The TCR-2 product employs the MIROC-Chem global chemistry transport model. The model has a T106 (1.1° × 1.1°) horizontal resolution with 32 vertical levels extending from the surface to 4 hPa. The TCR-2 meteorological fields are produced using the MIROC-AGCM (Watanabe et al., 2011) general circulation model with the model simulation nudged toward 6-hourly ERA-Interim reanalysis fields (Dee et al., 2011). The chemical mechanism in the model consists of 92 chemical species and 292 reactions. A priori anthropogenic emissions in the model are from the HTAP version 2 inventory (Janssens-Maenhout et al., 2015). NO_x emissions from soils are based on the Global Emissions Inventory Activity (GEIA) (Graedel et al., 1993). Biomass burning emissions are from the Global Fire Emissions Database version 4 (GFED4) (Randerson et al., 2018) and biogenic emissions are from version 2.1 of the Model of Emissions of Gases and Aerosols from Nature (MEGAN2.1; Guenther et al., 2006, 2012).

TCR-2 used a Local Ensemble Transform Kalman Filter (LETKF) (Hunt et al., 2007) to assimilate NO₂ observations from the SCanning Imaging Absorption spectroMeter for Atmospheric CartograpHY (SCIAMACHY), the Ozone Monitoring Instrument (OMI), and GOME-2; ozone observations from the Tropospheric Emission Spectrometer (TES) and the Microwave Limb Sounder (MLS); CO data from MOPITT; HNO₃ observations from MLS; and SO₂ data from OMI. The assimilated observations were used to optimize the atmospheric mixing ratio of ozone, NO_x, HNO₃, pernitric acid (HNO₄), dinitrogen pentoxide (N₂O₅), peroxyacetyl nitrate (PAN), and peroxymethacryloyl nitrate (MPAN). Emissions of NO_x (from the surface and lightning), SO₂, and CO were also optimized in the assimilation. Although the model has 92 chemical species, the state vector in the assimilation consisted of a smaller subset of 35 species as well as the NO_x, CO, and SO₂ emissions. The emissions were optimized using a state augmentation approach that used the background error covariance, determined from the forecast ensemble, to link the emissions and the atmospheric concentrations in the optimization.

The Copernicus Atmosphere Monitoring Service reanalysis (CAMSRA) product (Inness et al., 2019) is produced using the Integrated Forecast System (IFS) of the European Centre for Medium-Range Weather Forecasts (ECMWF). The product has a horizontal resolution of T255 (≈ 0.75°), with 60 vertical levels from the surface to 0.1 hPa. The chemical mechanism in IFS is an extended version of the CB05 (Yarwood et al., 2005) chemical mechanism for the troposphere, which consists of 55 chemical species and 126 reactions. Explicit chemistry is not included in the stratosphere. Instead, stratospheric ozone chemistry is parameterized using the “Cariolle-scheme” (Cariolle and Déqué, 1986; Cariolle and Teyssèdre, 2007). Anthropogenic emissions in CAMSRA are from the MACCity inventory (Granier et al., 2011), with modifications to increase wintertime road traffic emissions over North America and Europe following the correction of Stein et al. (2014). Biogenic emissions are from MEGAN2.1 (Guenther et al., 2006, 2012) driven by meteorological fields from the Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2; Gelaro et al., 2017). Soil and oceanic emissions are from the Precursors of Ozone and their Effects in the Troposphere (POET) database for 2000 (Olivier et al., 2003; Granier et al., 2005). Biomass burning emissions are from the Global Fire Assimilation System, version 1.2 (GFASv1.2; Kaiser et al., 2012).

The CAMSRA assimilation scheme is an incremental four-dimensional variational (4D-Var) data assimilation scheme (Courtier et al., 1994) with 12 h assimilation windows from 09:00–21:00 and 21:00–09:00 UTC. Background error covariances are diagonal so that chemical species in the control vector are optimized independently. CAMSRA optimizes ozone, CO, NO₂, and aerosol mass mixing ratios using observations of ozone from SCIAMACHY, the Michelson Interferometer for Passive Atmospheric Sounding (MIPAS), MLS, OMI, GOME-2, and the Solar Backscatter Ultraviolet Radiometer (SBUV/2), together with CO observations from MOPITT and NO₂ observations from SCIAMACHY, OMI, and GOME-2. Ozone precursor emissions are not optimized in the assimilation.

