As a meteorological driver, we used the regional climate model RegCM version 4.2 (hereafter referred to as RegCM4.2) developed by The International Centre for Theoretical Physics. Although the up-to-date version of RegCM is 4.5 (June 2015), the development of the modeling tools for this study started earlier when the newest version was 4.2. RegCM4.2 and its evolution from RegCM3 is fully described by . Its dynamical core is based on the hydrostatic version of the NCAR-PSU Mesoscale Model version 5 (MM5) . The radiation is solved within the Community Climate Model version 3 (CCM3) . The large-scale precipitation and cloud processes are calculated following and for convection parameterization we use the Grell scheme using the closure assumption in this study. RegCM4.2 includes two land-surface models: Biosphere–Atmosphere Transfer Scheme (BATS) originally developed by and the CLM3.5 model . In this study, the BATS scheme is activated. The single layer urban canopy model coupled to RegCM4.2 introduced by was not applied assuming that the urban-meteorological influence on the emissions impact will be minor and furthermore, to meet the computational demand of long climate simulations.

The chemistry simulations were carried out with the chemistry transport model CAMx (version 5.4). CAMx is a Eulerian photochemical dispersion model developed by ENVIRON Int. Corp. (http://www.camx.com). CAMx includes the options of two-way grid nesting, multiple gas phase chemistry mechanism options (CB-IV, CBV, CBVI, SAPRC99), evolving multi-sectional or static two mode particle size treatments, wet deposition of gases and particles, plume-in-grid (PiG) module for sub-grid treatment of selected point sources, Ozone and Particulate Source Apportionment Technology, mass conservative and consistent transport numerics and parallel processing. The ISORROPIA thermodynamic equilibrium model is implemented in CAMx to calculate the composition and phase state of an ammonia-sulfate-nitrate-chloride-sodium-water inorganic aerosol system in equilibrium with gas phase precursors. A detailed description of the model (the version used here) can be found at http://www.camx.com/files/camxusersguide_v5-40.pdf.

To achieve the goals of the study, a coupled system was designed consisting of RegCM4.2 and CAMx (denoted RegCMCAMx4) following the technique of online access coupling defined by . It represents an interactive two-way coupled modeling framework where chemistry is driven by the climate model and the calculated concentrations of the radiatively active gases and aerosols are fed back to the climate model's radiation code.

RegCMCAMx4 is a more advanced version of the original RegCMCAMx couple described by . The update interval for the meteorology from RegCM remained at 1 h which is sufficient . However, the original update interval for the species in the radiation code of 6 h was too coarse for describing the diurnal species evolution, therefore it has been reduced to 1 h as well. The original RegCMCAMx considered only the direct effect of sulfates and primary organic and black carbon. RegCMCAMx4 introduces the indirect effect of secondary inorganic aerosols (both sulfates and nitrates). For sulfates, it follows the work of where the cloud droplet concentration and effective droplet radius is modified according to the aerosol concentration. For nitrates both direct and indirect radiative effects are computed with the same method as for sulfates but with slightly modified optical properties following the works of and .

RegCMCAMx4 further replaces the method for calculating the coefficients of vertical turbulent diffusion (which is required by CAMx) with the newer scheme (as used in CMAQ model), which provides better agreement of model results with measurements, as shown by in a CAMx application over the same region and at similar horizontal resolution like in this study.

The added value of using an online coupled climate–chemistry modeling system is the possibility to calculate radiative feedbacks and impacts on temperature (and climate in general). We also assumed, based on previous validation studies involving RegCM and CAMx that the capability of these models reproducing the state of the atmosphere (both meteorology and chemistry) will not change significantly if coupling them online with respect to the case when they are coupled offline e.g..

The period of 2001–2010 was chosen to analyze the present-day impact of urban emissions on the air quality over central Europe. Calculation with RegCMCAMx4 were carried out on 10km×10km horizontal resolution domain centered over Prague, Czech republic of 160×120×24 (in x, y, and z direction) gridboxes for the climate model up to 50 hPa, while the chemistry model was integrated only on the lowermost 16 levels (approximately up to 300 hPa or 9000 m). The integration time step for the climate model was 30 s and 10 min for the chemistry model.

