The GEM model used in this study solves the hydrostatic primitive equations cast in spherical coordinates (with a rotated equator plane in order to minimize distortions over the uniform resolution area), with a terrain-following, hybrid pressure-type vertical coordinate comprising 28 levels, of which 8 may fall into a well-developed boundary layer. The top of the model domain is located at 10 hPa, whereas the height of the lowest atmospheric level is approximately 40 m. The model uses a semi-Lagrangian time discretization with a semi-implicit approximation of terms that give rise to fast gravitational modes – a feature crucial for its multiscale applications as well as for integration using variable-resolution meshes. This scheme permits using time steps several times longer than in, for instance, a split-explicit method. With a few exceptions, choices of physical parameterizations made for this study follow the 15 km version of the Canadian Regional Forecast System as documented by . The turbulence parameterization is based on a turbulent kinetic energy budget with inclusion of statistical subgrid-scale cloudiness and the Bougeault–Lacarrere specification of the length scale ; surface energy budget is modelled with the force–restore equation . Gravity wave drag effects are taken into account using a modified McFarlane parameterization . Condensation processes are handled with the Kain–Fritsch deep convection scheme , the so-called Kuo-transient shallow convection parameterization (see , ) and the Sundqvist scheme for non-convective clouds. Solar and infrared radiation are modelled using and schemes, fully interactive with clouds.

The gas-phase chemistry package of GEM-AQ describes 116 chemical and 19 photolysis reactions among 50 species or groups of species and is based on the second version of the Acid Deposition and Oxidation Model with extensions for free-tropospheric chemistry. For a complete list of species and reactions, see . A simplified aqueous-phase reaction module allows for oxidation of SO2 to sulphate. Aerosols are modelled with a sectional module CAM (Canadian Aerosol Module; , ) with five aerosols types – sulphate, black carbon, organic carbon, sea-salt and soil dust – size-segregated into 12 logarithmically spaced bins. The aerosol module includes parameterizations of nucleation, condensation, coagulation, sedimentation and dry deposition, in-cloud oxidation of SO2 and scavenging and below-cloud scavenging of aerosol species by rain and snow.

For this study the GEM-AQ model was run on the global variable grid with rotated equator and with the resolution of 0.2∘×0.2∘ over the European continent. The number of grid points was set to 288×264 on the globe and 197×190 in the core part (Fig. 1). In the vertical, 28 sigma-hybrid layers extending to 10 hPa were used. The simulation was performed from 1 January to 31 December 2006, as a set of 30 h forecasts with a 6 h overlap. The integration time step of 600 s was used.

Emission data were prepared for the experiment by TNO (TNO, Utrecht, the Netherlands) using MACC (Monitoring Atmospheric Composition and Climate) methodology (Pouliot et al., 2012). Anthropogenic emissions included primary gaseous pollutants such as O2, NOx, CO, NMVOC, NH3, CH4 and particle pollution of the fine and coarse mode for individual SNAP (standardized nomenclature for air pollutants) sectors. Hourly biogenic emissions provided by the AQMEII project were used. Emissions outside the area provided by AQMEII were compiled using EDGAR 2.0 (Emission Database for Global Atmospheric Research, for 1990 base inventory year) and GEIA (Global Emissions Inventory Activity) global inventories (Olivier et al., 1999; Olivier and Berdowski, 2001). Anthropogenic emissions were distributed within the four lowest model layers (up to ∼630m) with different injection height profiles for each of the SNAP sectors. Temporal profiles modulating annual and diurnal variation of emission fluxes for each SNAP were used. Surface anthropogenic and biogenic emission fluxes were applied as a bottom boundary condition in the vertical diffusion equation.

Analysis of ozone concentrations variability was based on daily maximum 8 h running average. Evaluation of the GEM-AQ model performance was done for 1386 stations available in the ENSEMBLE database .

