CMIP6 Earth System Models Project Greater Acceleration of Climate Zone Change Due To Stronger Warming Rates

Warming climate and precipitation changes induce notable shifts in climate zones. In this study, the latest generation of global climate models from the Coupled Model Intercomparison Project Phase 6 (CMIP6) and the previous generation CMIP5 under high‐emission scenarios are used together with observations and applied to the Köppen‐Geiger climate classification. The aim is to shed light on how projected warming and changes in precipitation will influence future climate zones and their associated ecosystems while revealing differences between the two model generations. Compared to CMIP5 models, CMIP6 models exhibit slightly improved performance in replicating the observed Köppen‐Geiger map for the historical (1976–2005) period and similar inter‐model agreement for the future. The models show major changes in climate zones with a range of projections depending on which ensemble subset is used: 37.9%–48.1% of the global land area is projected to change climate zone by the end of the century, with the most pronounced changes expected over Europe (71.4%–88.6%) and North America (51.2%–65.8%). CMIP6 models project a higher rate of areal climate zone change (km2/year) throughout the 21st century, which is mainly driven by their greater global land warming rates. Using a likely equilibrium climate sensitivity subset of CMIP6 models that is consistent with the latest evidence constrains the climate zone shifts, and their projections better match the results of CMIP5 simulations. Although the high warming rates of some CMIP6 models are less credible, the risks associated with them are greater, and they heighten the need for urgent action to preserve terrestrial ecosystems.

Climate classification schemes are simple but effective tools for attributing the impacts of climate change on ecosystems by combining the states of temperature and precipitation into one quantity (Feng et al., 2014). The Köppen climate classification was the first quantitative classification that empirically described the biome distribution of the Earth (Cui et al., 2021;Köppen, 1936). With modifications (Trewartha & Horn, 1980) and updates (Kottek et al., 2006;Peel et al., 2007), the Köppen climate classification has been used in numerous studies to show observed changes in the climate (C. Beck et al., 2005;Chan & Wu, 2015;D. Chen & Chen, 2013) and expected future changes based on the outputs of global climate model (GCM) projections (H. E. Beck et al., 2018;Feng et al., 2014;Mahlstein et al., 2013;Navarro et al., 2022).
Several of the latest generation of GCMs from Phase 6 of the Coupled Model Intercomparison Project (CMIP6; Eyring et al., 2016) show higher equilibrium climate sensitivities (ECS; representing the expected global warming following a doubling of atmospheric CO 2 compared to pre-industrial levels), than their predecessors in CMIP5 (Andrews et al., 2019;Forster et al., 2020;Gettelman et al., 2019;Golaz et al., 2019;Swart et al., 2019;Voldoire et al., 2019). The increase in the ECS in CMIP6 models is primarily attributed to positive cloud feedbacks and enhanced aerosol-cloud interactions (Meehl et al., 2020;Zelinka et al., 2020). The likely range of ECS (66%-100% probability) has been estimated at 1.5°C-4.5°C in the IPCC AR5 report (IPCC, 2013). However, the range has narrowed to 2.5°C-4°C in the latest AR6 report (IPCC, 2021), and the 5%-95% range is now estimated at 2.3°C-4.7°C (Sherwood et al., 2020). Some CMIP6 models with high ECS values have been criticized for simulating too strong warming that is unsupported by historical evidence (Tokarska et al., 2020;Zhu et al., 2020). Therefore, different approaches have been implemented to account for the so-called "hot model problem" (Hausfather et al., 2022), such as constraining the projections based on historical observations (Ribes et al., 2021;Tokarska et al., 2020) or weighting the models based on their performance and independence (Brunner et al., 2020;Liang et al., 2020). Although some models in CMIP6 may be less plausible due to their high ECS values outside the likely range, state-dependent cloud feedbacks could lead the climate system to a higher sensitivity state, making those higher ECS values plausible (Bjordal et al., 2020). Therefore, it is essential to evaluate the consequences of both the likely and less plausible but high-risk outcomes based on CMIP6 future climate projections, and to quantify their differences from expectations based on the previous generation of models in CMIP5.
In this study, we use observed global temperature and precipitation data along with CMIP5 and CMIP6 projections and apply the Köppen-Geiger climate classification (Peel et al., 2007) to analyze both observed and projected changes in climate zones from the beginning of the 20th century until the end of the 21st century and to highlight differences between the two model generations. We use historical CMIP5 and CMIP6 data to evaluate and compare their ability to replicate observed Köppen-Geiger climate classification maps (Gnanadesikan & Stouffer, 2006). We also use high-emission scenario future projections, the Shared Socioeconomic Pathway 5-8.5 (SSP5-8.5; Riahi et al., 2017) for CMIP6 and the Representative Concentration Pathway 8.5 (RCP8.5;van Vuuren et al., 2011) for CMIP5, to compare projections of future climate zone distributions both globally and at the continental level. We split the model generations into two subsets to address some CMIP6 models' tendency toward excessive warming for future climate state projections: the first subset contains models with better historical performance (without making any particular choice based on the ECS of the models) and the second subset contains the models that are consistent with the declared likely ECS range of 2.5°C-4°C (Hausfather et al., 2022). Finally, we look at the differences between their future rate of climate zone change projections and investigate the models' global mean land warming rates as the principal reason for the difference between the two model projections.

