Emissions from land transport, industry, and shipping contribute largely to global budgets of trace gases like NOx and O3, thereby impacting air quality and climate (e.g. ; and ). To quantify the impacts of these emissions, typically source–receptor relationships are calculated using perturbation or source apportionment methods (e.g.; and ). Many studies quantifying the influence of anthropogenic and natural emission sources (e.g. land transport emissions or lightning) on the ozone budget exist, but the uncertainties of such analyses are large. There exist three main sources of uncertainty: (1) the emission inventories, (2) the model biases/errors, and (3) the resolutions of the models and/or emission inventories. The influences of the first two factors, emission inventories and model biases, have been investigated by multi-scenario and/or multi-model analyses (e.g.; and ). Even though the influence of the model and emission inventory resolutions on simulated ozone mixing ratios is well known (e.g.; and ), the impact of the third factor – the model and emission inventory resolutions – on the simulated contributions of specific emission sources to ozone has not yet been systematically investigated in detail. It is important to investigate this third factor, as source apportionment studies focusing on climate usually use rather coarsely resolved global climate models (e.g.; and ), while air-quality-related studies use more finely resolved regional models (e.g.; and ). Therefore it is unclear if the results from global and regional models are comparable and how large potential errors, caused by the coarse resolution of global models, are. The present study is a first attempt to investigate the influences of the model and the emission inventory resolutions on the ozone contributions. More precisely, we investigate the influences of four different modelling aspects on source apportionment results for ozone; these aspects are as follows:

the applied model,

We apply the MECO(n) MESSy-fied ECHAM and COSMO models nested n times; e.g. model system together with a detailed source apportionment method tagging;. This model system couples, during runtime, the global chemistry-climate model EMAC ECHAM5/MESSy for Atmospheric Chemistry; with the regional chemistry-climate model COSMO-CLM/MESSy , which consists of the COSMO-CLM model equipped with the MESSy Modular Earth Submodel System; infrastructure. Due to the MESSy infrastructure, we apply identical submodels for calculating the chemical processes and the same source apportionment method in the global and regional model instances. In addition, the global model instance provides consistent boundary conditions for the source apportionment to the regional model instances, allowing a detailed intercomparison of the source apportionment results on different scales. Therefore, this model system is, to our knowledge, the first available model system allowing a seamless contribution analysis from the global to the regional scale. With this model chain we can directly compare the results at the regional and global scale, which allows us to estimate uncertainties of the contribution analyses caused by the model, the model resolution, and emission inventory resolution.

This paper is organised as follows. First, Sect.  gives an overview of the model system and discusses the investigation strategy and the performed simulations. In Sect.  we present a brief evaluation of the model results compared against ground-level and ozone sonde observations as well as a comparison of the ozone production rates simulated by EMAC and COSMO-CLM/MESSy (Sect. ). In Sect.  the differences of the ozone contributions caused by differences of model and emission inventory resolutions are analysed in detail. We provide a more detailed quantification of the differences in specific regions and a further discussion in Sect. .

We apply the MECO(n) model system, which couples the global chemistry-climate model EMAC during runtime (i.e. online) with the regional chemistry-climate model COSMO-CLM/MESSy . Both models, EMAC and COSMO-CLM/MESSy, calculate the physical and chemical processes in the atmosphere and their interactions with oceans, land, and human influences. They use the second version of MESSy to link multi-institutional computer codes . The core atmospheric model for EMAC is the fifth-generation European Centre for Medium-Range Weather Forecasts, Hamburg (ECHAM5), general circulation model . The core atmospheric model used in COSMO-CLM/MESSy is the COSMO-CLM model , a regional atmospheric climate model that is based on the COSMO (Consortium for Small-scale Modelling) model and jointly further developed by the CLM-Community. In the model system's acronym, MECO(n), “n” denotes the number of COSMO-CLM/MESSy instances nested into the global model framework. The initial and boundary conditions, which are required for each of these nested regional model instances, are provided by the next coarser model instance. This model instance can either be EMAC or COSMO-CLM/MESSy. Due to the online coupling, the boundary conditions for the regional model instances can be provided at every time step of the driving model instance. This is especially important for resolving short-term variations in chemically active species. As EMAC and COSMO-CLM/MESSy calculate both atmospheric dynamics and composition, the meteorological and chemical boundary conditions are as consistent as possible. In addition, the same chemical solver and kinetic mechanism are applied, leading to highly consistent chemical boundary conditions. Therefore, there is no need for lumping (i.e. treating different chemical species with similar chemical formulas as one species), scaling of boundary conditions for specific chemical species, or taking boundary conditions from different models.

