The last two generations of climate system models developed at LASG-IAP are FGOALS2 and FGOALS3, including two parallel subversions in CMIP5 (FGOALS-g2 and FGOALS-s2) and three parallel subversions in CMIP6 (FGOALS-g3, FGOALS-f3-L and FGOALS-f3-H). A detailed description of the model configurations can be found in related literature (Bao et al., 2013; Li et al., 2013, 2020; Zhou and Song, 2014; Guo et al., 2020; Zhou et al., 2020). Only four subversions are analyzed in the present study, as outputs of FGOALS-f3-H are currently not available on the CMIP6 data portal. Each version of the FGOALS models configures a similar coupling framework including oceanic, sea-ice, and land components, with differences mainly in the atmospheric models. Detailed information on these components is provided in Table 1. Note that the FGOALS-g3 (FGOALS-f3-L) model is the updated version of the FGOALS-g2 (FGOALS-s2) model.
CMIP Model Component Ocean Sea ice Atmosphere Land CMIP5 FGOALS-g2 LICOM2 CICE4 GAMIL2 CLM3 360 × 196 128 × 60 30 levels 26 levels FGOALS-s2 LICOM2 CSIM5 SAMIL2 CLM3 360 × 196 128 × 108 30 levels 26 levels CMIP6 FGOALS-g3 LICOM3 CICE4 GAMIL3 CLM4 360 × 218 188 × 80 30 levels 26 levels FGOALS-f3-L LICOM3 CICE4 FAMIL CLM4 360 × 218 288 × 180 30 levels 32 levels
Table 1. Model components and corresponding horizontal resolutions of FGOALS models in CMIP5 and CMIP6. The components of all the models are the Finite-volume Atmospheric model (FAMIL), the Spectral Atmospheric Model of IAP/LASG (SAMIL), the LASG/IAP Climate system Ocean Model (LICOM), the Community Land Model (CLM), the Community Sea Ice Model (CSIM), and the Los Alamos sea ice model (CICE). The version number of the component models are labeled after the acronyms.
The equilibrium climate sensitivity (ECS), defined as the equilibrium temperature under doubled CO2 forcing, is 2.1–4.7 K in CMIP5 models and 1.8–5.6 K in CMIP6 models (Zelinka et al., 2020), which is mainly due to stronger positive cloud feedbacks from decreasing extratropical low-cloud coverage and albedo in models from CMIP6 than CMIP5. In contrast, the ECS is about 3.7 K in FGOALS-g2 and 4.5 K in FGOALS-s2, but decreases to 2.84°C for FGOALS-g3 and 2.98°C for FGOALS-f3 (Zhou et al., 2013, 2020), which might be associated with the differences in model biases and model-simulated internal variability, Arctic climate, and ocean circulation responses.
The ability of climate models in reproducing the climatology of observations, which is measured by the model bias (i.e., deviation from observation), is an essential metric for evaluating model performance. Figure 1 shows the spatial distribution of annual mean climatology biases in SST and 2 m air temperature (TAS) in FGOALS models for 1979–2005. The ERSST.v5 (Huang et al., 2017) and high-resolution (0.5° × 0.5°) CRU TS4.04 (Harris et al., 2020) data are referenced as observations to calculate the model biases. Generally, the FGOALS models display similar bias patterns in SST and TAS in both CMIPs, with large cold SST bias in the North Pacific, warm SST bias along the eastern coast of subtropical ocean basins and the Southern Ocean, and large cold TAS bias in Eurasia and North America. The maximal and minimal model biases in both SST and TAS also change insignificantly from CMIP5 to CMIP6. However, the global root-mean-square error (RMSE) and global mean value of SST biases are reduced in FGOALS from CMIP5 to CMIP6, with noticeable reduction in the North Pacific, south of Greenland, Southern Hemisphere (SH), eastern subtropical oceans, and Southern Ocean. For TAS bias, there is also a prominent decrease in its magnitude over western Eurasia, North America and Australia. Besides, the updated FGOALS models significantly reduce the TAS biases of their predecessors over regions along the Rocky and Andes Mountains and Himalayas, suggesting an improvement in resolving the effect of topography in CMIP6 models. As a result, the global RMSE and global mean value of TAS biases also decrease, especially for the latter. The evaluation of model biases in FGOALS models suggests that there are significant improvements and changes in the performances of FGOALS-g3 and FGOALS-f3-L from their predecessors.
Figure 1. Spatial distribution of annual mean biases in SST (units: °C) and 2 m air temperature (TAS; units: °C) in (a, e) FGOALS-g2, (b, f) FGOALS-g3, (c, g) FGOALS-s2, and (d, h) FGOALS-f3-L, during 1979–2005. The ERSST.v5 and CRU TS4.04 data are referenced as the observations. The maximum value (Max), minimum value (Min), RMSE, and mean value (Mean) of global biases are labeled at the top of each panel.
