Assessing Global and Local Radiative Feedbacks Based on AGCM Simulations for 1980–2014/2017

We examine radiative feedbacks based on short‐term climate variability by analyzing atmospheric general circulation model (AGCM) simulations, including Atmospheric Model Intercomparison Project within CMIP phase 6 (AMIP6) with known effective radiative forcing (ERF) for 1980–2014 and one with zero ERF for 1980–2017. We first verify the Kernel‐Gregory feedback calculation by showing that both clear‐sky radiative fluxes and all‐sky radiative feedbacks from the kernel method agree with model simulations. We find that global‐mean net feedback for 1980–2017/2014 is −2 W m−2 K−1, about twice the feedback estimated for long‐term warming (4 × CO2) experiments. This difference is mainly caused by a near‐zero global‐mean net cloud feedback for 1980–2017/2014. We show that the lapse rate feedback for 1980–2017/2014 is the largest contributor to the amplified temperature change over the three poles (Arctic, Antarctic, and Tibetan Plateau), followed by surface albedo feedback and Planck feedback deviation from its global mean. Except for a higher surface albedo feedback in Antarctic, other feedbacks are similar between Arctic and Antarctic.


Introduction
The long term global-mean surface air temperature (SAT) response to CO 2 doubling, that is, equilibrium climate sensitivity (ECS), has been a long-standing and fundamental research topic (Charney et al., 1979;Intergovernmental Panel on Climate Change, 2013). A classical forcing-feedback framework (Hansen et al., 1997) is commonly used to explore this topic using the formula ΔR = ΔQ + λΔT s , where ΔR is net radiation imbalance at the top of the atmosphere (TOA), ΔQ is the effective radiative forcing (ERF), and ΔT s is the SAT change. λ is the radiative feedback parameter, including Planck response and feedbacks associated with lapse rate, water vapor, surface albedo, and clouds (e.g., Bony et al., 2006). Gregory et al. (2004) developed a simple yet practical method for estimating global mean λ from general circulation model (GCM) simulations in response to a constant forcing (e.g., 2 × CO 2 ).
Horizontal and vertical warming features, such as the tropospheric warming and polar amplification, have long been studied using both GCM simulations (e.g., Manabe & Wetherald, 1975;Po-Chedley et al., 2018) and observations (e.g., Fu et al., 2004Fu et al., , 2011Fu & Johanson, 2005;Santer et al., 2005Santer et al., , 2018Zhang et al., 2019). The forcing-feedback framework can also be applied at regional scales to provide local perspectives on climate feedbacks (Armour et al., 2013;Feldl & Roe, 2013). For example, Pithan and Mauritsen (2014) used this formalism to quantify feedbacks and identify the dominant factors in polar amplification based on the Coupled Model Intercomparison Project phase 5 (CMIP5) simulations (Taylor et al., 2012). Attempts have also been made to estimate climate feedbacks using observed short-term climate variations during the recent past decades (e.g., Dessler & Forster, 2018;Kolly & Huang, 2018;Kramer & Soden, 2016;Trenberth et al., 2015;Zhang et al., 2018). This provides an alternative method to the traditional method of estimating feedbacks based on long-term coupled atmosphereocean GCM (AOGCM) simulations (e.g., Caldwell et al., 2016), especially for determining cloud feedbacks (Dessler, 2010;Yue et al., 2019;Zhou et al., 2016), which are by far the largest source of uncertainty in ECS quantification (Ceppi et al., 2017). Zhang et al. (2018) quantified individual local feedbacks over the Arctic for the first time by examining short-term climate variations based on reanalysis products, satellite observations, and atmospheric GCM (AGCM) simulations, using the combined Kernel-Gregory method. Zhang et al. (2018) noted that the quantity ERF (ΔQ) is not observable, and its impact on the diagnosis of radiative feedback parameter has not been examined. Webb et al. (2012), Forster (2016, Skeie et al. (2018), and Gregory et al. (2019) also noted that the ERF term is an important source of uncertainty in estimating climate feedbacks, especially when using present-day climate data. With little knowledge on the ERF, there were two questions that were not addressed by Zhang et al. (2018): (1) What is the difference between model simulated clear-sky ΔR and the reconstructed clear-sky ΔR using the radiative kernel method? (2) What is the difference between the simulated total all-sky radiative feedback and the sum of derived individual all-sky radiative feedbacks using the combined Kernel-Gregory method? More importantly, incomplete knowledge of the ERF would produce uncertainty in quantifying the feedbacks based on short-term climate variability (Zhang et al., 2018).
