We use output from the historical simulations of the three latest generations of CCMs: CCMVal-2 refB1 , CCMI-1 refC1 , and CCMI-2022 refD1 . These are atmosphere-only historical simulations that use observations as boundary conditions for anthropogenic and natural emissions and prescribe sea surface temperature (SST) from observations. The quasi-biennial oscillation (QBO) is relaxed (nudged) towards observational winds in most simulations, but some models have an internally generated QBO. The time periods considered for these simulations (over which most of the models provided output) are 1960–2000 for CCMVal-2, 1960–2010 for CCMI-1, and 1960–2018 for CCMI-2022. We highlight that the different time periods covered by the generations implies that it is not possible to have a long overlap period that also covers the most recent decades when high quality observations are available. This implies that the periods are not exactly the same in the different datasets compared. We tried to minimize the impact of this caveat by averaging over long (20 years or more) time periods, adjusted to the different diagnostics and datasets.

The list of models used for each generation is provided in Tables , , and , along with the variables used for each model. The main characteristics and reference for the models are given in Appendix A. We consider all available output from all models in each generation unless the output has known issues. Note that this means that the number of models differs between generations and for different diagnostics, which can affect the spread of the multi-model figures. The alternative was to include only the models that contributed to all three intercomparisons, but this would have greatly reduced the number of models for some diagnostics, as few models output all variables across all three intercomparisons. In addition, given that not all models output all metrics, the different diagnostics are evaluated for a different set of models. The eight model families which participated in the three phases are identified in the right column of Tables –, these are models that are considered developments of the same model. In case a model provides several realizations (members) for a given simulation, the ensemble mean is considered.

As stated above, the figures in the main text show the multi-model median and spread for the three generations. The median is used instead of the mean because it is less sensitive to outliers and thus more representative of the central value of the distribution. The inter-model spread is computed as a symmetric interval of width equal to the median absolute deviation (MAD), which is the median of the absolute deviations from the median. The MAD is multiplied by a correction factor dependent on the number of simulations included to make it a good estimate of the standard deviation . Having a symmetric representation of the spread around the median is not appropriate when the size of the sample is small. Hence, for variables in which we have only a limited number of model simulations (∼ 7), we include the spread as the 15th and 85th percentiles of the distribution. This interval encloses a number of simulations similar to considering one standard deviation in a Gaussian distribution, and for a small sample of ∼ 7 models, it typically leaves 1–2 outliers out of the envelope.

Table presents an overview of the diagnostics, together with the variables needed and the corresponding observational dataset used for evaluation.

A particularly useful diagnostic to quantify the strength of stratospheric transport is the mean age of air (AoA), which measures the average time of residence of an air parcel since it entered the stratosphere through the tropical tropopause . AoA is easily represented in models and can be estimated from long-lived tracer concentrations, and it is thus useful to evaluate models against observations. The AoA includes the effects of advection by the overturning circulation and those of mixing. The tropical leaky pipe model TLP; provides a simplified picture of the stratosphere as a tropical region or “pipe” with ascending air, and descending air outside the pipe, with leaky barriers in between that allow for moderate amounts of mixing. The TLP model provides a useful tool for process understanding of the effects of individual processes on age of air. For example, showed that mixing across the subtropical barriers results in air reentering the tropical ascent region, thus recirculation of air masses, leading to an overall increase of mean age. Diagnostics based on the TLP model formulation were later applied to CCM and CTM data, confirming the aging effect of mixing on age in three-dimensional models . Likewise, showed theoretically that horizontal mixing serves to increase the overall AoA while keeping the difference between AoA in the tropics and extratropics constant. Computing the time of residence that would be due exclusively to the advection by the residual circulation (RCTT), and subtracting it from the AoA gives the mean aging by mixing. However, this aging by mixing also is affected by the residual circulation through the recirculation speed. Thus, a better measure of the effects of mixing is the diagnostic of mixing efficiency, which is computed as : 1ϵ=AoA-RCTT(RCTT-Tcorr)α+1α where Tcorr=H1wT‾*(z)-1wT‾*(zT) is the time of tropical ascent from the tropopause to a level z, with w‾*(z) the corresponding tropical upwelling, and α is the ratio of tropical to extratropical mass. Here, following , the mixing efficiency is calculated as the best fit of Eq. (1) to the tropical profiles of AoA and wT‾*(z) from the model data over a certain height range. This metric is a gross estimation of bulk mixing properties and does not represent either vertical or temporal variations. In addition, it includes effects of vertical diffusion .

The advective component of the BDC is typically quantified with the residual circulation in the Transformed Eulerian Mean (TEM) framework . The simple Eulerian zonal mean meridional circulation does not accurately represent the actual zonal mean transport because the zonal asymmetries (eddies) dominate in the stratospheric midlatitudes. The TEM residual circulation accounts for the effects of the eddies and thus is a good approximation of the zonal mean Lagrangian circulation, while keeping the advantages of an Eulerian calculation. The residual circulation components are defined as : 2v‾*=v‾-∂∂pv′θ′‾∂θ‾/∂p3ω‾*=ω‾+1acos⁡φ∂∂φv′θ′‾cos⁡φ∂θ‾/∂p

We convert the pressure velocity (Pa s⁻¹) provided by models and reanalyses into vertical velocity (m s⁻¹) by assuming a constant scale height of H=7 km. To compute the net upward mass flux of the overturning circulation, the upwelling must be integrated over the entire region between turnaround latitudes (with w‾*>0), which shift seasonally towards the summer hemisphere. Making use of mass continuity, a common way of computing this is by subtracting the maximum from the minimum values of the residual circulation streamfunction e.g.: 4MF=2πaΨ‾*(φTASH)-Ψ‾*(φTANH) where φTASH and φTANH are the southern and northern turnaround latitudes, respectively, where the residual streamfunction maximizes (and w‾*=0). We first compute the streamfunction Ψ‾* by integrating the meridional component of the residual circulation (v‾*) vertically from each level to p=0, assuming zero velocity at the top. We checked that the results remain qualitatively similar if the upwelling mass flux is computed by integrating the density-weighted vertical residual velocity between the turnaround latitudes, although the numerical values change slightly because both approaches are affected by numerical errors (not shown).

