The historical global anthropogenic emissions are taken from the Community Emissions Data System (CEDS) inventories , and the historical biomass burning emissions are obtained from the biomass burning emissions for CMIP6 (BB4CMIP; here, CMIP6 refers to the Coupled Model Intercomparison Project Phase 6) inventory . The emissions after 2014 are taken from the SSP3-7.0 scenario . The aircraft emissions are taken from the anthropogenic emission inventories as well (both historical and future scenarios), after applying the corrections presented in . The historical runs generally cover the period from 1994 to 2017 (from 2001 to 2017 for OsloCTM3), providing robust climatologies that are compared with aircraft observations over the same period. In order to assess the model's ability to simulate the mean UTLS composition, it is important to provide simulations with the most realistic transport conditions. This is why the runs from the CCMs are nudged by horizontal winds taken from a reanalysis, as indicated in Table . Three models are CCMs and are nudged with ERA-Interim (EMAC and LMDZ–INCA) or ERA5 (UKESM1.1), and one model is a CTM forced by the ECMWF OpenIFS product (OsloCTM3), similar to ERA-Interim. The simulation from another CTM participating in the ACACIA project (MOZART3, forced by ERA-Interim) has been included to present the model behaviour during the period from 1997 to 2007, but the dynamical field is a cyclic repetition of the year 2007, which removes the interannual variability in the atmospheric transport and, thus, cannot be treated as a model assessment, except to comment on the seasonality.

In this section, Table  summarizes the key model characteristics. In the next subsections (Sect. –), further detail is given for each model, notably the tropospheric and stratospheric chemical schemes implemented in the models are described.

The IAGOS research infrastructure (http://www.iagos.org, last access: 1 November 2022) provides long-term routine in situ observations of chemical species aboard a fleet of several passenger aircraft. Its predecessors – MOZAIC Measurements of water vapor and OZone by Airbus In-service airCraft; and CARIBIC Civil Aircraft for the Regular Investigation Based on an Instrument Container; – relied on the same principle. The MOZAIC measurements began aboard five equipped aircraft measuring ozone and water vapour in August 1994. CO observations began in December 2001, and NO_y measurements were operational on one aircraft between April 2001 and May 2005. CARIBIC has sampled a wide variety of atmospheric species since 1997 from one single aircraft, including those measured by MOZAIC. Since the mergence of the two programmes in 2008, their respective databases are referred to as IAGOS-CORE and IAGOS-CARIBIC. In the present study, we consider them as a single database called IAGOS hereafter, an approach validated by for ozone and CO. The period that we are analysing spreads from August 1994 until December 2017, hence the 1994–2017 (or 2001–2017) period covered by the models' run. The most sampled altitudes are between 180 and 310 hPa, and the vertical distribution of sampling varies geographically. The methodology used for the models' assessment using the IAGOS data is described in Sect. .

The main characteristics of the IAGOS instruments relevant to this study are indicated in Table . Concerning the IAGOS-CORE instruments, further information is available in for ozone; in and for CO; in , , , and for water vapour; and in and for NO_y. Note that the IAGOS-CORE water vapour measurements have an accuracy of 6 % RHL (relative humidity with respect to liquid water) in the vicinity of the midlatitude thermal tropopause (). Due to a moist bias in the IAGOS-CORE H2O observations for the driest air masses (RHL <5%) which are encountered in the lower stratosphere, this study does not quantify the model H2O biases in regions other than the upper troposphere. Concerning the IAGOS-CARIBIC instruments, further information is available in for ozone, in for CO, in and for water vapour, and in and for NO_y. More precisely, the latter has a total measurement uncertainty of 6.5 % (8 %) for a measured mixing ratio of 1 ppb (0.5 ppb). For both programmes, the time response of the water vapour sensors decreases with the measured water content.

