/ethz-b-000568941.
the trade-wind-cumulus topped MBL on its way toward the deep tropics. Hereby, different modes of mesoscale cloud aggregation are producing regional differences in cloud cover.
The horizontal and vertical circulations underlying the clouds shown in Figure 2-from the large-scale tropical overturning HC down to small-scale convective MBL circulations-are all represented explicitly on the model grid, even though many of the circulation features are resolved only at the coarse end of the spectrum. In the following section, we are going to evaluate the simulation and compare it to the CMIP6 historical runs.

Evaluation of the CTRL Simulation
We start the evaluation at the large-scale with the analysis of the meridional structure of the HC. Afterward, we look at the spatial structure and annual cycle of individual cloud regimes. Figure 3 shows the meridional distribution of clouds and surface precipitation along the HC-CS domain for the annual mean as well as for the 3-month-periods with southernmost (February-April) and northernmost (July-September) extent of the marine ITCZ. The ERA5 record (Figures 3a-3c) indicates that the annual mean cloud fraction and precipitation have their peak at 4°N. The peak shifts to 8°N during boreal summer, and splits into a primary peak persisting at 3°N, and a secondary peak at about 2°S during boreal winter. Comparing the 4-year-long (Figures 3a-3c) and the 30-year-long (Figures 3d-3f) ERA5 cross-sections shows that the climatological distributions of cloud and precipitation are well represented by the 4-year-long simulation period used in CTRL (see Section 2.1). The comparison of surface precipitation between ERA5 and GPM IMERG indicates a close agreement between these two reference data sets (Figures 3a-3c).

The Hadley Cell
The zonal mean precipitation is well reproduced in CTRL with respect to ERA5 with the exception of a slightly underestimated annual mean peak (red lines in Figures 3g-3i). The CMIP6-EM captures the northward shift of the ITCZ during boreal summer, but largely overestimates and misplaces the boreal winter secondary peak in the southern hemisphere (Figures 3j-3l). The latter is a manifestation of the double ITCZ problem ( Figure 3j). Besides an overestimation of precipitation and clouds in the southern hemispheric deep tropics, the double ITCZ also results in too frequent subtropical high clouds. We further show the cross-sections for the MPI-ESM model individually (Figures 3m-3o), as it is used to compute the climate delta for the PGW simulation. The double ITCZ problem is more pronounced in the MPI-ESM model than in the CMIP6-EM and results in a bimodal annual mean precipitation distribution that is almost symmetric about the equator.
In ERA5, the cloud field at the ITCZ consists of (a) a concentration of low-level clouds, (b) a secondary liquid cloud maximum at around 5 km, and (c) the deep-convective anvil clouds between 8 and 15 km altitude. In the subtropics, the free troposphere contains virtually no clouds below 10 km as a result of the stable and dry conditions in the downward branch of the HC. At the surface, low clouds are topping the MBL. The MBL is deeper south of the equator than north of the equator. In the former case, the MBL is located further off the coastal upwelling regions of Africa, and thus experiences warmer SST favoring the development of a deep MBL (e.g., Bretherton & Wyant, 1997). Beyond 25°S, clouds of extra-tropical origin penetrate into the subtropical atmosphere, in particular at high altitudes, where they contribute to the subtropical high-cloud fraction.
The annual mean and seasonal structure of ITCZ clouds in CTRL corresponds well with ERA5 (Figures 3a-3c and 3g-3i). The main difference is that in the annual mean and during Feb-Apr, the extra-tropical clouds reach less far into the subtropics in CTRL. Overall, the differences are not fundamental and we conclude that CTRL simulates the cloud field along the HC consistent with ERA5. In contrast, the CMIP6-EM does not reproduce the vertical cloud structure at the ITCZ as seen in ERA5 and CTRL (Figures 3d-3f and 3j-3l). Instead, many of the analyzed CMIP6 members simulate a too dense cloud field throughout the tropical tropospheric column which also penetrates too high up into the tropical tropopause layer in some models. Although the focus of this study lies on marine clouds, the structure of the HC over land is shown in Figure S7 of Supporting Information S1. While the vertical cloud structure in CTRL is comparable to ERA5, the high-cloud fraction and surface precipitation is significantly larger than in ERA5 and GPM IMERG. These quantities are both related to deep convection and thus indicate that deep convection at the continental ITCZ may be overestimated in CTRL, as will be shown later.