In addition to the residual circulation metrics that can only be compared with reanalyses, we include other metrics that help evaluate the strength of the advective component of the BDC against satellite observations, including an estimate of the mean global overturning circulation from the difference between age of upwelling and downwelling age and an estimate of tropical upwelling from the phase propagation of the water vapor tape recorder .

A set of diagnostics is used to evaluate mixing between the tropics and extratropics. We examine the subtropical transport barriers (STBs) based on the probability density function (PDF) of N2O across the tropics and extratropics. N2O is a long-lived tracer widely used in stratospheric transport studies e.g.,. The PDFs allow distinguishing the tropical and extratropical N2O typical values (two peaks in the frequencies), and the strong gradients formed at the edge of the surf zones, which produce a valley in the PDF between the two maxima. The separation of the two peaks and the depth of the minimum are compared to satellite results. For these calculations, the N2O monthly mean output on longitude, latitude, and pressure is first interpolated onto a common grid with potential temperature as the vertical coordinate. We checked with one model (CESM1-WACCM) that subsampling the daily model output to match the sampling pattern of the satellite observations did not modify the results substantially, except in the case of ACE-FTS due to its coarse sampling (Supplement, Fig. S13). We also checked with that model that using daily mean output instead of monthly mean does not significantly modify the results (Supplement, Fig. S12). Thus, based on these tests we decided to use the monthly mean output to reduce the computation burden. For satellite observations, we use level-2 data, as the limited sampling does not allow the monthly mean output to be representative. From the PDFs, we then identify the position of the STBs by finding the latitude that corresponds to the minimum in the PDFs, following the approach of and . This is the only quantitative metric we present from the N2O PDFs, but qualitatively the shape of the PDF (the sharpness of the peaks and separation between them) is linked to the circulation strength and the strength of mixing. For the polar regions we include diagnostics of the polar downwelling from the residual circulation and from the N2O tendency and vertical structure. In addition, the wave activity entering the stratosphere (eddy heat flux at 100 hPa averaged over 45–75° N/S) and the final warming date are evaluated against reanalyses.

We begin by showing the annual mean climatology of AoA in the three model generations and the two satellite datasets from in Fig. . Note that throughout the paper AoA is normalized by subtracting the value at the equator and at 100 hPa in each dataset, to remove differences due to the definition of the level of zero age (some models use the tropopause, some a fixed level in the upper troposphere, some at/near the surface). It is clear that all the CCMs underestimate the AoA, especially in the extratropical middle and upper stratosphere (compare Fig. a–c with d–e). Also, the newest generation of models exhibits the youngest median values. Figure f shows the latitudinal structure of the annual mean AoA at 50 hPa, which is a classical diagnostic in stratospheric transport evaluation studies, commonly referred to as the “wing plot” due to its characteristic shape e.g.,. This level is typically chosen because it maximizes the available in situ (balloon and aircraft) observations. Although not included in this study, we note that the in situ measurements are broadly consistent with the satellite observations and lay on the upper edge of the CCMI-2022 envelope, as shown in Fig. 10 and Fig. 11.

In all generations the spread in the mean age values across models increases with latitude, reaching up to 1.5 years in polar regions for CCMI-2022. In the tropics, the MMMed of the three generations gives an AoA of around 1 year at 50 hPa, which is in good agreement with the ACE-FTS satellite estimate. At higher latitudes, the agreement with observations decreases, and the models have a young bias throughout the extratropics. This bias is largest in the newer generation for which the observations fall on or above the upper edge of the spread. The mean age averaged globally and over 100–15 hPa is 2.3, 2.1 and 1.9 years in CCMVal-2, CCMI-1 and CCMI-2022, respectively, and 3.1 and 2.7 years in MIPAS and ACE-FTS, respectively. The young bias is more severe at upper levels, as shown for 10 hPa in Fig. g. This is true for all three generations, but especially again for the newest generation, which has the youngest mean age across all latitudes, with values in the polar regions that are 1.5 years younger than the observational estimate. A view of globally-integrated AoA as a function of altitude (Fig. S3 in the Supplement) confirms that the MMMed bias and the size of the inter-model spread increase with altitude. Because the mean age is an integrated transport diagnostic, the increase in bias with altitude and latitude could reflect the accumulation of transport biases as air masses ascend and move poleward through the stratosphere. It is important to note that a different number of models and different families are included in each generation. In particular, only three models provide the mean age across the three generations (CESM-WACCM, CMAM and GEOSCCM). Looking at the mean age of air wing plots for individual models across generations (see plot for common models in Fig. S2 of the Supplement) it becomes clear that there is no consistent change for the different models, with CMAM AoA staying approximately unchanged, GEOSCCM getting subsequently younger in each generation and CESM-WACCM getting younger from CCMVal-2 to CCMI-1 and then slightly older in CCMI-2022. On the other hand, the UK model family (F7 in Tables –) provided mean age ouput in CCMVal-2 and CCMI-2022, and the value changed from very old (UMUCKA-UKCA) to very young (UKESM-StratTrop). Similarly, the EMAC model participated in the last two intercomparisons (EMAC-L90MA and EMAC-CCMI2), and the newer generation also shows younger age. Thus, the participation of different models in different generations does not explain on its own why the mean age has gotten younger in the new generation. Note that the shift toward a younger mean age in GEOSCCM models was recently highlighted and investigated in .

