Natural variability contributes to model–satellite differences in tropical tropospheric warming

Significance

Climate models have, on average, simulated substantially more tropical tropospheric warming than satellite data, with few simulations matching observations. It has been suggested that this discrepancy arises because climate models are overly sensitive to greenhouse gas increases. Tropical tropospheric temperature trends from a large ensemble of simulations performed with a single climate model span a wide range that is solely due to natural climate variability. A subset of these simulations have warming rates in accord with satellite observations. Simulations with diminished tropical tropospheric warming due to climate variability exhibit characteristic patterns of surface warming that are similar to the observed record. Our results indicate that multidecadal variability can explain current model–observational differences in the rate of tropical tropospheric warming.

Abstract

A long-standing discrepancy exists between general circulation models (GCMs) and satellite observations: The multimodel mean temperature of the midtroposphere (TMT) in the tropics warms at approximately twice the rate of observations. Using a large ensemble of simulations from a single climate model, we find that tropical TMT trends (1979–2018) vary widely and that a subset of realizations are within the range of satellite observations. Realizations with relatively small tropical TMT trends are accompanied by subdued sea-surface warming in the tropical central and eastern Pacific. Observed changes in sea-surface temperature have a similar pattern, implying that the observed tropical TMT trend has been reduced by multidecadal variability. We also assess the latest generation of GCMs from the Coupled Model Intercomparison Project Phase 6 (CMIP6). CMIP6 simulations with muted warming over the central and eastern Pacific also show reduced tropical tropospheric warming. We find that 13% of the model realizations have tropical TMT trends within the observed trend range. These simulations are from models with both small and large climate sensitivity values, illustrating that the magnitude of tropical tropospheric warming is not solely a function of climate sensitivity. For global averages, one-quarter of model simulations exhibit TMT trends in accord with observations. Our results indicate that even on 40-y timescales, natural climate variability is important to consider when comparing observed and simulated tropospheric warming and is sufficiently large to explain TMT trend differences between models and satellite data.

Pronounced tropical tropospheric warming is a consistent feature of general circulation model (GCM) simulations of historical climate change. Although observations show a significant increase in the temperature of the midtroposphere (TMT) over the last four decades (1), satellite-based estimates of the rate of tropical tropospheric temperature change are substantially smaller than the multimodel average (2). This puzzling and controversial discrepancy has persisted through multiple generations of increasingly sophisticated GCMs and different versions of the satellite datasets.

Spaceborne observations of tropospheric temperature have been recorded by microwave sounding unit (MSU) instruments since 1978 (3). By the late 2000s, the multimodel average tropical tropospheric warming from phase 3 of the Coupled Model Intercomparison Project (CMIP3) was two to six times larger than in satellite observations, depending on the time period, atmospheric layer, and dataset considered (4). In the subsequent generation of climate model simulations (CMIP5), despite updates to satellite datasets (5–7) and efforts to better isolate the tropospheric warming signal from stratospheric cooling in model–satellite comparisons (2), the multimodel average tropical TMT trend continued to exhibit two to three times more warming than observational products (2, 8). A recent analysis of CMIP6 simulations indicates that most models significantly overestimate the rate of tropospheric warming within the tropics and globally (9).

Several nonmutually exclusive explanations have been proposed to explain these model–observation differences. One explanation is that GCMs are too sensitive to increases in the atmospheric concentration of greenhouse gases (9–11). Natural climate variability and systematic model-forcing errors are also possible explanations. Recent studies note that the divergence between modeled and observed tropical tropospheric warming began in the early 2000s (12, 13). During this period, internal climate variability contributed to a slowdown in observed tropical tropospheric warming (14). The timing of such decadal variations in climate is random and should not be captured by coupled atmosphere–ocean model simulations, except by chance. Deficiencies in external forcing over the 2000s also contributed to exaggerated warming in climate models (15–17). It is likely that some combination of these factors, in addition to substantial observational uncertainty, explains the apparent differences in TMT change in models and observations (16, 18, 19).

Our focus here is on natural variability. We use all available historical simulations from CMIP6 (the newest generation of GCMs) and a large initial condition ensemble to determine whether natural internal climate variability can explain longstanding differences between model simulations and satellite observations.