More details about the MECO(n) model system are presented in a set of publications including a chemical and meteorological evaluation (; and ). The set-up of the simulation applied in the present study is very similar to that described by . Therefore, we only present the most important details of the model set-up. The complete namelist set-up is part of the Supplement.

A MECO(2) set-up with one COSMO-CLM/MESSy instance over Europe with a resolution of 0.44∘×0.44∘ (≈ 50 km) and one instance covering Germany, with a resolution of 0.1∘×0.1∘ (≈ 12 km), was applied (see Fig.  for the computational domains). For simplicity, we name these two model instances hereafter CM50 and CM12. EMAC, CM50, and CM12 are running simultaneously in the same way as in externally coupled earth system models, with the different earth compartment models running in parallel (see Fig. 2 in for the details of the data exchange between the nested model instances). Both COSMO-CLM/MESSy instances use 40 vertical model levels (terrain following) with geometric height as the vertical coordinate. The height of the uppermost model level is at ≈22 km; the damping zone starts at 11 km. The thickness of the lowest model layer is ≈ 20 m. The boundary conditions for CM50 are provided by EMAC, which is operated in the T42L31ECMWF resolution, i.e. with a spherical truncation of T42 (corresponding to a quadratic Gaussian grid of approx. 2.8∘×2.8∘ in latitude and longitude) and 31 hybrid pressure levels in the vertical direction up to 10 hPa (corresponding to ≈ 30 km over Europe). The thickness of the lowest model layer corresponds to ≈ 60 m over Europe. The boundary conditions for CM12 are provided by CM50. The applied MESSy version is a modified version of MESSy 2.50, including ECHAM 5.3.02 and COSMO 5.00. All changes are included in MESSy 2.51. To facilitate a one-to-one comparison with observations, EMAC is “nudged” by a Newtonian relaxation of the temperature, the divergence, the vorticity, and the logarithm of surface pressure towards ERA-Interim reanalysis data for the years 2007 to 2010. Sea surface temperature and sea ice coverage are prescribed as boundary conditions for the simulation set-up from ERA-Interim as well.

Due to the MESSy infrastructure, the same diagnostics or chemical process descriptions are applied in all of the model instances. Following the modular structure of MESSy, each diagnostic or process description is coded as a so-called submodel. The applied submodels are listed in Table . Besides the name of the submodel and their reference, a short description provides general information on the process or diagnostic represented by the respective submodel. Most importantly, the same kinetic solver MECCA; and same TAGGING submodel are applied in each instance.

The chemical mechanism used by MECCA considers the chemistry of ozone, methane, and odd nitrogen. While alkynes and aromatics are not considered, alkenes and alkanes are considered up to C4. We use the Mainz Isoprene Mechanism MIM1; to model the chemistry of isoprene and some non-methane hydrocarbons (NMHCs). The mechanisms of MECCA and of the submodel calculating the scavenging of trace gases by clouds and precipitation SCAV; are part of the Supplement. The TAGGING submodel calculates the contributions of different emission sources to ozone and the relevant precursors. More details of this tagging approach are given in Sect. .

