Editing IPCC:AR6/WGI/Chapter-7 (section)

=== 7.5.6 Considerations on the ECS and TCR in Global Climate Models and Their Role in the Assessment ===

<div id="h2-24-siblings" class="h2-siblings"></div>

Coupled climate models, such as those participating in CMIP, have long played a central role in assessments of ECS and TCR. In reports up to and including the IPCC Third Assessment Report (TAR), climate sensitivities derived directly from ESMs were the primary line of evidence. However, since AR4, historical warming and paleoclimate information provided useful additional evidence and it was noted that assessments based on models alone were problematic ( [[#Knutti--2010|Knutti, 2010]] ). As new lines of evidence have evolved, in AR6 various numerical models are used where they are considered accurate, or in some cases the only available source of information, and thereby support all four lines of evidence (Sections 7.5.1–7.5.4). However, AR6 differs from previous IPCC reports in excluding direct estimates of ECS and TCR from ESMs in the assessed ranges ( [[#7.5.5|Section 7.5.5]] ), following several recent studies ( [[#Annan--2006|Annan and Hargreaves, 2006]] ; [[#Stevens--2016|Stevens et al., 2016]] ; [[#Sherwood--2020|Sherwood et al., 2020]] ). The purpose of this section is to explain why this approach has been taken and to provide a perspective on the interpretation of the climate sensitivities exhibited in CMIP6 models.

The primary consideration that led to excluding ECS and TCR directly derived from ESMs is that information from these models is incorporated in the lines of evidence used in the assessment: ESMs are partly used to estimate historical and paleoclimate ERFs (Sections 7.5.2 and 7.5.3); to convert from local to global mean paleo temperatures ( [[#7.5.3|Section 7.5.3]] ); to estimate how feedbacks change with SST patterns ( [[#7.4.4.3|Section 7.4.4.3]] ); and to establish emergent constraints on ECs ( [[#7.5.4|Section 7.5.4]] ). They are also used as important evidence in the process understanding estimates of the temperature, water vapour, albedo, biogeophysical, and non-CO <sub>2</sub> biogeochemical feedbacks, whereas other evidence is primarily used for cloud feedbacks where the climate model evidence is weak ( [[#7.4.2|Section 7.4.2]] ). One perspective on this is that the process understanding line of evidence builds on and replaces ESM estimates.

The ECS of a model is the net result of the model’s effective radiative forcing from a doubling of CO <sub>2</sub> and the sum of the individual feedbacks and their interactions. It is well known that most of the model spread in ECS arises from cloud feedbacks, and particularly the response of low-level clouds ( [[#Bony--2005|Bony and Dufresne, 2005]] ; [[#Zelinka--2020|Zelinka et al., 2020]] ). Since these clouds are small-scale and shallow, their representation in climate models is mostly determined by sub-grid-scale parametrizations. It is sometimes assumed that parametrization improvements will eventually lead to convergence in model response and therefore a decrease in the model spread of ECS. However, despite decades of model development, increases in model resolution and advances in parametrization schemes, there has been no systematic convergence in model estimates of ECS. In fact, the overall inter-model spread in ECS for CMIP6 is larger than for CMIP5; ECS and TCR values are given for CMIP6 models in Supplementary Material 7.SM.4 based on [[#Schlund--2020|Schlund et al. (2020)]] for ECS and [[#Meehl--2020|Meehl et al. (2020)]] for TCR (see also Figure 7.18 and FAQ 7.3). The upward shift does not apply to all models traceable to specific modelling centres, but a substantial subset of models have seen an increase in ECS between the two model generations. The increased ECS values, as discussed in ( [[#7.4.2.8|Section 7.4.2.8]] , are partly due to shortwave cloud feedbacks ( [[#Flynn--2020|Flynn and Mauritsen, 2020]] ) and it appears that in some models extra-tropical clouds with mixed ice and liquid phases are central to the behaviour ( [[#Zelinka--2020|Zelinka et al., 2020]] ), probably borne out of a recent focus on biases in these types of clouds ( [[#McCoy--2016|McCoy et al., 2016]] ; [[#Tan--2016|Tan et al., 2016]] ). These biases have recently been reduced in many ESMs, guided by process understanding from laboratory experiments, field measurements and satellite observations ( [[#Lohmann--2018|Lohmann and Neubauer, 2018]] ; [[#Bodas-Salcedo--2019|Bodas-Salcedo et al., 2019]] ; [[#Gettelman--2019|Gettelman et al., 2019]] ). However, this and other known model biases are already factored into the process-level assessment of cloud feedback ( [[#7.4.2.4|Section 7.4.2.4]] ), and furthermore the emergent constraints used here focus on global surface temperature change and are therefore less susceptible to shared model biases in individual feedback parameters than emergent constraints that focus on specific physical processes ( [[#7.5.4|Section 7.5.4]] ). The high values of ECS and TCR in some CMIP6 models lead to higher levels of surface warming than CMIP5 simulations and also the AR6 projections based on the assessed ranges of ECS, TCR and ERF (Box 4.1 and FAQ 7.3; [[#Forster--2020|Forster et al., 2020]] ).