The GEOS-Chem reanalysis (Qu et al., 2020a) was produced using version v35k of the GEOS-Chem adjoint model (Henze et al., 2007). The model is driven by MERRA-2 (Gelaro et al., 2017) meteorological fields at a horizontal resolution of 2° × 2.5° with 47 levels from the surface to 0.01 hPa. The model has detailed tropospheric ozone chemistry, with 43 transported species and 297 chemical reactions. Stratospheric ozone is parameterized based on the linearized ozone scheme of McLinden et al. (2000). A priori anthropogenic emissions in the model are from the HTAP version 2 inventory (Janssens-Maenhout et al., 2015). Biomass burning emissions are from GFED4 (Randerson et al., 2018), NO_x emissions from soils are based on Yienger and Levy (1995), and biogenic emissions are from MEGAN2.1 (Guenther et al., 2006, 2012)

GEOS-Chem uses a 4D-Var data assimilation scheme to assimilate NO₂ slant column densities (SCD) from OMI to optimize NO_x emissions. The GEOS-Chem reanalysis presented here assimilated OMI NO₂ retrievals from version 3 of the NASA standard product OMNO2 (Krotkov et al., 2017). The GEOS-Chem reanalysis assimilated only NO₂ as the focus of Qu et al. (2020a) was on the impact of discrepancies in OMI NO₂ retrievals on inferred NO_x emissions and posterior ozone simulations. The assimilation minimizes a cost function that is the sum of observation-error-weighted differences between the modeled and retrieved SCDs and departures of the emission scaling factors from the prior estimates weighted by the prior emissions error. In constructing the cost function, the modeled NO₂ vertical column densities (VCDs) are converted to SCDs using scattering weights from the OMI retrievals. Additional details of the GEOS-Chem assimilation are discussed in Qu et al. (2020a).

The CMAQ regional reanalysis (Kumar et al., 2025) uses version 5.3.2 of the CMAQ model driven by meteorological fields from version 4.1 of the Weather Research and Forecasting (WRF) model (Skamarock and Klemp, 2008). The reanalysis was conducted at a horizonal resolution of 12 km × 12 km over the CONUS with 35 vertical levels from the surface to 50 hPa. The meteorological initial and boundary conditions for WRF are from the ERA-Interim reanalysis, while chemical initial and boundary conditions for CMAQ are from the Whole Atmosphere Community Climate Model (Marsh et al., 2013; Gettelman et al., 2019). The chemical mechanism in the model is based on the Carbon Bond 6 version r3 scheme for gas-phase chemistry with the AERO7 aerosol module for aerosol processes, including secondary organic aerosols (Appel et al., 2021). The chemical mechanism has 147 species and 343 reactions. A priori anthropogenic emissions in CMAQ are based on the US EPA National Emissions Inventory (NEIv2) for 2011 (https://www.epa.gov/air-emissions-inventories/2014-national-emissions-inventory-nei-data, last access: 5 May 2025). WRF meteorology was used with the Sparse Matrix Operator Kernel Emissions (SMOKE) (de Almeida Albuquerque et al., 2018) to produce meteorology-dependent anthropogenic emissions for 2011, 2014, and 2017, which were adjusted in time using EPA reported annual state-wise trends. Biomass burning emissions are from the Fire Inventory from NCAR (FINN) version 2.2 (Wiedinmyer et al., 2023). Biogenic emissions are specified using the Biogenic Emission Inventory System (BEIS) (Vukovich and Pierce, 2002).

CMAQ assimilated standard Level 2 Collection 6.1 AOD from the Moderate Resolution Imaging Spectroradiometer (MODIS) and Version 8 of the MOPITT CO multispectral retrievals using a three-dimensional variational (3D-Var) data assimilation scheme. The CO chemical data assimilation capability in Kumar et al. (2025) was a recent addition to the previously existing MODIS AOD data assimilation capability. The assimilation optimized total aerosol mass per mode (Aiken, accumulation, and coarse) and CO mixing ratios. Background error covariance matrices generated for January and July are used seasonally to represent wintertime and summertime background error covariances, respectively. The wintertime background error covariance matrix is used for assimilating observations between November to March, while the summertime background error covariance matrix is used during the other months.