The first two columns of Figure 4 show the large-scale overturning motion of the HC in terms of the meridional and vertical mass flux along the HC-CS domain. Air rises at the surface of the ITCZ and diverges above an altitude of 10 km toward the poles. The poleward (i.e., the elevated southward) branch of the HC converges with the  . The three columns represent multi-year averages (left panels) during the entire year, as well as (center panels) during February-April and (right panels) July-September when the marine intertropical convergence zone reaches its southernmost and northernmost extent, respectively. The cloud fraction is obtained from the respective model output, except for COSMO where grid points with a specific cloud liquid plus ice water content ≥0.01 g m −3 are considered cloudy while the remaining grid points are considered cloud-free. As a complementary reference observation, Global Precipitation Measurement IMERG precipitation is shown as a red dashed line in panels (a-c).
northward branch of the Ferrel cell at 15°S in ERA5 (Figure 4a) which sets the latitude of strongest subtropical subsidence ( Figure 4b). The HC is closed by the trade winds at the surface that are largely confined to the MBL and converge at 5°N (Figure 4a).
Compared to ERA5, the poleward branch of the HC in the southern hemisphere reaches further south in CTRL and the CMIP6-EM (Figures 4a, 4e, and 4i). Consequently, the subtropical subsidence extends further south in CTRL and the CMIP-EM than in ERA5 (Figures 4b, 4f, and 4j). This dynamical difference between CTRL and ERA5 is consistent with the differences in the high-cloud fraction of extra-tropical origin at around 20°S between CTRL and ERA5 (Figures 3a and 3g). A further difference between CTRL and ERA5 is related to the ITCZ outflow in the lower troposphere. ERA5 shows a pronounced shallow circulation between 2 km < z < 6 km ( Figure 4a) which is much weaker in CTRL (Figure 4e). Thus, while the HC in ERA5 has a pronounced dual circulation structure (i.e., a deep circulation and a shallow circulation), the shallow circulation is almost absent in CTRL and the deep circulation is stronger (Figure 4m). In line with that, the subtropical subsidence profile in ERA5 shows a pronounced maximum at 3 km (Figure 4b), while it is more uniform throughout the troposphere in CTRL (Figure 4f). Differences in subsidence can result from differences in the radiative cooling rate or temperature stratification. The weaker subtropical subsidence at low levels in CTRL appears to be due to weaker radiative cooling rate compared to ERA5 (see Figure S8 in Supporting Information S1). The meridional outflow of the ITCZ in the CMIP6-EM (Figure 4i) is more evenly distributed over the free-tropospheric column compared to CTRL and ERA5, in line with the evenly distributed cloud fraction (Figures 3j-3l). Further, the imprint of the double ITCZ is well visible in the bias of the vertical wind field, showing an anomalous upward motion south of the equator compared to ERA5 (Figure 4r).
The third and fourth columns of Figure 4 show the thermodynamic structure of the HC along the HC-CS domain. Temperature in CTRL does not deviate from ERA5 by more than 1 K except in the subtropical lower troposphere (Figure 4o). The differences in relative humidity between CTRL and ERA5 ( Figure 4p) are also small except for altitudes above 15 km where temperatures are very low and small differences in the amount of deep-convective outflow have a large effect on the relative humidity. Overall, the subtropical troposphere is slightly drier in CTRL than in ERA5, and (as for temperature) the differences are largest in the lower troposphere. The tradewind inversion in CTRL is more elevated than in ERA5 which explains the lower temperature and enhanced humidity in CTRL in between (i.e., between 1 and 2 km). Above the inversion, the differences may be related to lower-tropospheric mixing which alters the moisture content of the free troposphere and thus modulates the clearsky radiative cooling rate. The drier free troposphere in CTRL (Figure 4p) may thus explain the weaker radiative cooling rate at these levels ( Figure S7 in Supporting Information S1), and consequently the warmer temperature ( Figure 4o) and weaker subsidence (Figure 4n) in the subtropical lower troposphere in CTRL. The biases in temperature and relative humidity in the CMIP6-EM (Figures 4s and 4t) are larger than in CTRL. This is expected since CTRL is driven by ERA5 at its boundaries while the CMIP6 simulations are global. Tropospheric temperature is lower in the CMIP6-EM than in ERA5 while stratospheric temperature is higher (Figure 4s). Further, the deep tropics in the southern hemisphere are much moister than in ERA5, due to the double ITCZ (Figure 4t).