The lightning-produced NOx emissions are only calculated in EMAC, using a parameterisation based on that is scaled to a global nitrogen oxide emission rate of ≈ 5 Tg N a-1 from lightning flashes. In CM50 and CM12 we use the emissions from EMAC (i.e. with same geographical, vertical, and temporal distribution), which are transformed online onto the grids for CM50 and CM12, respectively. This approach was chosen as the calculation of lightning-produced NOx is strongly coupled to the convection parameterisation e.g.. In different models and/or at different model resolutions convection occurs at different places and/or times and lightning emissions can differ largely. Our approach was chosen to allow for an easier comparison between the results of different model instances.

The calculation of emissions from soil NOx and biogenic isoprene (C5H8) is performed by the MESSy submodel ONEMIS described as ONLEM by. Following the parameterisations of and , the respective emissions depend on the meteorological conditions. In contrast to the lightning NOx emissions, the soil NOx and biogenic emissions are calculated separately by EMAC and CM50. This leads to differences in the soil NOx and C5H8 emissions (see Fig. S17 in the Supplement), influencing the calculation of the contributions. We have chosen this approach because the land–sea masks differ between models and model resolutions. If the emissions calculated by EMAC are used in the COSMO-CLM/MESSy model instances, some of the emissions would occur over the sea (or vice versa). This could lead to artificial errors in the contribution analyses. In EMAC, the isoprene emissions calculated by ONEMIS are scaled by a factor of 0.6 following and in COSMO-CLM/MESSy by a factor of 0.45 following.

To evaluate the performance of the different model instances and the different simulations, we compare the model results with ground-level observations of ozone and measurements from ozone sondes. For the evaluation we use observations by the European Monitoring and Evaluation Programme (EMEP)EMEP; http://www.emep.int; last access: 20 January 2020; and ozone sonde data from the world ozone database (WOUDC; http://woudc.org; last access: 20 January 2020). The methodology is described in detail by . In comparison to , however, we focus here on average values for JJA 2008 instead of values for June and December 2008. A list of the observation data used is part of the Supplement (Sect. S4).

For a quantitative evaluation we chose the metrics RMSE (root mean square error) and MB (normalised mean bias error). The definition of both quantities is given in Appendix A. Table  lists the RMSE and MB of the EMAC and CM50 instances for each simulation. As the EMAC set-up is identical in all simulations, the model results do not change. Generally, the results from the models are in agreement with the measurements. The RMSE has a range of around 19 to 26 µgm-3 and the MB has a range of 13 % to 21 %. These deviations from the measurements are in the range of comparable model systems e.g.. As already noted by , CM50 exhibits a larger positive ozone bias than EMAC. This bias is mainly caused by a more efficient vertical mixing in COSMO-CLM, as well as by a less stable boundary layer during the night. The latter is a common problem of many models and leads to diurnal cycles with ozone values that are too large during the night, resulting in an overall ozone bias e.g.. The results of CM12 are not presented here, as the domain only covers Germany, and therefore fewer stations can used for evaluation. The RMSEs and MBs for CM50 and CM12 are given in the Supplement (Table S2) and take into account the measurements at all stations located in the region covered by the CM12 domain.

In general, CM50 simulates larger ozone mixing ratios than EMAC over the continent (see Fig. ). This ozone bias in the case of CM50 compared to EMAC is caused neither by the finer resolution of the emissions nor by the different biogenic emissions compared to EMAC, because the positive ozone bias for CM50 compared to EMAC is also apparent in the results of ET42 and EBIO. Only over the Mediterranean Sea, lower ozone values are simulated by CM50 compared to EMAC. These lower ozone mixing ratios can be partly attributed to the coarser resolution of the emissions in EMAC compared to CM50, as the difference is lower in the ET42 simulation (Fig. b). The simulated ozone mixing ratios in CM50 are up to 7.5 nmolmol-1 larger (in JJA 2008) in ET42 compared to REF. Averaged over the area of the Mediterranean Sea the increase in ozone is around 3 nmolmol-1. The application of the soil NOx and biogenic emissions calculated by EMAC in CM50 (EBIO) leads to an increase in the ozone mixing ratios of 1 to 3 nmolmol-1. The differences are largest over south-eastern Europe, the Mediterranean Sea, and the Iberian Peninsula (Fig. c). Overall, the differences in the results of CM50 between REF, EBIO, and ET42 are small compared to the bias between EMAC and CM50. In particular, the positive ozone bias over Serbia and Bulgaria cannot be attributed to different biogenic emissions or the coarser resolution of the emission inventories in EMAC compared to CM50.