The Interpol-IAGOS software aims to facilitate the assessment of the model output with the IAGOS data by deriving two respective products that are directly comparable. It consists of a projection of the scattered IAGOS data onto the regular model grid, day by day, followed by a monthly average. For a given model, the subsequent gridded IAGOS product is then denoted as the “IAGOS–DM–model”, with the “–DM” suffix referring to the distribution onto the model grid. For further simplicity in this study, we refer to it as “IAGOS–model”, and the IAGOS–DM–model products in their ensemble are called “IAGOS–DM products”. Concerning the model output, a daily mask is applied with respect to the IAGOS sampling , excluding the non-sampled daily grid points. This way, the subsequent monthly products are representative of the same grid points and the same days. As in , their whole name is “model–M” (with the “–M” suffix referring to the IAGOS mask); except in this section where there can be confusion, we refer to them simply using the model name. The monthly average is calculated for each layer separately. This implies that a monthly average is calculated for both layers for each grid cell included in both the UT and LS during a month. Finally, the seasonal and annual climatologies are then derived from the monthly means with the same method and filtering as in . For each model, the IAGOS–DM–model and model–M product pairs are, thus, representative of the same time period as well.

For each model, the tropopause is defined dynamically as the isosurface of 2 PVU (potential vorticity units) derived from the model output. The UT spreads from 400 hPa up to the tropopause level but excludes the top grid cell in order to avoid the strongest mixing zone, directly impacted by both layers e.g.. The LS corresponds to all of the sampled grid cells above the 3 PVU isosurface. In order to limit the impact of errors in the modelled PV on the evaluation, we exclude cases in which the modelled PV value and the daily average of the observed ozone mixing ratios are very likely to represent different layers. As in previous studies , a daily grid cell in the UT (LS) is filtered out when the average ozone from IAGOS is more than 140 ppb (less than 60 ppb). Finally, the non-separated UTLS represents all of the measurement points above 400 hPa, and it has been added to this paper in order to include the output from the MOZART3 model without a PV field. In the tropics, the tropopause altitude does not allow the aircraft to sample stratospheric air masses. Consequently, only the UT is represented between 25° S and 25° N, and it includes all of the measurements above 300 hPa, although the uppermost part of the tropical troposphere remains higher than cruise altitudes. In the framework of the CCMI modelling experiment, reports that the nudging process tends to enhance the transport variability between the CCMs and can generate artificial transport in regions of strong gradients. This issue might be partly addressed by our methodology, notably the definition of the layers, which enhances the isolation between them, and the exclusion of the grid cells with an inconsistent PV value with respect to the ozone observations. The mean pressure differences between observations and the model 2 PVU tropopause shown in Supplement (Figs. S1–S5) do not exhibit noticeable variations between the models: mostly, they are less than 5 hPa (except in winter and spring for ozone and NO_y), and they are always less than 10 hPa. It might still be problematic for water vapour in the LMS, as the vertical gradient from the tropopause is particularly high, including 2 orders of magnitude , but, as described above, the model skills do not cover water vapour in the LMS. For each species and layer, the distance between the sampled grid cells and the tropopause does not vary enough across the models to play a significant role in the inter-model discrepancies, as there is no visible correlation with the chemical tracers.

As in , the chosen metrics are the modified normalized mean bias (MNMB), the fractional gross error (FGE), and Pearson's correlation coefficient, all defined in Supplement (Eqs. S1–S4). As the models have different PV fields and some of them have a different time period, each model is compared to its own IAGOS–model reference product. This study is divided into comparisons in the extra-tropics and in the tropics. The results in the extra-tropics are generally represented with several sets of metrics (quantiles, biases, and linear regression metrics). On the contrary, as the sampling in the tropics is more heterogeneous, three regions have been chosen, as in , and are represented by the mean zonal cross sections: western Atlantic–South America (called South America hereafter), Africa, and South Asia. As a compromise between an efficient sampling and spatial uniformity for the observed species, their respective zonal cross sections are averaged through the following longitude bands: 60–15° W, 5° W–30° E, and 60–90° E. For each of these regions, the year is divided into seasons that depend on the mean position of the Intertropical Convergence Zone (ITCZ). As in , the tropical season definitions were based on the months with the northernmost and southernmost position of the ITCZ (located with the observed horizontal winds and water vapour), as an extension of , who focused on Africa. They are summarized in Table , taken from .