Tropical Cloud Regimes
We continue with a more detailed evaluation of clouds and precipitation. Figure 5 shows the annual mean spatial pattern of the TOA albedo, surface precipitation, and TOA outgoing longwave radiation (OLR). The low-cloud albedo is substantially overestimated in CTRL (Figure 5c). Surface precipitation in CTRL is overestimated over land in comparison to the GPM IMERG data set (Figure 5g). The precipitation of the marine ITCZ is well represented to the East, but its westward extent is underestimated. OLR is far too low over land (Figure 5k), in line with the precipitation bias. Over sea, the underestimation of OLR in the deep tropics is smaller, but subtropical OLR is overestimated. The CMIP6-EM bias patterns of these three variables (Figures 5d, 5h, and 5l) reveal the two well-known deficits of GCMs over low-latitude oceans: The double ITCZ problem and the underestimation of stratocumulus clouds. Also note that over land, the CMIP6-EM shows much smaller biases than CTRL.
We continue with the evaluation of the annual cycle of clouds on the four marine analysis domains ATL, ITCZ, TRD and STC. Figure 6 shows the mean annual cycle of TOA albedo in CERES EBAF, CM SAF, CTRL, ERA5, and the CMIP6 mean and ensemble spread. The annual cycle is expressed as deviations from the annual mean.  , 2004-2014 and 2004-2010, dashed lines) to assess the effect of inter-annual variability. The annual cycle of the 4-year-period is very similar as in the extended period, indicating that the former is representative of the long-term conditions. Table 1 shows that the marine albedo in CTRL is overestimated by approximately 3.5%. However, the timings of annual maximum and minimum cloud cover as well as the amplitude of the annual cycle are much improved in CTRL compared to the CMIP6-EM on the ATL (Figure 6a) and the TRD (Figure 6c) domains, where CTRL even outperforms the ERA5 record. On the ITCZ (Figure 6b) and the STC (Figure 6d) domains, on the other hand, similar (though mitigated) deficiencies as in the CMIP6-EM are visible, that is, an overestimation and underestimation of the ITCZ albedo during boreal summer and winter, respectively, as well as an underestimation of the annual cycle on the STC domain. Figure 7 shows the annual cycle of surface precipitation over the marine analysis domains. The annual cycle is overall well represented in CTRL with the largest deviations found over the ITCZ domain (Figure 7b), where the simulated timing of maximum precipitation lags 1 month behind the observed due to an overestimation of the boreal summer precipitation (similar as for the albedo in Figure 6). The relative difference in precipitation amount between the different domains is well simulated in CTRL, but with slightly more precipitation on the TRD and the STC domains compared to GPM IMERG (Table 1). In the CMIP6-EM, the precipitation amount over the TRD is overestimated more strongly due to the double ITCZ problem.
The annual cycle of OLR is shown in Figure 8. Unlike for the albedo, there is a surprisingly large difference between CERES EBAF and CM SAF of about 6 W m −2 (Table 1). ERA5 is closer to CERES EBAF. The amplitude of the annual cycle of OLR in CTRL is overestimated on the three small analysis domains (ITCZ, TRD and STC; Figures 8b-8d). This appears to be mainly due to an overestimated high-cloud fraction originating at the ITCZ during the first half of the year and resulting in too much downwelling longwave radiation. We further see a signal of too high tropospheric water vapor content (not shown) originating from the African ITCZ which contributes to the opacity of the atmosphere over the Atlantic. Similar as for precipitation, the error of CTRL is largest over the ITCZ domain (Figure 8b).

