Surface Air Temperature Trend Over the Tibetan Plateau in CMIP6 and Its Constraint in Future Projection

The surface air temperature (SAT) trend on the Tibetan Plateau (TP) was 3.45°C 100 years−1 from 1961 to 2014. The multi‐model ensemble (MME) of 33 coupled models participated in the Coupled Model Intercomparison Project phase six (CMIP6) was about 1°C 100 years−1 lower than the observation. Although MME generally shows better skill in reproducing the distribution of SAT trend over TP than most of the CMIP6 models, its performance is greatly degraded by a small group of models, about 12% on average, with large biases. In this paper, the constrained multi‐model ensemble (CMME) based on a certain observation‐based threshold is used to constrain future projections of the SAT trend over TP. Compared with the MME results, the improvements in CMME are mainly over the eastern plateau in historical simulation and are relative to the reduction of the model biases to carbon dioxide (CO2) forcing. Under the high‐emission SSP5‐8.5 scenario, SAT increases significantly over the entire TP. The constraint of CMME on the MME is mainly over the eastern plateau with a difference of 0.5°C 100 years−1, about 6% of the MME results. Under the intermediate‐emission scenario SSP2‐4.5, the effect of CMME is relatively smaller, but the corresponding spatial distribution is similar to that under the SSP5‐8.5 scenario. The CMIP6 models tend to underestimate the warming trend projections over the water source regions in the northeastern plateau and should be noticed.


Introduction
The Tibetan Plateau (TP), the so-called "third pole," is the highest plateau on the Earth with the largest store of ice.Under global warming, the warming trend over TP is even more remarkable.Since the mid-20th century, most areas of TP have experienced significant warming, and the temperature increase is generally greater than that in the Northern Hemisphere and the corresponding plateau latitudes (e.g., Duan et al., 2006;Liu & Chen, 2000;Xu et al., 2017).In contrast to the global warming hiatus and the cooling trend in the other parts of China from 1998 to 2013, a warming acceleration was observed over TP during the same period (Duan & Xiao, 2015).The remarkable warming over TP may be attributed to the positive "snow albedo-temperature" feedback, latent heat release related to water vapor condensation, surface radiation flux, and aerosol effect (Pepin et al., 2015;Rangwala et al., 2010).The warming-induced extensive glacial retreat and permafrost degradation have a great impact on the ecosystem across TP (Yao et al., 2007(Yao et al., , 2012a(Yao et al., , 2012b)).
The warming trend over TP is of great importance for the climate over Asia and even the globe by the thermal and dynamic processes (e.g., Liu et al., 2020;Wu et al., 2015;Yanai et al., 1992;Ye & Gao, 1979).The TP enhances the coupling between the lower and upper tropospheric circulations and between the subtropical and tropical monsoon circulations, resulting in an intensification of the East Asian summer monsoon and a weakening of the South Asian summer monsoon (Wu et al., 2012).A reduction in upward surface sensible heat flux over the TP will result in a reduced thermal contrast between the Eurasian continent and adjacent oceans, thereby weakening the Asian summer monsoon circulation (Liu et al., 2012).Atmospheric heating induced by the rising TP temperatures can enhance East Asian subtropical frontal rainfall (Wang et al., 2008).Besides the Asian summer monsoon, the influences of the TP on the atmospheric circulation are also evident over upstream regions, including West Asia, North Africa, Europe and North Atlantic (e.g., Hoskins & Karoly, 1981;Wang et al., 2019;Zhao et al., 2012).There may be also interactions between the thermal status of the TP and oceanic circulation, such as the Atlantic Meridional Overturning Circulation (Fallah et al., 2016;Yang & Wen, 2020).Climate models are important for studies on climate change attribution and future projections.The decreased surface sensible heat flux over TP is evident in ground-based observation (Yang et al., 2011).By numerical simulations, the reductions in surface sensible heat flux over the TP are responsible for the weakened the monsoon circulation and postponed the seasonal reversal of the land-sea thermal contrast in East Asia (Duan et al., 2013).Duan et al. (2006) found that climate models forced by observed CO 2 in the 20 th century could reasonably reproduce the main characteristics of the Tibetan Plateau warming.That is, the main reason for the recent climate warming over TP is the increase of anthropogenic greenhouse gas emissions.Ramanathan et al. (2007) suggested that atmospheric brown clouds might be also responsible for the retreat of Himalayan glaciers.Based on coupled climate simulations and an optimal fingerprinting detection, the attribution study in Zhou and Zhang (2021) found that current generation of the state-of-the-art GCMs tend to underestimate the SAT responses over the TP to anthropogenic forcing.An exacerbated warming is expected over the TP in future projections.
The latest phase of the Coupled Model Intercomparison Project (CMIP6) has the largest number of participants with more than 40 model institutes (Eyring et al., 2016).Most of the CMIP6 models can reasonably describe the spatial distribution of SAT over TP, but the systematic cooling bias in the CMIP5 simulations still exists (Zhu & Yang, 2020).Besides the climatology states, the change of SAT, especially the SAT trend, is more important for the climate and ecological system over TP and the surrounding area.The traditional multi-model ensemble mean (MME) is often used to reduce the influence of inter-model uncertainty on climate change analysis, where the contribution of each model is equally weighted.However, enormous biases from a single model or a small number of models in some specific regions may skew the ensemble mean results far away from reality.Some heavily biased simulations may diminish the real signals or lead to unreliable results, especially on a regional scale.In this study, we evaluate the ability of the CMIP6 climate model in reproducing SAT over TP and try to calibrate its projections under different shared socio-economic paths by a new ensemble averaging method (CMME) introduced in Zhang et al. (2023).
The rest of the paper is organized as follows.Section 2 introduces the observation and model data, as well as the constraint method used to calibrate future projections.Section 3 evaluates the ability of the CMIP6 models in SAT trend reproduction over TP, the advantage of CMME, and the impacts of model biases in response to CO 2 -forcing on model performances.Section 4 analyzes the influence of CMME on the SAT trend projections under SSP5-8.5 and SSP2-4.5 scenarios over TP.A discussion is given in Section 5. Section 6 summarizes the study and its findings.