It is generally difficult to determine which information enters the formulation and development of parametrizations used in ESMs. Climate models frequently share code components, and in some cases entire sub-model systems are shared and slightly modified. Therefore, models cannot be considered independent developments, but rather families of models with interdependencies ( [[#Knutti--2013|Knutti et al., 2013]] ). It is therefore difficult to interpret the collection of models ( [[#Knutti--2010|Knutti, 2010]] ), and it cannot be ruled out that there are common limitations and therefore systematic biases to model ensembles that are reflected in the distribution of ECS as derived from them. Although ESMs are typically well-documented, in ways that increasingly include information on critical decisions regarding tuning ( [[#Mauritsen--2012|Mauritsen et al., 2012]] ; [[#Hourdin--2017|Hourdin et al., 2017]] ; [[#Schmidt--2017a|Schmidt et al., 2017a]] ; [[#Mauritsen--2020|Mauritsen and Roeckner, 2020]] ), the full history of development decisions could involve both process-understanding and sometimes also other information such as historical warming. As outlying or poorly performing models emerge from the development process, they can become re-tuned, reconfigured or discarded and so might not see publication ( [[#Hourdin--2017|Hourdin et al., 2017]] ; [[#Mauritsen--2020|Mauritsen and Roeckner, 2020]] ). In the process of addressing such issues, modelling groups may, whether intentionally or not, modify the modelled ECS.

It is problematic and not obviously constructive to provide weights for, or rule out, individual CMIP6 model ensemble members based solely on their ECS and TCR values. Rather these models must be tested in a like-with-like way against observational evidence. Based on the currently published CMIP6 models we provide such an analysis, marking models with ECS above and below the assessed ''very likely'' range (Figure 7.19). In the long-term historical warming (Figure 7.19a) both low- and high-ECS models are able to match the observed warming, presumably in part as a result of compensating aerosol cooling ( [[#Kiehl--2007|Kiehl, 2007]] ; [[#Forster--2013|Forster et al., 2013]] ; [[#Wang--2021|Wang et al., 2021]] ). In several cases of high ECS models that apply strong aerosol cooling it is found to result in surface warming and ocean heat uptake evolutions that are inconsistent with observations ( [[#Golaz--2019|Golaz et al., 2019]] ; [[#Andrews--2020|Andrews et al., 2020]] ; [[#Winton--2020|Winton et al., 2020]] ). Modelled warming since the 1970s is less influenced by compensation between climate sensitivity and aerosol cooling ( [[#Jiménez-de-la-Cuesta--2019|Jiménez-de-la-Cuesta and Mauritsen, 2019]] ; [[#Nijsse--2020|Nijsse et al., 2020]] ) resulting in the high-ECS models in general warming more than observed, whereas low-sensitivity models mostly perform better (Figure 7.19b); a result that may also have been influenced by temporary pattern effects (Sections 7.4.4 and 7.5.4). Paleoclimates are not influenced by such transient pattern effects, but are limited by structural uncertainties in the proxy-based temperature and forcing reconstructions as well as possible differences in equilibrium sea surface temperature patterns between models and the real world ( [[#7.5.4|Section 7.5.4]] ). Across the LGM, MPWP and EECO (Figure 7.19c–e), the few high-ECS models that simulated these cases were outside the observed ''very likely'' ranges (see also [[#Feng--2020|Feng et al., 2020]] ; [[#Renoult--2020|Renoult et al., 2020]] ; [[#Zhu--2020|Zhu et al., 2020]] ). Also the low-ECS model is either outside or on the edge of the observed ''very likely'' ranges.