Sensitivity to Climate Change
We continue with the analysis of the climate change signal obtained from the PGW simulation (see Section 2.2). Figure 9 shows the annual mean spatial distribution of surface precipitation change between CTRL and PGW, and between HIST and SCEN. Atlantic ITCZ precipitation mostly decreases except for parts of the central tropical Atlantic where PGW simulates a pronounced precipitation increase (Figure 9c). Averaged over the ITCZ and ATL domains, and scaled by the temperature change, precipitation increases by 0.2% K −1 and decreases by 4.5% K −1 , respectively. The temperature change for this computation was evaluated at 1 km altitude which roughly corresponds to the cloud base ( Figure 3g). Consistent with the CMIP6-EM (Figure 9f), the precipitation increase in COSMO (Figure 9c) is located in the center of the Atlantic, rather than along the West-African coastline (as, e.g., in the MPI-ESM; Figure 9i). Also consistent is the southward propagation of the precipitation maximum, that is, the most pronounced increase is located to the South of the precipitation maximum in CTRL/HIST (see also Figure 11). However, unlike the CMIP6-EM, COSMO simulates a pronounced precipitation reduction in the deep-tropical East Atlantic while the precipitation reduction in the West Atlantic Trades is less systematic (e.g., precipitation locally strengthens in the North Atlantic Trades), and precipitation over land is reduced instead of increased. Finally, while the precipitation changes in the CMIP6-EM and the MPI-ESM associated with the ITCZ are relatively symmetric about the equator as a result of the double ITCZ, this is not the case in COSMO which does not show a double ITCZ. Figure 10 shows the change in the thermodynamic structure of the HC. The temperature change PGW−CTRL (Figure 10b) is similar to SCEN−HIST of the MPI-ESM (Figure 10j) although slightly weaker overall. The similarity is expected since the latter is the climate delta used to derive the PGW boundary conditions. Tropospheric relative humidity is projected to decrease in the CMIP6 models (Figures 10h and 10l) which is a reflection of the overall drying of the tropics (e.g., Lau & Kim, 2015). COSMO projects a qualitatively similar humidity change pattern (Figure 10d) as the CMIP6-EM, but weaker humidity changes than the MPI-ESM. One thing to note is that the drying of the upper deep-tropical atmosphere is more pronounced in COSMO than in the CMIP6 models. Note that the relative humidity increase in the tropopause layer in all models appears to be associated with a comparably weak temperature increase due to enhanced longwave radiative cooling (Shine et al., 2003) and enhanced vertical moisture transport at high levels given the expansion of the troposphere (e.g., Lau & Kim, 2015). Figure 11 shows the simulated changes in the cloud field along the HC-CS domain. The signal PGW−CTRL ( Figure 11c) shows a rise of the anvil clouds at the ITCZ accompanied by a net increase in the vertically integrated anvil cloud fraction in the north Atlantic deep tropics (<10°N) and a reduction in the north Atlantic subtropics (>10°N). In the CMIP6-EM (Figure 11f), the rise of the high clouds is barely visible in the cloud field change, but will be visible in the meridional wind change (see Figure 12f). In contrast to COSMO, both the CMIP6-EM (Figure 11f) and the MPI-ESM (Figure 11i) exhibit a pronounced deep-tropics squeeze, that is, a reduction of the cloud fraction at the poleward margins of the annual mean ITCZ. Note that in the CMIP6 models, this reduction is visible at both instances of the double ITCZ (the real one north of the equator and the spurious one south of the equator) while in COSMO it is mainly visible at the northern edge of the ITCZ and only in the high-cloud fraction. Also note that when considering this factor with seasonal resolution (not shown), an equatorial intensification is always present when the ITCZ overlaps with the equator. As a result of the deep-tropics squeeze, the ITCZ precipitation in SCEN and PGW (Figures 11b, 11e, and 11h) is slightly more concentrated around the equator than in HIST and CTRL (Figures 11a, 11d, and 11g). Finally, we note that the change PGW−CTRL (Figure 11c) in trade wind clouds exhibits an opposite sign in the North and South Atlantic, unlike in SCEN− HIST (Figures 11f and 11i) where shallow cloud cover decreases in both hemispheres.