Observational and Model Data
The observed monthly mean SAT from 1961 to 2014 collected by China Meteorological Administration is used for model evaluation.The station locations over TP are shown in Figure 1.A gridded monthly SAT observation data set on a 0.25° latitude by 0.25° longitude resolution over China is also employed (CN05.1;Wu & Gao, 2013).It is based on ground-based observation of more than 2,000 stations and has been widely used for research on climate change, agriculture, hydrology, and ecology analysis in China (e.g., Leng et al., 2015;Zhou et al., 2014;Zhou et al., 2016).The station cover indicates that the CN05.1 data is more reliable in the east of TP than in the west.We used the outputs of 33 climate models participated in CMIP6.All the models have completed the historical, 1pctCO2, and the 21st century climate projections under the SSP5-8.5 and SSP2-4.5 scenarios with the variant label r1i1p1f1.The 1pctCO2 experiment is forced by a 1%yr −1 CO 2 increase, starting from its 1850 value as prescribed in pre-industrial control simulation.The SSP5-8.5 experiment is a future climate projection, high enough to produce a radiative forcing of 8.5 W m −2 in 2100.The SSP2-4.5 is a future climate scenario with emissions to produce a radiative forcing of 4.5 W m −2 in 2100.SSP5-8.5 and SSP2-4.5 update the RCP8.5 and RCP4.5 in CMIP5, representing the high-end and medium part of the range of Shared Socioeconomic Pathways (SSP).The historical and 1pctCO2 are the entry cards in CMIP6 that all the models participated in CMIP6 should conduct these experiments.The SSP5-8.5 and SSP2-4.5 are the tier 1 experiments of ScenarioMIP that have the highest priority and therefore have the most participating climate models (O'Neill et al., 2016).We choose these four experiments because the ensemble averaging will be more robust with larger ensembles.
To facilitate model evaluation and further analysis, all model data were bilinear interpolated to a horizontal resolution of 0.25° as CN05.1.We choose the resolution of CN05.1 to retain more observed signals since we focus on regional SAT changes in this study.Table 1 provides brief information about the 33 climate models.Notice that, we only use one member for one simulation, which contains large internal variability.The SAT bias, especially at regional scales could also be induced by internal variability.Therefore, all the model output and the observation data are smoothed by the 3-point 1-2-1 low-pass filter to remove fluctuations less than 50 years.