Here we have performed an AGCM simulation using observed sea surface temperatures (SSTs) for 1980-2017 with prescribed annually repeating forcing agent (thus the ERF is zero). We have also analyzed a set of newly available multiple AGCM simulations for a similar period  with known ERFs. The known ERF avoids the impact of the ERF uncertainty including in spatiotemporally varying ERF on the feedback estimate. We first assess the consistency between model simulations and independent kernel derivations for both clear-sky radiative fluxes and all-sky radiative feedbacks. We then investigate the radiative feedbacks on a global scale and regionally for the three poles: the Arctic (60-90°N), the Antarctic (60-90°S), and the Himalayan-Tibetan Plateau (HTP, 25-45°N and 65-105°E) based on our AGCM simulation and Atmospheric Model Intercomparison Project within CMIP phase 6 (AMIP6) multimodel data (Eyring et al., 2016). The three poles are all cold cryospheric regions on Earth, and all are highly sensitive to climate change (e.g., Box et al., 2019;Post et al., 2019;Yao et al., 2018). With known external forcing and verified methodology, we have higher confidence on the derived feedback based on short-term climate variability since 1980.

Model Experiments, AMIP6 Data Sets, and Methodology
We use the Community Atmosphere Model, version 5.3 (CAM5.3) at 0.9°× 1.25°horizontal resolution with a finite volume dynamical core and 30 vertical levels (Neale et al., 2010) to perform an AGCM simulation for 39 years . The model was configured to include full atmosphere-land coupling, but time-varying monthly SSTs and sea ice concentrations (SICs) were prescribed based on observations (Hurrell et al., 2008). In contrast to the standard AMIP model configuration, all external forcing agents (i.e., the greenhouse gases, aerosols, ozone, volcanic eruption, solar radiation change, and land use) were prescribed at the year-2000 values to eliminate the ERF term in the feedback calculation. This CAM5 experiment is hereafter referred to as CAM5FF. In addition to the CAM5FF simulation, we also use simulations for 1980-2014 from 13 AMIP6 (r1i1p1f1) climate models (Table S1 in the supporting information), which are different from CAM5FF by utilizing time-varying external forcings. These AMIP6 model outputs are regridded to have the same horizontal resolution (0.9°× 1.25°) as in CAM5FF.
We use the combined Kernel-Gregory method (Gregory et al., 2004;Kramer, Soden, & Pendergrass, 2019;Soden & Held, 2006) that was described in detail in Zhang et al. (2018) for radiative feedback calculations. We first compute monthly anomalies (denoted by Δ) with respect to monthly mean climatology over the appropriate historical time period (1980-2017for CAM5FF and 1980-2014 and then calculate the radiative feedbacks by regressing the corresponding component of ΔR onto ΔT s over the entire simulation span. The schematic diagram of this method is illustrated in Figure 1. The historical ERF for AMIP6 models can be derived from a pair of experiments, piClim-histall and piClim-control, as specified in the Radiative Forcing Model Intercomparison Project (RFMIP, Pincus et al., 2016). Both experiments are AGCM simulations with the prescribed preindustrial monthly SST and SIC climatology. The piClim-control is a time-slice simulation (30 years) with constant preindustrial forcing agents, whereas piClim-histall is a transient simulation (1850-2100) with time-varying external forcings. Thus, the time-varying monthly ERF at each grid can be calculated as the difference in TOA radiation between piClim-histall and monthly climatology of the 30-year piClim-control simulations (Gregory et al., 2019). Unlike the calculation of total and cloud feedbacks in CAM5FF shown in Figure 1, the total feedback in AMIP6 is performed as the regres- where the superscript 0 represents clear-sky conditions. The difference in these feedback calculations between CAM5FF and AMIP6 is that CAM5FF does not consider the effect of ERF because the forcing agents in CAM5FF are prescribed at the year-2000 values. Other feedbacks (i.e., Planck, lapse rate, water vapor, and surface albedo feedback) for the AMIP6 simulations are derived using the same method as that for the CAM5FF. We use the monthly radiative kernels at the TOA  that are derived from the Community Earth System Model Large Ensemble (CESM-LE) simulations (Kay et al., 2015). The radiative kernels represent the response of TOA radiation to a 1% increase in surface albedo (albedo kernel), a 1-K warming at each vertical level including the surface (temperature kernel), and an increase in atmospheric specific humidity caused by 1-K warming with constant relative humidity (water vapor kernel) Soden et al., 2008).