Observations and Reanalysis
This study uses two gridded observational data sets as the basis for the calculations: the Climatic Research Unit gridded Time Series v. 4.04 (CRU) monthly temperature data set (Harris et al., 2020) and the Global Precipitation Climatology Center (GPCC) monthly precipitation data set (Schneider et al., 2014), both of which are at 0.5° grid resolution. These two data sets, spanning 1901 to 2019 (GPCC extends back to 1891), are used to produce the past and present-day Köppen-Geiger climate classification maps for the 30-year periods 1901-1930 and 10.1029/2022EF002972 3 of 23 1990-2019, to evaluate GCM performance for the historical period 1976-2005 and to correct the mean biases of the GCMs. Throughout the study, Antarctica is excluded from the calculations due to the lack of sufficient observational data.
Analyses presented in the Results section use CRU temperature and GPCC precipitation as observational reference data sets. However, to investigate the uncertainty that might exist in these reference data sets and to measure the sensitivity of GCM performance to the reference data set, we have performed additional analyses utilizing other data sets as benchmarks, while their results are given in Figures S3-S9 in Supporting Information S1. Two additional temperature and precipitation data set couples are used. The first couple consists of the Global Historical Climatology Network version 2 and Climate Anomaly Monitoring System (GHCN-CAMS) monthly gridded temperature data set, covering the period from 1948 to 2019 (Fan & van den Dool, 2008), and the CRU monthly precipitation from 1901 to 2019 (Harris et al., 2020) data set (both at 0.5° resolution). The second couple is the ERA5 global reanalysis temperature and precipitation data (Hersbach et al., 2020) between 1959 and 2019 at 0.25° resolution. The ERA5 data have been regridded onto the common observational data grid resolution of 0.5° using bilinear interpolation.
Overall, the CRU temperature data set has good agreement with other observational and reanalysis data sets, such as CRUTEM4.6, UDEL, and JRA-55, with some differences in long-term trends (Harris et al., 2020). Even though the GPCC precipitation data set has large differences compared to the reanalysis data (e.g., ERA-40 and ERA-Interim;Schneider et al., 2014), it has very good agreement with the GPCP V2.2 precipitation data. Accordingly, in this study CRU and GPCC data sets are selected as the main observational reference for temperature and precipitation, respectively.

CMIP5 and CMIP6 GCMs
We use historical and future scenario simulations of monthly temperature and precipitation from 35 CMIP5 models (Taylor et al., 2012) and 33 CMIP6 models (Eyring et al., 2016); a complete list of the models can be found in Tables 1 and 2, respectively. Historical simulations span 1850 to 2005 (1850-2014) for CMIP5 (CMIP6), while future simulations cover the periods 2006-2100 (2015-2100) for CMIP5 (CMIP6). In this study, simulations from 1901 to 2100 have been used in the analyses. Only one variant (mostly r1i1p1 for CMIP5 and r1i1p1f1 for CMIP6) is considered for each model, as is the case for many previous studies (e.g., H. E. Beck et al., 2018;di Virgilio et al., 2022;Grose et al., 2020). All climate models have been downscaled onto the common grid of 0.5° resolution using bilinear interpolation prior to analysis. Historical simulations of the models have been used for evaluation against observations and for bias correction, while scenario simulations have been used to create the future Köppen-Geiger climate classification based analyses. For both sets of GCMs, high-emission scenarios have been used: the RCP8.5 (van Vuuren et al., 2011) for CMIP5 and the SSP5-8.5 (Riahi et al., 2017) for CMIP6. The SSP scenarios cover multiple baseline worlds facing different socioeconomic development challenges in the absence of new climate policies along with the climate forcing aspect of the RCPs (Riahi et al., 2017). Although both scenarios reach 8.5 W/m 2 radiative forcing by the end of the 21st century, disparities in their greenhouse gas and aerosol pathways lead to some differences in their projections that cannot be attributed to model behavior (Grose et al., 2020). However, the study of Tokarska et al. (2020) scaled the RCP8.5 warming by the SSP5-8.5 to RCP8.5 forcing ratio to estimate the CMIP5 response to the SSP scenario and found that the results are very similar to the original CMIP5 projections, suggesting that the stronger warming levels in CMIP6 are not substantially driven by differences between the scenarios.
The effective climate sensitivities for CMIP5 and CMIP6 GCMs given in Tables 1 and 2 are adopted from the study of Zelinka et al. (2020) to create likely model ensemble subsets (see Table 4) based on the IPCC's likely ECS range of 2. 5°C-4°C (IPCC, 2021). The adopted values are generally for r1i1p1 (CMIP5) and r1i1p1f1 (CMIP6) variants. However, for the cases where the variants selected by the study of Zelinka et al. (2020) and our study do not match, the sensitivity values calculated in their study are assumed to be representative for the variant used here (there are only two models out of 68 for which the variants do not match; see Tables 1 and 2). Effective climate sensitivity values given by Zelinka et al. (2020) were calculated following Gregory et al. (2004)'s method using 150-year simulations, which has been shown to underestimate actual ECS (Rugenstein et al., 2020;Zelinka et al., 2020). Nevertheless, effective climate sensitivity is referred to as ECS throughout the paper, as it is common to use the terms interchangeably for simplicity.
Of the 35 (33) CMIP5 (CMIP6) models, there are 16 (11) models that fall within the likely ECS range (see asterisks marked in Tables 1 and 2). In this study, we have used these models, that is, 16 (11) CMIP5 (CMIP6) models, as likely model ensembles (Table 4) to investigate the projections representing more plausible future conditions.  Zelinka et al. (2020) than the member we used are indicated in parentheses. The models that fall within the likely ECS range of 2.5°C-4°C (IPCC, 2021), and hence selected for the likely ensemble subset, are marked with an asterisk (*).