The Constrained Multi-Model Ensemble Method (CMME)
Climate simulations vary across climate models due to differences in the representation and approximation of climate system and processes by plausible solutions to the governing equations with current climate knowledge.Multi-model ensemble mean approach (MME) can improve the accuracy of a climate simulation by allowing model errors to cancel each other out and models that poorly performed to be down-weighted.However, in comparison with the Climatic Research Unit gridded data (CRU; Harris et al., 2020), Zhang et al. (2023) found that the CMIP6 MME has a relatively low capability in reproducing the SAT trend in the United States and Asia.Future projections in CMIP6 MME may overestimate the SAT-rising risk over North America but underestimate the risk over Asia.
Results in Zhang et al. (2023) indicate the relatively low capability of CMIP6 models in reproducing SAT trend over TP, which may be partly due to the relatively low resolution and uncertainty of the CRU data.Based on finer and more precise station observations in China, the constrained multi-model ensemble method (CMME) introduced by Zhang et al. (2023) will be used to calibrate future SAT trend projections over TP in this study.The observed SAT trend is calculated as the SAT trend in CN05.1 from 1961 to 2014 at each grid point over TP.The SAT projections in the second half of the 21st century (2050-2100) under the SSP5-8.5 and SSP2-4.5 scenarios are then calibrated by CMME.The CMME method is introduced as follows.
1. Calculate the annual mean observed SAT trend at each grid point.2. Define the observation based threshold (OBT), which is a certain threshold range centered on the observed SAT trend.3. Calculate the ensemble of models that can pass the threshold, that is, CMME.Each model only contributes to the statistics of CMME at the grid point where it meets the observed-based threshold.4. The cluster of models in CMME will then be used to calibrate the SAT projections.
The CMME, to some extent, inherits and develops the MME.It tries to eliminate the contributions of poorly performed simulations and further improves climate simulations.

The SAT Trend Over TP in the Period From 1961 to 2014
The high-altitude region of TP is mainly located in the southwestern part of the plateau with many high mountains, such as the Kunlun Mountains, Tanggula Mountains, and Himalaya Mountains.The terrain in the east of the plateau slopes more gently with the important source of freshwater (the "Three-River Source Region") in the 10.1029/2023JD039527 4 of 16 northeast and Hengduan Mountain in the southeast (Figure 1).Due to the complex terrain distribution of the plateau and a suite of physically interlinked processes, the SAT simulation, especially its trend, has always been a big challenge for climate models (Chen et al., 2017;Zhou & Zhang, 2021).