CMIP6 Simulations and Satellite Observations

We begin by comparing model and observational tropical tropospheric temperature time series (Fig. 1A). Since standard midtropospheric temperature (i.e., TMT) products include a small contribution from the cooling of the stratosphere, the TMT data analyzed here are adjusted to remove stratospheric influence (Materials and Methods). The average observed tropical TMT anomaly time series is within the range of CMIP6 coupled atmosphere–ocean GCM simulations, but warms less than the multimodel mean in the early 21st century. This timing is consistent with research that shows a deceleration in observed warming in the early 21st century due to multidecadal variability (14, 20, 21).

Fig. 1.

(A) Time series of monthly tropical ($\mathrm{20}°$S–$\mathrm{20}°$N) tropospheric temperature anomalies for the average of four observational datasets (purple), the full range of CMIP6 AMIP models (teal), and the mean (black) and full range (gray) of CMIP6 coupled atmosphere–ocean models. All anomalies are with respect to the monthly climatology over 1979 to 1988. (B) Histogram of tropical TMT trends (1979–2014) for 482 coupled GCM simulations. Models with 10 or more realizations are color-coded in order of ensemble average trend (from blue to red; see key). All results from models with less than 10 historical realizations are included in gray. The observed range of tropical TMT trends is denoted with purple shading, and the teal line shows the probability density function of trends in AMIP simulations performed as part of CMIP6 (in units of [K$\cdot {\mathrm{d}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{d}\mathrm{e}}^{-1}$]⁻¹).

The observed tropical TMT time series is well captured by atmosphere-only simulations from the Atmospheric Model Intercomparison Project (AMIP), which use observed sea-surface temperatures (SSTs) as a boundary condition (Fig. 1A). The close agreement between satellite observations and AMIP simulations occurs because tropical tropospheric temperature is closely coupled to tropical SST (22–26). In 118 of the 163 AMIP simulations, simulated tropical TMT trends are within the range of satellite trends over 1979–2014 (Fig. 1B and SI Appendix, Table S1). If we employ alternate observed SST datasets as boundary conditions in a single model, the range of simulated tropical TMT trends expands (SI Appendix, Fig. S1). Two conclusions can be drawn from these results: The latest AMIP simulations are consistent with satellite-observed tropospheric warming estimates, and structural uncertainty in observed SST datasets is sufficiently large to affect model–data trend consistency.

Unlike the AMIP simulations, CMIP6 coupled atmosphere–ocean simulations typically have larger tropical TMT trends compared to the observations (Fig. 1B). The CMIP6 multimodel mean tropical TMT trend over 1979 to 2014 is 0.30 K$\cdot {\mathrm{d}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{d}\mathrm{e}}^{-1}$, which is approximately 1.5 to 3.3 times larger than observed MSU trends (0.09 to 0.20 K$\cdot {\mathrm{d}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{d}\mathrm{e}}^{-1}$). This is consistent with results from earlier CMIP5 models: The CMIP5 multimodel mean trend over the same time period (extended using the Representative Concentration Pathway 8.5 experiment; SI Appendix) is 0.32 K$\cdot {\mathrm{d}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{d}\mathrm{e}}^{-1}$ (2). Since systematic errors in volcanic and solar forcing in CMIP5 models in the early 21st century (15, 17) were largely removed in CMIP6 simulations, these specific deficiencies in external forcing should play a smaller role in the model-versus-observed differences in tropical TMT trends assessed here. We find that 13% of CMIP6 historical simulations are within the range of current observational tropical TMT trend estimates (61 of 482 simulations in 14 of the 55 CMIP6 coupled GCMs considered; SI Appendix, Table S2). If we extend a subset of the historical simulations through 2018, model–observational consistency does not improve substantially (SI Appendix, Table S2).

Other recent research analyzing tropical midtropospheric temperature trends in CMIP6 reported that every model simulation warmed more than the average satellite observed trend (across four observational datasets) (9). As noted above, we find greater model–observational agreement. There are three reasons for this. First, we consider the range of uncertainty arising from different observational TMT datasets, rather than the observational average. Second, we analyze a much larger set of CMIP6 output (482 simulations from 55 models compared to 38 simulations from 38 models). Third, we remove the influence of stratospheric cooling from TMT in both models and observations. Over the satellite era, stratospheric temperature trends are primarily driven by human-caused ozone depletion. Systematic model errors in specifying changes in stratospheric ozone or ozone-depleting substances can have substantial impact on lower stratospheric temperature, and hence on model-versus-observational comparisons of “raw” TMT trends (2).