Köppen-Geiger Climate Classification
We use the updated version of the Köppen-Geiger climate classification as explained in detail in Peel et al. (2007). The calculation scheme for obtaining the respective climate classes is given in Table 3. Throughout the study, Köppen-Geiger climate zones are calculated using 30-year time slices of monthly temperature and precipitation data.

Evaluation of GCM Performance
Comparison of the climate model simulations against observations can be utilized as a diagnostic tool to evaluate the change in model accuracy between the model generations (Flato et al., 2013). Several previous studies conducted at regional scales have reported incremental improvements in the CMIP6 models' capability T cold = Temperature of the coldest month, P dry = Precipitation of the driest month, MAP = Mean annual precipitation, MAT = Mean annual temperature, T hot = Temperature of the hottest month, P sdry = Precipitation of the driest month in summer, P wwet = Precipitation of the wettest month in winter, P wdry = Precipitation of the driest month in winter, P swet = Precipitation of the wettest month in summer, T mon10 = Number of months where the temperature is above 10, P threshold depends on conditions: P threshold = 2 × MAT if 70% of MAP occurs in winter, P threshold = 2 × MAT + 28 if 70% of MAP occurs in summer, P threshold = 2 × MAT + 14 otherwise. Summer (winter) is the warmer (colder) of the two 6-month periods, October-March or April-September.
to represent the observed climatology of mean temperature and precipitation (Bağçaci et al., 2021;Grose et al., 2020;Lun et al., 2021;Ortega et al., 2021), as compared to CMIP5. In this study, we compare the accuracy of Köppen-Geiger climate classification maps obtained using the two generations of climate models, CMIP5 and CMIP6, against observation-based maps for the reference periods (1976-2005, 1961-1990, and 1990-2019) to evaluate their performance.
Having all model and observational data at the same spatial resolution (0.5°), the climate models' skill in replicating the reference period Köppen-Geiger climate classification map is measured in terms of the correctly assigned percentage of total land area based on corresponding grid cells: where n a is the total number of grid cells over land where the climate model and observations agree on the assigned climate class, n t is the total number of grid cells over land areas and A i is the area of each corresponding grid cell.
As a result of this evaluation process, the 25 CMIP5 and 25 CMIP6 models with the highest accuracies have been selected to use as the other ensemble subset in addition to the "likely model" ensemble (models are listed in Tables 1 and 2; selection criteria are listed in Table 4). Here, the number 25 is arbitrarily selected to ensure ensemble size-based sampling uncertainties for CMIP5 and CMIP6 are the same and to retain a higher number of models than those eliminated.

Production of the GCM-Based Köppen-Geiger Climate Classification Maps
GCM data contain systematic biases, which manifest as deviations from the historical observed climatology (Mehrotra & Sharma, 2015;Papadimitriou et al., 2017). Here, to correct the mean biases in the climate projections, a simple mean bias correction (Maraun, 2016) has been applied for CMIP5 and CMIP6 historical and future temperature and precipitation simulations (corrected historical GCMs were only used for the change rate calculations and not for performance evaluation). During the bias correction calculations, the 1976-2005 monthly temperature and precipitation observations were taken as the reference (for details on the bias correction method, please see Text S1 in Supporting Information S1).
By using the bias-corrected climate projections of CMIP5 and CMIP6 models, a separate Köppen-Geiger climate classification map has been created for each model for each time period of interest. Then, four different ensemble subsets have been created, for which the details are given in Table 4. For each subset, an additional map has been created by selecting the most frequent climate class for each grid cell among the climate classifications of all models (H. E. Beck et al., 2018). Consequently, a total of four global Köppen-Geiger climate classification maps were created based on these four model ensemble subsets. These form the basis of the climate classification analyses representing the ensembles.
To determine the historical accuracy of the GCMs based on the Köppen-Geiger climate classifications, two maps have been created for the 1976-2005 period using the most frequent climate class approach for the CMIP5 and CMIP6 ensembles, and their accuracies have been calculated using Equation 1. Grid cells that are assigned The methodology used to create the "ensemble maps" also enables measurement of the level of consistency between the GCMs for the future period (2071-2100) (H. E. Beck et al., 2018). Level of consistency (L.o.c.) is defined for a grid cell as: where N f is the frequency of the most frequent climate class for that grid cell on the climate model ensemble and N t is the total number of models in the ensemble. Then, global level of consistencies for CMIP5 and CMIP6 are calculated by taking area-weighted averages of level of consistencies for all grid cells over the land areas.