We now explore in more detail the advective component of the BDC, the mean meridional overturning circulation. As described in Sect. , we compute the net upward mass flux to account for the fact that the region of tropical upwelling is not fixed but moves seasonally towards the summer hemisphere.

Figure a shows the vertical profile of the annual mean upward mass flux from the tropical tropopause to the middle stratosphere. The mass flux in our observation-based references, the reanalyses, decreases rapidly with height due to the exponential decay of air density, with values around 10 × 109 kg s⁻¹ at the tropopause decreasing to 1.5 × 109 kg s⁻¹ at 10 hPa. However, there is a clear difference between the reanalyses, with ERA5/5.1 exhibiting larger values than the other two reanalyses below 20 hPa. The CCMs are generally able to capture the mean vertical structure of the overturning mass flux seen in the reanalyses. In the range 50–20 hPa, where ERA5 shows larger values than the other two reanalyses, the models agree better with MERRA-2 and JRA-3Q. Below 50 hPa, in contrast, the models agree better with the higher values found in ERA5/5.1. In this lower stratosphere region, MERRA-2 and JRA-3Q are outside the inter-model spread range of the models for the last two generations (CCMI-1 and CCMI-2022). The oldest generation (CCMVal-2) shows a wider spread, which includes all reanalyses.

Figure b shows the seasonal cycle in the net upward mass flux in the lower stratosphere (70 hPa). The difference between ERA5/5.1 and the other reanalyses is concentrated in the boreal winter months (November to March), overall leading to a larger amplitude in ERA5/5.1's seasonal cycle. Across the three generations, the models MMMed show a larger seasonality (∼ 40 %) compared to MERRA-2 and JRA-3Q (∼ 30 %), although still smaller compared to ERA5 (∼ 50 %). The inter-model spread is largest throughout the year in the oldest generation (CCMVal-2), and notably reduced in the newest generation (CCMI-2022), although the MMMed is very similar to that from the previous generation (CCMI-1). As will be shown, the reduced spread is a common feature of the newest model generation found across several metrics (but not in all). At higher levels, the annual cycle is reduced (with the semi-annual variability becoming dominant, e.g., ) and the agreement between model generations and reanalyses increases as the magnitude becomes smaller (Fig. S5 in the Supplement).

As mentioned in Sect. , an alternative way to derive the mean meridional overturning circulation that enables direct comparison with satellite data is based on age and the TLP model. This method has a the advantage of being an estimate of the overturning circulation that is independent from the residual circulation mass flux estimate in Fig. . On the other hand, there are important caveats associated with this observational estimate, as will be discussed below. Previous studies have shown that the difference in mean age between the tropics and the extratropics in isentropic coordinates is inversely proportional to the mean strength of the overturning circulation, and does not depend on mixing .

Figure shows this age gradient metric computed from the mean age output for the three generations of models and estimated from the satellite-based mean age data . The metric is computed on isentropic levels, but approximate corresponding pressure levels are also given for reference. While the mass flux values are broadly consistent with those in Fig. , this simplified estimate shows a substantially larger spread (e.g., 3 × 109 kg s⁻¹ near 70 hPa for CCMI-2022, which corresponds to almost 50 % of its magnitude). The newest generation shows larger values throughout the altitude range, which is not observed in Fig. a. These differences between the mean meridional overturning circulation results are not due to the fact that different models provide residual circulation and mean age output (Figs. S1 and S4). Rather, they are likely due to the very different approaches, variables and calculations involved in each method. A comparison between residual circulation and age gradient mass flux was carried out in , but was based on variability, not climatology. The observational age gradient mass flux estimates suggest that the models' overturning circulation is too fast above ∼ 40 hPa, and too slow below that level. This result is also at odds with that derived from reanalysis mass flux (Fig. ), which suggests an overestimation only in the lower stratosphere. However, there are important observational caveats to consider. As shown by (see their Fig. 11), at 50 hPa, the difference in mean age between tropics and extratropics is too small for the age derived from SF6 with the mesospheric sink correction for both MIPAS and ACE-FTS estimates, compared to in situ observations. This is because the MIPAS mean age is biased high (old) in the tropics relative to in situ observations and agrees well in the extratropics, while the ACE-FTS mean age is biased low (young) in the extratropics, and agrees well in the tropics. This leads in both datasets to a lower meridional gradient in mean age and thus an overestimated overturning circulation. The ACE-FTS non-corrected for SF6 approximates better the gradient suggested by in situ measurements Fig. 11 of , and using these values, the curve is more consistent with the model output in the lower stratosphere (below ∼ 30 hPa, see Fig. ). However, at the upper levels, this results in lower values that fall outside the model range, as expected, since the influence of the mesospheric sink of SF6 on the age calculation should increase with altitude. Hence, we conclude that this metric should be considered with care, given the observational limitations. An additional potential source of discrepancy is the temperature and static stability biases in the models, which result in shifted isentropic levels and their thickness relative to reanalyses. Overall, our analyses show that the overturning mass flux estimate from the age gradient should not be directly compared to residual circulation-based diagnostics.

An additional estimate of tropical upwelling in the lower stratosphere can be obtained from the tape recorder signal imprinted by the seasonal cycle in cold point temperature . The dry and wet anomalies imprinted by the larger dehydration in boreal winter as compared to summer propagate upward with the residual circulation, and thus the phase lag between different levels can be used as an estimate of the upwelling strength. The tape recorder signal is captured by most models, though with very different amplitudes and upward propagation rates (Supplement Figs. S5–S7). Figure a–d shows the MMMed tape recorder signal in the three generations, and the SWOOSH satellite dataset results for comparison.