<div id="_idContainer068" class="Basic-Text-Frame"></div>

[[File:82b743adf7edf9bb6503d75269b423b1 IPCC_AR6_WGI_Figure_7_19.png]]

'''Figure 7.19''' '''|''' '''Global mean temperature anomaly in models and observations from five time periods. (a)''' Historical (CMIP6 models); '''(b)''' post-1975 (CMIP6 models); '''(c)''' Last Glacial Maximum (LGM; Cross-Chapter Box 2.1; PMIP4 models; [[#Kageyama--2021|Kageyama et al., 2021]] ; [[#Zhu--2021|Zhu et al., 2021]] ); '''(d)''' mid-Pliocene Warm Period (MPWP; Cross-Chapter Box 2.4; PlioMIP models; [[#Haywood--2020|Haywood et al., 2020]] ; [[#Zhang--2021|Zhang et al., 2021]] ); '''(e)''' Early Eocene Climatic Optimum (EECO; Cross-Chapter Box 2.1; DeepMIP models; [[#Zhu--2020|Zhu et al., 2020]] ; [[#Lunt--2021|Lunt et al., 2021]] ). Grey circles show models with ECS in the assessed ''very likely'' range; models in red have an ECS greater than the assessed ''very likely'' range (&gt;5°C); models in blue have an ECS lower than the assessed ''very likely'' range (&lt;2°C). Black ranges show the assessed temperature anomaly derived from observations ( [[IPCC:Wg1:Chapter:Chapter-2#2.3|Section 2.3]] ). The historical anomaly in models and observations is calculated as the difference between 2005–2014 and 1850–1900, and the post-1975 anomaly is calculated as the difference between 2005–2014 and 1975–1984. For the LGM, MPWP and EECO, temperature anomalies are compared with pre-industrial (equivalent to CMIP6 simulation ‘piControl’). All model simulations of the MPWP and LGM were carried out with atmospheric CO <sub>2</sub> concentrations of 400 and 190 ppm respectively. However, CO <sub>2</sub> during the EECO is relatively more uncertain, and model simulations were carried out at either 1120ppm or 1680 ppm (except for the one high-ECS EECO simulation which was carried out at 840 ppm; [[#Zhu--2020|Zhu et al., 2020]] ). The one low-ECS EECO simulation was carried out at 1680 ppm. Further details on data sources and processing are available in the chapter data table (Table 7.SM.14).

As a result of the above considerations, in this Report projections of global surface temperature are produced using climate model emulators that are constrained by the assessments of ECS, TCR and ERF. In reports up to and including AR5, ESM values of ECS did not fully encompass the assessed ''very likely'' range of ECS, raising the possibility that past multi-model ensembles underestimated the uncertainty in climate change projections that existed at the times of those reports (e.g., [[#Knutti--2010|Knutti, 2010]] ). However, due to an increase in the modelled ECS spread and a decrease in the assessed ECS spread based on improved knowledge in multiple lines of evidence, the CMIP6 ensemble encompasses the ''very likely'' range of ECS [2 to 5] °C assessed in ( [[#7.5.5|Section 7.5.5]] . Models outside of this range are useful for establishing emergent constraints on ECS and TCR and provide useful examples of ‘tail risk’ ( [[#Sutton--2018|Sutton, 2018]] ), producing dynamically consistent realizations of future climate change to inform impact studies and risk assessments.

In summary, the distribution of CMIP6 models have higher average ECS and TCR values than the CMIP5 generation of models and the assessed values of ECS and TCR in ( [[#7.5.5|Section 7.5.5]] . The high ECS and TCR values can in some CMIP6 models be traced to improved representation of extratropical cloud feedbacks ( ''medium confidence'' ). The ranges of ECS and TCR from the CMIP6 models are not considered robust samples of possible values and the models are not considered a separate line of evidence for ECS and TCR. Solely based on its ECS or TCR values an individual ESM cannot be ruled out as implausible, though some models with high (greater than 5°C) and low (less than 2°C) ECS are less consistent with past climate change ( ''high confidence'' ). High climate sensitivity in models leads to generally higher projected warming in CMIP6 compared to both CMIP5 and that assessed based on multiple lines of evidence (Sections 4.3.1 and 4.3.4, and FAQ 7.3).

<div id="7.5.7" class="h2-container"></div>

<span id="processes-underlying-uncertainty-in-the-global-temperature-response-to-forcing"></span>