Temperature Anomalies
Global mean surface temperature (land and ocean) between 2001 and 2020 has been observed to be 0.99°C higher than the average between 1850 and 1900 (90% confidence interval: 0.84 to 1.10), with larger increases over land areas than oceans (IPCC, 2021). We have calculated the temperature anomalies over each land climate zone by subtracting the area-weighted average temperatures (because grid cells have different areas over the different locations) between the two time frames: where Δ is the temperature difference over the climate zone, μ(T i ) is the average temperature over the climate zone, i is the accompanying climate zone (30 in total, from Af to EF) of the base period, time slice2 is the latter time period of 30 years, and time slice1 is the former time period of 30 years.
To calculate future temperature anomalies, we obtain two bias-corrected multi-model ensemble means (one for CMIP5 and one for CMIP6) for the end of the century (2071-2100) using the 25 models with the highest accuracies (these ensemble means are different than the previously mentioned ensemble maps, as the frequency approach cannot be used here).

Rate of Climate Zone Change
We define the rate of climate zone change as the total land area changing its climate zone per year between two consecutive 30-year time frames. To calculate the rate of climate zone change, we have obtained global Köppen-Geiger climate classification maps starting from the beginning of the 20 th century until the end of the 21st century, using 30-yearly time slices of monthly climate data sets as input for each map. Maps are created by shifting the 30-year input window at 5-year intervals (e.g. , 1901-1930, 1906-1935, 1911-1940, …, 2071-2100). The change rate of the climate zones over the Earth's land area is defined as: where n c is the total number of grid cells over land areas changing their climate zones between two consecutive 30-year time frames (e.g., climate zones of 1901-1930 vs. 1931-1960), ∆ is the time difference between the middle years of the two periods (30 years in this case), and A j is the area of the corresponding grid cell. The calculated change rate is assigned to the middle year of the second time frame (1945 for this example).
Loss rates of major climate zones were obtained in a similar way to the change rates of climate zones, but the total lost area of any given major climate zone was divided by the total area of that climate zone in the base period to obtain a percentage change. Finally, previously obtained 30-year Köppen-Geiger global maps with 5-year intervals were also used to obtain the areal coverage of each major climate zone as a time series with 5-year time steps.
To determine total land area changing its climate zone as a function of global mean land warming, we first brought all individual models to the same global land warming threshold levels (e.g., 1°C, 1.5°C, 2°C) relative to the baseline period observations . Then, using the ensemble models obtained with the frequency approach for each warming level, we calculated the total land area for the grids that have changed their climate zones ( ∑

=1
) relative to the baseline period observations.
Throughout the calculations of the rate of change, loss rates, and total land areas, bias-corrected historical and future simulations of CMIP5 and CMIP6 models have been used  with the observational data (1901-2019).

Confidence Intervals
Confidence intervals were estimated by applying a non-parametric bootstrap method (Efron, 1979), sampling across models creating 1000 sets of model ensemble realizations. After choosing the relevant number of ensemble members (e.g., 11 models for CMIP6 likely ensemble, 16 models for CMIP5 likely ensemble, or 25 models for CMIP5 and CMIP6 ensembles), each set of ensemble realizations has been obtained by random selection of models allowing for replacement. Then, all analyses in the study have been repeated with those samples (by obtaining a most frequent climate class-based estimate for each realization). Finally, using all 1000 realizations, 2.5% and 97.5% quantiles were taken as the upper and lower bounds for the 95% confidence interval. Figure 1 shows the global distribution of the climate zones for the past , observed shifts in present-day climate (1990-2019) from the past, and projected future (2071-2100) shifts from the present-day climate based on CMIP5 and CMIP6 ensembles. Overall, the change between the observed climate classification maps during the early 20th and 21st centuries (Figure 1b) is considerably less than the projected change between the early and late 21st centuries for both CMIP5 and CMIP6 ensembles (Figures 1c and 1d). The projected changes in climate zones by the end of the century are more extensive in the CMIP6 ensemble than in CMIP5, especially over the cold climate zones in the high latitudes of the Northern Hemisphere ( Figure 1e; and global maps showing the distributions of the climate zones for GCM ensembles and observational data are shown in Figure S1 in Supporting Information S1).