The analysis of model performance was undertaken for 492 stations available in the ENSEMBLE database for PM10 concentration. The 24 h averages were calculated based on hourly measurements and model results.

For rural background stations (472 sites) the GEM-AQ model underestimated ozone concentrations during the cold season (January–mid-March and November–December). During the summer, underestimation of peak values ranges from 30 to 40 µgm-3. Although the correlation coefficient is high (0.92), MBE is -7 µgm-3 and MAGE is 18 µgm-3 (Table 1). For suburban (391 sites) and urban stations (527 sites), the characteristics of variability range and agreement with measurement is comparable. In January, March, September and October the model slightly overestimated ozone concentrations. In June and July during pollution episodes, modelled concentrations are lower than observed by 10 to 20 µgm-3. Results for other months show very good agreement in terms of average concentration levels and short-term variability, which is confirmed with a very high correlation coefficient – 0.93 for both types of stations. MBE is positive for suburban stations (1.9 µgm-3) and for urban stations (2.5 µgm-3), with MAGE 17.2 and 17.4 µgm-3 respectively. In contrast to , who reported higher correlation coefficient for ozone concentrations for rural sites than for “locally influenced sites”, the GEM-AQ model performance in terms of the Pearson correlation coefficient was almost at the same level for all types of stations (0.92 for rural sites, 0.93 for suburban and urban sites).

The evaluation results reported by for CAMx and CHIMERE models showed that for both models, ozone concentrations at rural stations are reproduced not as well as at suburban and urban stations. This has been confirmed by the GEM-AQ model results. While MAGE is relatively constant (18.2 for rural, 17.2 and 17.4 for suburban and urban respectively), there are significant differences in MBE values for different types of stations.

Analysis of seasonal variability of model performance was undertaken by using results from the CAMx model for January and July. Modelled ozone concentrations were systematically underpredicted. CAMx performance was the worst at urban stations in January while in July the bias was lower. Also, reported a strong underestimation of ozone concentrations during the first part of the year for CAMx and CHIMERE models. CMAQ results, described in , indicate that model performance for the daytime ozone varied seasonally. In the winter, CMAQ overestimated ozone concentration by 8 % and in the spring and summer ozone levels were underestimated on average by ∼4 and 2 % respectively. In autumn model performance was worst, with significant overestimation by 30 %. Results from the GEM-AQ model also show seasonality in MBE and the Pearson correlation coefficient. However, MAGE is relatively constant for all seasons – on average 17.5±0.8µgm-3. For winter and autumn the model slightly overestimated, with MBE 0.93 and 0.09 µgm-3 respectively. Although MAGE and MBE are smallest in autumn, the correlation coefficient is lower than in winter (0.78 in autumn, 0.87 in winter). In the spring and summer the model tends to underestimate. However, with nearly the same MAGE, model results for summer are noticeably better with MBE of 1.4 µgm-3 and correlation coefficient 0.92, while in spring the bias is larger at -3.2 µgm-3 and the correlation coefficient is lower at 0.75.

Analysis of model performance for different regions showed differences between models. The best performance for CAMx and CHIMERE models was for southern Europe, while north-western and eastern Europe were characterized with a negative bias in the range of 10–30 %, especially in winter. In terms of the correlation coefficient, the CHIMERE model performed well in southern, north-western and eastern Europe, while CAMx showed better results in eastern Europe. The CMAQ model overestimated daytime ozone concentrations in the south-western part of the domain and underestimated the north-eastern part, including the British Isles, in winter. The largest overestimation occurred over northern Italy (Po Valley), while the largest underestimation was over the Czech Republic and Poland.

The GEM-AQ model underestimated daytime ozone concentrations; the highest values were found in northern Europe (MBE -16.8 µgm-3) especially during the cold part of the year. One possible reason relates to uncertainties of NOx emissions in Scandinavia, as well as over British Isles and the north-western part of Europe, which may contribute to overestimated titration processes in the model. Another source of possible underestimation is too-weak transport of ozone from the upper troposphere in high latitudes. However, such an analysis is beyond the scope of the presented study, which is focused on surface air quality and does not include the analysis of the vertical structure of the atmosphere.