Observation and Simulations
The annual mean SAT trend during 1961-2014 in the gridded observation CN05.1 is first evaluated against the station observation over the station locations (Figure 2).The CN05.1 can reasonably capture the relatively weak SAT warming in southeastern TP and the relatively stronger SAT trend in northeastern and southern TP.The major differences are the spatially smoother LSAT trend in CN05.1.Eliminating the two stations with the largest and smallest SAT trends, the correlation between the station observation and CN05.1 at the rest 117 station sites is significant at 1% level with correlation coefficient of 0.61.The CN05.1 may be more suitable for in our study since its resolution is already higher than all the CMIP6 models.And it is beyond the models' capability to reproduce the SAT trend in the station observations precisely.
Spatially, the SAT trend is statistically significant at the 1% level over the whole TP in CN05.1 (Figure 3a).The mean warming amplitude is 3.45°C 100 years −1 .The warming is pronounced around the Himalayan Mountains and Qaidam Basin with warming higher than 4°C 100 years −1 .The warming is moderate with the warming below 2.5°C 100 years −1 over the Pamir Plateau in northwestern TP and the Hengduan Mountains in southeastern TP.In the CMIP6 MME, the warming rate over TP is 2.31°C 100 years −1 , which is lower than CN05.1 by more than 1°C 100 years −1 (Figure 3b).The CMIP6 models systematically underestimate the amplitude of SAT increasing rate over the hinterland of the plateau, but slightly overestimate the SAT rate over the northwestern and southeastern parts of the Plateau (Figure 3c).As a result, the spatial gradient of SAT trend in the MME is much smaller than the observation.
The Taylor Diagram in Figure 4b evaluates model performances in reproducing the distribution of historical SAT trend over TP statistically.As a result, correlation coefficient between the simulation of each model and the observation (CN05.1) is less than 0.65, and the spatial variability of more than 70% of the models is less than the observation.The correlation between the MME and the observation is higher than 94% of the models, although the spatial variability in MME is 0.39 of that in the observation.
Using the constraint method introduced in Section 2.2, we examine the relationship between skills of SAT trend reproduction and the number of selected models under different observation-based thresholds (Figure 5).As shown, pattern correlation coefficient increases with the narrowing of the OBS-based threshold.The correlation coefficient increases more rapidly when the range of threshold is relatively large.For example, the correla- tion coefficient increases by 0.07 (from 0.69 to 0.76) when the threshold changes from 0.1-1.9 to 0.2-1.8,but increases only by 0.01 (from 0.97 to 0.98) when the threshold changes from 0.8-1.2 to 0.9-1.1.The relationship between the range of threshold and the number of selected models, however, shows the opposite.
Figure 6 makes this same point more visually by showing the spatial distributions of model numbers under different thresholds.Under the threshold of 0.3-1.7 times the observed trend, the mean number of selected models in the plateau area is 28.9 and the correlation with the observation increased from 0.58 in MME to 0.82.That is, about 12% (4 models on average) of the 33 models with large biases significantly reduced the model capability of CMIP6 models.Under the threshold of 0.5-1.5, the correlation reaches up to 0.92, and the number of selected models is 22.0, about 2/3 of the total model set.Under the threshold of 0.6-1.4,the mean number of selected models decreases rapidly to about half of the models (16.9) with the correlation coefficient increased by only 0.02.The number of selected models can generally represent model capability in reproducing SAT trend.The spatial distributions of the number of selected models indicate that model capability is relatively high over the eastern Tibetan Plateau but relatively low over the western hinterland of the Plateau.
We choose the threshold of 0.5-1.5 times of the observed trend in this study due to its relatively high correlation with the observation and the number of selected models.As shown in Figure 7, although the constrained results still underestimate the warming trend in the historical period, the biases reduced all around the plateau in comparison with MME (Figure 7b).The differences between CMME and MME are similar to the MME bias but in opposite sign (Figure 7c).The evolution of regional mean SAT over TP in CMME gets closer to the observation than that in MME (Figure 4a).The statistical results for CMME as shown in the Taylor Diagram are also better than MME (Figure 4b).We further examine the SAT evolutions in the three stations along the same latitude around 32°N over TP in station observation, CN05.1, MME, and CMME (Figure 8).The differences among these sites may be mainly attributed to the differences of altitude.The observations show that the SAT trend at higher altitudes is generally larger than that at lower altitudes, which is consistent with previous studies (Pepin et al., 2015;Qin et al., 2009;Salama et al., 2012).The SAT evolution in CN05.1 are similar to the station observations.The SAT trends in MME and CMME are smaller/larger than those in CN05.1 at high/low altitudes sites.That is, the SAT trends are spatially smoother in MME and CMME, which may be related to the relatively low resolution of CMIP6 models.The horizontal resolutions of CMIP6 model are generally larger than 100 km and are hard to describe the topography of the TP precisely.The SAT trend biases in MME are relatively large at high elevation stations.Evaluations on the ERA-40 reanalysis also demonstrated the growing temperature bias with elevation (Frauenfeld et al., 2005).The SAT trend biases in CMME are generally smaller than those in MME at all the three sites.In comparison with the CN05.1 at all the 119 station sites (Figure 9), the SAT trend in MME is 0.65°C 100 years −1 smaller than in CN05.1, whereas the CMME is 0.42°C 100 years −1 smaller.CMME has a relatively smaller root mean square error (RMSE, 0.59) and higher correlation coefficient (COR, 0.94) than MME (0.95 and 0.60).The ratio of SAT trend in CMME to that in CN05.1 also closer to 1.0 with narrower range (0.7-1.5) than that in MME (0.4-2.2).