The MMMed of the three generations shows a consistent tape recorder signal with weaker amplitude compared to SWOOSH, but there is a very large spread across individual models. The MMMed amplitude is larger and thus more realistic in the new generation. For a quantitative comparison of the MMMed and the inter-model spread, Fig. e shows the phase of the tape recorder as a function of altitude relative to the maximum amplitude of the tape recorder anomalies (near the cold point tropopause) up to 10 km above the peak amplitude (corresponding to ∼ 25 hPa). Compared to the observations, the models generally capture the speed of the ascent above the tropopause, although most models tend to carry the anomalies upward too fast (shorter phase lag for a given level), especially higher than 3 km above the maximum tape recorder amplitude (∼ 66 hPa). This estimate thus provides independent qualitative support for the result obtained with the upward mass flux of the residual circulation, that upwelling is overestimated in the models, although the tape recorder diagnostic suggests that this bias extends to higher levels. It is important to keep in mind that this metric does not provide a clean measure of advection by tropical upwelling, as it includes the effects of vertical and horizontal mixing and diffusion . We will discuss the information on horizontal mixing that can be extracted from the tape recorder referring to Fig. f in the next section. Note that the newest model generation has the smallest spread and closest values to the observations. Finally, we note that the results might also depend on the rate of methane oxidation and its increase with altitude, which might be different across models.

The amount of water vapor that enters the stratosphere is controlled by the temperature of the cold point tropopause (CPT). Figure a shows that the seasonal cycle of the CPT temperature is overall captured by the models, consistent with the reasonable MMMed tape recorder signal in Fig. , but with a very large spread across models in the absolute values (up to 5 K). Note that ERA5/5.1 has a colder CPT compared to the other reanalyses (by ∼ 1 K), consistent with the faster tropical upwelling seen in Fig. . In addition, there is a warm bias of 1–2 K in the newest generation multimodel median, although the seasonality is better captured than in previous generations. The lapse rate tropopause is too high in the models, especially in the extratropics, where there differences across models are largest (Fig. b). This issue has thus persisted since the CCMVal-2 intercomparison by . In the tropics, the newest generation has the highest tropopause.

The other major component of tracer transport, in addition to advection by the mean meridional overturning circulation, is mixing. As mentioned in the Introduction, strong quasi-horizontal mixing occurs in the surf zone in the extended winter of each hemisphere and is the main transport mechanism in that region, while the slow ascent is dominant in the tropical stratosphere. The transition between these two different regions and transport regimes can be visualized by plotting the probability density functions (PDFs) of a long-lived tracer such as N2O across tropics and midlatitudes.

Figure shows the PDFs of N2O as a function of potential temperature for various satellite datasets (MIPAS, MLS, ACE-FTS; top panels) and the individual CCMs from CCMI-2022 in boreal spring, following . The altitude range corresponds approximately to 80–1 hPa. The PDFs show two separate branches that are distinguishable between ∼ 450 to 1200 K, which correspond to the tropical and extratropical mixing ratios (the tropical pipe and the surf zone), which can be considered well-mixed. Between these two modes, there is a minimum in the PDF which corresponds to the steep gradients created at the equatorward edge of the surf zone, which reveals the presence of a subtropical mixing barrier. The three observational datasets present strong similarities, although the details are different. The most notable difference is that MLS presents an unrealistic step-like behavior below 800 K (∼ 10 hPa), likely due to a localised high-bias in N2O at this level compared to these other instruments . The results for the SH are similar, with models maintaining the sign of the bias in both hemispheres and across generations (Supplement Figs. S14–S18). For example CCSRNIES-MIROC has too small a distinction between the branches and GEOSCCM has too large tropical values (Supplement Fig. S18). In order to compare the distributions more quantitatively, Fig. a,b shows the PDFs at the level of 600 K (∼ 30 hPa). The N2O concentrations have been normalized by removing the mean for each dataset to facilitate analysis of the key features beyond the mean bias, which are peak-to-peak separation and the depth of the minimum.

The double peak structure is apparent in all observational datasets, but ACE-FTS shows substantial deviations from the other two datasets, both in the magnitude and location of the peaks. Keeping in mind that this satellite has a much sparser sampling, as well as the agreement between the other two datasets, we compare the models against MLS and MIPAS. Note that removing the mean takes care of the MLS shifted behavior in the lower stratosphere seen in Fig. . In both hemispheres, the MMMed of CCMI-2022 tends to place the maxima too close together, as identified in most models in Fig. . On the other hand, the depth of the minimum is generally well captured by the MMMed in all model generations, although it is slightly less pronounced, which could reflect a smoother transition from the tropical to the extratropical regimes. From the minimum in the N2O PDF, we derive the position of the subtropical mixing barriers (Supplement Fig. S21). We find that the three multimodel medians are remarkably close to the observations, and indeed, the intergenerational spread in the median values is smaller than the difference between the two observational datasets (MIPAS and MLS). The spread is larger in the NH (originating in JJA, not shown), and is reduced in the newest generation.

Figure c, d shows the climatological N2O contours for the MMMed of the three generations and MIPAS. There is an overall low bias in CCMI-2022 N2O mixing ratios compared to MIPAS (isopleths are located at lower levels), as mentioned regarding Fig. . In the tropics, this bias maximizes near 20 hPa, such that the decline of N2O with altitude is too steep in the lower stratosphere (Fig. S20 in the Supplement). This could reflect overestimation of horizontal mixing. The sharp transition between tropical and extratropical N2O concentrations is smoothed out in the models' MMMed compared to the observations, consistent with the PDF behavior (the two peaks are closer together and the minimum is less pronounced in Fig. a–b). Note that the models represent the midlatitude plateau (surf zone) and the tropics-extratropics transition more accurately in the SH spring than in the NH spring.