Characteristics of Suppressed Tropical Tropospheric Warming

Fig. 1B shows that individual CMIP6 coupled models can produce a wide range of satellite-era tropical TMT trends through multidecadal internal variability alone. To further elucidate the role of internal variability, we employ the Community Earth System Model Large Ensemble (CESM1 LENS; ref. 27). This consists of 40 individual realizations of historical climate change, each performed with the same model starting from different atmospheric initial conditions. Differences between realizations are solely due to differences in the evolution of internal climate variability. As in the CMIP6 multimodel ensemble, the CESM1 LENS ensemble-mean tropical TMT trend exceeds the range of tropical TMT trends in satellite observations (Fig. 2A), but the ensemble also contains three individual realizations that are within the observed range.

Fig. 2.

(A) Time series of tropical-average ($\mathrm{20}°$S–$\mathrm{20}°$N) tropospheric temperature anomalies for four observational datasets (purple) and for the mean (black) and full range (gray) in 40 CESM1 LENS simulations. All anomalies are with respect to the monthly climatology over 1979 to 1988. Also shown are the time series of LENS realization 13 (red) and realization 8 (blue), which represent the ensemble members with the largest and smallest Ni$\tilde{\mathrm{n}}$o 3.4 region SST trends (respectively). The linear trend (in parentheses) and least-squares fit are shown for each time series. (B) Map for the intraensemble regression of tropical-average $-$TMT trend (1979–2018) onto the local SST trend (Materials and Methods). The black box denotes the Ni$\tilde{\mathrm{n}}$o 3.4 region. (C) As in B, but using equatorial ($\mathrm{5}°$S–$\mathrm{5}°$N) ocean potential temperature instead of SST. Stippling in B and C denotes statistical significance (P $<$ 0.05). (D) Tropical TMT trend versus the SST trend in the Ni$\tilde{\mathrm{n}}$o 3.4 region. Ensemble members 8 and 13 are colored blue and red, respectively. The purple box shows the range of observed trends.

To isolate the pattern of ocean-temperature change accompanying tropical TMT trends that are suppressed due to internal climate variability, we regress the local SST and ocean potential temperature trends against the tropical-average $-$TMT trend across the 40 CESM1 LENS members (note the minus sign). This calculation yields the pattern of ocean-temperature change associated with reduced tropical tropospheric warming (Fig. 2 B and C). Principal component analysis of the 40 SST trend maps (SI Appendix, Fig. S2 A and B) produces a similar pattern to the regression map in Fig. 2B, indicating that the dominant mode of tropical variability in patterns of ocean-surface warming is closely related to variability in the tropospheric temperature trend.

The regression pattern in Fig. 2 is similar to the pattern of temperature anomalies associated with the negative phase of the El Ni$\tilde{\mathrm{n}}$o–Southern Oscillation (ENSO) (typically referred to as La Ni$\tilde{\mathrm{n}}$a) and the negative phases of the Interdecadal Pacific Oscillation (IPO) and the Pacific Decadal Oscillation (28–31). LENS realizations with the imprint of this La Ni$\tilde{\mathrm{n}}$a-like mode of multidecadal variability contain smaller than average tropical TMT trends. These realizations exhibit enhanced warming in the equatorial western Pacific at a depth of $\sim$150 m and reduced warming (or cooling) across the central and eastern Pacific sea surface and near-surface (Fig. 2 B and C).

The central Pacific Ni$\tilde{\mathrm{n}}$o 3.4 region (the black box in Fig. 2B) is a common metric of ENSO variability. Reduced warming of SSTs in this region accompanies smaller than average tropical-mean SST warming (SI Appendix, Fig. S2C) and reduced tropical TMT trends (Fig. 2 B and D). The scaling between warming in the central Pacific and tropical average TMT (and SST) trends is consistent with a large body of research indicating that ENSO is the main driver of large interannual variations in tropical tropospheric temperature. Sea-surface warming in the central tropical Pacific drives an exchange of energy from the ocean into the atmosphere, resulting in atmospheric convection and heating (32, 33). In turn, atmospheric warming is rapidly propagated throughout the tropical free troposphere, resulting in remote tropical surface warming via moist convective processes (34).