Impacts of Model Responses to CO 2 -Forcing on Model Biases
In the context of global warming, the increasing SAT over TP is also attributable to the increasing concentration of greenhouse gases, mainly CO 2 .Here we use the 1pctCO2 experiment to investigate the deviations of model response to CO 2 and discuss the impact of the model deviations on the historical SAT trend biases over TP.
The 1pctCO2 experiment can be understood as a historical experiment under ideal conditions.In the 1pctCO2 experiment, CO 2 was the only anthropogenic external forcing, rising from the pre-industrial condition in 1850 at a rate of 1% per year to about 500ppmv in 1910.As shown in Figure 10a, the SAT trend is significant in MME all over TP with similar spatial distribution to that in the historical experiment (Figure 3b).It confirms the leading role of CO 2 concentration in the warming over TP.The highest warming area is located over the southwest of the Plateau, with the highest warming exceeding 4.2°C 100 years −1 , and the lowest warming area is located over the Hengduan Mountain in the southeast of the Plateau.Similar to the differences between CMME and MME in the historical period (Figure 7c), the warming amplitude of CMME in the northeast (i.e., the main water source area)/southeast of the plateau is stronger/weaker than that of MME in 1pctCO2 (Figure 10c).
However, different from that in the historical experiment, the warming intensity in the southwest Plateau in CMME was weaker than that in MME.That is, in addition to the CO 2 -forcing, other factors such as the internal variability of the climate system may also contribute to the SAT changes over the southwest Tibetan Plateau, where the altitude is relatively high.The sparse coverage of observations may also contribute to the uncertainty in SAT trend simulations over the southwest of Tibetan Plateau.Therefore, we suggest that the impact of CMME is mainly over the northeast and southeast of TP, and is related to the model response error to CO 2 -forcing.

The Constraint of CMME on SAT Projection and Its Further Implication
The improvement in CMME in the historical simulation is because by construction the CMME is closer to the observed trends.Considering the major role of CO 2 -forcing response in SAT simulation, the CMME constraint may, to a certain extent, be able to reduce the SAT trend biases over TP, especially in future scenarios with CO 2 as the main external forcing.Figure 11 analyzes the SAT trend projections over TP from 2050 to 2100 in MME and the corresponding results in CMME under the high emission scenario SSP5-8.5 and the moderate emission scenario SSP2-4.5.Under the SSP5-8.5 scenario, the warming trend in CMME is 8.16°C 100 years −1 , about 0.16°C 100 years −1 higher than that in MME.The constraint of CMME on SAT projection is mainly over the eastern part of TP.The differences between CMME and MME are similar to that in 1pctCO2.The SAT trend over the northeastern part of TP in CMME is about 0.5°C 100 years −1 higher than that in MME, which is about 6% of the warming range of MME.Since the rising trend of CO 2 concentration under the SSP2-4.5 scenario is much smaller than that under SSP5-8.5 (2.0 ppmv year −1 v.s.11.7 ppmv year −1 ), the SAT trend in SSP2-4.5 is less than 1/3 of that under SSP5-8.5.The differences between CMME and MME are also larger under SSP5-8.5.This may be due to the more sensitive CO2-forcing response in CMME.As shown in Figure 3, the SAT trend is underestimated over almost the whole TP.The warming response to CO2-forcing tends to be stronger in CMME to reduce the cooling biases in MME.Despite of the strength differences, the spatial distributions of the influence of CMME are similar under both scenarios.
It is worth noting that, the warming intensity of CMME is higher than that of MME in the main water sources in the northeast TP.Correspondingly, snowmelt runoff will also change.The change in snowmelt runoff will have an important effect on the seasonal and annual runoff in this region.During 1957-2000, SAT in the "Three-River Source Region" increased, and the snowmelt runoff started earlier as observed in the main hydrological stations (Lv et al., 2009).There is a positive correlation between the beginning time of snowmelt runoff and annual runoff in the headwaters of the Three Rivers.That is, the earlier the beginning time of snowmelt runoff, the less the annual runoff will be.Smith et al. (2017) also found that the end of the snowmelt season is trending almost universally earlier and the length of the snowmelt season is thus shortening in many regions of high mountain Asia including TP.Therefore, considering the SAT projections under the SSP5-8.5 scenario, we may also need to pay attention to the runoff projections in the northeast Tibetan Plateau by CMIP6 models.