The CAQRA product (Kong et al., 2021) is a regional reanalysis for Asia that employs the Nested Air Quality Prediction Modeling System (NAQPMS) chemical transport model (Wang et al., 2000). The model is driven by meteorological fields from WRF at a horizontal resolution of 15 km × 15 km. The meteorological initial and boundary conditions for WRF are from the NCAR-NCEP reanalysis, while chemical boundary conditions are from the Model for Ozone and Related Chemical Tracers (MOZART; Brasseur et al., 1998; Hauglustaine et al., 1998) model. The chemical mechanism in CAQRA is the carbon bond mechanism Z (Zaveri and Peters, 1999), with aqueous-phase chemistry and wet deposition based on the Regional Acid Deposition Model (RADM) mechanism from version 4.6 of CMAQ and inorganic aerosol processes represented by ISORROPIA1.7 (Nenes et al., 1998). Anthropogenic emissions in the model are from the HTAP version 2.2 inventory with a 2010 base year (Janssens-Maenhout et al., 2015). Emissions of VOCs are from the MEGAN-MACC model (Sindelarova et al., 2014). Biomass burning emissions are from GFED4 (Randerson et al., 2018), soil NO_x emissions are from the Regional Emission Inventory in Asia (Yan et al., 2003), and oceanic emissions are from the POET database (Granier et al., 2005).

CAQRA uses an LETFK data assimilation scheme to assimilate surface observations of ozone, CO, NO₂, SO₂, PM_2.5, and PM₁₀ to optimize the atmospheric concentration of these constituents. In constructing the background error covariance, inter-species correlation is neglected. Thus, in the assimilation, each chemical species is optimized using only observations of that species. Additional details of the CAQRA assimilation configuration are available in Kong et al. (2021).

To evaluate the reanalyses, we use ozonesonde observations from the TOAR-II Harmonization and Evaluation of Ground-based Instruments for Free Tropospheric Ozone Measurements (HEGIFTOM) effort (https://hegiftom.meteo.be/datasets/ozonesondes, last access: 5 May 2025). The measurement precision for the ozonesondes is better than 3 %–5 % with an accuracy of about 5 %–10 % (Smit et al., 2007). The HEGIFTOM ozonesonde data were harmonized to remove systematic biases and to provide an uncertainty estimate for every measurement. The database contains time series observations from 43 sites. Here we use data from 39 sites that each had a minimum of 48 measurements between January 2003 and December 2022.

Surface ozone observations from the TOAR-I data set (Schultz et al., 2017) were used for evaluation of the reanalyses. The dataset consists of a consistent long-term record (1990–2014) of surface ozone observations that have been harmonized and processed with consistent quality control. We used the 2° × 2° gridded monthly mean observations for the period 2003–2014 for comparison with the reanalysis. All of the reanalyses were regridded onto the ozone 2° × 2° grid for the evaluation. The ozone data are available for both rural and urban sites. However, since the lowest resolution of the available reanalyses is 2° × 2.5°, which cannot reliably distinguish between urban and rural locations, we follow the approach of Huijnen et al. (2020) and Sekiya et al. (2025) and use only rural TOAR-I observations in the evaluation of the reanalyses.