Previous studies have proposed a metric for the tropics-extratropics mixing strength derived from the TLP model. Specifically, the vertical gradient in tropical mean age can be related with the mixing strength . Qualitatively, this can be understood by considering the fact that mixing from the extratropics into the tropics makes tropical air older, and so the vertical gradient is related to mixing. There are two main caveats about this; the same caveats about satellite data biases as with the meridional age difference above and the fact that the metric treats the extratropics as well mixed, leading to apparent negative mixing efficiencies at higher levels. Nevertheless, we show this metric in Fig. .

We note that there is non-negligible observational uncertainty, as evidenced by the difference between the two satellite-derived products. The models generally agree across generations within the large uncertainty given by the intermodel spread, with the newest generation having the smallest spread. The model values generally agree with the observational estimates below ∼ 500 K (∼ 40 hPa), although there is evidence that their mixing is too weak in the lower stratosphere (below 450–500 K, depending on the observational dataset considered). Between 500–550 K the models overestimate mixing, but note that above this level, the well-mixed assumption invoked by the mixing strength theory in breaks down and observed estimates for mixing efficiency are negative: these points have been removed. The models' mixing efficiency remains positive throughout the range of isentropes presented in Fig. , suggesting that the models remain well mixed through a deeper depth of the stratosphere than the observations. To summarize, while there is overall agreement in the lower stratosphere, with some suggestion of too weak horizontal mixing below ∼ 50 hPa, the observational uncertainty is too large to constrain the models. Finally, similar to Fig. , this could instead be reflecting too large vertical diffusion .

An additional metric for the horizontal mixing can be provided by the damping of the water vapor tape recorder signal, as this damping is due to mixing with extratropical air. Figure f shows the peak-to-peak amplitude of the tape recorder signal as a function of height. There is a substantial spread in the damping of the tape recorder across models, which is reduced in the newer generation. Overall, it appears that most models in CCMI-2022 damp the signal too slowly with altitude around 5 km above the tropopause (∼ 50 hPa), which implies that they tend to underestimate the mixing with the extratropics. However, this result is inconsistent with the N2O analyses, which show too smooth tropics-extratropics transitions and too strong vertical gradients in the tropics, both compatible with strong mixing. The weak tape recorder damping could also be caused by an excessive vertical diffusion in the models extending the anomalies upward; another possibility could be too fast methane oxidation in that region.

In summary, different metrics of mixing do not provide a consistent picture of mixing biases in models. While there is some evidence from the tape recorder and the vertical age gradient metrics of an underestimation of mixing between the tropics and extratropics in the lower stratosphere in CCMs, which is broadly consistent with the low mixing efficiency associated with the young bias in mean age of air (Fig. ), the N2O structure suggests an overestimation of mixing. These discrepancies could be reflecting other effects such as excessive vertical diffusion or methane oxidatoin in the models, or could be due to limitations in the tracer representation from satellite observations.

We now turn to the polar regions, which feature a distinct transport regime characterized by the isolation from the middle latitudes by the polar night jet and the polar downwelling of the deep branch of the mean meridional overturning circulation. Figure  shows the annual cycle of total polar downwelling mass flux in the two hemispheres at 30 hPa, computed from the maximum of the streamfunction for the NH and the minimum for the SH.

The most outstanding feature is that in the SH, the models have too weak downwelling compared to reanalyses, especially in spring (September–October). In December, the bias is reversed, and all models overestimate downwelling, which is at its minimum in the reanalyses. This behavior in December could reflect the delayed breakdown of the polar vortex that will be discussed below. The NH downwelling shows better agreement and a smaller inter-model spread, with a slight underestimation of the maximum downwelling in December–January, especially compared to ERA5. These results are consistent across model generations.

Polar downwelling is responsible for the downward transport of air masses and tracers in the polar stratosphere. In particular, for a tropospheric-origin long-lived tracer such as N2O, polar downwelling reduces its concentrations in the lower stratosphere . Consistently, the N2O tendencies are negative in the extended winter, reflecting the influence of polar downwelling. Figure a–b shows the seasonal cycle of N2O monthly tendencies in both hemispheres in the model results and satellite observations from MIPAS and MLS. In the SH, both observational datasets show a pronounced minimum in September-October, which is not captured by the models. This missing negative tendency is most likely related to the weak downwelling in these months found in Fig. . In the NH, the winter negative tendencies peak in December and are generally better captured by the models compared to the SH.

In spring and summer, high latitude N2O increases as the polar vortex barrier disappears, allowing for enhanced mixing with N2O-rich air from lower latitudes. The positive tendencies in spring and summer are reasonably well represented in the model results. In the NH, MIPAS suggests an earlier increase of N2O compared to MLS and to the model results, placing the maximum tendency one month earlier (March instead of April). To test if this difference could be due to the different periods covered, which includes years with sudden stratospheric warmings, we repeated the analysis for the common period, but obtained similar results. Another possibility is the different coverage of the polar region by both instruments. The seasonality in the model results agrees best with the MLS seasonality, which has a better spatio-temporal coverage. In both hemispheres, the inter-model spread is comparable to the observational uncertainty.