For the rest of the domain the agreement between modelled and observed ozone concentrations is good. However, summer episodes are underestimated. During winter and spring months the MBE distribution shows significant overestimation of ozone concentrations over the Alps, which indicates too intensive downward mixing in the mountain regions, where in the cold part of the year concentrations are highest. In spring and autumn there is a systematic overestimation for suburban and urban stations in Benelux, Germany and France that results in positive MBE 0.4 for Region II, while for other regions MBE is negative. In general, best model performance was achieved in the summer.

In winter and autumn, topography plays a very important role in the distribution of higher ozone concentration levels, although modelled concentrations seem to be overestimated. Higher concentrations are over the southern part of Europe (with the exception of the Po Valley) and the Mediterranean Sea. In spring, concentrations are significantly higher in southern, central and eastern Europe, while in western Europe and the British Isles the increase is not that significant due to the inflow of relatively clean Atlantic air masses. In summer there is further increase of ozone concentrations over land, with the maximum over Region IV. Elevated ozone concentrations are also over the Mediterranean Sea and the Black Sea. In contrast to ozone distribution patterns in spring, ozone concentrations are lower over north-eastern and eastern Europe. The reason for this effect might confirm the hypothesis of the advective nature of ozone episodes in eastern Europe and the role of high-pressure-system blocks during summer months . However, due to the lack of measurements, model results cannot be evaluated in this region. In autumn, concentrations decrease to lowest values in the north-eastern part of Europe.

Surface PM10 concentrations at rural background stations (119 sites) were systematically overestimated in all seasons. In January and February the model captured quite well periods with high PM10 concentrations, but peak values were overestimated. Best model performance was from mid-February to May. From June to September the systematic bias was relatively constant at ∼10-15µgm-3. In autumn, MBE is largest and the model overestimated significantly for all periods with observed higher concentrations. The correlation coefficient is high (0.72): MBE is 10.8 µgm-3 and MAGE is 18.2 µgm-3 (Table 2).

At suburban stations (110 sites) model performance is very good with MBE 4.7 µgm-3 and MAGE 18.4 µgm-3. From January to April the agreement of modelled and observed time series, averaged over all suburban stations in the domain, is good except for a short winter episode. At the beginning of May the model underestimated PM10 concentrations by about 10 to 20 µgm-3. From June to September, modelled and observed concentrations at suburban stations averaged over the domain vary in the range of 20–40 µgm-3. There is a small systematic overestimation of modelled PM10 – on average 5 µgm-3. As in the case of rural stations, starting in October the discrepancies between model and measurements increase. The model generated three PM10 episodes from October to December which were not observed. In spite of discrepancies, the correlation coefficient is relatively high (0.75), which confirms good agreement in terms of average concentration level and short-term variability.

For urban stations (263 sites) the agreement between modelled and observed PM10 concentration levels is good. MAGE is comparable to the value obtained for suburban stations (18.3 µgm-3), MBE is small (1.3 µgm-3) and the correlation coefficient is 0.72. The model underestimates observed PM10 concentration during two severe pollution episodes at urban stations in January and at the beginning of February. From the second week of February concentrations decrease and the agreement between modelled and observed daily average PM10 levels was good. In May the model did not capture an increase of PM10 concentrations. From June to September the model tends to overestimate PM10 levels by 5 to 10 µgm-3. In October and December the model showed significant overestimation, while in November the agreement was good.

Daily mean PM10 concentrations averaged over suburban and urban stations show good agreement of the modelled and observed concentrations averaged over the domain for values below 40 µgm-3. At suburban stations, modelled PM10 concentrations in the range of 60 to 100 µgm-3 were overestimated, while for urban stations the highest concentrations, above 60 µgm-3, were underestimated.

