High resolution air quality simulation over Europe with the chemistry transport model CHIMERE

. A modiﬁed version of CHIMERE 2009, including new methodologies in emissions modelling and an urban correction, is used to perform a simulation at high resolution (0.125 ◦ × 0.0625 ◦ ) over Europe for the year 2009. The model reproduces the temporal variability of NO 2 , O 3 , PM 10 , PM 2 . 5 better at rural (RB) than urban (UB) background stations, with yearly correlation values for the different pollutants ranging between 0.62 and 0.77 at RB sites and between 0.52 and 0.73 at UB sites. Also, the fractional biases (FBs) show that the model performs slightly better at RB sites than at UB sites for NO 2 (RB = − 33.9 %, UB = − 53.6 %), O 3 (RB = 20.1 %, UB = 25.2 %) and PM 10 (RB = − 5.50 %, UB = − 20.1 %). The difﬁculties for the model in reproducing NO 2 concentrations can be attributed to the general underestimation of NO x emissions as well as to the adopted horizontal resolution, which represents only partially the spatial gradient of the emissions over medium-size and small cities. The overestimation of O 3 by the model is related to the NO 2 underestimation and the overestimated O 3 concentrations of the lateral boundary conditions. At UB sites, CHIMERE reproduces PM 2 . 5 better than PM 10 This pri-marily the result of


Introduction
Regional chemistry transport models (CTMs) are powerful tools widely used to assess and investigate air quality issues. The application field of CTMs is broad: understanding the atmospheric physico-chemical processes Hodzic et al., 2010;Jiménez-Guerrero et al., 2011), assessment of emission control scenarios Kiesewetter et al., 2014), past and future regional air pollution trends (Colette et al., 2011), chemical weather forecast cooperation (Balk et al., 2011;Kukkonen et al., 2011;Copernicus (http://www.copernicus.eu/)) and natural hazard emergency response (Colette et al., 2011;Matthias et al., 2012). A review of the major air quality modelling forecast systems operating in Europe is given by Menut and Bessagnet (2010).
CTMs were used initially to simulate gas phase tropospheric pollutant concentrations within the lower troposphere; a coarse horizontal resolution was sufficient to achieve this objective. During the past decade, air quality legislation has focused more and more on particulate matter (PM) concentrations and CTMs have been equipped with Published by Copernicus Publications on behalf of the European Geosciences Union.

22
E. Terrenoire et al.: High-resolution air quality simulation over Europe aerosol modules. High PM concentrations are usually observed in urban areas (EEA, 2012), leading to a need for CTM applications at urban scale. A large number of studies highlight the need to perform simulations at high resolution in order to assess the urban air pollutant concentration patterns, especially for PM compounds (Queen and Zhang, 2008;Stroud et al., 2011;Wolke et al., 2012;Fountoukis et al., 2013).
A list of long-term European model evaluation studies that took place during the past decade can be found in Pay et al. (2012a). However, the simulations performed in these studies adopted a coarse (0.25 to 0.5 • ) horizontal resolution (Tarrason et al., 2006;Matthias, 2008;Stern et al., 2008;Kim et al., 2011;Sollazo et al., 2012;Pay et al., 2012a;Pirovano et al., 2012;Zhang et al., 2013) and the evaluation was generally done using a limited set of measurement stations. Examples of model validation can also be found for higher resolution applications (0.1 • × 0.1 • ) but over limited domains, covering an urban area or a country (Hodzic et al., 2005;Tombette and Sportisse, 2007;Flaounas et al., 2009;Chemel et al., 2010;Pay et al., 2012b). Therefore, there is a clear need for evaluating high-resolution simulations (< 0.1 • × 0.1 • ) performed over large domains, such as continental areas.
To fill this gap, a high-resolution air quality simulation was performed over most of Europe and evaluated for the year 2009 using the CHIMERE model (Menut et al., 2013). For this study, the CHIMERE model was improved in order to simulate air quality at the urban scale. The evaluation process was conducted over a comprehensive spatial and temporal data set in order to effectively quantify the model accuracy. The aims of the study are threefold: (i) to get an accurate picture of air quality at the urban scale using air quality modelling; (ii) to test an original set of emissions and an urban meteorological correction; (iii) to evaluate comprehensively the simulation using the largest set of monitoring stations available over Europe.
The paper is organized as follows. Section 2 is devoted to the description of CHIMERE and the methodology used to prepare the anthropogenic emissions as well as the set of observations selected for the evaluation. Section 3 describes and analyses comprehensively the ability of the model to reproduce the concentrations of the different selected pollutants. Finally, Sect. 4 summarizes the main findings raised by the study and gives suggestions for future work related to high-resolution regional modelling.