In order to further explore the impact of the weak polar downwelling in the SH spring, Fig. c shows the vertical profile of polar N2O mixing ratios in SON in model results and satellite observations. The mixing ratios of N2O in the lower stratosphere (below ∼ 60 hPa) are substantially larger in the model results than in the observations, which is consistent with the weak polar downwelling found in all generations in Fig. . Note that in late spring, the polar vortex can break and increase substantially horizontal mixing, which increases the N2O mixing ratios in the polar region. Nevertheless, we will show below that this increased mixing happens too late in the models on average, and thus it is likely not contributing to the larger N2O mixing ratios in SON.

The downwelling of the residual circulation is associated with the Coriolis torque exerted by the wave drag from the dissipation of planetary Rossby waves (resolved by the models) and gravity waves (parameterized in the models). The wave drag also decelerates the polar vortex. A common diagnostic of the resolved upward wave activity flux entering the stratosphere (which will then dissipate at some point and drive the residual circulation) is the midlatitude eddy heat flux at 100 hPa, as it is the dominant term of the vertical component of the Eliassen-Palm flux . Figure  shows this diagnostic applied to the model results and reanalyses.

There is good agreement for the NH eddy heat flux between the model and reanalysis results, although the models tend to slightly underestimate the mid-winter maximum, especially in CCMVal-2, and overestimate the summer minimum. In the SH, the reanalysis climatology peaks in October, when the climatological polar night jet has weakened with respect to its very strong midwinter values that inhibit wave propagation. The models are generally able to capture the seasonality, although the inter-model spread is even larger than for the NH (the spread is up to 2/3 of the mean value). For CCMVal-2, the MMMed delays the peak by about one month. For CCMI-1 and CCMI-2022, the MMMed follows the reanalysis values more closely, although the October peak wave activity is underestimated. Note that the individual models show very different timing and amplitude of the wave activity maximum (Supplement Fig. S26). The underestimation of the spring maximum in resolved wave activity could be related to the underestimation of polar downwelling seen in Fig. . It is important to note that gravity wave drag plays a large role in driving polar downwelling. However, only a small set of models output the gravity wave drag and thus we do not examine it here.

A key process in the polar stratosphere, crucial for terminating the ozone depletion season, is the breakdown of the polar vortex, i.e., the date when the polar vortex breaks down and the summer circulation begins. There is a long-standing cold bias in the polar lower stratosphere in models, which is connected to an excessively long-lived polar vortex and is thought to be due to insufficient wave drag. This is likely from unresolved gravity waves e.g.,. We examine the final warming date (FWD) as a function of altitude in both hemispheres in the three generations of CCMs and the reanalyses in Fig. . Note that these are computed from the monthly mean output through linear interpolation to daily values assuming that the monthly mean corresponds to the 15th day, following . We have verified with the reanalyes that this method gives very consistent results to those using the daily mean field. The thresholds to define the FWD are selected to allow for all models to reach the conditions in the lower stratosphere, since most models have too strong winds, as will be discussed below.

It is evident that in all model generations, the vortex breakdown occurs too late in both hemispheres, and this delay is larger in the SH. In the newest generation, the FWD bias is even more pronounced, being up to 3 weeks for the MMMed in the SH lower stratosphere and up to 1 month considering the intermodel spread. For instance, at 50 hPa, the reanalyses place the FWD at the beginning of November, and CCMI-2022 places it almost at the end of the month. In the NH, the late bias is more modest (∼ 1–2 weeks), and similar in the three generations, with reanalyses placing the FWD at 50 hPa in mid-April (with larger interannual variability than in the SH, not shown) and models at the end of April-beginning of May. It is noticeable that the bias extends throughout the stratosphere, and in the NH it reverses sign in the upper stratosphere.

In summary, the models have too weak downwelling, especially in SH spring (September–October). On average the models tend to underestimate the peak in resolved wave activity entering the stratosphere, and have difficulties in capturing its timing in the SH. In addition, the polar vortex lasts too long, especially in the SH, in all model generations, which results in too large downwelling in December.

After discussing the main transport processes, we turn to the representation of the stratospheric ozone climatology, as capturing the evolution of the ozone layer is one of the key motivations for the development of CCMs. We examine partial ozone columns (in Dobson Units, DU) computed from the models' output volume mixing ratios (in mol mol⁻¹) over two vertical layers: the lower stratosphere, computed between 100–20 hPa in the tropics and between 150–20 hPa in the extratropics, and the upper stratosphere, computed over 10–1 hPa. Because ozone is an output available for all the models, we included only the common models, which are the 8 model families introduced in Tables – (the corresponding figures for all models are shown in the Supplement, Figs. S26 and S27).

The two satellite datasets show some differences, with generally lower ozone values for SWOOSH when compared to SAGE-CCI-OMPS+ (∼ 10–15 DU). This observational uncertainty is consistent with the typical uncertainty for satellite data at these altitudes e.g.. Overall, all model generations overestimate ozone in the lower stratosphere, especially in the SH midlatitudes and in the tropics (Fig. b, f), which results in a globally integrated (90° S–90° N) lower stratospheric ozone bias of about ∼ 15 DU (Fig. c). In contrast to previous generations, the latest generation (CCMI-2022) overestimates SH polar ozone decline in October–November by about 25 DU (∼ 25 %; Fig. a), also when considering all the models (Supplement Figs. S29 and S31). Both the depth and duration of ozone decline are overestimated in CCMI-2022, with the observations falling on the high end of the inter-model spread. The large ozone decline in the latest generation is likely related to the particularly large delay in the vortex breakdown date (Fig. a). A long-lived polar vortex can imply maintained low polar cap temperatures that facilitate heterogeneous chemical ozone destruction. Moreover, the low ozone could be in turn contributing to the cold pole bias, since the low ozone mixing ratios absorb less solar radiation, resulting in a colder polar stratosphere, which induces a stronger polar vortex to maintain thermal wind balance.