Consistency of the Clear-Sky Radiative Fluxes at the TOA
Given the linear decomposition of the total feedback (i.e., λ ≈ ∑ λ x ), it is important to test the closure of radiative fluxes when we use the radiative kernel method . The appropriateness of the kernel method can be verified by the clear-sky linearity test (Soden et al., 2008). In the present study, we compare the clear-sky TOA flux anomalies extracted directly from the model output (ΔR 0 ) against the sum of individual clear-sky radiation terms (ΔR 0 x ) associated with temperature (T), water vapor (q), and surface albedo (α). The sum of ΔR 0 x can be calculated using the clear-sky radiative kernels (K 0 x ) and the corresponding change (Δx) with the formula ∑ K 0 x Δx, where x = T, q, and α. Figures 2a1 and 2a2 show the time series of global mean clear-sky radiative flux anomalies based on the CAM5FF monthly output for shortwave (SW) and longwave (LW) radiation, respectively. Here positive fluxes are defined as downwelling for both SW and LW. It is evident that the clear-sky fluxes derived from radiative kernels (red lines) agree well with the true values simulated by the CAM5FF (blue lines) for both SW and LW. Thus, our linearity test suggests that the net radiative fluxes at the TOA (NET = SW + LW) are internally consistent with clear-sky estimates. We refer to this feature of the forcing/feedback estimates by saying the estimates are "closed" or satisfy a "closure test." Note that the trends of time series of SW and LW fluxes, especially after 2010, are opposite.
We further examine the spatial patterns between CAM5FF and kernel-derived clear-sky radiative flux anomalies at the TOA. Figures 2b1-2b3 and 2b4-2b6 show the closure test using the annual mean of 2017 as an example for SW and LW radiation, respectively. The patterns of clear-sky anomalies from model output (Figures 2b1 and 2b4) are very similar to the patterns calculated using the kernel method (Figures 2b2 and 2b5). We can see that the clear-sky flux anomalies are mostly positive for SW and negative for LW radiation, which have been indicated by the trends in Figures 2a1 and 2a2. Nonclosure is measured by the difference between kernel-derived and CAM5FF fluxes in Figures 2b3 and 2b6. Most of the SW disagreements occur at high latitudes where the surface albedo feedback is important and somewhat more nonlinear, while kernels assume linearity . On global average, the relative difference of SW (0.3%) is smaller than that of LW (6.7%). Thus, the relative difference between kernel-derived and CAM5FF clear-sky net fluxes, 3.9%, is considerably less than the reference value (15%) used in Caldwell et al. (2016). Our results also show that the seasonal patterns of simulated clear-sky TOA fluxes can be reconstructed by radiative kernels as well (see Figure S1). After verifying the clear-sky radiation closure, we next perform the radiative feedback closure test.