Model description
CHIMERE is a regional CTM that has undergone several extensive evaluations Vautard et al., 2007a;Van Loon et al., 2007). It calculates the concentrations of the main chemical species (SO 2 , NO 2 , O 3 , PM 10 , PM 2.5 ) that are involved in the physico-chemistry of the lower troposphere. CHIMERE has been described in detail in several papers: Schmidt et al. (2001) for the dynamics and the gas phase module;  for the aerosol module and Vautard et al. (2005Vautard et al. ( , 2007b for further model developments including the introduction of a mineral dust emission module and a convection scheme into CHIMERE. The version used in this study is CHIMERE 2009 with specific modifications described in the following paragraphs of this section. The aerosol model species are the primary particle material (PPM), secondary inorganic aerosol (SIA; sulfate, nitrate and ammonium) based on the ISORROPIA thermodynamic equilibrium model (Nenes et al., 1998), secondary organic aerosol (SOA, whose formation is represented according to a single-step oxidation of the relevant anthropogenic and biogenic precursors), sea salt and dust (non-African mineral dust is not included). The particles' size distribution ranges from 39 nm to 10 µm and the particles are distributed into eight bins (0.039, 0.078, 0.156, 0.312, 0.625, 1.25, 2.5, 5 and 10 µm). Vertically, the domain is divided in eight hybrid-sigma layers from the ground to 500 hPa. Gas phase tropospheric chemistry is represented using the reduced MELCHIOR chemical mechanism (120 reactions and 44 gaseous species) and the dry and wet depositions are taken into account. For the study, a nested domain (328 × 416 grid boxes) that covers most of Europe from 10.43750 • W to 30.43750 • E in longitude and 35.90620 to 61.83375 • N in latitude with a horizontal resolution of 0.125 • × 0.0625 • (8 km × 8 km) is designed (Fig. 1). Boundary conditions for the mother domain were derived from monthly mean climatology based on the Laboratoire de Météorologie Dynamique Zoom -Interaction avec la Chimie et les Aérosols (LMDz4-INCA3) model for gaseous species (Hauglustaine et al., 2004) and the Goddard Chemistry Aerosol Radiation and Transport (GOCART) model for aerosols (Ginoux et al., 2001). A complete and high-resolution set of both biogenic and anthropogenic emissions are needed in order to perform the CHIMERE computations. Emissions of six biogenic CHIMERE species (isoprene, α-pinene, β-pinene, limonene, ocimene and NO) were calculated using the Model of Emission of Gases and Aerosols from Nature (MEGAN v.2.04, Guenther et al., 2006). Wildfire emissions are also taken into account and issued from the Global Fire Emissions Database version 3 (Kaiser et al., 2012). Modelled concentrations for comparison with observations are extracted from the lowest vertical level, which extends from the ground to about 20 m. A detailed description of CHIMERE can be found in Menut et al. (2013).