The mean ozone distribution at the surface from the five reanalyses for 2006–2016 and the differences between the individual reanalyses and the ensemble mean are shown in Fig. 1. The global mean distribution (Fig. 1a) consists of a band of high ozone across the northern subtropics and a minimum in the tropics. In North America, there are high values over the southeastern US and the mountain west. In Asia, high ozone values are located over the Tibetan plateau. Figure 1b shows that GEOS-Chem is high everywhere compared to the ensemble mean, with differences exceeding 10 nmol mol⁻¹ in Asia and western North America. Figure S1 shows that the largest relative differences in GEOS-Chem are between 20 %–40 %. In contrast, CAMSRA is low over all continental regions, with the largest difference of about 10–15 nmol mol⁻¹ over central Africa and parts of Asia. The largest relative differences in CAMSRA are in central Africa and western South America, where the reanalysis is more than 40 % lower than the ensemble mean (Fig. S1 in the Supplement). TCR-2 is slightly lower than the ensemble mean at the high latitudes and about 5–10 nmol mol⁻¹ higher over tropical South America and central Africa. The high ozone over tropical South America and central Africa in TCR-2 correspond to differences exceeding 40 % relative to the ensemble mean in these regions. For the regional reanalyses, CAQRA is about 5 to 15 nmol mol⁻¹ lower (less than 40 %) across much of Asia, with small positive differences over Thailand and Myanmar, whereas CMAQ had small differences of less than 5 nmol mol⁻¹ across much of the US, with positive differences in the Pacific Northwest and negative differences in the southeastern US. Examination of the seasonal differences between the individual reanalyses and the ensemble mean (Fig. S2) reveals that the pattern of differences shown in Fig. 1b is relatively consistent seasonally in all reanalyses, with the exception of CMAQ, although the magnitude of the differences is generally larger in the winter hemisphere. CMAQ has positive differences across much of the US in December–February (DJF), and negative differences in June–August (JJA). However, the high ozone in GEOS-Chem relative to the ensemble mean could be because the GEOS-Chem reanalysis did not assimilate ozone observations. Sekiya et al. (2025) found that ozone in the lower troposphere was high in the GEOS-Chem control (the model without assimilation) relative to the TCR-2 and CAMSRA controls, and that data assimilation increased ozone in CAMSRA and TCR-2, but had a minimal impact on ozone in GEOS-Chem.

In the middle troposphere, the mean ozone distribution, shown in Fig. 2a, is similar to that at the surface, with a band of high ozone across the northern subtropics and mid-latitudes, but with fewer small-scale features than at the surface. Note that only the global reanalyses are evaluated in the free troposphere since only surface fields are available from the regional reanalyses. Examination of the differences between the individual global reanalyses and the ensemble mean (in Fig. 3a) shows that the reanalyses are all closer to the mean in the middle troposphere. The relative differences are shown in Fig. S3. The major differences with respect to the ensemble mean are mainly in the tropics and subtropics, with TCR-2 and GEOS-Chem exhibiting an opposite pattern of differences. The differences in the individual reanalyses are generally less than 5 nmol mol⁻¹, except for localized regions in the tropics where the differences are larger. TCR-2 has a large difference with respect to the ensemble mean over the tropical Atlantic, northern South America, and the Indian Ocean, where the relative differences are 10 %–20 %, with the largest relative difference exceeding 20 % over northern South America (Fig. S3). The GEOS-Chem relative differences are 10 %–20 % and are opposite in sign to those of TCR-2 over the tropical Atlantic, northern South America, and the Indian Ocean. CAMSRA has small positive differences over the Pacific and small negative differences over Africa and the southern tropical Atlantic and Indian Oceans (Fig. 3a). The pattern of the differences in the reanalyses is consistent seasonally (see Fig. S4) but with a seasonally-dependent latitudinal shift. In GEOS-Chem and TCR-2 the difference pattern is shifted into the tropics of the summer hemisphere in a manner similar to the seasonal shift of the Intertropical Convergence Zone (ITCZ). The spatial distribution of the differences in GEOS-Chem and TCR-2 in the tropics and subtropic is similar to the spatial pattern of the multi-model spread in ozone in the middle troposphere reported by Miyazaki et al. (2020a) in their multi-model data assimilation analysis using four different CTMs, including GEOS-Chem and MIROC-Chem. Indeed, they found that the largest multi-model spread was over northern south America, where we see large differences in TCR-2 and GEOS-Chem. Miyazaki et al. (2020a) suggested that the large multi-model spread over northern South America could be due to the background errors not accurately capturing rapid error growth associated with deep convection and biomass burning. They also suggested that discrepancies in isoprene emissions and the isoprene oxidation chemistry could be another factor contributing to the large multi-model spread in the region.