The circulation changes along the HC-CS domain are shown in Figure 12 in terms of the meridional and vertical mass fluxes. COSMO simulates an upward shift (maxima rise from approximately 12-14 km) and a weakening of the upper-level meridional outflow of the ITCZ in PGW compared to CTRL (Figures 12a-12c). This change pattern qualitatively agrees with the CMIP6-EM (Figures 12d-12f) and the MPI-ESM (Figures 12g-12i), but the change in magnitude is slightly stronger compared to the CMIP6-EM, and substantially stronger compared to the MPI-ESM. Along with the change in the meridional wind, the upward motion at the ITCZ in COSMO  TRD, andSTC Shown for CTRL (2007-2010), ERA5 (2007ERA5 ( -2010, CMIP6-EM HIST (1985,

and the Satellite Observations (CERES EBAF and CM SAF for Albedo and OLR, and Global Precipitation Measurement IMERG for Precipitation)
Albedo (Figures 12j-12l) reaches higher levels (i.e., becomes stronger beyond 10 km) and intensifies across all levels at the equator (i.e., at the southern flank of the annual-mean ITCZ in CTRL). This response of the ITCZ to warming in COSMO qualitatively agrees with the response of the CMIP6 models (Figures 12m-12r), yet there are some distinct differences: First, the intensification of the ITCZ in COSMO is more pronounced in the northern hemi-  sphere and much stronger near the equator, unlike in the CMIP6 models (compare Figure 12l and Figures 12o  and 12r) where it is much broader and weaker. In the CMIP6 models the equatorward shift of the secondary ITCZ (i.e., the one south of the equator) fosters upward motion in the southern hemisphere near the equator in SCEN. This difference between COSMO and the CMIP6 models is consistent with the absence of a double ITCZ in CTRL. Second, while the weakening of the ITCZ north of the equator in the CMIP6-EM is particularly pronounced at low levels, it is more pronounced in the mid troposphere (between 6 and 10 km) in COSMO (compare Figure 12l and Figure 12o). Third, above 10 km, COSMO shows a substantial intensification of upward motion across the whole meridional extent of the CTRL ITCZ such that overall, the ITCZ intensifies at high levels. With respect to subtropical subsidence, COSMO simulates a similar response as the CMIP6 models, but with a slightly more pronounced strengthening above 10 km (consistent with the stronger upward motion at the ITCZ at high levels).

Evaluation of CTRL
In Section 3, we discussed the realism of the ERA5-driven CTRL simulation in comparison to the CMIP6-EM and found significant differences. As the two underlying simulation strategies differ strongly, it is not feasible to disentangle effects due to computational resolution (3 km vs. 50-200 km) and simulation setup (ERA5 driven atmospheric simulations vs. free-running coupled simulations). The main purpose of the following discussion is thus to summarize the differences between the CTRL simulation and the CMIP6 ensemble, and to determine whether the ERA5-driven simulations are credible enough to serve as the basis of climate-change simulations using the PGW approach.
The improved representation of the annual cycle of the albedo, in particular on the TRD analysis domain (representative of shallow cumulus clouds), as well as the accurate vertical structure and meridional position of the ITCZ (i.e., no double ITCZ) are perhaps the most promising improvements compared to the CMIP6-EM. The prescribed SST obtained from ERA5 likely has a beneficial impact on the properties of the MBL and the position of the ITCZ in CTRL. For instance, the double ITCZ problem of the CMIP6-EM is thought to be related to air-sea interaction, among other factors (Li & Xie, 2014;Lin, 2007). It would therefore be interesting to test if for instance a coupled model setup at kilometer-resolution or a GCM-driven kilometer-resolution simulation, or even a coarse-resolution reanalysis-driven simulation, were to suffer from the double ITCZ problem. The results by Segura et al. (2022) suggest that such simulations may result in a double ITCZ. Under the assumption that the improved representation of the ITCZ in our limited-area CTRL simulation is due to the forcing from ERA5, our application demonstrates one benefit of the PGW approach compared to conventional downscaling, that is, that GCM circulation biases are not propagated to the limited-area simulation. We argue that this realistic representation of the ITCZ location is a good starting point to study its climate change signal.