Despite the mean biases, the models are generally able to correctly capture the lower stratospheric ozone seasonality, with correlations between the observational and model seasonal cycles above 0.95 (see Taylor diagrams in Supplement Figs. S31–S35). The lowest correlations (less accurate representations of the seasonality) are found in the SH polar region.

In contrast with the lower stratosphere, ozone in the upper stratosphere is underestimated by the models in the global mean (about ∼ 10 %; Fig. c). The underestimation is found in all regions, and is most severe in the newest generation (CCMI-2022), also when considering only common models (Supplement Fig. S30). The reason for the opposite model biases in the global lower and upper stratosphere is unclear. The seasonality is reasonably captured in general (correlation above 0.9, see Taylor diagrams in Supplement Figs. S36–S40), with highest correlations found in midlatitudes, and the lowest correlations in the polar regions. Note that the agreement between the datasets is also weakest in the polar caps, possibly due to sparser coverage of these regions.

Figure a shows the 1980–1999 trends for the three model generations in global mean AoA. There is a good intergenerational agreement in the MMMed, with negative trends peaking in the lower stratosphere at around -6 % per decade. At higher levels, the newer generations show -5 % and -4 % per decade for CCMI-2022 and CCMI-1, nearly constant with height, and linearly decreasing to -3.5 % per decade for CCMVal-2. There is considerably large spread, especially in the older generations, and surprisingly reduced in the newest one for the lower stratosphere.

Previous studies have shown that SH ozone depletion has contributed significantly to enhancing the BDC trends over the last two decades of the 20th century e.g.,. This is due to the acceleration of the residual circulation forced by the dynamical effects of ozone depletion in the SH polar lower stratosphere . In order to explore this effect, Fig. b and c show the AoA trends in CCMI-2022 over the ODS increase period (1980–1999) and the ODS decline period (2000–2018), which we refer to as ozone depletion and recovery periods, respectively. In order to remove the effect of natural variability on these short-term trends, we have applied a multiple linear regression to the CCMI-2022 AoA using as regressors the Multivariate ENSO Index Version 2 (https://www.psl.noaa.gov/enso/mei, last access: 18 April 2026), two QBO indices corresponding to the two principal components of the equatorial mean wind between 70 and 10 hPa , the solar 10.7 flux (https://psl.noaa.gov/data/correlation/solar.data, last access: 18 April 2026), and the Global Space-based Stratospheric Aerosol Climatology Version 2.0 . We apply the regression to the newest generation over the period 1980–2018 and compute the trends for the residuals over the shorter periods. The results are qualitatively consistent with those obtained with the full field, in particular the 1980–1999 trends remain practically unchanged (Fig. S43 of Supplement). Consistent with previous literature, the mean age trends are substantially reduced in the latter period compared to the former. The change in the global mean AoA trend from the first to the second period is even larger than found before for longer periods (from 6 % per decade to 1 % per decade). In addition, the trends are much more robust in the first period, with all models presenting negative trends (Supplement, Fig. S42). An additional proof that the Antarctic ozone hole is behind the large trend before 2000 is that the negative AoA trends are larger in the SH hemisphere over the earlier period. This is consistently seen in the three generations (Fig. S44 of Supplement). In the post-2000 period the much smaller trends still display an asymmetry, being more negative in the SH than in the NH, consistent with previous studies e.g.,.

Figure a–c shows the trends of the annual mean upward (Fig. b) and downward (Fig. a, c) net mass flux of the residual circulation for the three generations and three reanalyses. Again, there is a very good quantitative agreement across generations for the MMMed trends, with a spread that increases with altitude and is reduced in the newest generation. The upward mass flux accelerates 1 %–2 % per decade over the lower and middle stratosphere (below ∼ 5 hPa), and the trends are much reduced and uncertain in the upper stratosphere, although positive for the MMMed in the three generations. Both hemispheres show strengthened downwelling in the lower and middle stratosphere, and the trends are larger and more significant in the SH, especially between ∼ 50 and 10 hPa. At higher levels, there is a larger spread across the model generations and within each generation. The reanalyses show very diverse results. The upward mass flux increases significantly by 4 %–6 % per decade in the lower stratosphere according to MERRA-2 and JRA-3Q, while it decreases up to -2 % per decade in this region in ERA5. This reanalysis discrepancy is consistent with . The downwelling trends also differ significantly across reanalyses. In the SH, the three reanalyses indicate an acceleration (albeit with very different magnitudes), while in the NH, even the sign is different across the reanalyses.

Figure d, e shows the trends of the vertical component of the residual circulation for the periods before and after 2000 in the CCMI-2022 simulations' multi-model median. The model results show an acceleration of tropical upwelling and extratropical downwelling during both periods, with considerably larger SH downwelling acceleration in the first period. While there is no attribution made in this analysis, these results are in line with the large impact of ozone depletion on the acceleration of the residual circulation found in previous studies. Figure f, g shows the average of the trends for the three reanalyses. The averaged reanalysis trends are noisier than the model trends, but display an overall consistent behavior, with larger tropical upwelling and polar downwelling in the SH in the ozone depletion period. Again, the results are highly consistent when no variability is removed (Fig. S46 of the Supplement).

One of the most robust atmospheric responses to increasing greenhouse gases is a rise in tropopause height, resulting from the combined effects of tropospheric warming and stratospheric cooling, which jointly shift the lapse-rate boundary between the two layers upward. This is captured by all models, as seen in the timeseries in Fig. a. The rise of the tropopause is similar in the three model generations (∼ -0.5 hPa per decade over 1980–2010), consistent with the value given by for CCMVal-2 models, and slightly lower than that in the three reanalyses (∼ -1 hPa per decade).