The annual climatologies of the model biases in the UT and the LS are shown in Figs. – for the four models with an available PV field. A climatology is also shown for one of the gridded IAGOS products (hereafter referred to as “IAGOS–DM products”; see Sect. ) in order to provide a view of the expected features, but each bias remains relative to a different IAGOS–DM product, notably due to the different time periods. As the IAGOS–DM climatologies are relatively similar through the simulations with the same duration (not shown), we chose only one of the IAGOS–DM climatologies with the longest time period (by default IAGOS–EMAC, i.e. the gridded IAGOS product on the EMAC model's grid, as explained in Sect. ). We note the sampling differences between the climatologies from OsloCTM3 (2001–2017) and the other three assessed models (1994–2017). Interannual variability is, therefore, likely to cause moderate differences in the observed climatologies for ozone and water vapour in OsloCTM3, the time period of which is 25 % shorter than for the other models. This does not apply to CO and NO_y, as their IAGOS-CORE observations started in 2001.

On a yearly average, we first notice common features between the models. The four models exhibit a geographical anticorrelation between ozone and CO biases in the LS (Figs.  and ). The O3/CO ratio shown in Fig.  summarizes this pattern well, with the MNMB decreasing with latitude. On the other hand, upper-tropospheric ozone is overestimated in the midlatitudes, whereas CO tends to be underestimated. These combined features suggest that the four models tend to overestimate the overall impact of stratosphere–troposphere exchanges across the extratropical tropopause. Most of the models also overestimate ozone in the tropics, which is analysed in detail in Sect. . We can notice that the models showing a more positive (negative) bias in ozone (CO) in the low-latitude LS have the same tendency in the tropical UT, possibly as an impact of isentropic cross-tropopause exchanges that can extend biases to the adjacent layer. Figure  shows that, contrary to the other species, each model NO_y climatology is simultaneously characterized by both low and high biases as well as a few grid cells with a weak bias. The map derived from the observations is heterogeneous too, with upper-tropospheric minima over Northwest America, the North Atlantic corridor, and near the Azores anticyclone (less than 400 ppt). A regional-scale maximum is visible over Northeast America (800–1400 ppt) and tropical Africa. The four models generally underestimate the magnitude of these geographical extrema, showing overly small variability. Finally, in Fig. , upper-tropospheric water vapour tends to be underestimated in the northern extra-tropics and, at least, in the northern tropics.

In this subsection, we compare and characterize the observed and modelled seasonal cycles together. We then show synthesizing metrics to assess the model geographical distributions in the extra-tropics. Figures – provide an overview of the seasonal climatologies in the UT and the LS, and Fig.  presents the same information for the non-separated UTLS. Please note that observed lower-stratospheric water vapour in Fig.  is displayed as an indication of the seasonality; however, for the reasons explained in Sect. , some of its values are probably overestimated, and it cannot be used for a bias quantification. The height of the box plots illustrates the geographical variability. Water vapour seasonality in the UT (shown in Fig. ) is well approximated in the simulations, with a wintertime minimum and a summertime maximum directly linked with convection and temperature, the latter of which governs saturation vapour pressure. The LS shows a similar pattern, although the contrast between the summertime water vapour maximum and the rest of the year is more pronounced than in the UT. This feature is consistent with the increased impact from the troposphere during this season and the extremely steep vertical gradient in water vapour. Based on CARIBIC measurements between 2005 and 2013, found that the summertime maximum was primarily due to shallow cross-tropopause mixing in the extra-tropics. This takes place during the whole year; thus, the summertime H2O maximum in the UT increases the upward moisture flux. The other two pathways for moisture transport into the LS, i.e. localized deep convection events and, higher in the LS, quasi-isentropic mixing with the tropical transition layer (TTL), mostly take place during summer (and early autumn for the latter), but they were found to have a lesser contribution to the summertime moisture maximum.

The models have more difficulties in reproducing CO seasonality, according to Fig. . The observed springtime peak has been explained by an accumulation of CO in the lower troposphere during winter followed by an increase in the convective transport during spring , allowing the lower-tropospheric CO reservoir to impact the UT. This springtime maximum is not visible in the simulations; its magnitude is underestimated as well, and contrary to the observations, the springtime distribution is similar to winter, a feature that extends up to the LS. The comparison with the realistic water vapour cycles in the UT tends to exclude convective transport from the causes of this discrepancy, except pyroconvection. It is possible that CO lifetime or CO emissions are underestimated, which reduces the wintertime accumulation in the lower troposphere and then the upward CO flux during spring. The lower values in the two tropospheric tracers (H2O and CO) in both the UT and LS with UKESM1.1 suggest an underestimation in the upward fluxes from the surface up to the UT, which favours an underestimation in the LS too.