The ERA Interim reanalysis was chosen as driving meteorological conditions while for the chemical model, chemical boundary conditions (CBC) were taken from a similar 10-year run performed by RegCMCAMx4 over a larger, 30km×30km domain covering the whole of Europe.

As already mentioned, the TNO 2005 emissions were chosen to cover the studied period. Over the focused region, their resolution is about 7km×7km, which is sufficient for a 10km×10km computational grid. TNO are sector based annual emission data which were first regridded into the model grid. Than for each sector, specific temporal disaggregation factors and NMVOC speciation profiles were used to decompose the annual sums into hourly emissions following the inventory . The temporal profiles they provide were compiled to describe typical central European human activity profiles regarding transport, combustion, production etc. Biogenic emission of isoprene and monoterpenes were calculated following .

A number of experiments was carried out to examine the effect of city emissions on the regional air quality. These are summarized in Table 1. The total impact of all city emissions is evaluated as the difference between experiments 05BASE and 05ZERO (the “05” means that the 2005 emission were used). We were also interested in the impact the emissions from all other cities have on a selected city. To achieve this goal, we performed a run where all city emissions are removed except those from Prague. Apart from the total impact, it is also of interest to see how the individual species emitted contribute to the overall impact. We therefore evaluate also the partial impact of major gaseous pollutants, namely NOx, NMVOC. As the interest of policy makers is to estimate the consequences of possible emission reduction in cities, we propose to evaluate this partial impact in a framework of a sensitivity test where the emissions of the above-mentioned pollutants will be reduced by 20 %. For the sensitivity runs, no radiative feedbacks were calculated and the same meteorological conditions were thus used as a driver for these simulations. The assumption made here was that the main driver for the air quality changes are emissions, as the meteorological impact of the online coupled ozone and aerosols are expected to be small, having small feedback to the chemistry.

In order to justify the model's applicability for the presented goal, a detailed quantitative validation is provided for surface concentrations of O3, NO2, SO2 and PM2.5. We also included a minimal validation of the meteorological results: the near surface temperature is compared with the European EOBS climate data set that is currently extended until 2014. A detailed validation for the meteorological output is planned in a follow-up study, which intends to present the meteorological feedbacks of the presented chemical perturbation induced by urban emissions. Figure  presents the comparison of model seasonal near surface temperatures averaged over the 2001–2010 period with the E-OBS gridded data set. It indicates a negative model bias, which is highest in spring (around -1 to -2 K over large areas) and lowest in summer (0 to -1.5 K). This is attributable to overestimated cloudiness in RegCM, as already concluded by , who used the same model and set-up. Over mountainous areas like the Alps, southeastern Carpathians, biases reach values from -5 to 5 K, however, this is most probably related to coarse model representation of specific terrain features (valleys, ridges etc.) that highly determine the surface temperatures. A striking feature is the overestimation of temperature on eastern edge of the domain. As the boundary-close cells of the domain are strongly influenced with the boundary conditions, this may indicate that the reanalysis used (ERA Interim) is somewhat warmer than the gridded observational data.

For the chemical validation, the AirBase version 8 data (http://www.eea.europa.eu/data-and-maps/data/airbase-the-european-air-quality-database-8) provided by the European Environmental Agency, are used. We selected only rural background stations which are more consistent with the model-provided value that represents a 10km×10km average. Further the stations ale filtered to exclude high-elevation stations (above 2000 m). In the end, 328 stations for O3, 280 for NO2, 200 for SO2 and 53 for PM2.5 were selected for comparison with model results. For the gaseous pollutants, hourly, daily and monthly averages are considered while for aerosols only daily and monthly data are considered. The validation is done separately for winter (DJF), summer (JJA) months and for the whole year. The statistical measures evaluated were the correlation coefficient (r), root mean square error (RMSE), normalized mean square error (NMSE), the ratio of standard deviations (σr, calculated as σobservation divided by σmodel) and fractional bias (FB), as defined by and adopted by . They identified these metrics as the most important in assessing air quality model accuracy. Choosing these metrics further eases the comparison of the RegCMCAMx4 model performance with its former version presented in who applied the same metrics. The experiment 05BASE gave the base for the validation.