Concerning the simulation of low clouds, the representation of the annual cycle of the albedo in CTRL is better on the TRD domain than on the STC domain. This discrepancy may relate to the type of clouds most prevalent on the two domains. The TRD domain is predominantly covered by trade-wind cumulus clouds while stratocumulus clouds are more frequent on the STC domain (Warren et al., 1988). The difficulty to represent the annual cycle of stratocumulus clouds in a kilometer-resolution model with 60 vertical levels is not unexpected since a firm representation of the stratocumulus-topped MBL with its very shallow inversion cloud layer is challenging even in LES (e.g., Stevens et al., 2005). Nevertheless, the fact that the COSMO simulations yield stratocumulus decks already at kilometer-resolution, notably without any shallow convection scheme, is very promising. In the tradewind cumulus regime clouds often aggregate into clusters that frequently exceed the kilometer-scale (e.g., Bony et al., 2020). The CTRL simulation indeed produces such clusters (see Figure 2) suggesting that some of the dominant mesoscale patterns of MBL circulations and clouds in the Trades are at least partially resolved. Similar results have been found in previous studies using kilometer-resolution models (Caldwell et al., 2021;Heim et al., 2021;Klocke et al., 2017). It is interesting to note that the annual cycle of albedo in CTRL on the TRD domain is actually better simulated than in ERA5. This result suggests that the improved representation of these clouds is not primarily a result of the prescribed SST, but portrays the added value of explicit convection and fine model resolution.
On the other hand, we find a mean bias in the low-cloud albedo in the CTRL simulation compared to satellite observations ( Figure 5). This bias was found to be caused by an overestimation of cloud water (i.e., cloud opacity) rather than cloud fraction (Heim et al., 2021). As shown by Liu et al. (2022), this bias of the COSMO model at kilometer-resolution can be reduced through systematic model calibration. The model version used here is still based on a set of empirical parameters that were calibrated for applications over continental regions of the mid-latitudes (Bellprat et al., 2016). We also find a bias in quantities related to deep convection at the continental ITCZ over Africa ( Figure 5). Compared to the well calibrated COSMO simulations in the mid-latitudes (e.g., Ban et al., 2021;Leutwyler et al., 2017;Vergara-Temprado et al., 2020;Zeman et al., 2021), the bias in precipitation and OLR is still quite substantial. The set of empirical parameters used in this study differs from other COSMO setups that have been used over Africa (Bucchignani et al., 2016;Sørland et al., 2021). A calibration effort similar as it was done for the tropical Atlantic in Liu et al. (2022), but for continental Africa would likely result in a simulation setup with less biased deep convection overall. Yet, with upcoming global KRM simulations in mind, it should be noted that a set of model parameters with global validity will be necessary. As a final note, it is possible that the poor representation of the continental ITCZ could affect the representation of the marine ITCZ via the lower-tropospheric mean easterly flow or via gravity waves (e.g., Leutwyler & Hohenegger, 2021).

Climate Change Signal PGW−CTRL
The changes SCEN−HIST in wind and humidity at the Atlantic HC compare qualitatively well to the global CMIP5 models (Lau & Kim, 2015). This agreement indicates that, despite the local computational domain employed, the obtained results may be indicative of the global patterns. Concerning the change signal in COSMO (PGW− CTRL), the tropospheric warming profile closely follows the climate delta (SCEN−HIST) of the MPI-ESM simulation ( Figure 10). This similarity is expected since the temperature change is a large-scale signal that enters the model at the lateral boundaries (see Section 2.2). The change signal PGW−CTRL for humidity shows a qualitatively similar change pattern as the CMIP6-EM and the MPI-ESM, however, with a slightly weaker drying of the tropical troposphere than in the MPI-ESM driving model ( Figure 10). The distribution of humidity is tied to the representation of deep convection and how it changes between CTRL and PGW (or HIST and SCEN).