Figure  shows that the ozone maximum in the upper troposphere takes place in summer, with a peak in photochemical activity, whereas the minimum occurs during winter. In the lowermost stratosphere, the ozone maximum takes place during spring due to the effects of the descending branch of the Brewer–Dobson circulation, which transports ozone-rich air masses down from the deeper extratropical stratosphere, whereas the minimum takes place during autumn. The models reproduce well these features and the dichotomy in the UT between high-ozone seasons (spring and summer) and low-ozone seasons (winter and autumn). Although most models are positively biased, the magnitude of the seasonal cycle is similar to the observations. In the LS, the ozone distribution is harder to reproduce during summer and autumn, with a geographical variability only spreading over the lower half of the observed distribution. It is characterized by a tendency to overestimate the low-latitude ozone minima and underestimate the high-latitude ozone maxima (e.g. Fig. ). Similarly to ozone, both observed and modelled NO_y mixing ratios show a springtime maximum in the LS and a summertime maximum in the UT. In the UT, the summertime maximum is linked to photochemical activity, enhanced lightning frequency, more intense boreal forest fires, and enhanced convection that uplifts diverse ozone precursors from the lower troposphere. This is consistent with the detailed individual NO_y species in Figs. –, where each of the models generally shows a summertime maximum in NO_x, PAN, and (especially) HNO3. The latter is notably affected by the conversion from NO_x via photochemical activity . In the LS, the impact of the Brewer–Dobson circulation coupled with HNO3 production from nitrous oxide decomposition in the stratosphere is reproduced, as shown in Fig. . The only exception is the UKESM1.1 model, which instead shows an upper-tropospheric seasonality in the LS, although the influence of the Brewer–Dobson circulation remains visible through more elevated springtime mixing ratios compared to the UT seasonal cycle. On the contrary, the OsloCTM3 model shows higher NO_y values in the LS. As ozone amounts are within the same range as the other models, it excludes stratospheric circulation from the possible causes of NO_y discrepancies. Thus, for UKESM1.1 (OsloCTM3), the lower (higher) HNO3 values in the LS (Fig. ) might be due to an underestimation (overestimation) of N2O flux into the stratosphere, an overestimated (underestimated) N2O lifetime in the stratosphere, or an underestimated (overestimated) HNO3 lifetime against stratospheric aerosol uptake. The latter is a possible contributor for OsloCTM3, as its mass density of particular sulfate and nitrate is 10 % lower than in the LMDZ–INCA simulation. Lower NO_y and HNO3 mixing ratios in UKESM1.1 are unlikely due to the different representation of the Brewer–Dobson circulation in the reanalyses, as the mean age of air in the northern LS is longer in ERA5 than in ERA-Interim , which would tend to convert more N2O into HNO3 in UKESM1.1. In the end, considering both ozone and NO_y in the LS, the similarities between observations and models, notably during spring, are encouraging with respect to the stratospheric chemistry and diabatic transport for all of the models.

As shown in Figs. – and , NO_y partitioning changes substantially between the models. In the UT, a higher proportion of NO_y is represented by HNO3 with OsloCTM3. As it is affected by wet scavenging, it can explain the lower NO_y mixing ratios with this model, combined with very low PAN quantities. On the contrary, the higher levels of NO_y in the UT with the EMAC model can be linked to the high proportion of PAN which is not soluble and has a chemical lifetime of several months e.g.. The higher amount of PAN might inherently be linked to the colder EMAC temperatures (∼-4 K), as it increases its lifetime against thermolysis. In the UT, the higher (lower) NO_x mixing ratios from the EMAC (LMDZ–INCA) model can also explain the higher (lower) ozone mixing ratios. The inter-model variability with respect to PAN in the UT might also have consequences for the air quality evaluation in the subsidence regions, as the PAN lifetime against thermolysis decreases drastically during descending motion, from months down to minutes when the temperature reaches 20 °C. Regarding this variability, surface ozone production due to PAN subsidence is, thus, likely to vary substantially across the models (at least in remote areas with a NO_x-limited regime), with a maximum for EMAC and a minimum for OsloCTM3. In the LS, PAN inter-model variability is the most noticeable of all NO_y species, with a factor reaching 12 between the median mixing ratio from EMAC and OsloCTM3 and an important difference in every couple of models. The low (high) amounts of PAN simulated with OsloCTM3 (EMAC) are, at least partially, related to the low (high) amounts in the UT as well.