Meteorology
Meteorological data needed by CHIMERE are derived from the fields of the Integrated Forecast System (IFS) runs at the European Centre for Medium-range Weather Forecasts (ECMWF). The choice of feeding CHIMERE directly with the IFS data stems from the results of a sensitivity analysis where the performance of the Weather Research and Forecasting (WRF) limited area model and the IFS was evaluated against observations (not shown here). The analysis shows evidence of a systematic overestimation of the wind speed by WRF, a feature confirmed by several other studies (Mass and Ovens, 2011;Jimenez and Dudhia, 2012;Vautard et al., 2012). Another reason for the direct use of ECMWF-IFS fields is that it avoids an ad hoc meteorological numerical weather calculation (e.g. WRF), thus reducing the computational cost. For more details about the comparison between ECMWF-IFS and WRF performance, the reader is referred to Miglietta et al. (2012).
IFS adopts a 0.25 • horizontal grid spacing from surface up to 0.1 hPa (91 levels in total). It provides standard meteorological variables (temperature, wind components, specific humidity, pressure, sensible and latent heat fluxes) that need to be vertically and horizontally interpolated on the CHIMERE grid (eight levels). Some additional variables are also diagnosed by CHIMERE from the available fields, such as friction velocity and vertical wind speed, which are used to complete the description of vertical transport and turbulent diffusion.
However, the main limitation of such data is that the IFS regional scale data cannot sufficiently represent the urban scale meteorology observed in the urban boundary layer. This is crucial, as the urban canopy affects the wind circulation and the urban energy balance (Sarrat et al., 2006), which directly impact the transport and the vertical diffusion of primary pollutants over cities (e.g. O 3 , NO 2 and PM). Usually, operational meteorological observations are performed outside urban areas (e.g. airport) for representativeness reasons. Several studies have shown large differences between urban and rural winds (Fisher et al., 2006), showing a wind speed ratio (rural/urban) of up to 2. Another study, focused on Lisbon (Portugal), showed that modelled wind speed ratio inside the canopy and at the top of the urban sub-layer was within the range 0.1 to 0.6 (Solazzo et al., 2010).
The vertical diffusion (diffusion coefficient, Kz) is used in CHIMERE to compute the vertical turbulent mixing in the boundary layer following the parameterization of Troen and Mahrt (1986). In order to integrate the influence of the urban canopy on meteorology, the wind speed, as well as the Kz, is modified within the CHIMERE version used for this study. The wind speed and the Kz are divided by a factor of 2 over the urban areas (urban cells) in the lowest CHIMERE layer as described in Bessagnet et al. (2012). The grid cells' land use classification was derived from the GlobCover (Bicheron et al., 2008) land cover database covering the period December 2004-June 2006 In order to estimate the potential impact of the urban correction, a sensitivity test was performed for January 2009. Figure 2 shows the results of the simulation without the urban correction for NO 2 , O 3 , PM 10 and PM 2.5 . Concentrations are normalized by the values obtained in the actual base case that include the urban correction. A rather large impact over all major European cities is observed for the four pollutants. For NO 2 , a decrease in concentrations ranging from a few % (suburban areas) up to 45 % (e.g. Paris, London, Lisbon and Glasgow) is observed when the urban correction is not applied. Conversely an increase of O 3 concentration is observed over both the large cities (10 to 120 % in the city centre) and medium-size cities (10 to 30 %), essentially due to a decrease of NO 2 titration. Likewise NO 2 , the concentrations of both PM 10 and PM 2.5 decrease from a few % to a maximum of 43 % in cities such as Paris, London and Katowice in the south of Poland. Finally, the yearly box-whisker plots time series (Figs. S1 to S4 in the Supplement) show that the median of the modelled concentrations increases for NO 2 (0.6 ppb), PM 10 (0.8 µg m −3 ) and PM 2.5 (0.4 µg m −3 ) at urban background (UB) stations when using the urban correction while it decreases for O 3 (0.8 ppb). At UB stations, on average, the impact of the urban correction is positive (reducing the bias) but small.

Anthropogenic emissions
Five main steps can be identified in the anthropogenic emission pre-processing: (1) the horizontal and vertical spatial regridding of the raw emissions to comply with the CHIMERE grid, (2) the temporal disaggregation, (3) the chemical speciation, (4) the hourly disaggregation and (5) the surface flux calculation within CHIMERE.