The SST trend in the Ni$\tilde{\mathrm{n}}$o 3.4 region is an accurate predictor of the tropical TMT warming across the CESM1 LENS (Fig. 2D). The SST trend in the Ni$\tilde{\mathrm{n}}$o 3.4 region varies due to internal climate variability, but also includes a contribution from externally forced warming (Fig. 3A), which is why the trend in this region is above zero for every ensemble member. Only 1 of the 40 ensemble members approaches the near-zero Ni$\tilde{\mathrm{n}}$o 3.4 trend in the observations. Ensemble members with reduced warming in the Ni$\tilde{\mathrm{n}}$o 3.4 region have larger increases in tropical ocean heat uptake (SI Appendix, Fig. S2D). Changes in other Pacific-based indices of tropical–extratropical climate variability, such as the trend in the IPO (29), also scale with the tropical TMT trends across the CESM1 LENS (SI Appendix, Fig. S2E).

Fig. 3.

SST and equatorial ocean potential temperature trend (1979–2018) for the CESM1 LENS ensemble mean (A), Ensemble Member 8 (with the smallest Ni$\tilde{\mathrm{n}}$o 3.4 region SST trend) (B), and the average observed SST trend across four observational datasets (Materials and Methods) and potential temperature change from Ocean Reanalysis System 5 (C). The Ni$\tilde{\mathrm{n}}$o 3.4 region is the black box in the SST trend maps.

As expected, the forced (i.e., ensemble average) SST and near-surface ocean potential temperature trend have relatively homogeneous warming patterns (Fig. 3A). Individual ensemble members exhibit more spatial structure. For example, ensemble member 8 has the smallest Ni$\tilde{\mathrm{n}}$o 3.4 trend, the La Ni$\tilde{\mathrm{n}}$a-like pattern signature seen in the regression maps in Fig. 2 B and C, and suppressed tropical TMT warming (Figs. 2D and 3B). The tropical TMT trend of realization 8 is 0.22 K$\cdot {\mathrm{d}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{d}\mathrm{e}}^{-1}$, which is $\sim$20% lower than the ensemble average trend of 0.27 K$\cdot {\mathrm{d}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{d}\mathrm{e}}^{-1}$.

The SST and ocean-temperature changes estimated from observations and ocean reanalysis are very similar to those in ensemble member 8 (Fig. 3 B and C). The pattern correlation between the observed SST change and the SST change in ensemble member 8 is 0.64. In contrast, ensemble members with Ni$\tilde{\mathrm{n}}$o 3.4 trends near the median or maximum in the CESM1 large ensemble have SST trend patterns that are dissimilar to or anticorrelated with the observed SST change (SI Appendix, Fig. S3).

The similarity in the pattern of ocean-temperature change in ensemble members with less-than-average warming and observations/reanalysis suggests that internal climate variability has markedly contributed to reduced tropical TMT warming over the satellite era. The overlap between simulated and observed tropical TMT trends demonstrates that multidecadal climate variability (as manifest in CESM1 LENS) is sufficiently large to explain the difference between tropical TMT trends in the LENS ensemble average and satellite data.

CMIP6 Tropical TMT Warming Modulated by Multidecadal Pacific SST Variability

As in the case of the CESM1 LENS, the between-realization variability of the SST warming pattern in individual CMIP6 models is dominated by an ENSO-like structure (SI Appendix, Fig. S4). CMIP6 simulations with reduced SST trends in the tropical central and eastern Pacific have reduced tropical TMT warming (Fig. 4 and SI Appendix, Fig. S5A). In accord with the CESM1 LENS results, Fig. 4 also shows that individual models can simulate appreciable variability in tropical central Pacific warming (horizontal lines), which translates into a wide range of tropical TMT trends for individual model ensembles (vertical lines in Fig. 4). Model realizations within the observed range of tropical TMT trends, however, do not always reproduce the observed pattern of SST changes (SI Appendix, Fig. S6).

Fig. 4.