In contrast to the GEM-AQ model results, reported that most models used in their analyses had difficulties in reproducing elevated PM concentrations during winter. Most models underestimated PM10 levels over Europe, although during summer the performance was better. In the Mediterranean region, the highest concentrations were in summer months. Results from the CAMx model were substantially underestimated and performance was poor for both January and July, with similar magnitude of error statistics. Also, the DEHM model (Brandt et al., 2012) underestimated PM10 concentrations.

The GEM-AQ model tends to overestimate PM10 daily mean concentrations. The model performance is clearly better for urban stations. With nearly the same MAGE, 18.3±8µgm-3, the highest overestimation was for rural background stations, 10.8 µgm-3. For suburban and urban stations the performance is better – 4.7 and 1.3 µgm-3 – and the correlation coefficients are 0.72 and 0.75 respectively.

The work by shows that the best performance of CAMx and CHIMERE models in terms of PM10 concentrations was in north-western Europe for rural-background stations. In southern and eastern Europe, model results were underestimated. reported that in winter the domain-averaged MBE is -21.5 µgm-3 for the CMAQ model. For other seasons the underestimation is lower – in the range of -11/-16µgm-3. The smallest bias was in northern France. In spring and summer the bias spatial pattern was similar to the winter case and the bias tends to improve in autumn.

In southern Europe, modelled concentrations were systematically underestimated for all seasons with the bias -9.41 µgm-3. However, for other regions GEM-AQ tends to overestimate PM10 daily mean concentrations. In winter in northern and western Europe the model overestimates, while in central and eastern Europe there was a strong underestimation. These lead to error compensation in terms of averaged MBE (0.3 µgm-3) for winter months and a relatively high MAGE of 22.5 µgm-3. In the spring model performance is the best in terms of MBE (1.9 µgm-3) and MAGE (14.8 µgm-3), but the correlation coefficient is lower than in other seasons (0.51). In autumn, the model overestimated over all regions but the correlation coefficient is relatively high (0.79). The worst performance in terms of correlation coefficient was over northern Europe (0.32) and southern Europe (0.56).

The distribution of MBEs for PM10 clearly shows that anthropogenic emission data were overestimated over Germany and the Czech Republic. MBE in this region (especially over the eastern part of Germany) was positive for all types of stations and for all seasons. This impacted the average model bias in Region II. In Region III modelled PM10 concentrations were underestimated during severe winter pollution episodes in January and February 2006. During this period, low temperatures were observed and differences between modelled and observed PM10 levels might be caused by too-low emission estimates that did not account for household heating. In addition, as the lowest model layer height is at ∼27m, the structure of the stable boundary layer over urban regions might not be fully reproduced. Underestimation of PM10 in southern Europe is most probably due to underestimated mineral dust emissions and transport from North Africa. However, this does not relate to AQMEII emission estimates but to an online dust emission module in the GEM-AQ model. Further work will be undertaken to revise geophysical fields describing soil properties that are used for dust uptake. The highest error was for rural stations, which clearly indicates that emissions in remote regions were overestimated. For urban stations in January, February and March the model systematically underestimated PM10 concentrations. The temporal variability of MBE for different types of stations is similar.

Seasonal variability of PM10 concentrations is dominated by the distribution and intensity of anthropogenic and natural sources. In winter, PM10 concentrations are highest over central and eastern Europe, although the model tends to overestimate over Germany. Also, elevated concentrations are present over the Northern Atlantic due to sea salt generation during winter storms. In spring, the winter maximum over the Atlantic dissipates and concentration over North Africa and southern Europe is increasing. The highest concentrations remain over central Europe. A similar pattern was calculated for summer months, with lower maximum but higher background values. In autumn, PM10 concentrations over the Atlantic Ocean and over the North Sea increase and over southern Europe decrease. In eastern Europe, PM10 background concentrations are also higher than in spring and summer months.