Closure of All-Sky Radiative Feedbacks at the TOA
Using the local radiative feedback framework described in Figure 1 (see also Zhang et al., 2018), zonal-mean feedbacks are calculated through the regression of zonal-mean radiative flux anomalies onto zonal-mean SAT anomalies, as shown in Figure 3. We use the traditional feedback decomposition (Zelinka et al., 2020) in which the temperature feedback is calculated by holding specific humidity fixed and the water vapor feedback is quantified using the TOA radiation change due only to temperature-mediated changes in specific humidity. The Planck feedback represents the divergence of outgoing LW radiation as the Earth warms, assuming a vertically uniform temperature change with the surface warming. It is negative everywhere and its magnitude has a slight decrease from the tropics to the high latitudes owing to warmer temperatures in the tropics (Feldl & Roe, 2013). The lapse rate feedback corresponds to the vertically nonuniform temperature departures from surface temperature change. The lapse rate feedback is negative in the tropics where the temperature profiles follow a moist adiabat but positive in the high latitudes as a result of temperature inversions there (Bony et al., 2006). Note that there is a steep slope in the lapse rate feedback over the Southern Ocean where the CMIP5 models have a pronounced spread in lapse rate and water vapor feedbacks . It's not surprising that the water vapor feedback is positive over all latitudes but stronger in the tropics where water vapor changes are larger than in middle and high latitudes. A stronger positive water vapor feedback is corresponding to a more negative lapse rate feedback, reflecting that the water vapor is strongly related to temperature (Held & Shell, 2012). The sum of lapse rate and water vapor feedback (see Figure S2) is positive everywhere and is strongest at the equator where the water vapor concentrations are the largest (Bony et al., 2006). The surface albedo feedback is largely determined by the distribution of cryospheric regions and thus is locally focused on the high latitudes. The peak in the surface albedo feedback across 30-40°N is likely associated with the third pole (i.e., HTP; see also Figure 2b2). The cloud feedback, which has the largest intermodel spread in CMIP5 and CMIP6 (Ceppi et al., 2017;Zelinka et al., 2020), is more complex than other radiative feedbacks. The zonal-mean pattern of SW cloud feedback is partially opposite to its LW component. The net cloud feedback is mostly positive in the tropics (except deep tropics) and slightly negative in the high latitudes from CAM5FF results, which is consistent with the cloud fraction decrease in the tropics and low clouds increase in the high latitudes under a warmer climate (Zelinka et al., 2012b; see also Figure S3). But it is not necessarily expected to see the same feedbacks operating during the AMIP period as in an equilibrium warmer climate due to disequilibrium and internal , SAT (T s ), temperature profile (T), water vapor profile (q), and surface albedo (α), are obtained by subtracting monthly climatological mean (1980-2017for CAM5FF and 1980-2014. A 12-month running mean is applied to smooth these anomalies. Finally, the radiative flux anomalies are regressed onto the SAT anomalies to obtain feedbacks (λ, λ x and λ c ) using the linear least squares (LLS) fit.

Geophysical Research Letters
variability (e.g., Armour, 2017;Marvel et al., 2018). The net cloud feedback from CAM5FF is negative in the midlatitudes of the Southern Hemisphere.
Based on above individual feedback diagnosis, we obtain two estimates of the total feedback: one is the linear regression coefficient (λ) by regressing CAM5FF ΔR onto the SAT anomalies ΔT s , and the other one is the sum of individual kernel-derived feedbacks (∑λ x ). The two agree well for both SW and LW. The residual (i.e., ε = λ − ∑ λ x ) is small over all latitudes (black lines in Figure 3) representing the contribution from higher-order Taylor expansion terms (i.e., λ¼ (Bony et al., 2006). By regressing the global average ΔR onto global average ΔT s , we can get the global-mean individual feedbacks and total feedback. On global average, the absolute SW residual feedback (0.123 W m −2 K −1 ) is larger than that of LW (0.025 W m −2 K −1 ). The relative difference of net total feedback between λ and ∑λ x is only 4.8% over the AMIP period. So far, the kernel method and linear decomposition of the total feedback have been confirmed by clear-sky radiative fluxes closure and all-sky radiative feedbacks closure tests. These tests along with the known ERF in both CAM5FF and AMIP6 models greatly enhance our confidence in estimating individual radiative feedbacks over the globe and three poles based on short-term climate variability.