Tropical TMT trend (1979–2014) versus the Ni$\tilde{\mathrm{n}}$o 3.4 region SST trend for 482 CMIP6 historical simulations. Models are color-coded based on ECS, with the lowest ECS values in dark blue and the highest ECS values in dark red. The range of observational values is denoted by the purple box. Horizontal lines indicate the range of Ni$\tilde{\mathrm{n}}$o 3.4 trends for individual models with five or more ensemble members. The legend also provides the number of ensemble members, n, and the correlation coefficient, r, between the Ni$\tilde{\mathrm{n}}$o 3.4 and tropical TMT trends for each model. Right Upper provides a similar color-coded display of the range of tropical TMT trends plotted against ECS (abscissa). Models with more than one ensemble member are displayed as vertical lines, and models with only one ensemble member are denoted with dots.

A basic metric of the climate’s response to external forcing is effective climate sensitivity (ECS), which is the global average surface-temperature response to a doubling of atmospheric carbon dioxide concentration (35). There is a tendency for models with larger ECS values to have larger tropical TMT trends (Fig. 4). The correlation coefficient between each model’s ensemble average tropical TMT trend and ECS is 0.59 (SI Appendix, Fig. S5B), consistent with suggestions that exaggerated model sensitivity to greenhouse gases is contributing to model–observational differences in the rate of tropical tropospheric warming (11). Indeed, models with larger climate sensitivity values tend to have fewer simulations in accord with satellite observations (Fig. 4). For models with an ECS value between 2 and 3 K, 23% of simulations are within the range of observed tropical tropospheric temperature trends. This value falls to 18% (8%) for models with ECS values between 3 and 4 K (4 and 5 K). In general, models with larger ECS tend to have greater tropical TMT trends and require a greater contribution from multidecadal internal variability to demonstrate consistency with satellite observed trends (SI Appendix, Fig. S5).

TMT trends are not determined by ECS alone, however. This is evident from the wide spread in TMT trends among realizations of a single model and from the fact that models with higher ECS values do not invariably have larger tropical TMT trends than models with lower ECS values. Other factors that likely influence simulated TMT trends include model representation of ocean heat uptake and the size and time evolution of anthropogenic aerosol forcing (36).

While it is conceivable that model biases in climate sensitivity are the most important cause of disagreement between modeled and observed tropical TMT warming rates, it is also possible that internal variability is partly masking the substantial warming that would be expected if the Earth had a large climate sensitivity value. Consistent with the latter interpretation, CMIP6 models with both low and high ECS values (i.e., ECS $<$ 2 K and ECS $>$ 4 K) have simulations within the range of observed tropical tropospheric temperature trends. As noted above, CESM LENS can also produce tropical TMT trends consistent with observed results, despite the relatively large climate sensitivity of that model (CESM1-CAM5, 4.1 K).

The preceding discussion pertained to the tropics. For global averages, a greater fraction of model simulations are in accord with satellite estimates of tropospheric warming: 24% of all CMIP6 simulations (118 of 482) yield trends within the range of global TMT observations. As in the case of the tropics, models with ECS values as low as 1.9 K and as large as 4.9 K produce warming in accord with observations (SI Appendix, Fig. S7). Taken together, the tropical and global results indicate that both internal climate variability and model ECS biases may contribute to disagreement between modeled and observed tropical TMT warming rates. Further research is required to determine the relative role of each of these factors. Our analysis does not support the conclusion that model ECS biases are the sole determinant of differences in modeled and observed tropospheric warming rates.

GCMs have widely varying representations of decadal variability in the tropical Pacific (Fig. 4). Some model ensembles simulate a large range of SST trends in the Ni$\tilde{\mathrm{n}}$o 3.4 region (e.g., CESM2), while others have a relatively small spread (e.g., INM-CM5). Since the real-world magnitudes of forced Ni$\tilde{\mathrm{n}}$o 3.4 warming and internally generated climate variability are unknown, it is difficult to deconvolve the influence of climate sensitivity and natural variability on observed tropical tropospheric warming. Despite this difficulty, our results strongly suggest that the observed warming of the tropical troposphere has been diminished due to the imprint of La Ni$\tilde{\mathrm{n}}$a-like multidecadal variability. Although only 13% of CMIP6 ensemble members have tropical TMT trends matching the observations, this number increases to 47% (35 of 74 simulations) when we consider CMIP6 ensemble members with a Ni$\tilde{\mathrm{n}}$o 3.4 trend that is at least one SD smaller than the multimodel average. Thus, even on timescales of 40 y, the effect of climate variability needs to be considered when comparing the CMIP6 average tropical TMT trend (which provides an estimate of the “pure” signal in response to external forcing) with satellite observations which comprise both signal and the noise of internal variability.