Since the intensification of deep-tropical convection near the equator is more pronounced in COSMO than in the MPI-ESM model, some differences in the humidity change are expected.
As expected by FAT (Hartmann & Larson, 2002) and the stability iris hypothesis , the anvil clouds rise and cover a (slightly) smaller area in PGW than in CTRL. In this respect, our simulation qualitatively agrees with high-resolution simulations of radiative-convective equilibrium in aqua-planet configurations, whereof a majority shows a reduction in the high-cloud fraction with warming (Wing et al., 2020). Yet, there is also a pronounced increase in the anvil cloud fraction around the equator such that locally, the high-cloud fraction increases which represents a different response compared to the CMIP6 models.
The circulation changes show distinct differences between the COSMO model and the CMIP6 models. Most importantly, the COSMO model simulates a narrow zone near the equator with a strongly intensified vertical mass flux. This is consistent with the increase in the anvil cloud fraction near the equator. Further, the overall narrowing of the annual mean ITCZ is less pronounced than in the CMIP6 models. Thus, overall, even though the annual mean ITCZ in COSMO becomes weaker at its margins (south of 0° and north of 10°N), the broad region of upward motion in CTRL (0-10°N) shows an intensification more than a weakening. This is a qualitative difference to the CMIP6 models which show two overall weakening ITCZ branches, which however both approach the equator and thus result in a local intensification of deep convection just at the equator.
An often discussed hypothesis on the change in the dynamics of the HC is the prominent deep-tropics squeeze, that is, the narrowing of the annual mean ITCZ, detectable in GCMs (e.g., Byrne & Schneider, 2016;Lau & Kim, 2015). In our CMIP6 ensemble, the squeeze is clearly evident in the form of a strengthening and narrowing of the deep-tropical convection and a corresponding reduction of cloud fraction at the edges of the ITCZ (Figures 11 and 12). However, this narrowing of the annual mean ITCZ seems to be enhanced by the fact that the CMIP6-EM projects a similar but mirrored change signal at both branches of the ITCZ (i.e., the one north of the equator, and the spurious one south of the equator-the double ITCZ). This perception is supported by the fact that the narrowing of the ITCZ in GCMs is associated mainly with a northward shift of the southern edge (Byrne & Schneider, 2016). As previously discussed, the deep-tropics squeeze can also be visually detected in the kilometer-resolution COSMO simulation but it does not seem to result in an overall weaker ITCZ, and-given the absence of a double ITCZ-the pattern is less symmetric about the equator than in the CMIP6 models. So, the question arises whether the narrowing of the deep tropics in the CMIP6-EM would be equally pronounced if it did not exhibit the double ITCZ in HIST. The double ITCZ was found to relate to the strength of the low-cloud feedback in GCMs (Tian, 2015) which was argued to be driven by differences in the lower-tropospheric stability depending on the strength of the double ITCZ (Webb & Lock, 2020). Whether and how the double ITCZ responds to warming and how this relates to radiative feedbacks is thus of high relevance for climate projections and requires further research.
There are some limitations of the model setup presented in this study. The COSMO model was originally designed as a weather prediction model, and aerosols and ozone are represented in a simplified manner compared to comprehensive climate models. Further, the one-moment microphysics scheme assumes a constant cloud-droplet number concentration. Changes in aerosol concentrations therefore do not directly alter the properties of the simulated clouds. Keeping ozone and aerosol concentrations constant between CTRL and PGW is thus a pragmatic choice for the given model configuration. Still, accounting for such effects might alter the simulated response to warming. For instance, the MPI-ESM shows an increase and slight upward shift of the ozone maximum between HIST and SCEN. Another simplification of the modeling setup in this study is the use of a limited-area model and the PGW approach. Given that the same weather enters the model domain at the boundaries in CTRL and PGW, large-scale circulation changes from the GCM may be restrained by the persistence of the weather phenomena at the lateral boundaries. Specifically, at the boundary between the subtropics and the mid-latitudes, it is unclear how the extension of the HC toward South with warming (e.g., Lau & Kim, 2015) is restrained by the fact that the mid-latitude frontal systems enter the PGW simulation at the same latitudes as in CTRL.