Köppen-Geiger Climate Classification Maps
A question might arise whether the results are sensitive to the choice of the reference data set (CRU and GPCC for temperature and precipitation, respectively). We have investigated the consistency of the climate maps with two additional data sets (GHCN-CAMS/CRU and ERA5/ERA5 for temperature/precipitation) for the present-day conditions ( Figures S6 and S7 in Supporting Information S1). Present-day observed Köppen-Geiger maps show some differences, especially over Siberia, mountainous regions, and at the boundaries of climate zones, depending on the observation data used. Overall, CRU/GPCC versus ERA5 areal overlap is 76.9%, CRU/GPCC versus GHCN-CAMS/CRU is 80.4%, and ERA5 versus GHCN-CAMS/CRU is 76.2%. Accordingly, CRU/GPCC has a slightly higher average overlap against other products (78.65%) than ERA5 (76.55%) and GHCN-CAMS/CRU (78.3%).
Of the 25 models in the CMIP6 ensemble (Table 2), 12 models exhibit higher ECS than the IPCC's likely range of 2.5°C-4°C (IPCC, 2021); hence, the climate classifications differ somewhat. Namely, for CMIP6 (CMIP5), the average ECS value of likely ensemble is 3.09 K (3.20 K), while for 25-member ensemble the average is 4.05 K (3.20 K) (Tables 1 and 2). Accordingly, we have determined the projections based on likely ensembles to better understand the effects of higher ECS models on climate zone distributions. The likely model projections for the Köppen-Geiger climate zones by the end of the century for CMIP5 and CMIP6 (Figure 2) are more consistent with each other than the 25-member ensembles (Figures 1c-1e): 90.55% of the land areas overlap for the likely model ensembles, whereas it is about 79.08% for the 25-member ensembles (global maps showing the climate zone distributions for likely GCM ensembles are shown in Figure S2 in Supporting Information S1). Also, the apparent differences observed over the higher latitudes between CMIP5 and CMIP6 ensembles almost vanish for the difference between likely ensembles (Figure 2d). However, some CMIP6 models in the likely model ensemble have low climate classification accuracies (e.g., FGOALS-g3 in Table 2), while some models with ECS values outside the likely range have high accuracies (e.g., HadGEM3-GC31-MM in Table 2). Therefore, a joint investigation of the distribution of the climate zones based on the accuracy-based ensemble (i.e., 25-member ensemble) along with the likely model ensemble is valuable. A better approach might be treating the CMIP6 ensemble as a lower likelihood but higher risk projection and the likely ensemble as the more plausible projection.

Performance of CMIP5 and CMIP6 Climate Classifications
The  Figure 3e. When the ensemble model maps are used, the accuracies of both model generations increase, but the relative performance for the correct classification of CMIP6 (65.34%) becomes only marginally better than CMIP5 (63.74%) (Figures 3a and 3b). This implies that although individual CMIP6 model simulations generally yield higher classification accuracies than CMIP5, the rate of improvement reduces when the ensemble information obtained from these model generations is used in the context of climate classifications. Nevertheless, the improved performance observed in CMIP6 versus CMIP5 is not sensitive to the choice of the reference data set or the period, as CMIP6 slightly but consistently performs better (Figures S8 and S9 in Supporting Information S1). These accuracy estimates are lower than those obtained in previous studies using CMIP5 simulations (Belda et al., 2015;Phillips & Bonfils, 2015), likely because we have used a larger number of climate classes to calculate the accuracy estimates than previous studies. Although the ability to represent the historical climate in the ensemble maps is marginally improved in CMIP6 models compared to CMIP5, the regions that are misidentified by CMIP5 and CMIP6 models for the reference period are very similar, such as the west coasts of North and South America, Amazonia, central Australia, southern Africa and Siberia, in addition to the regions of boundaries between climate zones (Figures 3a and 3b). CMIP5 simulations have been reported to have temperature biases over high elevations of the Himalayas, Greenland, and over ocean upwelling regions off the west coasts of South America and Africa (Flato et al., 2013). Also, most of the CMIP5 models have been shown to overestimate precipitation over regions with complex topography (Mehran et al., 2014) and some underestimate rainfall over Amazonia (Yin et al., 2013). Biases in the ensemble mean temperature and precipitation of CMIP6 are lower than for CMIP5 for the reference period 1976-2005 (6.44% reduction in 30-year mean temperature bias and 3.29% reduction in 30-year mean precipitation bias), while previously mentioned regional bias patterns remain (Figures 3c and 3d). The models' regional temperature and precipitation biases against observations lead to misclassifications in their Köppen-Geiger climate classification maps. Therefore, not surprisingly, regions with large temperature or precipitation biases, such as the west coasts of America, Himalayas, Amazonia, and Siberia (CMIP6: Figures 3c and 3d, CMIP5: Figure S10 in Supporting Information S1), have misclassified climate zones and lower climate classification accuracies (Figures 3a and 3b). Nevertheless, some biases in the models do not lead to misclassifications in the Köppen-Geiger maps, such as the cold bias over a large portion of Greenland (Figure 3c). This is because of the threshold-based nature of the Köppen-Geiger climate classification; colder simulated temperatures for a region already in the polar frost (EF) climate zone do not affect its classification. On the other hand, for some regions of Earth, ensemble mean biases and climate classifications based on the frequency method might not be correlated, such as over Australia. Most of the bias patterns are shared across the two model generations, but there are also regional differences. Some examples are reduced warm bias over Northern Canada and aggravated wet bias over the western coasts of Peru and Ecuador in CMIP6 versus CMIP5 (detailed differences between the historical simulations of the two model generations can be seen in Figure S11 in Supporting Information S1).
Inter-model consistency between the future projections of the Köppen-Geiger climate classification maps can be interpreted as an estimate of their future confidence levels. Future (2071-2100) CMIP6 climate classification level of consistency (global average of 79.18%; Figure S12b in Supporting Information S1) is slightly lower than that for CMIP5 (80.06%; Figure S12a in Supporting Information S1), with similar patterns to those observed in the historical period. Moreover, consistencies become lower toward the higher latitudes, as also found in the H. E. Beck et al. (2018). Although the overall scores and general patterns of the future level of consistencies for the two model generations are similar, some regional differences exist, as is the case with historical biases. For example, CMIP6 models show a higher future level of consistency than CMIP5 in central Russia but a lower level of consistency in northern Russia.