Discussion

Multiple factors influence the observed satellite-era tropical TMT trend, including the magnitude of external forcing changes, the climate response to external forcing, ocean heat uptake, and internal climate variability. Recent research has found that the global surface-temperature change since the late 1970s scales with the transient climate response in GCMs. Constraining this relationship with observations of surface or tropospheric temperature change allows analysts to estimate the transient and equilibrium climate response (9, 37, 38). These emergent constraints require that aerosol forcing uncertainty and multidecadal variability have little effect on global surface-temperature change.

Consistent with previous work that considered shorter timescales (14, 19), we have shown that multidecadal variability has a substantial impact on 40-y tropical tropospheric temperature trends. Observations of suppressed warming over the central and eastern tropical Pacific Ocean, taken together with the appreciable influence of model-simulated internal variability in this region (21, 39, 40), suggest that the muted observed trend in Ni$\tilde{\mathrm{n}}$o 3.4 SSTs contributed to reduced warming in observations of tropical tropospheric temperature change. Individual model realizations are capable of replicating the observed muted Ni$\tilde{\mathrm{n}}$o 3.4 SST trend and its impact on tropical TMT trends, but the average of a large ensemble or multimodel ensemble will primarily reflect the forced response. It is reasonable to ask whether the observed trend lies within the model-generated trend distributions, but it is not reasonable to require that the model average trend matches the single realization of forced response and internal variability in the observations.

We note that multidecadal variability can also result in apparent model underestimates of observed climatic changes. Both surface and radiosonde records contain periods in which observed warming is at the upper bound of model-simulated warming (19, 41). In addition to its impacts on temperature, multidecadal variability has also contributed to rapid, observed Arctic sea-ice loss (42) and to changes in the strength of the Pacific Walker Circulation (43, 44). The presence of such large multidecadal climate variability must be accounted for when using observational data to construct emergent constraints on transient or equilibrium climate sensitivity.

Disentangling and quantifying the competing influences of forcing, response to forcing, and multidecadal variability on tropospheric temperature change is an important challenge, particularly in view of the nontrivial uncertainties in each of these factors. In addition to natural internal variability, systematic model-forcing biases may contribute to model–observational differences in the rate of tropical tropospheric warming (36, 45). A further complication is that satellite observations of TMT changes are likely to contain residual biases, which could improve or diminish model–observational agreement (7, 8).

Model biases in the forced climate response may also play a role. GCMs tend to weaken the zonal gradient in equatorial Pacific SST, whereas observations show a strengthening of this gradient on multidecadal and century-long timescales, suggesting that the model forced response may be biased (46–48). Such model biases would contribute to model–observational discrepancies of tropical tropospheric warming. We note, however, that recent studies make use of large model ensembles to show that internal variability has a substantial impact on changes in the equatorial Pacific SST gradient since the mid-19th century (40, 49). Some large ensembles have a number of realizations in accord with the observed changes in the equatorial Pacific SST gradient (40). In other ensembles, model–observational agreement is marginal, pointing toward model deficiencies in simulating the pattern of forced SST changes. Nevertheless, future reversals of Pacific decadal climate variability may result in enhanced tropical tropospheric warming in the coming decades (40).

CMIP6 models tend to exhibit larger ECS values compared to CMIP5 models (35). Our study indicates that models with larger climate sensitivity values tend to have greater tropical TMT trends. To be consistent with high ECS values, observed tropical TMT trends would therefore need to include an appreciable contribution from multidecadal climate variability to explain large ECS values (and/or substantial negative radiative forcing from anthropogenic aerosols).