Spatial regridding of anthropogenic emissions
The raw emission data of the main air pollutants (NMVOC, NO x , CO, SO 2 , NH 3 , PPM) are provided per activity sector, according to level 1 of the Selected Nomenclature for Air Pollution (SNAP). They come from three different databases: The large point sources from the fine-scale TNO-MACC emission data for 2007 were added to surface emissions in order to deal with only gridded emissions.
The spatialization of the anthropogenic emissions depends on the SNAP sector. The TNO-MACC emissions were used as a proxy variable to regrid EMEP 0.5 • × 0.5 • emission data (Kuenen et al., 2011) for the SNAP activity sectors 3, 7, 8, 9 and 10. For SNAP sectors 1, 4, 5 and 6, emissions were distributed over artificial land use corresponding to the European Pollutant Emission Register (EPER) industries (eea.europa.eu/data-and-maps/ data/eper-the-european-pollutant-emission-register-4). For PM 2.5 , the GAINS national emissions totals were considered more reliable (Z. Klimont, personal communication, 2012) and  Netherlands, Serbia and Montenegro and Turkey. Moreover, additional factors were applied on Polish regions (x4 or x8) to compensate for the lack of PM 2.5 emissions due to the domestic heating activity in the available inventory. Finally, emissions from SNAP 2 were disaggregated according to population density. The goal of this approach is to better capture the spatial variability of the SNAP 2 sources. The population data were provided by the Joint Research Centre (JRC) over a regular grid at 0.083 • × 0.083 • horizontal resolution. For the elaboration of the SNAP 2 emissions, a distinction is made between gaseous and PPM species to better reallocate the anthropogenic biomass burning emissions (SNAP 2) over the rural areas. According to the French National Spatialised Emission Inventory (INS), (Ministère de l'Ecologie et du Développement Durable, 2009), available at municipality level and derived using the bottom-up approach, there is clear evidence that PPM 2.5 (PPM with an aerodynamic diameter ≤ 2.5 µm) emissions per inhabitant sharply decrease when the population density increases (Fig. 3). This is due to the increase of the relative contribution of wood burning in the fuel mixture moving from urban centres to rural areas (e.g. due to an increase in domestic fireplaces). This effect is noticeable only for PPM 2.5 , because biomass burning sources contribute to a very large fraction of PM emissions in SNAP2, while they have less influence than other fuels on gas phase pollutants (INS, Ministère de l'Ecologie et du Développement Durable, 2004).
The vertical repartition of the emissions into the different levels of the CTM is known to be of great importance (Bieser et al., 2011). It was calculated for each SNAP sector following the calculation of Bieser et al. (2011) (Table 1). A new layer (0-20 m) was added to the current implementation compared to the original EMEP configuration. For SNAP 2, 6, 7, 8 and 10 all the emissions are released into the first level of the model.
As an illustration, Fig. 4 shows the spatial distribution of the total annual PPM 2.5 emissions from SNAP 2 derived from the hybrid emission inventory used for this study. Compared to the original method, SNAP 2 emissions around the medium-size and large cities increase when the population proxy is used. This is because when the Land Use (LU) proxy  is used, emissions from each type of LU cell have the same weight, thus giving rise to a flatter distribution than using population density. Considering for example an EMEP cell (0.5 • × 0.5 • ) including a big city as well as a small town, emissions are modulated in the same way over urban cells of both areas if the LU approach is followed, whereas most of the emissions are allocated just in the big city if population is used.

Emission temporal modulation
Time disaggregation was done on the basis of the Generation and Evaluation of Emission (GENEMIS) project data using monthly, daily and hourly coefficients depending on the activity sector (Society et al., 1994).
However, for SNAP 2, a new temporal profile derived according to the degree day concept is used. The degree day is an indicator used as a proxy variable to express the daily energy demand for heating (Verbai et al., 2014;Quayle and Diaz, 1980). The degree day for a day "j " is defined as: whereT is the daily mean 2 m temperature. Therefore, a first guess daily modulation factor could be defined as: where N = 365 days.
Considering that SNAP 2 emissions are not only related to the air temperature (e.g. emissions due to production of hot tap water), a second term is introduced in the formula by means of a constant offset C. To better assess the influence of this offset, C can be expressed as a fraction ofD (degree day annual average). Considering: where A is defined by user (A = 0.1 for this application), we can express and The daily modulation factor (F j ) is therefore defined as: Note that F j is mass conservative over the year and replaces the original monthly and daily modulation factors. The choice of A is left to the user to express the relative weight of hot water production with respect to heating. For this application, we set A = 0.1 in order to replicate as much as possible the original CHIMERE temporal profiles during the warmer season.
As an illustration, Fig. 5 shows the 2009 daily modulation factors applied to the SNAP 2 emissions at three locations that are both geographically and climatically different: Katowice (Poland), Paris (France) and Madrid (Spain). The highest factors for the three locations occur during the winter period and the lowest ones during the summer, as expected. This means, for example, that during the cold periods the emission from SNAP 2 can be up to three times more intense (e.g. beginning of January for Madrid or end of February for Paris) than during the spring or the autumn periods. It is interesting to note that in Katowice during the beginning of the year the factors are relatively lower than at the two other locations, meaning that over this period the difference between the daily mean and the annual mean temperatures is lower in Katowice than at the two other locations. This does not mean that temperature in Katowice during January was higher than Madrid and Paris, but simply that the latter ones experience a higher variability than Katowice between January and the summer season. Indeed, it is worth noting that in Katowice the modulation factor during the summer is often greater than 0.1, indicating that in several cases, the daily temperature in Katowice is lower than 20 • C. Conversely, at the end of the year, all locations experienced a cold outbreak of the same intensity relative to their local annual mean temperature.