CMIP5 and CMIP6 Future Temperature Anomalies
We calculate the observed temperature anomalies between 1990-2019 and 1901-1930 over each Köppen-Geiger climate zone: Cwc has experienced the least amount of warming (0.21°C), and Dsd has experienced the greatest amount of warming (1.86°C), with the global mean land surface temperature increase of 0.95°C ( Figure S13 in Supporting Information S1). Projections of the least and the most severe warming climate zones by the end of the century are consistent across different climate model ensembles: Cfc and Dfd ( Figure S14 in Supporting Information S1). However, the range of the projected warming changes sharply across the ensembles. The CMIP5 RCP8.5 ensemble mean projects a global mean land warming of 4.29°C [95% confidence interval: 3.93-4.66] by the end of the century relative to the present-day ( Figure S14a in Supporting Information S1). On the other hand, mean warming reduces to 3. 82°C [3.49-4.16] when the likely CMIP5 model subset is used ( Figure S14b in Supporting Information S1). This difference between the two ensemble subsets becomes more prominent for CMIP6: the ensemble mean projects 5. 64°C [5.11-6.20] of warming ( Figure S14c in Supporting Information S1), while the likely model ensemble mean projects 4. 45°C [4.09-4.80] by the end of the century relative to present-day observations ( Figure S14d in Supporting Information S1). The sharper decline in the mean projected warming and uncertainty range from the CMIP6 ensemble to the CMIP6 likely ensemble compared to the difference between the CMIP5 ensembles is due to the preponderance of the high sensitivity models in the CMIP6 ensemble (Hausfather et al., 2022). Although the levels of warming differ distinctively across the ensembles, colder high latitude climate zones are projected to experience more severe warming based on all ensemble projections, causing stronger shifts in the climate zones toward the poles.

Distribution of Climate Zones
The percentage of global land area covered by each major Köppen-Geiger climate zone is shown in Figure 4. In the past , the zone of tropical climate (A) covered 22.51% of the land area, expanding to cover 23.18% in the present  The observed and projected changes in areal coverage percentages of all 30 Köppen-Geiger climate classes are shown in Figure 5 (The exact area coverages of each climate class for past, present, and future are given in Figure   Figure 4. Percentage land area coverages of the major climate zones for past, present, and future periods. Error bars represent the 95% confidence interval for model ensemble projections. S15 in Supporting Information S1). Similar to previous studies (H. E. Beck et al., 2018;Hanf et al., 2012), there is a projected expansion in the tropical savannah zone (Aw): land area coverages for past (Figure 5a), present (Figure 5b), future CMIP5 ensemble (Figure 5c), future CMIP5 likely ensemble (Figure 5d), future CMIP6 ensemble (Figure 5e), and future CMIP6 likely ensemble (Figure 5f) (Figure 5g), they become marginal when the projections of the likely ensembles are compared (Figure 5h).

Global and Continental Climate Zone Changes
Global past, present, and future Köppen-Geiger maps (Figures 1 and 2) indicate that some continents are projected to experience greater changes in their present-state climate zones than others. Since the beginning of the 20th century, the Earth has already experienced 14.77% of its land area changing its climate classification, with the most extensive changes observed over North America, Europe, and Oceania ( Figure S16 in Supporting Information S1). Furthermore, the CMIP5 ensemble simulates a world where 37. 94% [35.27-41.53] of the climate zones will change by the end of the 21st century compared to present-day climate zones (Figure 6a), while the CMIP5 likely ensemble projects 39. 66% [38.36-43.46] of global change (Figure 6b). These are lower than the 46.3% level obtained by Feng et al. (2014) but higher than the 20% level obtained by Mahlstein et al. (2013). The lower level found in the latter study is because Mahlstein et al. (2013) used linearly detrended time series to eliminate short-term climate variability, whereas we retain this information so as not to risk unintentional elimination of real signal. The change is even more pronounced in CMIP6, as the global percentage area of future classification change is projected to be 48. 09% [45.36-50.97] based on the CMIP6 ensemble (Figure 6c), and 41. 92% [40.29-45.62] based on the CMIP6 likely ensemble (Figure 6d). In all four CMIP5 and CMIP6 ensembles, Asia, Europe, and North America have higher percentage of areas projected to change, while Africa, Oceania, and South America have values below the global mean ( Figure 6).  (Figure 6a), CMIP5 likely ensemble (Figure 6b), CMIP6 ensemble (Figure 6c), and CMIP6 likely ensemble (Figure 6d) Figure 7a shows the rate of areal climate zone change (10 6 km 2 /year), in the observations and model ensembles.