Figure  shows scatter plots comparing observed and simulated ozone monthly mean concentrations at all considered stations of the EMEP network. The simulated concentrations in all model instances and simulations lie, with one outlier, within a factor of 2 of the measurements. As already discussed, the simulated ozone concentrations at most stations show a positive ozone bias. The simulated ozone concentrations are lower than the measured ozone concentrations only at a few stations . The ozone bias is very similar in all CM50 simulations; EBIO and ET42 show almost the same bias as REF. Only the simulation EVEU shows a slightly lower positive ozone bias. Accordingly, the change in the anthropogenic emission inventory has a larger impact on the model results than the influence of the emission inventory resolution and the geographical distribution of the biogenic emissions.

To evaluate the simulated ozone mixing ratios in the free troposphere, the model results are compared to ozone sonde data (see Sect. S4 in the Supplement for a list of considered stations). In total, 510 individual ozone sonde launches are considered for the year 2008. To compare the ozone sonde data with the model results, the vertical ozone profiles simulated by the model were sampled online at every time step of the model at the location where the ozone sonde was launched. Drifts of the ozone sonde by winds are not taken into account. For every launched ozone sonde, we averaged the simulated vertical profiles in time over the measurement period (usually some hours). These vertical profiles of simulated ozone mixing ratios are compared to the measurements of the ozone sonde data. As the main focus of this comparison is the free troposphere, we restrict this analysis to all data in the pressure range of 600 to 200 hPa.

The probability density functions (PDFs) for the measured and simulated vertical ozone distributions are displayed in Fig. . The results show that in the free troposphere both model instances (EMAC, CM50) simulate very similar vertical ozone distributions. Accordingly, the positive ozone bias of CM50 compared to EMAC is confined to the boundary layer. Further, in general a positive ozone bias is apparent, which is will known for EMAC e.g..

Figure  shows the absolute and relative contributions of different emission sources to the European ozone column up to 850 hPa as simulated by EMAC and CM50 for the REF simulation (see Table S1 in the Supplement for a detailed definition of the tagging categories). The largest absolute and relative ozone contributors are the anthropogenic non-traffic and the biogenic categories, both with contributions of more than 1 DU, corresponding to more than 15 %. Both model instances simulate similar absolute ozone contributions from the categories anthropogenic non-traffic (≈ 1.0 DU), land transport (≈ 0.7 DU), shipping (≈ 0.5 DU), and biomass burning (≈ 0.4 DU). For the biogenic category, CM50 calculates slightly larger absolute contributions compared to EMAC (see Sect. ), but the differences are small compared to the temporal variability in the contributions. Further, CM50 calculates larger absolute contributions from the lightning and stratosphere categories. This mainly affects the categories land transport, anthropogenic non-traffic, shipping, and biomass burning, where EMAC simulates 0.1 to around 1 percentage point larger relative contributions compared to CM50. At the same time, the increased vertical mixing in CM50 leads to an increase in the relative contributions from the categories stratosphere, lightning, and aviation compared to EMAC. Here, the differences are in the range of 0.1 to around 1.5 percentage points.

The positive ozone bias of CM50 compared to EMAC indicates vertical mixing that is too efficient in CM50 (see Sect. ). Therefore, the larger contributions from the categories stratosphere and lightning in CM50 compared to EMAC are likely an artefact of this too efficient vertical mixing. However, this could partly be a feature of the increased resolution, as individual stratosphere–troposphere exchange (STE) events are better represented in CM50 compared to EMAC due to the increased resolution . Generally the correct representation of STE events poses a big challenge in most models (e.g.; and ), and our results suggest a large difference in the contribution of STE to ground-level ozone between the results of different models.