Discussion: Comparison With the "Emergent Constraint"
The "emergent constraint (EC)" is a famous approach to constrain future quantities of interest and has been widely used since IPCC AR5 (Allen & Ingram, 2002;Brient & Schneider, 2016;Cox et al., 2018;DeAngelis et al., 2015;Hall & Qu, 2006;Tsushima et al., 2016).In EC, a suitably large ESM ensemble is used to identify the relationship between current or past climate and future climate.The EC then tries to reduce uncertainty in climate projections by the model-derived relationship and contemporary observed measurements.The earliest work by EC is in the study of the snow-albedo feedback (Hall & Qu, 2006).Inter-model variations in snow-albedo feedback in the seasonal cycle are highly correlated with the feedback in context of climate change.Eliminating the model errors by observed estimates of snow-albedo feedback strength in the seasonal cycle will directly reduce the spread of feedback strength in climate change.The EC approach has also been used to constrain future warming in CMIP6 at the global scale (Tokarska et al., 2020).The constrained CMIP6 median warming in the SSP5-8.5 is over 16% lower by 2050 and 14% lower by 2100 compared to the raw CMIP6 median.
The CMME approach in our study also tries to reduce the uncertainty in future projections by contemporary observations as the EC technique.However, we use observations to eliminate models with large biases in recent SAT trend reproduction.It is under the premise condition that models simulate the observable SAT trend better will also project future SAT trend more reasonably.Physical explanation is the key role of CO 2 -forcing response in SAT trend simulations, which also works at the regional scale.Therefore, the merit of CMME is that it can be used for SAT trend calibration at regional scales.
The EC approach have mainly tended to focus on constraining globally aggregated quantities or elements at a relatively large scales, such as constraints on equilibrium climate sensitivity, cloud feedback (e.g., Brient & Schneider , 2016;Tsushima et al., 2016), hydrological cycle (e.g., Allen & Ingram, 2002;DeAngelis et al., 2015), carbon cycle (e.g., Cox et al., 2013;Wenzel et al., 2016).It is a challenge for EC technique in regional climate constraint since a robust relationship may be hard to identify at specified areas.
For CMME, there are some complications for the potential applications to the other variables.One major challenge is that it may be not possible to find a specific physical explanation responsible for the biases.Regional biases may be attributed to model errors far away by teleconnections.Another challenge comes from the uncertainties in observations, especially for the variables with sparse spatial coverage and raw quality.