Multicentury paleoclimate reconstructions of ENSO demonstrate substantial multidecadal variability (50), making it difficult to rule out large values of ECS. A recent assessment of Earth’s climate sensitivity suggests that ECS is likely between 2.6 and 3.9 K (51); CMIP6 models with ECS values in this range routinely produce tropical TMT trends in accord with satellite observations (20% of realizations; SI Appendix, Table S2). Our results clearly indicate that multidecadal variability is an important factor that needs to be accounted for when evaluating how well current climate models perform in reproducing satellite observations of tropospheric warming.

Materials and Methods

GCM Data and Analysis.

We utilize coupled atmosphere–ocean model simulations of the historical record (1979–2014) from CMIP6 (52). The simulations are driven by specified historical changes in natural and anthropogenic external forcings. To complement data from CMIP6, we also use utilize simulations from the CESM1 large ensemble (1979–2018). The CESM1 LENS consists of 40 simulations using the same model (CESM version 1 with Community Atmosphere Model version 5) and identical time-varying external forcings (27). In order to compare data from these numerical experiments with satellite observations, we use a local weighting function approach to calculate synthetic satellite brightness temperatures (2).

To determine patterns of SST change ($\beta$) that correspond to suppressed tropical TMT trends in CESM1 LENS, we regress the tropical-average $-$TMT trend (1979–2018) on the local SST trend such that $\mathrm{\Delta }\mathrm{S}\mathrm{S}\mathrm{T}\left(\mathrm{i},\overrightarrow{x}\right)=\beta \left(\overrightarrow{x}\right)\cdot -\mathrm{\Delta }\mathrm{T}\mathrm{M}\mathrm{T}\left(\mathrm{i}\right)$. This regression uses the 40 tropical-average $-$TMT trend values, $-\mathrm{\Delta }\mathrm{T}\mathrm{M}\mathrm{T}$, and the SST trend, $\mathrm{\Delta }\mathrm{S}\mathrm{S}\mathrm{T}$, for each ensemble member, i, at each location, $\overrightarrow{\mathrm{x}}$ (the ensemble average is removed from $\mathrm{\Delta }\mathrm{S}\mathrm{S}\mathrm{T}$ and $\mathrm{\Delta }\mathrm{T}\mathrm{M}\mathrm{T}$ before the regression is performed). We perform the same procedure on ocean potential temperature to recover the analogous pattern of ocean-temperature change.

We use model ECS values derived from 150-year ${\mathrm{C}\mathrm{O}}_2$ quadrupling experiments (35); values are calculated by using the Gregory method (53). Reported ECS values are from an update to Zelinka et al. (2020) (35).

Observations.

Satellite observations of tropospheric temperature are based on passive microwave soundings from more than a dozen satellites. Multisatellite merged datasets in Fig. 1A are from Remote Sensing Systems, the National Oceanic and Atmospheric Administration (NOAA) Center for Satellite Applications and Research, the University of Washington (5), and the University of Alabama at Huntsville (6, 7, 54). Microwave-based tropospheric measurements include an influence from the stratosphere. We remove stratospheric contamination from both the satellite observations and the GCM synthetic brightness temperatures using a standard regression approach (55).

We rely on four primary SST datasets. These are from the Program for Climate Model Diagnosis and Intercomparison (56), the Centennial In Situ Observation-Based Estimates of the Variability of SST and Marine Meteorological Variables version 2 (57), the NOAA Extended Reconstructed SST version 5 (58), and the Hadley Center Sea Ice and SST dataset version 1 (59). For SST trend maps, we use the average of these four records. We also analyze 100 realizations of the Hadley Center SST Dataset version 3 to estimate the observational range in Ni$\tilde{\mathrm{n}}$o 3.4 SST warming (60, 61). We use the average of five realizations from the Ocean Reanalysis System version 5 to estimate changes in ocean potential temperature (62).

Data Availability

All datasets used here are publicly available. Synthetic satellite temperature time series calculated from model simulations are provided at Program for Climate Model Diagnosis & Intercomparison (https://pcmdi.llnl.gov/research/DandA/). CMIP6 data is available from Earth System Grid Federation (ESGF) (https://esgf-node.llnl.gov/projects/cmip6/). CESM LENS data is available at Community Earth System Model (CESM) (https://www.cesm.ucar.edu/projects/community-projects/LENS/).