Previous studies have shown that tropical upwelling trends are significantly different when viewed from a tropopause-relative coordinate . Figure b shows the trends of the upward mass flux over 1980–1999 (similar to Fig. b) in tropopause-relative coordinates. This is computed by regridding the mass flux timeseries at each level into a tropopause-relative log-pressure altitude grid (using the tropopause pressure averaged over 20° S–20° N) and then computing the trends. It is clear that the trends decrease substantially at the tropopause level, where the models simulate near-zero trends in tropopause-relative coordinates, and ∼ 1 %–2 % per decade on pressure coordinates. On the other hand, at higher levels, the trends are only slightly lower than those in Fig. b (trend values differ less than 0.5 % per decade). In any case, we note that both the decrease in tropopause pressure and the acceleration of tropical upwelling largely arise from the same primary underlying driver, namely tropospheric warming. Therefore, the rising tropopause should not be considered a cause of tropical upwelling acceleration, but rather both are a common response.

Coupled radiative-dynamical responses to climate-forcing affect tropopause temperature, which in turn determine the transport of water vapor into the stratosphere. Figure c shows that the cold point tropopause is cooling in the three reanalyses between 1980 and 2010 (around -0.29 ± 0.14, -0.54 ± 0.16, and -0.13 ± 0.15 K per decade in ERA5/5.1, JRA-3Q and MERRA-2, respectively). On the other hand, the CCM multi-model medians of the three generations show a near-zero trend over this period (0.040 ± 0.25, 0.060 ± 0.082 and 0.035 ± 0.098 K per decade for CCMIVal2 – until 2000 –, CCMI-1 and CCMI-2022, respectively). The trend in the reanalyses is partially explained by a strong drop in cold point temperature after the year 2000, which was a random feature not necessarily to be expected in the model simulations, but which exacerbates the underlying negative trend. Note that future climate simulations project a cold point warming , and a warming has been detected in observations since 2000 .

Changes in AoA are not only related to changes of advective transport by the residual circulation, but also to mixing efficiency . show that the contribution of changes in mixing efficiency to the future changes in AoA in CCMI-1 refC2 simulations present a large spread across models, varying between 10 % and 30 %. Moreover, consistent with the results for the climatology obtained by and captured here in Fig. , they show that the spread in mixing efficiency changes is closely correlated with the spread in the magnitude of the AoA decrease over the 21st century. To test this behavior, Fig.  shows the scatter plot of changes between the two decades centered in 1980 and 1999 (1975–1984 and 1994–2003) in mean age versus RCTT and mixing efficiency.

The results show overall higher correlations of the changes in mean age with changes in mixing efficiency than with changes in RCTT, consistent with the results for future climate simulations described above. However, this result is less robust than for the climatology (Fig. ), because for CCMI-1 the correlation between mean age and RCTT is notably higher than for the other generations. Also, the correlation with mixing efficiency is low in CCMI-2022 (0.25). Nevertheless, for any given generation (including CCMI-1), and for all generations together, the correlations are higher with mixing efficiency than with RCTT. This suggests that the mean age changes are more strongly connected with changes in mixing or diffusion than with the changing strength of the residual circulation.

Ozone depletion over Antarctica has impacted not only the strength of the residual circulation but also the polar vortex. Specifically, it is known that ozone depletion cooled the polar lower stratosphere in SH spring and strengthened the polar vortex, leading to a delay of the final warming date. Figure  shows the trends of the FWD over the period 1980–1999 for the SH polar vortex, computed as the day when the zonal-mean zonal wind reaches a value of 30 m s⁻¹ at 50 hPa.

The reanalyses show a FWD delay of ∼ 8 d per decade, with very good agreement across them and large error bars due to the large interannual variability. The MMMed of the three generations captures this trend very well. Nevertheless, there is a very large spread across the individual models, with some models even producing negative trends. The newest generation performs best in capturing these trends: although it exhibits the largest interquartile range, it produces fewer extreme values, and almost all models show a positive trend.

We lastly examine the temporal evolution of the ozone columns analyzed in Sect. .

Figure shows the annual mean ozone anomalies evolution in the lower stratosphere from the three generations of models compared to the SWOOSH and SAGE-CCI-OMPS+ datasets. All three model generations simulate a steady decline of lower-stratospheric ozone from 1960 to the end of the 1990s, with close agreement among models except for the SH polar latitude band (85–60° S), where CCMI-2022 shows a notably larger decline. This difference could be partly due to updated ODS and GHG forcings. However, these are not very different over the historical period , and thus it is likely that the newer models are more sensitive to the external forcings. The larger decline in the SH is consistent with the lower climatologically averaged ozone in Fig. a in October and November, which is in turn related to the longer-lasting polar vortex in this generation (Fig. a). Note that the faster decline in CCMI-2022 is partly driven by the UKESM-StratTrop model (see Supplement Fig. S47), which presents ozone anomalies of 60 up to over 100 DU in the 1960s to 1980s. This model indeed severely overestimates ozone decline (by ∼ 40 DU, see Supplement Fig. S31). However, the other models also show larger ozone column anomalies compared to the previous generations, with most models starting above 20 DU in 1960, while staying below that value in previous generations.

The equivalent plot for the upper stratosphere is shown in Fig. . Both satellite and models show a sharp decline from the 1960s up to the end of the 1990s, followed by slightly upward trends, consistent with the statistically significant ozone recovery detected in the upper stratosphere e.g.. There is a general agreement across model generations over the satellite era (after 1985). Before then, the depletion is slightly stronger for the oldest generation than the following ones. There is close agreement between the models and the satellite records, with smaller interannual variability than in the lower stratosphere.