Rate of Climate Zone Change
The rate of change in the GCMs for the historical period is substantially lower than that of observations until the 1990s, probably because internal climate variability induces most of the climate zone shifts until this time, and our frequency-based ensemble approach quells internal variability to some degree. However, the rate of change accelerates after the 1990s in both GCM generations for all the ensemble combinations, and in observations, Model Intercomparison Project Phase 6 (CMIP6) ensemble and present-day observations, (f) the difference between future CMIP6 likely ensemble and present-day observations, (g) the difference between future CMIP6 and CMIP5 ensembles, (h) the difference between future CMIP6 and CMIP5 likely ensembles. The land area change percentages represent the absolute change in the areas of climate classes (e.g., a −3% projected decrease in a climate class that covers 7% of the present-day land area means 4% projected future land area coverage). Error bars represent the 95% confidence interval for model ensemble projections.
when the climate warming signal starts to predominate over internal variability. Greater acceleration of change is projected in the CMIP6 ensemble with higher change rates than in the CMIP5 ensemble throughout the 21st century. The ECS constrained CMIP6 likely ensemble still projects a higher rate of climate zone change (reaching 10 6 km 2 /year by the end of the century relative to the 2041-2070 climatology) than the corresponding CMIP5 likely ensemble, although the difference between the two generations is smaller than that of the 25-member ensembles.
The smaller difference between the likely ensembles of the two generations signals that the steeper °C/year warming rates might be the culprit of the greater acceleration of climate zone change observed in the CMIP6 ensemble.
To estimate whether the difference emerges due to different global mean land warming rates or spatial pattern differences between the two model generations, we examine the total land area changing its climate zone at different global mean land warming levels relative to the 1901-1930 observations (Figure 7b). Although the changing land area as a function of global land warming is also slightly greater for the CMIP6 ensemble than the CMIP5, the results suggest that the greater change rates observed in the CMIP6 are dominated mainly by the models' stronger warming rates rather than the spatial pattern differences between the two model generations. For a comparison of the change rates as a function of global land warming level for the individual and ensemble models with mutual rescaling, and calculation details please see Text S2, Figure S17 in Supporting Information S1 and Tables S1-S3 in Supporting Information S2. A previous study using CMIP5 simulations has also found that the pace of change in climate zones is dominated by the increasing global temperature rather than the changes in precipitation (Mahlstein et al., 2013).
We note that our CMIP5 and CMIP6 change rates do not lie within the 95% confidence intervals at the beginning of the calculation period in Figure 8a. That is probably because of the reason stated above: an individual model's internal variability is higher than the variability of our most frequent climate class-based approach for that period. Some models appear more than once for each bootstrap realization, and it limits the elimination of internal variability to some degree. Therefore, our CMIP5 and CMIP6 change rates remain lower than their confidence intervals for the earlier periods. However, after the 1990s, the climate warming signal becomes the dominant factor for the change rates, and our GCM-based change rates fall within their confidence intervals thereafter. Figure 8 shows the loss rates of the five major climate zones (A-E). Loss rates of climate zones C, D, and E accelerated after 1980 in the observation period. Higher loss rates of these three climate zones are projected to remain higher in the future than the other two (A and B), with the polar (E) climate zone having the highest rate. Although both CMIP5 and CMIP6 simulations project an increasing trend in the loss rates of these three zones similarly, the magnitudes of the losses are more pronounced in CMIP6 compared to CMIP5 (Figures 8a and 8c). Yet, the gap between the two generations becomes smaller in the likely model subgroups (Figures 8b and 8d). Figure 9 shows the total land areas of tropical (A) and arid (B) climate zones, which are projected to expand continuously throughout the 21st century for all ensembles of CMIP5 and CMIP6, despite the fluctuations that appeared in the 20th century. Importantly, having a constant area does not necessarily mean that the climate zone has not changed its geographical location (Mahlstein et al., 2013). There is no distinct trend in the area of the temperate (C) zone since it is projected to balance its lost areas by generally expanding toward the north, into the mid-latitudes of present-day cold (D) climate regions in the Northern Hemisphere, while it is harder to define a general trend over the Southern Hemisphere (Figures 1c, 1d, 2b, and 2c). Finally, there is a constant reduction in the projections of areal coverages of both cold (D) and polar (E) climate zones, with more explicit reductions in their areas projected in CMIP6. Once again, the gap between the two likely ensembles is smaller than that between the two 25-member ensembles, but the CMIP6 likely ensemble still projects greater reductions than CMIP5. The continuous dramatic reduction in the area coverage of the polar climate zone is because it cannot recover its lost areas by expanding into other locations.