The above-mentioned statistical measures are collected in Table  for O3, NO2, SO2 and PM2.5. 3-3-3 columns are dedicated for hourly, daily and monthly data averaged over the whole year, DJF and JJA, respectively.

The average monthly and hourly cycles for DJF and JJA were selected, which provide a measure of the model's ability to capture the basic chemical climatology of key-species concentrations. Figure  plots the average monthly variation (left column) of the gaseous species O3, NO2 and SO2. The middle and right column provide the average diurnal cycle of these species for DJF and JJA, respectively.

Further, the monthly mean values of PM2.5 and its major components were compared to observations (Fig. ) distinguishing between DJF and JJA. For PSO4 and PNO3, measurements from the already mentioned AirBase database were used. For carbonaceous aerosol, the monthly data from the BC/OC measurement campaign data described by covering July 2002–June 2003 were used with the assumption that the basic climatology of these data is similar to the 2001–2010 average. These measurements considered BC/OC from PM10 aerosol (size < 10 µm), our model (CAMx) that uses a two bin approach (fine and coarse particles), calculates them as fine particles (size < 2.5 µm). We applied the factor of 0.8 to the measured values to estimate the PM2.5 fraction. This value is compiled from , and as an average of different seasons and character of the measurement site (cold vs. warm season and urban vs. rural).

We assessed the model bias further for urban stations as well, although it is well accepted that these stations are not suitable for standard chemistry transport model evaluation (at such resolution as in this study). We selected 10 urban background stations for Berlin, Budapest, Frankfurt, Katowice, Ljubljana, Milan, Munich, Prague, Vienna and Warsaw. The same statistical metrics were chosen as above but only for hourly (daily) averages for gases (particle matter). The results are collected in Table .

The impact of urban emissions from large cities on the regional air quality is evaluated in terms of selected air quality measures. For quantifying the exposure of the ecosystems, particularly crops, to elevated ozone levels, a widely used measure, the accumulated exposure over the threshold (AOT), introduced by , can be used. In this study, we evaluated the present AOT for the threshold of 40 ppbv for crops and forests (AOT40crop/forest) where the integration is done from May to July and from April to September, respectively. Further, the number of exceedances above a certain threshold is evaluated for daily maximum 8 h running ozone mean, the hourly NO2 and SO2 and the daily SO2 values. Finally, the mean JJA O3, mean DJF and annual SO2 and the mean annual PM2.5 surface concentrations are considered. These measures are established in the EC Directive on ambient air quality and cleaner air for Europe (2008/50/EC) and are implemented also in the Czech legislation. These are summarized in Table .

Further in the paper, we will present the spatial distributions of the (1) absolute change of the chosen metrics by the introduction of city emissions (calculated as experiment 05BASE minus 05ZERO) and the (2) relative change which is calculated differently for ozone and for other pollutants. In the case of ozone, which (as we will see further in the paper) both increases and decreases, the change is shown relative to the no-urban emission case (05ZERO). For all other species, the change is shown relative to the all-emission case (05BASE), i.e. we are interested in the relative contribution. We also calculated the all-city-average of the maximum impact which coincides with the location of the cities themselves, summarized in Table .