At 250 hPa the mean ozone distribution shown in Fig. 2b reveals the influence of the extratropical lower stratosphere with high ozone at the high-latitudes. In the tropics there is a minimum in ozone over the warm pool of the tropical western Pacific, which reflects the influence of convective transport (e.g., Pan et al., 2017). The mean differences between the individual reanalyses and the ensemble mean (Fig. 3b) are small, with the largest differences mainly in the southern midlatitudes. As a result of the minimum in ozone over the tropical western Pacific, the largest relative differences in TCR-2 and CAMSRA are found in this region (Fig. S5). In addition to large relative differences over the western Pacific, GEOS-Chem also has large relative ozone differences in the southern midlatitudes, where it is high relative to the ensemble mean. The spatial pattern of the differences in TCR-2 is opposite that in GEOS-Chem. The high ozone in GEOS-Chem relative to the ensemble mean is likely due to the fact that GEOS-Chem did not assimilate ozone observations. The high ozone in GEOS-Chem extends across the southern extratropics in DJF and MAM (see Fig. S6), whereas it is confined to the southern midlatitudes in JJA and SON. In TCR-2, the differences in the southern midlatitudes are largest in JJA and SON. CAMSRA has high ozone relative to the ensemble mean in the northern high-latitudes in JJA and low ozone in the southern high-latitudes in DJF.

Daily maximum 8 h average (MDA8) ozone is a daily metric widely used for air quality standards and for ozone exposure studies (e.g., Turner et al., 2016; Fleming et al., 2018; Lyu et al., 2019; Chen et al., 2024). In particular, it is used for evaluating the mortality risks associated with short-term ozone exposure (Fleming et al., 2018). We include MDA8 in our evaluation since the ozone reanalyses could be useful for ozone exposure studies where local surface air quality observations are lacking. The ensemble mean MDA8 ozone distribution and the differences between the individual reanalyses and the ensemble mean are shown in Fig. 4. The spatial pattern of differences in MDA8 (Fig. 4b) is similar to that shown for ozone in Fig. 1b, except for the CMAQ reanalysis. In CMAQ, MDA8 ozone is high relative to the ensemble mean over the western and central US, in contrast to the pattern of differences in mean ozone shown in Fig. 1b. Overall, we find that the regional differences in MDA8 ozone between the individual reanalyses and the ensemble mean are generally smaller than the differences shown for surface ozone in Fig. 1b. The exception is the TCR-2 reanalysis which has larger MDA8 differences with respect to the ensemble mean over South America and central Africa.

The seasonal variations in regional mean ozone are shown in Figs. 5–7 for the nine regions defined in Table 2 (and shown in Fig. S7). The regions are based on those in Miyazaki and Bowman (2017). At the surface (Fig. 5), the seasonal cycle is consistent across all the reanalyses. In all regions, CAMSRA has the lowest ozone mixing ratios. For example, in the US, ozone in CAMSRA is about 5 nmol mol⁻¹ lower than in TCR-2. In contrast, GEOS-Chem ozone is 5–10 nmol mol⁻¹ higher than in TCR-2. Indeed, GEOS-Chem has the highest ozone concentrations in all regions except for South America. In South America, TCR-2 has higher mean ozone than GEOS-Chem in all months except for August. In China, CAQRA is fairly consistent with CAMSRA.

In the free troposphere, at 500 hPa (Fig. 6), the seasonality of the global reanalyses is generally consistent in the extratropics, which is in agreement with the results shown in Fig. 3a. In the tropics, there are larger discrepancies between the reanalyses over South America, North Africa, and the Middle East. Over South America, for example, TCR-2 has higher ozone concentrations between January–July and October–December, similar to the discrepancies observed at the surface (Fig. 5). Over the US, ozone concentrations peak about three months later in GEOS-Chem than in the other reanalyses. And over China, GEOS-Chem has a broader ozone maximum than CAMSRA and TCR-2. In the UTLS (Fig. 7), TCR-2 and CAMSRA are generally in agreement in terms of the ozone concentrations and variability in all regions, probably due to the assimilation of ozone measurements from MLS (Sekiya et al., 2025). Over South America and southern Africa, GEOS-Chem exhibits a larger seasonal cycle than CAMSRA and TCR-2. The discrepancy could be attributed to the lack of assimilation of ozone measurements in GEOS-Chem.