The lower-stratospheric features described above are generally visible in the UTLS as well, as illustrated in Fig. , for the species with a strong positive vertical gradient. Notably, the springtime maximum is well represented by every model for ozone and almost all of the models for NO_y, including MOZART3, which confirms that all of the models catch the seasonality of the Brewer–Dobson circulation. The water vapour and temperature maxima in summer are also visible in the simulations. For the five models, the NO_y seasonal cycle is characterized by a springtime maximum in HNO3, due to the Brewer–Dobson circulation, and by a summertime maximum in both NO_x and PAN, due to convection, photochemistry, and lightning emissions.

In Figs. –, we note that, when they are noticeably biased, the sign of the annual mean bias is generally representative of all of the seasons, although its magnitude is not. With this perspective, the annual means shown below still provide relevant information. Figure  synthesizes some model skills in terms of annual averages in the extra-tropics, and the inter-model ranges are indicated more precisely in Table . The ozone mixing ratio is generally more difficult to model in the UT than in the LS, in terms of the geographical distribution (rUT(O3) =0.45–0.74, compared to rLS(O3) =0.75–90) as well as in terms of the mean biases, with the latter being essentially positive for most models in the UT (MNMB =0.005–0.36 and FGE ∼ MNMB for three models) and weak in the LS (MNMB =-0.08–0.006, with a maximum FGE at 0.16, which is particularly low). On the contrary, the CO correlation coefficient is higher in the UT (rUT(CO) =0.63–0.78) than in the LS (rLS(CO) =0.11–0.70), probably reflecting the difficulty in mapping the effects of cross-tropopause exchange. In the UT, ozone and CO biases tend to be positive and negative, respectively. This difference can be linked to overestimated cross-tropopause exchange and/or overestimated photochemical activity, thus more ozone production and more CO destruction. In the UT, both surface tracers (CO and H2O) also show good correlations (rUT(H2O) ∼0.95 and, for most models, rUT(CO) ∼0.8). The skill difference between the two species can be explained by either uncertainties in CO emissions in each region or an underestimation of the detrainment altitude from pyroconvection, consistent with the negative biases in CO in the UT. Interestingly, the bias magnitudes for both species are higher for EMAC and UKESM1.1. Concerning EMAC, the systematically negative temperature biases (-3.7 K on average) could be another factor controlling lower water vapour amounts in the UT via saturation, but this would not be consistent with the combination of lower temperatures (-4.0 K on average) with more water vapour in the LS, compared to the other models. A comparable cold bias with EMAC has been diagnosed in with a similar simulation set-up; this prior study also identified a wet bias compared to the observations from the Halogen Occultation Experiment HALOE; at 200 hPa in the extra-tropics. They concluded that an overestimation of lower-stratospheric water vapour would cause excessive radiative cooling and, thus, a cold bias – a relation that had already been shown in previous studies. This moisture overestimation in the LS is confirmed in Fig. , showing higher H2O amounts in EMAC compared to the observations, given that the latter are probably overestimated. Concerning UKESM1.1, it is worth reiterating that the low mixing ratios in H2O and CO in both layers (see Figs.  and ) suggest an underestimation in the upward fluxes from the surface up to the UT. Finally, NO_y shows the largest variability in the MNMB in the UT and the lowest correlation coefficient among the four chemical species. This could be due to the important inter-model variability in the spatial distribution of lightning emissions e.g. or to the washout of HNO3, the latter of which depends on the cloudiness representation and NO_y partitioning.