Chemical speciation
Annual NO x emissions were speciated into NO, NO 2 and HNO 2 using the coefficients recommended by the International Institute for Applied Systems Analysis (IIASA) (Z. Klimont, personal communication, 2012, Table 2). For NMVOC, a speciation was performed over 32 NMVOC National Acid Precipitations Assessment Program (NAPAP) classes (Middleton et al., 1990). Real NMVOC species were aggregated and assigned to the model species following Middleton et al. (1990).

Observation data
Observed data come from two different databases. selected as we wanted to avoid stations with complex terrain. Figure 1 displays the spatial distribution of the AirBase (green squares for rural background (RB) and blue dots for UB) and EMEP network stations (red triangles) used for the evaluation. The spatial distribution of the stations is homogeneous over the populated region of western Europe, while several gaps are seen in Eastern Europe and in the Balkan countries. A total of 1009 UB AirBase, 560 RB AirBase and 85 EMEP stations are used for the evaluation. However, fewer stations are available for PM 2.5 both at UB sites (267) and especially at RB sites (92). The EMEP network database includes fewer sites than AirBase, but it is the only European network providing PM 10 speciation data, which are crucial to investigate the model performance deeper. At EMEP stations, high-volume Whatman quartz filters or tapered element oscillating microbalance (TEOM) samplers are use to perform PM 10 measurements. The measured quantities are analysed mainly with the gravimetric method; however, the micro balance technique is used in some countries. One should be aware that the reactivity of NO 3 and the volatile character of NH 4 NO 3 results in an underestimation of the measured concentrations, especially during the summer. Hence, according to Putaud et al. (2004), an uncertainty of ±15 % should be considered for major SIA species. Details about the station type classifications and the different measurement techniques are available through the AirBase and EMEP websites.

Data analysis methodology
In this paper, an "operational evaluation" is performed (Dennis et al., 2010). The evaluation techniques include statistical and graphical analyses applied in order to determine the degree of agreement between the model and the observations in an overall sense. We selected different statistical indicators for their ability to diagnose the model performance from different perspectives including temporal correlation, bias and the absolute error between observation and modelled values. Therefore, along with the observed (OM) and modelled (MM) mean concentrations, the observed (σ _obs) and modelled (σ _mod) standard deviation, the correlation coefficient (R), the root mean square error (RMSE), the fractional bias (FB) and the fractional error (FE) are calculated. Details about the calculation of the statistics performed using the Atmospheric Model Evaluation Tool software (AMET) can be found in Appel et al. (2011). The performance evaluation was based on yearly and seasonal statistics using the daily mean values of all stations available for the given typology (UB and RB).