The validation showed that the modeling system captures the observed annual cycle of ozone with negative bias encountered in each month except late summer and autumn. The chemical boundary conditions used by our model were taken from a 10-year simulation from a larger domain, which however was forced with time invariant, spatially constant boundary conditions (40 ppbv) and this artificial constraint could propagate to the inner domain. showed that the final ozone levels greatly depend on the imposed boundary conditions. This constant constrain is also evident in the underestimation of the standard deviation for each averaging period, especially for DJF. The diurnal cycle underlines the monthly model bias, giving lower hourly values in DJF and better agreement in JJA, but with an underestimation of the JJA daily maximum values. The lower afternoon values of ozone in JJA can be attributed also by higher afternoon and evening NO2 values, as a result of the NO + O3 reaction. A very similar result is provided by both in terms of monthly and hourly variation.

Recently, were investigating the impact of CBC on the simulated ozone concentrations using the same two models as in our study (an offline couple of models RegCM and CAMx). They found a clear improvement in the correlation coefficient when using global chemistry model (ECHAM5/MOZART) based CBC. Their correlation of monthly ozone values using time/space invariant CBCs is 0.74 that compares very well to our value (0.77). Introducing the MOZART based CBCs, the correlations often increased by more than 0.1.

Compared to , our modeling system performs better in terms of correlation, with r=0.6 and r=0.67 for hourly and daily values against 0.51 and 0.53 in the later study which used an earlier version of the RegCMCAMx coupled system. This improvement probably lies in the duration of the data of comparison: in a 10-year time frame, the main drivers of the variability are the diurnal and monthly variations which contribute to an overall correlation in a significant way . , who used the same models in offline couple for also a 10-year experiment over Europe, achieved similar values of correlations. In general, our model performs for ozone better during JJA, when photochemistry is more intensive. This is true also for RMSE, NMSE and FB. and came to the same conclusion.

Very low correlations are achieved in case of NO2, especially for the hourly values. In general, it is difficult to achieve a higher degree of agreement in case of precursor species for at least two reasons. The driving meteorology, which greatly influences the hour-to-hour evolution of the NO2 concentrations, is a result of a 10-year climate model run. The climate model does not need to accurately reproduce the hour-to-hour, day-to-day weather pattern; however it has to reconstruct the climate close to reality in terms of averaged quantities and capability to capture extremes . Further, the emission decomposition into hourly values is based on numerous assumptions about the typical temporal evolution of a certain activity sector and the actual emissions may differ for a particular hour. At last, ozone precursor species are modeled always with a higher degree of uncertainty with great differences between models and set-ups (including chemistry mechanism) while they give very similar results in terms of final ozone concentrations .

Another striking feature is the underestimation of the modeled NO2 values. , who encountered a similar negative model bias, concluded that one reason lies in the overall underestimation of emissions and in the suppressed NO to NO2 conversion due to volatile organic compounds in CB-IV mechanism. Our configuration invoked the CB-V chemistry mechanism which was to remove this erroneous feature in the earlier version of the CB mechanism . However, the negative bias persists in our simulations, which in consequence could mean that the emissions are probably underestimated for the region modeled. Further, seen in the hourly plots, the underestimation mainly occurs during night-time similarly as in . Many other studies argued that chemistry in air quality models performs less biased during daylight . At last, biomass burning emissions were not accounted for in our simulations while it is an important contributor to NO2 burdens, especially for southern stations .

The strong DJF overestimation of the SO2 levels in was attributed to inadequate treatment of emissions in considering them as only area sources. An important improvement in the RegCMCAMx4 model against its earlier version was the treatment of part of SO2 emissions as elevated source which better compiles with the reality. However, our results suggest a minor improvement in DJF and the positive bias, although smaller, remained. On the other hand, summer encounters a clear negative bias. This could indicate that both the incorrect monthly disaggregation of the annual SO2 emissions and the overestimated conversion to sulfate aerosol (in JJA) play a role here as well. Reduced deposition can contribute to SO2 overestimation as well, as concluded by who applied the same deposition scheme as in this study.