Figures – provide further information on the annual geographical distribution mentioned above for each species, layer, and model in the northern extra-tropics. The particularly high correlation for water vapour (r=0.95) shown in Fig.  is characterized by a well-reproduced meridional structure, notably with strong variability in the lowest latitudes (orange and red dots) due to dry subsiding and moist convective regions. This is notably characterized by a linear regression slope close to 1 for EMAC, LMDZ–INCA, and OsloCTM3. All of these features are representative of all of the seasons, with the highest correlation and linear regression slope during summer, when the tropospheric humidity reaches its maximum. These features are encouraging with respect to the modelled impact of meteorological systems on the extratropical UT in terms of geographical variability, despite the negative mean biases present in most models.

In the lowermost stratosphere, Fig.  shows that ozone geographical variability is relatively well reproduced by the models with a distinct northward gradient. This gradient tends to be underestimated because of a positive bias in the lowest values (in the subtropics) for most models and a negative bias in the highest values in the subpolar regions. The northward gradient is also visible for NO_y (Fig. ), with an underestimated regression slope and a substantially lower correlation coefficient. Contrary to ozone, this is characterized by poor correlations inside each zonal band. It suggests that the NO_y correlation in the LS is mostly due to the northward gradient and that the smaller scales are hardly captured by the models for this variable.

Concerning the UT in Fig. , the observed ozone climatology does not show any latitude gradient, with yearly means ranging between ∼50 and 80 ppb, independent of the latitude bin, except for some subtropical locations that are poorer in ozone and can reach ∼35 ppb. The EMAC model reproduces this feature relatively well, although with a systematic overestimation. The other models do not differentiate between the northernmost two bins (45–55 and 55–65° N), but they tend to make a distinction between 25–35, 35–45, and 45–65° N. LMDZ–INCA and OsloCTM3 tend to simulate a northward gradient. As this is a characteristic of the LS, it would be consistent with overestimated stratosphere–troposphere transport. Concerning UKESM1.1, a significant portion of the subtropical ozone values are higher than those in the high latitudes. This might be a consequence of the lower-stratospheric biases in the UT, via cross-tropopause exchange; as in the LS, subtropical ozone in the UT is particularly overestimated and high-latitude ozone is less abundant than in the other models. For NO_y in the UT (shown in Fig. ), we also notice that the models simulate more NO_y in the high latitudes and less in the subtropics (as for the LS), which is not consistent with the observations and also suggests overestimated cross-tropopause exchange.

It has to be noticed that these diagrams present average values through the whole sampled extra-tropics in the Northern Hemisphere. An assessment based on region-specific characteristics could provide more information on specific processes, such as tropopause folds during the Middle Eastern summer or isentropic transport from the tropical troposphere into the extratropical lowermost stratosphere . Another limitation of this approach is that the tropopause altitude decreases with the latitude, whereas the cruise altitude does not depend on latitude. Consequently, the subtropics are more sampled in the UT than in the LS; conversely, the high latitudes are more sampled in the LS than the UT. Thus, the scores shown for the different layers are not completely representative of the same geographical area.

The comparison with previous model assessment studies provides complementary information. First, most of the CCMVal-2 free-running models assessed in underestimated the vertical stability in the northern midlatitudes (40–60° N), especially for the semi-Lagrangian models and the models with the lowest vertical resolution. As a consequence, they generally underestimated ozone and HNO3, whereas they overestimated water vapour at 200 hPa, which is included in the lowermost stratosphere at these latitudes. Although all of the CCMVal-2 models did not have a specific tropospheric chemical scheme and only the EMAC model is involved in both studies, our results tend to confirm the ability of the models to reproduce (1) the seasonality of the Brewer–Dobson circulation through ozone and HNO3 tracers and (2) the overestimation of cross-tropopause mixing, notably the effect of the tropospheric influence on lower-stratospheric ozone that maximizes during summer and autumn. On this last point, the effect of the vertical resolution is visible on lower-stratospheric ozone with the lower-resolution model (LMDZ–INCA) showing the lowest vertical gradient in ozone. Still, it does not seem to be the most controlling factor for water vapour, as the EMAC model is one of the most highly resolved models but has the weakest water vapour vertical gradient, unless the nudging makes this inter-model hierarchy less evident. Concerning the impact from the Brewer–Dobson circulation on the LS, a better understanding of the simulations' biases could be established by exclusively assessing the dynamical behaviour; hence, adding other variables, like the stratospheric age of air or the zonal momentum, would be relevant for a more complete model evaluation e.g..