The time series of the regional mean ozone concentrations are shown in Figs. 8–10. As can be seen in Fig. 8, the lower ozone mixing ratios at the surface in CAMSRA are present in all years. The differences in the mean concentrations between CAMSRA and the other reanalyses is particularly pronounced in the three African regions. Across most regions, with the exception of South America, the three global reanalyses have similar interannual variability in surface ozone, with a high ozone in GEOS-Chem and a low ozone in CAMSRA. In South America, ozone in TCR-2 is generally higher than 20 nmol mol⁻¹ throughout the year, while CAMSRA ozone is lower than 10 nmol mol⁻¹ during the summer months. For the regional reanalyses, we find that in the US, surface ozone variability in CMAQ is similar to that in GEOS-Chem, whereas in China, CAQRA is similar to CAMSRA. In the northern extratropical middle troposphere, over the US, Europe, and China (Fig. 9), the variability in the three global reanalyses is remarkably similar. The worst agreement is found over Northern Africa, where the interannual variability is significantly different in each of the three global reanalyses. Over South America and central Africa the simulation of the ozone maxima in TCR-2 is consistent with the other reanalyses, but TCR-2 significantly overestimates the ozone minima. In the UTLS (Fig. 10), Northern Africa is the region with the greatest disagreement between the reanalyses in their simulation of the ozone variability. The South American, Northern African, and Central African regions are all strongly influenced by convective transport, which is highly variable and will pose a challenge for the reanalyses. These regions are also strongly influenced by biomass burning emissions and biogenic ozone precursor emissions, which are poorly constrained in models.

The linear trend in surface ozone in the reanalyses is shown in Fig. 11. There are large regional differences between the reanalyses in the trends. For example, in Asia TCR-2 has positive trends (that are as much as 0.4 nmol mol⁻¹ yr⁻¹), whereas CAMSRA has large negative trends in East Asia (exceeding -0.4 nmol mol⁻¹ yr⁻¹) with small positive trends in South Asia. In North America, the pattern of the trends is similar in CAMSRA, GEOS-Chem, and CMAQ, with negative trends over the southeastern US and the mountain west, and positive trends over northern and eastern Canada, although the negative trends in GEOS-Chem are not statistically significant. In TCR-2 there are negative trends over much of North America. The main statistically significant trends in GEOS-Chem are in the southern mid-latitudes, where there are negative trends that are consistent with the negative trends that TCR-2 exhibits across the southern extratropics. However, in this region CAMSRA has positive trends.

The mean surface ozone observations from TOAR-I for 2006–2014 and the differences between the individual reanalyses and the TOAR observations are plotted in Fig. 12. The relative biases are shown in Fig. S8. The evaluation only extends to 2014 because observations in the TOAR-I database are not available after 2014. In addition, the CAQRA reanalysis was not included in this evaluation because of the short temporal overlap between CAQRA and the TOAR-I database and the limited number of TOAR-1 observations over Asia. As can be seen in Fig. 8b, GEOS-Chem, TCR-2, and CMAQ are biased high relative to the TOAR observations, with GEOS-Chem exhibiting the highest global bias of about 14 nmol mol⁻¹ (47 %). The mean bias for CAMSRA and TCR-2 is 1.6 nmol mol⁻¹ (7 %) and 5.5 nmol mol⁻¹ (20 %), respectively. Over the US, CMAQ has a mean high bias of 6.4 nmol mol⁻¹ (22 %). The CAMSRA global mean bias reflects the compensating influence of negative biases over the central and western US and over central and eastern Europe, and positive biases everywhere else. Examination of the seasonality of the bias (see Fig. S9) reveals that in GEOS-Chem the largest global mean bias with respect to the TOAR observations is 16.6 nmol mol⁻¹ in JJA and the smallest mean bias is 10.5 nmol mol⁻¹ in DJF. In TCR-2 and CAMSRA the largest mean bias is 9.8 nmol mol⁻¹ and 5.3 nmol mol⁻¹, respectively, in JJA and the smallest mean bias is 1.7 nmol mol⁻¹ and -1.6 nmol mol⁻¹, respectively, in DJF. In CMAQ the largest total mean bias is 10.4 nmol mol⁻¹ in DJF, while the smallest total mean bias is 4.3 nmol mol⁻¹ in JJA.