To understand the model performance concerning the PM2.5 levels, we have to look at the comparison of the main fine particle matter components. In DJF, PM2.5 is largely underestimated: the main contributors to this bias is the underestimation of nitrate aerosol and both black and organic carbon. The DJF sulfate aerosol is in acceptable agreement with the observations, which, given that SO2 is overestimated, means that the SO2 to SO4 conversion is underestimated, leading to fair observation-model agreement. During JJA on the other hand, probably too strong SO2 to SO4 transition occurs, resulting in (1) SO2 concentrations even more negatively biased and (2) an overestimated sulfate aerosol. achieved better agreement for PSO4 than for SO2 arguing that, again, precursor species are often simulated with lower accuracy than secondarily formed pollutants. , who used CAMx with the same chemistry mechanism and aerosol module (ISORROPIA) arrived at the same conclusion. This is however not true for nitrate aerosol in our experiments, which shows a reasonable agreement (at least in terms of range of simulated values) with observations for JJA, but DJF model values are greatly underestimated. encountered the opposite situation, they had larger difficulties to capture JJA values than those during DJF. In our case, the DJF underestimation is probably connected with less NO2 simulated during this season. Similar underestimation occurs in the study of , especially for high concentrations. In general, the secondary inorganic aerosol model biases can be attributed to overall difficulties in simulating heterogeneous and aqueous phase processes .

Black carbon is usually underestimated in both DJF and JJA, in a similar extent than in . obtained comparable values as well and attributed this negative model bias to deficiencies in describing coating processes which are burdened by large uncertainties and directly determine the BC lifetime (BC has to become hydrophilic to get washed out by the wet deposition). Even more striking underestimation occurs for the organic carbon, although this bias is reduced compared to . This is probably due to different emission data used here for primary OC. However, the largest source for low modeled OC values probably lies in (1) in modeling the gas-to-particle partitioning that is affected with uncertainty with a large number of tunable parameters , (2) disregarding biomass burning aerosol that occurred in 2003 in eastern Europe affecting the measurements of .

As expected, for selected urban stations, our modeling system is, in general, less accurate, especially in terms of correlations. Urban stations are often influenced by local or nearby emissions sources, far below the models spatial resolution. The relatively coarse input emissions data cannot resolve the variability of these sources leading to much worse observation-model agreement compared to rural background stations. In case of ozone, the model over urban stations is positively biased in summer, which can be explained by the instant dilution of concentrated urban emission into the 10 km model grid which tends to overestimate ozone production . The opposite holds for the modeled NO2 concentrations which, due to instant dilution, are negatively biased in the model within urban environment. The model performance for SO2 and PM2.5 is worse as well over urban stations compared to rural ones, probably for the same reasons as for ozone and NO2.

Generally, the impact of city emissions on ozone is characterized by two main features: over cities, all examined metrics decreased. Enhancements are encountered further from urban centers and are of lower magnitude than the decreases. performed regional chemistry simulations over Istanbul and Athens and arrived at similar results: decrease of O3 over urban areas due to reaction with NO and, as a consequence of NMVOC transport, a smaller production of O3 at downwind areas due to increasing NMVOC / NOx ratio. Previously, showed that regional transport plays and important role in carrying urban pollution to larger distances leading to O3 formation downwind from cities. focused on both Istanbul and Athens and found up to 27 and 5 ppbv decreases of ozone due to urban emissions from these cities. Although they are not covered by our domain and are characterized by warmer climate, the changes are consistent with our JJA mean changes, especially in the case of Athens, which is affected by higher background pollution as is typical for cities over our domain as well.

Our results further suggest that while the enhancement of average ozone further from cities (as a result of downwind transport) is relatively small (up to 0.5 ppbv), however, a much larger increase is detected when considering metrics describing accumulated and extreme ozone values. In conclusion, the importance of cities' impact on ozone levels lies in higher potential for extreme ozone pollution over downwind areas during favorable meteorological conditions rather than in an increase of average levels. This is seen especially in changes in the number of exceedances, but the AOT40 levels show often large enhancements around cities as well.