In addition to the model assessment, the model intercomparison of background CO, ozone, and NO_x in the UT and the LS (Figs. , , and , respectively) can provide a further understanding of each model's ozone sensitivity to aircraft NO_x emissions, as the critical NO mixing ratio separating net production and net destruction of ozone depends on these three parameters . Although it ignores the behaviour of lots of non-measured VOCs and methane, it still provides a comparison of several factors controlling the sensitivity of the net ozone production to NO_x emissions. In the UT, we can expect the most different ozone responses between the EMAC and LMDZ–INCA models, as EMAC shows higher NO_x and ozone values and lower CO values, contrary to LMDZ–INCA. In the LS, it can also be expected that the LMDZ–INCA model maximizes the ozone response, as NO_x (CO) is relatively low (high), and this difference can be enhanced during summer and autumn with relatively lower ozone values. The two models showing the highest NO_x values (EMAC and especially OsloCTM3, with more than twice the median compared to LMDZ–INCA and UKESM1.1) can be expected to have a lower ozone response.

The zonal cross sections shown in Figs.  and  compare the reference runs with the observations in three tropical regions: South America–Atlantic (called South America hereafter), Africa, and South Asia. First, we present a brief summary of some observed patterns that have been investigated in , notably based on , , and for the three respective regions from west to east. In a second step, we present an overview of model skill. The mean pressure is represented as well, in order to identify the changes in the observed variable that can be associated with changes in the sampling mean altitude. This case occurs at the edge of the sampled regions, notably for NO_y during December–February above South America and during June–October above Africa. The IAGOS–OsloCTM3 profiles are represented as well, as their sampling period is shorter than the other three models (2001–2017 instead of 1995–2017), which causes important differences only in the IAGOS–OsloCTM3 ozone transects in July–August in South America and water vapour transects in June–October above Africa.

Most of the observed features have been investigated in , such as the impacts of wet and dry seasons, linked to shifts in the ITCZ. In the two western regions (South America and Africa), where the zonal cross sections cover most of the tropical latitudes, the wet season is characterized by a co-located maximum in water vapour and a minimum in ozone, both linked to the intense convection of humid surface air. The latter is rather rich in fresh pollutants and becomes enriched in NO_x emitted by lightning during convective uplift. In the upper branch of both Hadley cells, ozone is produced by photochemistry during its poleward transport.

Above Africa, the seasons with the northernmost and the southernmost ITCZ (June–October and December–March, respectively) are particularly visible in the IAGOS observations. During the meridional transport in the upper branch of the strongest Hadley cell, CO accumulates in the areas where zonal wind shear is greater , visible in December–March for NO_y, reaching a maximum at approximately 10° from the ITCZ. Using a method based on the FLEXPART Lagrangian dispersion model SOFT-IO;, found that these CO peaks originated from intense biomass burning in the dry season. A sensitivity test regarding biomass burning with the LMDZ–INCA model found similar conclusions. In the monsoon season (June–October), both studies agree on a major biomass burning contribution to the southward shift in the peak of CO. Finally, the Asian summer monsoon (right panels in Fig. 12) is characterized by the warmest and most humid air masses, as expected from the strongest convective system. During this season, the subtropical jet stream and its subsequent stratospheric intrusions are confined to the northern side of the Himalayas , thereby ensuring a weak stratospheric influence in this season .

The analysis of the modelled local and seasonal features provides interesting information about the representation of the convective systems and their outflows. All of the models capture the location of the peaks in water vapour in the African upper troposphere, irrespective of the season; the maximized water vapour amounts (as well as temperature, not shown) above the Asian summer monsoon (Fig. , right column); and the strongest CO peaks above Africa. More precisely, above Africa, in the December–March (June–October) season, the models capture the southward (northward) shift in the ITCZ relatively well, as characterized by an ozone minimum co-located with the water vapour maximum. This agreement among the models and with the observations highlights a realistic representation of the strongest convective systems. It is probably improved by the nudging, by the use of a common surface temperature field based on observations, and by a common (or similar) inventory for biomass burning emissions.