An interesting extension of this study would be to repeat the analysis using PGW simulations derived with climate deltas of different GCMs to test the sensitivity of the change signal PGW−CTRL to the climate delta. The role of SST warming patterns appears to be of particular interest here. Given the importance of the SST pattern on changes of the ITCZ (Huang et al., 2013), it would not be surprising to find differences in the change PGW−CTRL in terms of structure and location of the ITCZ for different climate deltas.

Conclusion
In this study, we conducted what is, to our best knowledge, the first application of the PGW approach on a marine tropical domain that contains the entire HC. We performed two 4-year-long simulations at 3.3 km horizontal resolution with the limited-area model COSMO over the tropical Atlantic. The analysis includes an evaluation of the structure of the HC and tropical clouds under current climate conditions (CTRL), and a comparison of the obtained climate change signal (PGW−CTRL) to that of a CMIP6 model ensemble (SCEN−HIST). The radiative feedback between CTRL and PGW will be analyzed in a follow-up study. The main analysis findings include: 1. An improved representation of the vertical structure and seasonal cycle (in terms of the meridional location) of the Atlantic ITCZ compared to the CMIP6 ensemble. In particular, our limited area simulation with explicit convection does not suffer from the double ITCZ problem. 2. An improved representation of the annual cycle of the TOA albedo compared to the CMIP6 ensemble, in particular in the trade-wind cumulus region where CTRL even outperforms the ERA5 reanalysis. This suggests that kilometer-resolution simulations are a suitable tool to study cloud feedbacks in the trade-wind region. Despite disabling the models shallow convection scheme, stratocumulus clouds are evident, albeit somewhat too frequent, and with an underestimated amplitude of the annual cycle. 3. The dynamics of the ITCZ in our kilometer-resolution simulation respond to warming in an overall similar way as in the analyzed GCMs. Yet, while the CMIP6 ensemble shows a clear narrowing of the ITCZ, (i.e., a prominent deep-tropics squeeze) and a strong reduction in the anvil cloud fraction, the kilometer-resolution simulation shows a weaker narrowing and a stronger central intensification of the ITCZ, as well as a local increase in the anvil cloud fraction.
Overall, our results demonstrate the merit of high-resolution climate simulations in a real-world configuration to compare against GCM projections. Kilometer-resolution models enable an unprecedented view on tropical clouds and circulations from the large-scale tropical overturning circulation down to small-scale convective MBL circulations and clouds. Even though global kilometer-resolution climate simulations are not yet feasible, our study demonstrates that downscaling strategies like the PGW approach allow to gain insights from these models already today. We presented one such simulation that largely agrees with the findings from the GCMs, but also produces remarkable differences in the climate-change response of the HC and the ITCZ. The realism of this response is difficult to assess as long as such simulations remain a rarity. We will analyze in more detail the cause of the response in upcoming work.

Data Availability Statement
The CERES EBAF TOA radiation data are available at https://ceres-tool.larc.nasa.gov/ord-tool/jsp/ EBAFTOA41Selection.jsp via https://doi.org/10.5067/TERRA-AQUA/CERES/EBAF-TOA_L3B004.1 (CERES EBAF, 2020). The CM SAF TOA radiation data are available at https://wui.cmsaf.eu/safira/action/viewProduk-tList?dId=3 via https://doi.org/10.5676/EUM_SAF_CM/TOA_GERB/V002 (CM SAF TOA Radiation, 2020). COSMO, CLM, and C2SM communities for developing and maintaining COSMO in climate mode, and the Federal Office for Meteorology and Climatology Mete-oSwiss, CSCS, and ETH Zürich for their contributions to the development of the GPU-accelerated version of COSMO. We acknowledge the World Climate Research Programme, which, through its Working Group on Coupled Modeling, coordinated and promoted CMIP6. We thank the climate modeling groups for producing and making available their model output, the Earth System Grid Federation (ESGF) for archiving the data and providing access, and the multiple funding agencies who support CMIP6 and ESGF. Open access funding provided by Eidgenossische Technische Hochschule Zurich.