The values which we have discussed so far, however, are averages on the continental scale. On the regional scale the differences can be much larger. Geographical distributions of the differences in the absolute and relative contributions as simulated by EMAC and CM50 are given in the Supplement (Figs. S3 and S4). Exemplarily, we want to focus on the categories land transport, an important anthropogenic emission source, and biogenic emissions. As discussed in Sect. , the biogenic emissions are calculated online by both model instances and depend on the meteorology and surface properties. While the total emissions are comparable, the geographical distribution and the area-averaged contribution differ (see Supplement Fig. S17 and Tables S2 to S10). As disparity between online-simulated emissions is a typical inter-model difference, a detailed investigation of the influence of these differences is of interest.

So far, the results indicate that with respect to average values on a continental scale the differences caused by the resolutions of the model/emission inventory are rather small. This confirms findings by , which reported only a small influence of the global redistribution of megacity emissions (which can be seen as a locally decreased emission resolution) on the global ozone budget.

To summarise and quantify these differences in more detail, Fig.  shows the (a) absolute and (b) relative contributions of O3tra to ground-level ozone averaged over the CM50 domain, as well as for the geographical regions defined in the PRUDENCE project . The results of EMAC are not analysed for these geographical regions because, due to the coarse resolution, some regions would only consist of a few grid points.

Figure  also shows that on the scale of smaller regions the absolute and the relative contribution of O3tra to ground-level ozone is only slightly influenced by the coarse resolution of anthropogenic emission inventories (ET42) as well as having a different geographical location and resolution of biogenic emissions (EBIO). This holds not only for the mean O3tra contributions, but also for the extreme values expressed by the 95th percentile. Further, the simulated differences in the biogenic and shipping categories, which are more affected by the differences in the emission inventories in the two simulations, are also rather small (see Figs. S13 and S14 in the Supplement). The largest simulated differences in the mean contribution of shipping emissions to ground-level ozone between the REF, EBIO, and ET42 simulations are around 0.5 nmolmol-1 and below 0.5 percentage points. The largest change (95th percentile) in the biogenic category in the Iberian Peninsula region is around 0.7 nmolmol-1 and 0.5 percentage point.

Compared to the differences between the contributions of O3tra in the REF, ET42, and EBIO, the differences caused by a changed emission inventory (EVEU) are larger. In the Mediterranean region, the mean and 95th percentile of the contributions of O3tra increase by 1 nmolmol-1 and 2 percentage points, respectively. In the Alps region, the increase in the mean and 95th percentile of the contributions is up to 1.3 nmolmol-1 and 3 percentage points, respectively. Similarly, for the contribution of shipping emissions the differences are largest with the changed emission inventory (up to 1.5 nmolmol-1 and 1 percentage point). Accordingly, changes in the resolution of the emission inventory or the biogenic emissions can affect the contribution from anthropogenic categories (such as land transport and shipping). However, on the regional scale the main drivers of uncertainties are clearly the anthropogenic emissions and differences caused by the model resolution and/or model differences. For example we found regional differences (see Sect. 4.1) in the contribution of O3tra to ground-level O3 between EMAC and CM50 of up to 20 % around the Naples region, which in this case can mainly be attributed to the coarse land–sea mask used in EMAC, leading to land transport emissions to occur over the sea.

The results of the model evaluation, however, are not very helpful in judging which of the two emission inventories are more realistic. Although EVEU shows a smaller ozone bias compared to REF, caused by reduced precursor emissions, it is unclear if lower anthropogenic non-traffic emissions in the VEU compared to MAC emission inventories are realistic.