A comparison of the global reanalyses with ozonesonde data in the middle troposphere and UTLS is shown in Fig. 13. The ozonesonde data were binned into the following three latitude bins: 20–60° S, 20° S–20° N, and 20–60° N. There were three ozonesonde sites used in the 20–60° S latitude bin, 11 in the 20° S–20° N bin, and 17 in the 20–60° N bin. At 500 hPa the reanalyses capture the variability in the ozonesonde data well, with correlations of about 0.8 and 0.9 in the southern and northern midlatitudes, respectively. The correlation is lower in the tropics, with values between 0.5–0.7. The higher correlation in the midlatitudes is expected since ozone transport in this region is dominated by large-scale synoptic processes that are well constrained by meteorological reanalyses. In contrast, transport in the tropics is dominated by convective processes, which are less well constrained by models and meteorological reanalyses. The mean bias across all three regions in the middle troposphere is less than 4 nmol mol⁻¹ (10 %). In the northern midlatitudes, the largest bias, which is in the GEOS-Chem reanalysis, is only about 2 nmol mol⁻¹ (4 %). The standard deviation is also low, less than 4 nmol mol⁻¹, in the tropics and northern midlatitudes. The standard deviation increases to about 6 nmol mol⁻¹ in the southern midlatitudes. However, it should be noted that there are only three ozonesonde sites in the southern midlatitude region.

In the UTLS (at 250 hPa) the correlations in the midlatitudes are just as high as at 500 hPa. However, the correlations in the upper tropical troposphere are slightly lower. In the northern midlatitudes, the largest mean bias is about 10 nmol mol⁻¹ (10 %). In the tropics, where the ozone mixing ratio is lower, the mean bias is smaller. The largest mean bias, exceeding 20 nmol mol⁻¹ (30 %) is found in GEOS-Chem in the southern midlatitudes. This large bias in GEOS-Chem in the southern midlatitudes could explain the high ozone in GEOS-Chem relative to the ensemble mean in the upper troposphere of the southern midlatitudes shown in Fig. 3b. The standard deviation in the UTLS is larger than at 500 hPa, particularly in the midlatitudes, which is expected given the higher ozone mixing ratios in the lower stratosphere and the small-scale meteorological processes that drive variability in the tropopause region. At 250 hPa the ozone distribution will be strongly influenced by variations in the tropopause and it is a challenge for models to reproduce the small-scale, dynamically-driven variations in ozone in this region.

As NO₂ is a key ozone precursor, examination of the NO₂ distribution in the reanalyses could provide insight in the source of the differences in surface ozone between the reanalyses. The ensemble mean NO₂ distribution and the differences between the individual reanalyses and the ensemble mean are plotted in Fig. 14. As shown in Fig. 14b, the ensemble mean reflects the combined influence of low NO₂ mixing ratios in GEOS-Chem and high NO₂ in CAMSRA. CMAQ has low NO₂ in the US, whereas CAQRA generally has high NO₂ in eastern China and low NO₂ in western China, northern India, and southeast Asia. The seasonal variations in regional mean NO₂ at the surface are shown in Fig. 15. Across all region, GEOS-Chem has low NO₂ seasonally compared to the other reanalyses. For most regions, except South America and China, CAMSRA has higher NO₂ than TCR-2. For example, over North America, India, and the Middle East, CAMSRA has much higher NO₂ than TCR-2. The large discrepancies in surface NO₂ between the reanalyses is due, in part, to the differences in the assimilation configuration employed in the reanalyses. For example, GEOS-Chem constrained only NO_x emissions, CAMSRA constrained NO₂ and ozone mixing ratios, TCR-2 constrained NO₂ and ozone mixing ratios as well NO_x emissions, and CMAQ constrained aerosol mass and CO mixing ratios. In addition, GEOS-Chem has the lowest spatial resolution at 2° × 2.5°. These differences will have a significant impact on the simulated surface NO₂ in the reanalyses since NO_x emissions represent a strong forcing on surface NO₂ mixing ratios in the assimilation, with implications for surface ozone. In their observing system experiments, Sekiya et al. (2025) found that assimilation of the satellite NO₂ data produced a larger increase in lower tropospheric ozone in TCR-2 than in GEOS-Chem, which they suggested could be due to differences in the sensitivities of ozone to changes in NO_x emissions in GEOS-Chem and TCR-2, as reported by Miyazaki et al. (2020a) in the multi-model data assimilation system. Sekiya et al. (2025) also noted that Qu et al. (2020a) reported lower posterior NO_x emission estimates in GEOS-Chem when assimilating the NASA standard OMNO2 NO₂ product compared to when assimilating the KNMI DOMINO product (Boersma et al., 2011) or the QA4ECV product (Boersma et al., 2018). The GEOS-Chem results presented here are based on the OMNO2 product.