Monthly outputs (1850–2100) of historical simulations and low warming scenarios (RCP2.6 and SSP1-2.6) from the FGOALS models are analyzed. Besides, an additional 15 CMIP6 models (Table 2) are used to compare with their family predecessors in CMIP5. The CMIP5 and CMIP6 multi-model ensemble mean (MME) results are calculated based on the 15 pairs of models without the FGOALS models. The pre-industrial control runs are also used to remove the effect of climate drift in the models. Near-surface air temperature, surface temperature (skin temperature), precipitation, and zonal winds are used in this study. Note that in CMIP outputs, surface skin temperature is equivalent to SST in ice-free ocean. All atmospheric variables are linearly interpolated onto a common grid of 2° latitude × 2° longitude for ease of comparison. Only one member of each model is utilized for the analyses.
CMIP5 CMIP6 1 BCC_CSM1.1(m) BCC-CSM2-MR 2 CanESM2 CanESM5 3 CESM1-CAM5 CESM2 4 CNRM-CM5 CNRM-CM6-1 5 GFDL-CM3 GFDL-ESM4 6 GISS-E2-R GISS-E2-1-G 7 HadGEM2-ES HadGEM3-GC31-LL 8 IPSL-CM5A-LR IPSL-CM6A-LR 9 MIROC5 MIROC6 10 MIROC-ESM MIROC-ES2L 11 MPI-ESM-LR MPI-ESM1-2-LR 12 MPI-ESM-MR MPI-ESM1-2-HR 13 MRI-CGCM3 MRI-ESM2-0 14 NorESM1-M NorESM2-LM 15 NorESM1-ME NorESM2-MM
Table 2. CMIP models used in the present study.
As RF first increases to a peak around the year 2045 and then decreases, we separately calculate the linear trends during the RF increasing stage (1850–2050) and RF decreasing stage (2050–2100) to investigate the climate responses during these two distinct periods. The separation point for the trend calculation (2050) is chosen to lead lag the RF inflection point (2045) by only five years because the time scale of the fast response in the ocean mixed layer is 3–5 years (Held et al., 2010). The AMOC index is defined as the maximum value of the meridional stream function at 35°N in the Atlantic.
Figure 4 shows the linear trend of surface skin temperature (∆TS) under RCP2.6 and SSP1-2.6 during 1850–2050 and 2050–2100. When RF increases, surface warming is generally large over land and polar regions, especially the Arctic. The reduced surface warming or even surface cooling in the NA (i.e., the so-called NA warming hole) and Southern Ocean is robust in all four models, consistent with the MME results (Figs. 5a and b), despite large differences in the warming magnitude and detailed spatial structure, like the location of the NA warming hole. The El Niño-like warming pattern, Indian Ocean Dipole-like warming structure, and reduced subtropical warming in the SH are also prominent in all four models, consistent with the MME results from CMIP5 and CMIP6. These tropical warming patterns also exist in FGOALS-g2 but are not well displayed owing to the small magnitude of warming. Besides, the surface warming is locally enhanced over high mountain regions like the Tibetan Plateau and Rocky and Andes Mountains in all FGOALS models, illustrating the effect of topography in shaping the surface warming structure.
Figure 4. Linear trend of annual surface skin temperature (units: °C), which is equivalent to SST over ice-free ocean, during (a–d) 1850–2050 and (e–h) 2050–2100 in the four FGOALS models. The white (black) dots indicate the trend is insignificant (significant) at the 95% level. The global mean value is labeled in the title of each plot.
Figure 5. MME linear trends of annual surface temperature (units: °C) during (a, b) 1850–2050 and (c, d) 2050–2100 in (a, c) CMIP5 and (b, d) CMIP6, along with (e–h) their SNRs, defined as the absolute value of MME change divided by the intermodel standard deviation. The magenta and black contours indicate SNRs of 1 and 3, respectively.