The impact on NO2 levels is important only over cities themselves indicating that in urban plume further from urban areas, the NOx ages to HNO3 decreasing the contribution to the total NO2 levels. The contribution in cities goes up to 50–70 % indicating that a large part (i.e. 30–50 %) of the urban NO2 pollution is of non-urban origin. This supports region- or country-wide emission control strategies as their emissions undergo regional transport. found for Athens and Istanbul an even larger contribution around 95–96 % over these cities. Earlier, calculated the eastern Asia megacities contribution to NOz levels and found values around 10–30 % over cities, however NOz contains species produced during plume aging further from city centers causing this lower contribution.

In terms of average quantities (annual and DJF mean), the urban sulfur dioxide contribution is similar to NO2 contribution. Urban SO2 emissions are responsible for up to 70–80 % of pollution over urban areas themselves. In other words, 20–30 % of urban SO2 pollution comes from other areas (rural and minor cities, villages) giving importance to the regional emission transport. found over 50–75 % contribution near megacities of eastern Asia and 10–30 % over large areas in eastern China far from megacities. Similar values are obtained in our simulations (around 10–20 %) over large parts of the domain. They argue that large contribution within the cities indicate that the industry is concentrated within urban areas. This is often the case for eastern European cities where we obtained the highest urban SO2 contributions. The urban contribution to SO2 daily and (mainly) hourly exceedances is slightly higher than the contribution to average values and much more resembles the emission pattern indicating the enhanced importance of local urban emission in high air pollution episodes (when these exceedances occur) compared to the inter-urban pollution transport.

The annual average PM2.5 impact calculated in our simulations (up to 10–15 µgm-3) is in line with values obtained for Istanbul and Athens from : 18 and 12 µgm-3, respectively. However, their relative contribution is higher due to probably lower background pollution (around 62 and 55 %, respectively) compared to ours (30–60 % contribution over cities).

The emission reduction sensitivity test showed that in cities, the chemical regime is NOx-saturated meaning it is dominated by the reaction of NO with O3. This causes that NOx-oriented emission reduction to actually worsen the ozone levels above cities and its close environment. This holds for the AOT40s and exceedance change as well, where decreases are modeled only far from cities, where the urban plumes undergo photochemical aging. The response to 20 % NMVOC emission reduction is less dominant and in terms of all metrics it leads to reduction of ozone. tested the ozone response to 30 % reduced NOx or NMVOC emissions and found the ozone concentrations more sensitive to NOx emissions than to VOC emission. They also showed that above and around cities, reduced (increased) NOx emissions led to enhanced (suppressed) ozone by up to 8–10 % (in both direction), which is similar to the percentage change extracted from our simulations giving about 5–8 % ozone increase due to 20 % NOx emission reduction. The smaller numbers can be partly due to the lower emission reduction scenario. In case of NMVOC emission reduction, our numbers (up to -2 %) are only half of the relative ozone reduction achieved in their study (up to -4 %). The small ozone decreases due to NOx reduction in terms of AOT40s and exceedances further from cities in our simulations are probably caused by less NOx available in urban plumes when it mixes with biogenic NMVOC emissions over downwind areas .

Interestingly, the simultaneous NOx + NMVOC reduction scenario led to a very similar ozone response than the NOx reduction alone. This can be explained by the dominating effect of ozone titration due to NOx emissions. Indeed, the NMVOC reduction caused only minor ozone changes. Consequently, the effective ozone reduction strategy depended on the targeted area. Urban emission reduction of NOx and NMVOC improves the air-pollution over rural areas, however over the cities themselves, this usually leads to worsening of the ozone levels. According to our simulations, the only effective emission reduction strategy to decrease ozone levels is to reduce NMVOC emissions significantly while changing the NOx emissions only slightly or not at all.

Compared to the impact of all (100 %) emissions, the 20 % NOx + VOC emission change approximately equals to the 1/5th of the total impact (mainly for JJA ozone and AOT40s). This justifies the 100 % emission perturbation approach instead of a smaller perturbation introduced to maintain linearity. A similar approach was used recently for aviation emission impact .