Spatio-temporal variability of the modelled concentrations
In this section, the 2-D annual mean concentration maps of NO 2 , O 3 , PM 10 and PM 2.5 ( Fig. 6) are analysed. On each map, the observed concentration values at each station are represented by a coloured dot for both RB and UB stations.
The winter (December-January-February) and the summer (June-July-August) seasonal means are also drawn to analyse the inter-seasonal variability of the modelled concentrations of NO 2 and O 3 (Fig. 7), PM 10 and PM 2.5 (Fig. 8), SOA, dust and sea salt (Fig. 9), SO 2 and sulfate (Fig. 10), and finally nitrate and ammonium species (Fig. 11). NO 2 concentrations are directly linked to emissions mainly coming from SNAP 2 (non-industrial combustion plants), 7 (road traffic) and 8 (other mobile source). Therefore, Fig. 6 shows that the highest annual mean concentrations are located over urban areas and along ship tracks (Atlantic Ocean, English Channel, and Mediterranean Sea). Specific areas with high concentrations are identified: the Po valley, Paris, the Benelux countries, London, southern Poland (e.g. Katowice), Athens, Madrid and Barcelona. For these specific areas, as well as over Europe, the concentrations are much higher during winter than during summer due to higher emissions and light vertical dispersion (e.g. shallow boundary layer height, stagnant conditions, thermal inversion) (Fig. 7). Ship tracks in the Mediterranean Sea, along the coast of Portugal and especially in the English Channel are characterized by rather high NO 2 yearly mean modelled concentrations, ranging between 2 and 12 ppb. The observation values represented by the dots show a systematic underestimation of the NO 2 concentrations over land, with a negative bias around 4 ppb. Moreover, two main areas located over the south of Poland (Katowice) and some parts of Romania (industrial hot spots) show an even larger underestimation (10-20 ppb). Conversely, some areas where the observations are overestimated are Paris, London, Madrid, Barcelona and Athens. For these areas, the methodology used to downscale the national annual emissions could be the reason for the overestimation.  Overall, CHIMERE slightly overestimates O 3 concentrations over Europe. Figure 6 shows that the highest annual mean concentrations are located below the 45 • latitude (from 30 ppb over the coasts to 48 ppb over the sea) where the strongest photolysis over Europe occurs. The O 3 maximum concentrations are modelled during summer while the winter seasonal mean is below 30 ppb over most of Europe, reaching values close to zero in the Benelux countries, the Po valley, Germany and Poland. The highest summer concentrations are calculated over and around the Mediterranean Sea, where lower boundary layer heights (as compared to continental planetary boundary layer heights), strong photolysis rates and low dry deposition constitute favourable conditions for the build-up of high O 3 concentrations (up to 56 ppb). During the warm periods (spring and summer), most of the capital cities across Europe are characterized by ozone decrements due to the ozone titration by NO x as well as along shipping routes in northern Europe and the Strait of Gibraltar (Fig. 7).
For PM 10 , the highest modelled concentrations (20-30 µg m −3 ) are located south of a line that goes from the south of Portugal to the north of Poland (Fig. 6). The annual mean concentrations show a slight underestimation over the continent. However, over cities such as Milan, Paris and Krakow, CHIMERE overestimates the observations, likely due to the methodology used to downscale the SNAP2 emissions from biomass burning, which are still too concentrated over the centre of the urban areas instead of being more distributed over the surrounding areas (Timmermans et al., 2014). In winter, the modelled concentrations are highest over the continent (10-30 µg m −3 ), with lower concentrations over central Spain (10 µg m −3 ). Hot spots are located over the large European urban areas and over industrial primary emissions areas (e.g. Romania and Bulgaria). In summer, the concentrations are strongly influenced by the amount of dust and boundary conditions in the south part of the domain (up to 36 µg m −3 ) and the urban PM 10 pattern is weak except for Katowice, Milan and Paris (28-50 µg m −3 ) (Fig. 8). The concentrations over the continent are a factor of 2 lower than the hot spot areas. In winter, much lower concentrations are modelled over the Mediterranean Sea In order to have a better insight into modelled PM composition over Europe, the modelled concentrations for dust, sea salt, SOA (Fig. 9) and SIA (Figs. 10 and 11) are examined. The highest dust concentrations occur in the summer over north Africa (up to 30 µg m −3 ) and southern Spain (up to 25 µg m −3 ). In these regions, dust is by far (90 %) the largest PM 10 component in the summer with a large contribution of the boundary conditions (largest of all seasons). The accuracy of the boundary conditions in this area is therefore of great importance in order to correctly model the PM 10 concentrations. For sea salt, the highest values occur during the winter and are located over the North Sea (up to 13 µg m −3 ), where it represents the major part of the PM 10 mass. Over land, a zonal gradient is observed, with a maximum concentration modelled over western Europe (5 µg m −3 over Ireland) and the minima over eastern Europe (< 1 µg m −3 ), where the influence of oceanic winds carrying sea salt particles is lower than over the west of Europe. For SOA, the contrast between summer and winter is striking. In winter, the concentrations are low (< 1 µg m −3 ), with a maximum in the Po valley (4 µg m −3 ), while during summer the concentrations over land range between 0.5 µg m −3 in the western part of Europe and 8 µg m −3 over the Balkans.
For sulfate (Fig. 10), during winter a sharp zonal gradient is observed, with modelled minima in western Europe (2 µg m −3 on average) and maxima in eastern Europe (up to 8 µg m −3 ). The highest concentrations are located near the main SO 2 industrial and urban emission areas in Romania, Bulgaria, Bosnia, Serbia, Hungary and south Poland that use sulfur-rich coal for energy production and domestic heating (e.g. in the Katowice region). During the summer, sulfate resulting from the gas oxidation of SO 2 occurs mainly over the Mediterranean Sea, where emissions from ships are high and intense photolysis allows for the production of oxidant radicals (e.g. OH radical). The stagnant meteorological con- ditions in summer also favour accumulation and recirculation of pollutants in the Mediterranean basin. For ammonium and nitrate (Fig. 11) a strong seasonal variability is modelled, with lower concentrations in summer than in winter. Nitrate concentrations correspond less to precursor emission pattern than sulfate, due to the influence of a more complex chemical transformation pathway (e.g. the thermodynamical equilibrium with ammonium). The highest modelled concentrations are seen during winter over the Po valley (up to 15 µg m −3 ), over the Benelux countries (6 µg m −3 ) and in southern Germany (8 µg m −3 ). Ammonium wintertime concentrations show a wide and rather homogeneous pattern, covering both western and eastern Europe, together with a hot spot covering the entire Po valley (up to 5.2 µg m −3 ), due to abundant emissions and frequent weak circulation conditions. Ammonium concentrations in western Europe are mainly driven by nitrate availability, whereas in Eastern Europe the ammonium spatial pattern is closely related to the sulfate pattern.