During 2050–2100, corresponding to the significant decreasing trend of GMST (Figs. 2a and b), the surface cooling pattern is prominent in FGOALS-g2 and FGOALS-s2. However, the cooling magnitude differs across regions and is mainly large in the tropics in FGOALS-g2 and in the Northern Hemisphere (NH) mid and high latitudes in FGOALS-s2. Despite the GMST change being negligible during 2050–2100 in these two CMIP6 models, there is still a significant cooling trend in the tropics and warming trend in the NH mid and high latitudes in FGOALS-g3, but overall weak temperature change in FGOALS-f3-L. Besides, there is a broad significant increasing trend in the Southern Ocean in both FGOALS-g3 and FGOALSf-3-L, which is missing in the CMIP5 FGOALS models. This suggest that ∆TS may evolve prominently even under weak GMST change, but with the pattern differing significantly across models. It is worth noting that the ∆TS patterns in the FGOALS models are generally consistent with the MME results during the RF increasing stage, with the global pattern correlation coefficients all exceeding 0.79 (Table 3). However, during the RF decreasing stage, the pattern consistency with the MME results drops dramatically, ranging from −0.11 to 0.5 in the four models, suggesting that there is large model uncertainty in the further changes of surface temperature after 2050. In the CMIP5 and CMIP6 MMEs (Figs. 5c and d), the surface temperature mainly cools over land and warms over the SH oceans, but with the magnitude much reduced compared to that in the FGOALS models during 2050–2100 (Figs. 4e and f). Over the Tibetan Plateau, the surface cooling during the RF decreasing stage is also locally enhanced in most of the FGOALS models (Figs. 4e–g), and the CMIPs’ MMEs (Figs. 5c and d), which is similar to the situation during the RF increasing stage and suggests that the surface temperature over that region displays robust responses to RF changes.
CMIP MME and models Global Tropical oceans 1850–2050 2050–2100 1850–2050 2050–2100 CMIP5 MME FGOALS-g2 0.86 −0.11 0.73 −0.21 FGOALS-s2 0.93 0.38 0.37 0.48 CMIP6 MME FGOALS-g3 0.79 0.50 0.80 −0.04 FGOALS-f3-L 0.93 −0.02 0.59 −0.10
Table 3. Pattern correlations between FGOALS models and their corresponding MME results. The bold values indicate that the correlations are significant at 95% confidence level.
To evaluate the role of model uncertainty, which is measured by the intermodel standard deviation (SD), in future projections, we further calculate the rate of model consistency in the sign of MME change and signal-to-noise ratio (SNR) in each grid cell. The former is measured by the rate of models displaying change with the same sign of the MME results and is shown in Figs. 5a–d, while the latter is defined as the absolute value of MME change divided by the intermodel SD and is shown in Figs. 5e–h. During the RF increasing stage, there is high model consistency in the sign of MME change across the globe, with a model consistency rate below 2/3 only appearing over a very limited area in the NA and Southern Ocean (white dots in Figs. 5a and b). The SNR is generally larger than 3 (black contours in Figs. 5e and f) over most regions during 1850–2050, indicating a robust MME change relative to its intermodel spread, and is only lower than 1 over the NA and Southern Ocean (magenta contours). During 2050–2100, the model consistency largely reduces over most regions, with consistent sign of change among models (black dots) mainly appearing over the Pacific subtropics and land regions with large surface cooling (Figs. 5c and d). Correspondingly, the SNR is much smaller than that during 1850–2050, with a very limited area displaying values larger than 1. The low model consistency and small SNR during the RF decreasing stage suggest that the intermodel spread is much larger than the MME change, which is inevitably associated with the intermodel differences in simulating the natural variability. During 2050–2100 (50 years), as the RF change is weaker and the length of time for trend calculation is much shorter than those during 1850–2050 (200 years), the interference of internal variability in the trend calculation rises consequently. Besides, the warming effect from the deep ocean slow warming also largely offsets the cooling effect from the RF decrease, especially over regions with strong ocean dynamics (Long et al., 2020). All these factors complicate the projections of the surface temperature changes during the RF decreasing stage. As a result, there are large intermodel differences in the ∆TS pattern during 2050–2100, hence lowering the reliability of the MME changes. Indeed, during 1850–2050, the CMIP5 and CMIP6 MME ∆TS patterns are highly similar, suggesting robustness of the projected surface temperature response to the increase in RF. During 2050–2100, despite the GMST change being insignificant in both CMIP5 and CMIP6, the ∆TS pattern diverges substantially over the Arctic, East Asia, North America and NA Ocean.
Generally, surface temperature responses under low warming scenarios are distinct between the RF increasing and decreasing stages, and vary substantially across models, especially during the RF decreasing stage. Given that the pattern formation mechanisms for ∆TS display large variations in space (Xie et al., 2010), we further investigate the ∆TS pattern in the tropics and three other regions with noticeable local changes (the Arctic, NA subpolar region and Southern Ocean) in detail.
|360 × 196||128 × 60|
|30 levels||26 levels|
|360 × 196||128 × 108|
|30 levels||26 levels|
|360 × 218||188 × 80|
|30 levels||26 levels|
|360 × 218||288 × 180|
|30 levels||32 levels|