In the present study, we are focusing on the question: are contributions of emissions to ozone a matter of scale? To answer this question we compare the influences of the model, the model resolution, the emission resolution, and the emission inventory on the results of ozone contribution analyses. For this we apply the MECO(n) model system which combines a global and a regional model by means of an online nesting technique. By applying the identical tagging diagnostics (source apportionment method) in the regional and global model with consistent boundary conditions, we are able to compare the results of model instances with different resolutions to investigate the influence of the model and emission inventory resolutions onto the diagnosed ozone contributions. Such analyses are important for quantifying uncertainties of ozone source apportionment studies, which arise due to limitations of the model and/or computational resources.

For the specific model set-up involving the global model EMAC and the regional model COSMO-CLM/MESSy our results show that simulated differences in ozone contributions on a continental scale (e.g. Europe) are rather small. The largest differences in the contribution of anthropogenic emission sources were up to 10 % for the contribution of land transport emissions to ground-level ozone. However, the contribution of stratospheric ozone to ground-level ozone calculated by EMAC and COSMO differs by up to 30 %. One main reason for this large difference in the contributions of stratospheric ozone between the two models is the existence of enhanced vertical mixing and larger convective up- and downdrafts in COSMO-CLM/MESSy compared to EMAC. Taking the comparison with the measurements into account, the vertical mixing in COSMO-CLM/MESSy and the enhanced stratospheric contribution are likely too large. On the regional scale, the differences between the contributions of anthropogenic emission sources simulated by COSMO-CLM/MESSy and EMAC are much larger. Here, we observed differences of up to 20 % for the contributions of land transport emissions to ground-level ozone. This difference is mainly caused by the coarse land–sea mask used in the global model instance, leading to emissions of land transport emissions over sea, different ozone dry deposition, and missing biogenic emissions. Taking the results of the same model instance (CM50) into account, the largest influence on the results is caused by different emissions inventories. Locally, however, coarsely resolved emission inventories and differences between the biogenic emissions can also lead to differences of up to 20 %. In addition, we showed how the differences in the source apportionment results between different model instances can help to explain model biases and the physical/chemical mechanisms causing these biases.

Apart from the many model specific findings of this study, its results have important implications for other modelling studies and modellers who are applying source apportionment methods. These implications are as follows:

First, our study shows that average continental contributions of anthropogenic emissions are quite robust with respect to the model and the model resolution used. This means that global models at coarse resolution can be used to perform ozone source apportionment in the global context.

Clearly, this study is only a first step in quantifying the driving sources of uncertainties and, particularly, the role of the model and emission inventory resolutions on the results of ozone contribution studies. Especially, as some processes like vertical diffusion or vertical transport can heavily alter the model results, follow-up studies need to take into account more (and different) models to better quantify the uncertainties due to differences in the meteorology simulated by different models. In addition, the two analysed anthropogenic emission inventories clearly do not reflect the whole spectrum of different emission estimates. Further, our analyses only focused on differences near the origin of the emissions. An increased resolution leads to a more realistic chemistry within the plumes downwind of the emission hotspots. This can affect the long-range transport from different precursors and might influence regions far away from the emission region. Especially, calculations of radiative forcings are very sensitive to ozone near the tropopause. In a coarsely resolved model, the overestimated absolute contributions might lead to a biased radiative forcing. This effect, however, is difficult to quantify and would require very finely resolved global chemistry-climate models or two-way nesting capabilities, which feed back information about the contributions from the fine grid back to the coarse grid. For a next step, a further increase in the model and emission resolution should be envisaged. Even if we found only small differences between the 50 and 12 km resolutions this step would be important, as even with a 12 km grid resolution emissions are diluted over large areas. A finer resolution could strongly reduce the dilution . Such an analysis, however, is hindered by two aspects. First, consistent emission inventories (anthropogenic and natural) with a resolution of 1 km over areas that are large enough to compare models on a regional and global scale must be available. Second, requirements with respect to the computational time of chemistry-climate models with ≈1 km resolution over large computational domains are very demanding, hindering the detailed quantification of the differences caused by the resolution over long integration periods.