Global and Local Radiative Feedbacks Over the Three Poles
For the global-scale analysis, we compare the global-mean radiative feedbacks in CAM5FF with those obtained from other data sets, including CMIP5, CMIP6, CESM-LE, and AMIP6 simulations. Twenty-eight CMIP5 long-term AOGCM simulations first analyzed by Caldwell et al. (2016) and then updated by Zelinka et al. (2020) were used to be consistent with the analysis of 27 CMIP6 models (see Table S1 and S2 in Zelinka et al. (2020) for details). We calculate radiative feedbacks in CESM-LE following Pendergrass et al. (2018) (using updated data sets available at https://github.com/ apendergrass/cam5-kernels). Thirteen AMIP6 short-term AGCM simulations are analyzed in this study, among which four models are also available to diagnose the ERF term and the subsequent total and cloud feedbacks (see section 2 and Table S1). The global individual feedbacks for CMIP5, CMIP6, and CESM-LE are also estimated using the radiative kernel method and the adjustment method for cloud feedback calculations (Soden et al., 2008). Zelinka et al. (2020) used Huang et al. (2017 radiative kernels and regressed the global-and annual-average of ΔR onto the corresponding ΔT s to get λ from

Geophysical Research Letters
the difference between CMIP5&6 abrupt quadrupling of CO 2 experiments (abrupt-4×CO 2 ) and preindustrial control (piControl) simulations. In Pendergrass et al. (2018), the climate change signal Δ was calculated as the difference of the 30-year average from each ensemble member between 2071-2100and 1976(Kay et al., 2015 and then λ equals to the ratio of ΔR to ΔT s . For temperature and water vapor feedbacks, both studies (CMIP5&6 and CESM-LE) only considered tropospheric levels. To be consistent, the calculation of global temperature and water vapor feedback in CAM5FF and AMIP6 (shown in Figure 4a) is vertically integrated from the surface to the tropopause, defined as 100 hPa at the equator and decreasing with cosine of latitude to 300 hPa at the poles following Pendergrass et al. (2018). Note that the absolute net residual feedback is slightly increased from 0.098 (see Figure 3) to 0.128 W m −2 K −1 (see Table S2) in CAM5FF results when integrating from the surface to tropopause rather than to the top model layer. Figure 4a shows that the residual term of CAM5FF is, however, still smaller than most CMIP5&6 models.
Overall, the net total feedback agrees well between CESM-LE and CMIP5&6 ensemble means and between CAM5FF and AMIP6. This is also true for individual feedbacks except for the cloud SW and LW feedbacks. The net total feedbacks from CAM5FF and AMIP6, however, are significantly smaller than those from CMIP5&6 and CESM-LE, which is largely caused by the differences in cloud feedbacks between CAM5FF/AMIP6 and CMIP5&6/CESM-LE. The consistency of Planck feedback (−3.2 W m −2 K −1 ) across model simulations is expected, and the robustness of Planck feedback has been noted in different feedback decompositions (Vial et al., 2013). While the lapse rate feedback in CAM5FF and AMIP6 is slightly lower than that in CMIP5&6 and CESM-LE, the water vapor feedback is comparable. Thus, the sum of lapse rate and water vapor feedbacks in CAM5FF and AMIP6 is smaller than that in CMIP5&6 and CESM-LE, which also contributes to the stronger negative total feedback in CAM5FF and AMIP6 than CMIP5&6 and CESM-LE. The surface albedo feedback in CAM5FF and AMIP6 is in close agreement to the CMIP5&6 ensemble mean but with a slightly smaller value. While the net cloud feedback is comparable between CMIP5&6 ensemble mean and CESM-LE, the disagreement in both SW and LW cloud feedback is significant. It is interesting to note that the slightly positive SW and somewhat negative LW cloud feedbacks in CAM5FF agree closely with that of CMIP6 ensemble mean for SW and CESM-LE for LW. Among the four AMIP6 models that are available for cloud feedback calculation, the estimated positive mean LW cloud feedback in AMIP6 is comparable to CMIP5&6 ensemble mean. However, the negative mean SW cloud feedback results in a near-zero net cloud feedback (−0.02 W m −2 K −1 ) in AMIP6. Zhou et al. (2016) pointed out that the decadal cloud feedback over the AMIP period resulting from the decadal cloud variations is smaller than the long-term ones. Overall, a stronger negative net total feedback (about −2 W m −2 K −1 ) based on short-term CAM5FF and AMIP6 simulations than that from long-term CMIP5&6 and CESM-LE simulations (about −1 W m −2 K −1 ) is primarily due to a near-zero net cloud feedback compared to a positive net cloud feedback associated with a long-term warming.