Discussion and Conclusions
This study presents the observed and projected differences between the high-emission scenario GCMs in CMIP6 (SSP5-8.5) and the previous generation models of CMIP5 (RCP8.5) based on the Köppen-Geiger climate classification. Individual CMIP6 models, on average, have higher accuracy than CMIP5, where the ensemble-based CMIP6 estimate is only slightly better at representing the observed Köppen-Geiger climate classes for the reference period 1976-2005. However, similar bias patterns remain, which may result from some long-standing errors, such as the persistence of a double intertropical convergence zone (ITCZ) and Pacific cold tongue biases (Tian & Dong, 2020), despite the improvements in the new generation of CMIP6 models. Considering that there are uncertainties in the simulated climate zone distributions due to deficiencies in the physics of the models, which particularly affect precipitation, further advancement in the parameterization of the physical processes of the models could lead to more accurate climate classifications in the future (Tapiador et al., 2019). Although CMIP6 models generally perform better during the historical period, their mean inter-model consistency in terms of assigning a common climate class to a grid for the future period (2071-2100) is slightly lower than that of CMIP5. On the other hand, such consistency among CMIP6 models is better in some regions in terms of temperature and precipitation climatology, for example, the Mediterranean (Bağçaci et al., 2021) and Sahel (Monerie et al., 2020).
We find that the subset of 11 CMIP6 models that lie within the likely ECS range as updated in the recent IPCC report (IPCC, 2021) projects greater warming over all 30 of the Köppen-Geiger climate zones by the end of the century than its CMIP5 counterpart (consisting of 16 models). However, the difference between the mean warming projections relative to the present-day observations over global land areas of the likely CMIP6 ( The increased warming trend observed in the CMIP6 projections does not alter the direction of changes in tropical, cold, and polar climates projected in CMIP5 but amplifies their magnitudes, whereas expansion of arid climate is projected nearly at the same level in both model generations. Moreover, the shift from cold summer to hot summer zones, such as Dfc to Dfa, in the mid-latitudes of the Northern Hemisphere, is projected to become more prominent in the CMIP6 projections versus CMIP5. Using the likely model ensembles does not change these conclusions, but projections of the percentage of land areas for the climate zones of CMIP5 and CMIP6 converge toward each other. As an important consequence of warming, this may result in greening in tundra areas and increased duration of the frost-free season in cold and polar climate zones. By the end of the 21st century, climate zones over 37. 94% [35.29-41.76 . Also, Europe, North America, and Asia are projected to experience greater changes in their present-day climate zones than the global average in all four of the ensembles by the end of the century. Europe is projected to have the greatest change from the current climate, with 65%-91% of its land area projected to be in a different climate zone by the end of the century. Especially, the expansion of the temperate climate zone (C) into the present-day regions of cold (D) climate in Europe is projected to become salient by the end of the 21st century. However, a possible reason for relatively smaller projected changes in some continents might be the nature of the Köppen-Geiger climate classification. Since it is threshold-based, further increases (reductions) in the climatic conditions of a region already above (below) a threshold will not change its climate class even though the climate regime of the region is adversely affected, such as further temperature increases in the tropical climates of Africa (Mahlstein et al., 2013). Therefore, redefinition of climate classes might be necessary to account for the expected changes in the future climate.
The rate of climate zone change is projected to accelerate throughout the 21st century in both generations of climate model simulations, with higher rates projected in CMIP6. The change rates of likely ensembles converge toward each other, but the CMIP6 likely ensemble still has a greater change rate than that of the CMIP5 likely ensemble. Since there is a correlation between Köppen-Geiger climate zones and major biome types (Rohli et al., 2015), it implies that if the trajectory of the CMIP6 SSP5-8.5 scenario is to be followed, whether based on the likely ensemble or the 25-member ensemble, some vulnerable species with lower dispersal abilities in the regions with higher rates of change (Loarie et al., 2009;Sandel et al., 2011) might have less time to adapt to changes than previously projected by CMIP5 RCP8.5. Nevertheless, one should be cautious when relating the changes in the climate zones directly to actual biomes since there might be a delay in the vegetation changes, and other factors might affect the relationship between vegetation and climate zones (H. E. Beck et al., 2018), for instance, soil properties, pests, human intervention, and fertilization from increasing atmospheric CO 2 levels.
Although some models with high ECS could not simulate the observed warming trend well (Tokarska et al., 2020), we find that some of these models have high historical climate classification accuracies; hence they cannot be completely discarded in an analysis based on climate classification. However, the converging projections of CMIP5 and CMIP6 likely ensembles against the diverging projections of the unconstrained model ensembles suggest that the difference in the climate zone projections between the two generations depends strongly on the ECS values of the models in the ensembles. We also find that the greater climate zone change rates simulated in CMIP6 are mainly dominated by the higher warming rates of these models. Since the plausibility of these high warming rate projections is a matter of debate, we recommend interpreting the likely model ensemble results as the more plausible future projections and the ensembles containing the models with high equilibrium climate sensitivities as the lower likelihood but potentially high-risk outcomes.