Summary
Based on the ground-based observation in the Tibetan Plateau, we evaluate the performances of CMIP6 climate models in reproducing the SAT trend from 1961 to 2014.The CMME method is introduced and employed to investigate the possible reasons for the model biases in SAT trend reproduction.Furthermore, the impacts of CMME on the SAT trend projections over TP under the SSP5-8.5 and SSP2-4.5 scenarios are also examined.The main findings are as follows.1.In the historical period from 1961 to 2014, the most significant warming over TP was mainly around the Himalayan Mountains and the Qaidam Basin.The ensemble mean of CMIP6 models (MME) underestimated the warming trend by more than 1°C 100 years −1 .The spatial correlation between MME and observation is only 0.65, although it is higher than most of the CMIP6 models.2. The CMME method aims to reduce the contributions of CMIP6 models with large biases.The CMME results indicate the great impact of model response to CO 2 -forcing on SAT trend reproduction over TP, especially over the eastern part of the Plateau.3.Under the SSP5-8.5 scenario, the mean SAT trend over TP in CMME is about 0.16°C 100 years −1 higher than that in MME and more pronounced over the northeastern Plateau.The warming intensity in the water source regions in the northeast of the Plateau may be underestimated in MME under the SSP5-8.5 scenario.Although the impact of CMME on the SAT trend projection under the medium emission scenario (SSP2-4.5) is much weaker, the spatial distribution of CMME impact is similar to that under the SSP5-8.5 scenario.

Figure 1 .
Figure 1.Topography of the Tibetan Plateau (color map, unit: m) and the spatial distribution of meteorological stations (black dots).The graylines denote the Yellow River and the Yangtze River on the Tibetan Plateau.The green triangles denote the three meteorological stations in western, central and eastern TP along 32°N.

Figure 2 .
Figure 2. (a) The SAT trend over TP from 1961 to 2014 in station observations.(b) Same as (a), but for the results in the nearest grid in CN05.1.Unit: °C 100 years −1 .

Figure 4 .
Figure 4. (a) The series of annual mean SAT anomaly over TP in relative to 1961-1970 mean.The 3-point 1-2-1 low-pass filter is used to remove fluctuations less than 50 years.The shading indicates the spread of SAT in the 33 CMIP6 models.(b) Taylor diagram for the SAT trend distribution over TP.The reference data is CN05.1.The results of MME and CMME are represented by blue and red dots, respectively.The results of each model are represented by gray circles.The nine models with negative spatial correlations with the reference data are not shown.

Figure 5 .
Figure 5.The regional averaged number of selected models over TP (abscissa) and the spatial correlation between the simulated and the observed SAT trend (ordinate) under different observation based (OBS-Based) thresholds.Different colors denote different thresholds.

Figure 6 .
Figure 6.The number of selected models under different OBS-Based threshold over TP.The spatial correlation with CN05.1 is marked at the top-right corner of each plot.

Figure 8 .
Figure 8. Observed and simulated SAT anomaly at three stations along 32°N over TP.The anomaly is relative to 1961-1970 mean.All series are smoothed by 3-point low-pass filter.OBS stands for the ground station observation.

Figure 9 .
Figure 9. Scatter plots of SAT trend at all the TP stations in CN05.1 (abscissa) versus those in (a) MME and (b) CMME (ordinate).Unit: °C 100 years −1 .NMB stands for normalized mean bias, RMSE for root-mean-square error, COR for correlation coefficient, and equation for the linear regression statistics.The gray dashed lines represent the factor of 0.5 and 2.0 of the SAT trend in CN05.1.

Figure 10 .
Figure 10.The SAT trend over TP from 1850 to 1910 in 1pctCO2 experiment, when the CO 2 concentration increased from pre-industrial concentration to about 500ppmv.(a) MME, (b) CMME, (c) Difference between CMME and MME.Unit: °C 100 years −1 .The meshed area in (c) marks the region with the same symbol as that in Figure7c, where the SAT trend is relative to the response to CO 2 and therefore has a certain predictability.

Figure 11 .
Figure 11.The SAT trend over TP from 2050 to 2100 under the SSP5-8.5 and SSP2-4.5 scenarios in (a) and (d) MME, (b) and (e) CMME.Trends significant at the 1% level are dotted.Panels (c) and (f) the difference between CMME and MME.The meshed area in (c) and (f) is the same as in Figure 10c.Unit: °C 100 years −1 .

Table 1
Information of the 33 CMIP6 Climate Models