Regarding the effects of convection, the variability between the models can be found in the peaks' intensity of water vapour and CO and in the CO peaks' location. In most cases, water vapour shows a small bias for LMDZ–INCA and OsloCTM3 and a dry bias for EMAC and UKESM1.1. The EMAC dry bias is possibly explained by a cold bias in the UT (∼-5 K, not shown) that lowers the saturation vapour pressure and/or the detrainment altitudes. The other models show particularly well-reproduced temperatures, and all of the modelled temperature profiles are well correlated with the observations (not shown).

The location and the width of the CO maximum depend on the model, notably above Africa: it is rather co-located with the ITCZ for the EMAC and UKESM1.1 models, whereas it is shifted 5–10° equatorward for the other models, in agreement with the observations. Concerning June–October, with the same observation dataset, showed a peak in the anthropogenic contribution to upper-tropospheric CO co-located with the ITCZ (as for EMAC and, to a lesser extent, UKESM1.1 and OsloCTM3) and a peak in the biomass burning contribution shifted 10° southward (as for the LMDZ–INCA model). OsloCTM3 seems to show a compromise between the two categories, with a flatter and wider maximum including both the ITCZ position and the observed CO peak. A sensitivity test using the LMDZ–INCA model that reproduces the CO peak during December–March and June–October well (with respect to its location and its magnitude, the latter of which is near 140 ppb) concluded that biomass burning is the main factor influencing the peak intensity, with a contribution reaching 30 and 45 ppb during DJFM and JJASO, respectively, but also with a southward shift in the CO peak during June–October. Thus, the negative CO bias, combined with the absence of the southward shift in the EMAC and UKESM1.1 simulations, is likely to reflect an underestimation of the impact of biomass burning in the tropical UT. As the dry bias in the water vapour peaks in these two models suggests a less intense convection, it implies a weaker Hadley circulation. Consequently, the lower-tropospheric entrainment into convective motions in the ITCZ has a reduced geographical extent and, thus, includes less air from the dry region. This could explain the lack of CO accumulation in the higher tropical latitudes with these two models. Inversely, the LMDZ–INCA model and, to a lesser extent, OsloCTM3 show another peak in CO in July–August above South America at relatively similar latitudes to the peak in Africa, although this is absent from the observation profiles. As it is mainly due to biomass burning for LMDZ–INCA and because LMDZ–INCA and OsloCTM3 show similar behaviours for CO, it suggests that both models overestimate the effects of the intercontinental connection with Africa during this season (with respect to duration and/or intensity) or the impact of local biomass burning emissions.

Contrary to CO, the peaks in NO_y observed in December–February (December–March) show an important negative bias in LMDZ–INCA and OsloCTM3, whereas the negative bias is lower with EMAC and UKESM1.1, especially above South America. Concerning the Asian summer monsoon, all of the represented models overestimate ozone and NO_y mixing ratios, possibly reflecting an overestimation of the lightning flash rate, as is known for LMDZ–INCA ; an underestimation of HNO3 uptake; and/or an overestimation of the entrained surface pollutants.

Most of the models tend to overestimate ozone, except for the LMDZ–INCA model that instead shows negative biases. The overestimation of ozone in the tropical UT from UKESM1.1 is consistent with a recent comparison based on ozone partial columns between UM–UKCA and the OMI–MLS satellite observations during the 2005–2018 period and in the 450–170 hPa pressure range. The overestimation is representative of the whole tropospheric ozone column in the tropics, and the main factor in the UT is probably an overestimation of the lightning NO_x emissions. The LMDZ–INCA model shows particularly low NO_x levels, with the lowest (highest) mean values near 18 ppt (195 ppt), whereas the other models have NO_x minimum (maximum) levels at 52–91 ppt (278–430 ppt). This first-order statement can be sufficient to explain most of the LMDZ–INCA lower ozone values as well as most of the higher CO values with longer photochemical lifetimes for CO. Each of these two factors favours ozone production efficiency from lightning and aviation. A similar diagnostic applies to HNO3. Although the stronger convection with LMDZ–INCA compared to EMAC theoretically produces more NO_x due to a higher lightning activity, it is possible that the LMDZ–INCA model overestimates NO_y removal by HNO3 wet scavenging, both with a more efficient conversion of NO_x into HNO3 and with further precipitation due to a stronger convection.