CAM5FF and AMIP6 can also be used to investigate the relative contribution of individual feedbacks to local temperature amplification over the three poles relative to the tropics (see Figure S4 for the SAT trends over four regions). The local energy budget equation can be written as ΔQ + λΔT s + ΔH = 0, where ΔH is the change of horizontal heat flux convergence to local system. By decomposing the feedback term λ following Pithan and Mauritsen (2014) λ P and λ P ′ are the global mean Planck feedback and local Planck feedback deviation from its global mean, respectively. Thus, we can assess the contribution of individual feedbacks to local temperature change by assuming λ P as the reference response. Figures 4b-4d show the scatter plot of individual feedbacks between the three poles, respectively, and the tropics (30°S to 30°N). Overall, the residual feedback in CAM5FF is smaller than that in AMIP6 over all four regions. Relative to the tropics, the lapse rate feedback, surface albedo feedback, and Planck feedback deviation in both CAM5FF and AMIP6 are positive and contribute to temperature amplification over the three poles differently, as measured by the distance from each symbol to the 1:1 line. The positive water vapor feedback is larger over the HTP than that over the Arctic and Antarctic. However, the sum of lapse rate and water vapor feedbacks is comparable (close to the 1:1 line) across four regions, especially between HTP and tropics in Figure 4d. The cloud feedback is positive over the HTP but negative over the Arctic and Antarctic in CAM5FF. The negative cloud feedback over the Antarctic is consistent with the CMIP5 ensemble mean results presented in Goosse et al. (2018), but 10.1029/2020GL088063

Geophysical Research Letters
CMIP5 results show a slightly positive cloud feedback over the Arctic (Pithan & Mauritsen, 2014), which is consistent with AMIP6 ensemble mean results. We further contrast the feedbacks among the three poles in Figures 4e-4g. Except for the surface albedo feedback, other local feedbacks are almost symmetrical between the Arctic and the Antarctic. Relative to both north and south poles, the water vapor and cloud feedbacks have comparable contributions to temperature amplification over the HTP in CAM5FF. Overall, except for the sign of cloud feedback, the AMIP6 ensemble mean (diamonds) is similar to the CAM5FF estimates (pluses) for the Planck feedback deviation, lapse rate, water vapor, and surface albedo feedback over the three poles.  Table S2.
(b-g) Scatter plot of individual feedbacks over the Arctic, the Antarctic, the HTP, and the tropics. Detailed values are summarized in Tables S3 and S4. The legends of symbols and colors are shown on the right-hand side.

Conclusions and Discussions
This study extends the work of Zhang et al. (2018) by evaluating TOA radiative fluxes and feedback decomposition using model simulations and kernel-based estimates. The decomposition is verified using a short-term CAM5 atmosphere-only simulation where ERF is constrained to be 0 (CAM5FF). The results show that the CESM-CAM5 radiative kernels from Pendergrass et al. (2018) can reconstruct the global pattern and seasonal cycle of TOA clear-sky radiation budget within an uncertainty of 10% or less over the AMIP period. An all-sky radiative feedback closure test shows that the combined Kernel-Gregory approach used in Zhang et al. (2018) can properly decompose the global and local total feedback into individual feedbacks (i.e., Planck, lapse rate, water vapor, surface albedo, and cloud feedbacks) with a residual of 5% (relative difference) for the global mean. We further compare the global radiative feedbacks from CAM5FF with other data sets (CMIP5, CMIP6, CESM-LE, and AMIP6) and then quantify the relative contribution of individual feedbacks to local temperature changes over the three poles (i.e., the Arctic, the Antarctic, and the HTP) where temperature changes are large. On the global scale, results of CAM5FF and AMIP6 show a nearly doubled negative net feedback reported in the long-term simulations (CMIP5&6 and CESM-LE), which is contributed by a slightly lower lapse rate and surface albedo feedbacks and mostly a near-zero cloud feedback during the AMIP period. The near-zero net cloud feedback in CAM5FF, 0.03 W m −2 K −1 , can be decomposed into a slightly positive SW cloud feedback (0.12 W m −2 K −1 ) and a negative LW cloud feedback (−0.09 W m −2 K −1 ). The AMIP6 ensemble mean also shows a near-zero global-mean net cloud feedback (−0.02 W m −2 K −1 ), resulting from the comparable negative SW (−0.31 W m −2 K −1 ) and positive LW (0.29 W m −2 K −1 ) components. While the net cloud feedback in CAM5FF and AMIP6 is far less than that of CMIP5&6 and CESM-LE (0.34-0.52 W m −2 K −1 ), the CAM5FF and CMIP6 SW cloud feedbacks are consistent while CAM5FF and CESM-LE LW cloud feedbacks are consistent. In contrast to a positive cloud feedback in CAM5FF over the HTP, the cloud feedback is slightly negative over the Arctic and the Antarctic. However, the cloud feedback in AMIP6 ensemble mean is slightly positive over the Arctic. Relative to the tropics (30°S to 30°N), the lapse rate feedback is the largest contributor among all feedbacks to temperature amplification over the three poles, followed by surface albedo feedback and Planck feedback deviation from its global mean. Interestingly, except for the surface albedo feedback, other feedbacks are almost the same between the Arctic and the Antarctic.
Due to the absence of a cloud kernel in CESM-CAM5 radiative kernels, we cannot directly verify the closure of all-sky TOA radiative fluxes between the true values simulated in the CAM5FF and kernel reconstruction. More specifically, the cloud feedback in this study is not directly calculated using the cloud kernel but instead uses the adjustment method (Soden et al., 2008) that requires TOA radiative fluxes from GCM simulations (see Figure 1). Observation-based cloud radiative kernels (e.g., Yue et al., 2016Yue et al., , 2019 might help to verify the closure of all-sky TOA radiative fluxes and further quantify cloud feedback difference between the direct cloud kernel method (e.g., Zelinka et al., 2012a) and the adjustment method. In this study, CESM-CAM5 radiative kernels are used in feedback calculations for an atmosphere-only simulation conducted with the same model (CAM5.3). Thus, it would be interesting to know the impact of different kernels, such as the kernels derived from other GCMs (Soden et al., 2008), reanalysis (Huang et al., 2017), and satellite observations (Kramer, Matus, et al., 2019), on the feedback calculation. Furthermore, our CAM5FF and AMIP6 results give a large negative global net radiative feedback, about −2 W m −2 K −1 , mostly due to the near-zero cloud feedback. This can be related to the shorter time scale or weaker warming over the recent AMIP period in the AGCM simulations than in the long-term strongly forced AOGCM simulations (e.g., CMIP5&6 abrupt-4×CO 2 ). Some studies have found a temporal variation of feedbacks and attributed it to the dependence of feedbacks on the spatial patterns of surface warming that evolve over time (e.g., Andrews et al., 2015;Armour, 2017;Dong et al., 2019;Marvel et al., 2018;Zhou et al., 2016).
In summary, we have developed an explicit kernel-regression method to calculate radiative feedbacks in short-term AGCM simulations with prescribed external forcings. This practical method has been verified using a closure test of both the clear-sky TOA radiation budget and all-sky radiative feedbacks. It is also the first time that AMIP simulations drawn from the CMIP6 activity are used to simultaneously investigate the relative contribution of individual feedbacks to the temperature change over the three poles showing very large environmental changes.