Model evaluation
Tables 4 and 5 display the different yearly and seasonal statistics for SO 2 , NO 2 , O 3 , PM 10 , and PM 2.5 at RB and UB Air-Base stations respectively. Table 6 displays the same metrics computed at EMEP monitoring network sites and also in-  UB (Fig. 6s) stations respectively, as well as sulfate, total nitrate and total ammonia calculated using the EMEP stations data (Fig. 7s), are available in the supplement.

Sulfur dioxide
The

Nitrogen dioxide
Throughout the year, CHIMERE accurately reproduces the temporal variability of NO 2 at the RB sites (R = 0.68), but with a large negative bias (FB = −33.9 %), particularly during winter (FB = −44.5 %). At UB stations, although the temporal variability is rather well reproduced (R = 0.61), the model underestimation is even larger than at RB sites, both over the entire year (FB = −53.6 %) and especially during the winter season (FB = −63.9 %). This poor performance could be explained by the general underestimation of urban NO x emissions, especially in the Eastern European cities. Moreover, the yearly mean UB diurnal cycle (Fig. 8s) shows a rather persistent negative bias throughout the day (4 ppb in mean). Nevertheless, the bias is largest during the morning (8 a.m.) and evening traffic (8 p.m.) peaks, indicating a very likely underestimation of the NO x traffic emissions over urban areas.
At both RB and UB stations, CHIMERE usually performs better in reproducing the temporal variability of the observed concentrations (e.g. standard deviation and correlation coefficient) than the mean values. Further investigation of the model behaviour over urban areas performed using the DELTA tool  indicates that the performance of CHIMERE is significantly better over large European cities' (e.g. capitals') UB stations than over the UB stations of medium-size and small cities. The bias between observed and modelled concentrations is reduced from 4.58 ppb when using all UB available stations to 1.31 ppb when focusing on the largest 30 cities of Europe. Conversely, the  adopted horizontal resolution which increases the dilution of the emitted pollutants is not able to accurately simulate the spatial gradient of the emissions over medium-size and small cities, giving rise to the underestimation of the observed concentrations.

Ozone
Similarly to NO 2 , the daily temporal variability of O 3 concentrations is well reproduced at both RB (R = 0.77) and UB sites (R = 0.73). The comparison of modelled and observed concentrations quantiles shows that the highest values are well reproduced while lower quantiles are overestimated (Figs. 5s and 6s). Throughout the year, the modelled values show a rather homogeneous bias which is higher at UB (FB = 25.2 %) than at RB sites (20.1 %) and is likely linked to the NO x underestimation that is larger at UB than at RB, thus limiting the O 3 titration by NO 2 . The yearly mean diurnal cycle (Fig. 9s) shows a larger positive bias (9 ppb) during the morning (7 a.m.) than during the afternoon ozone peak (5 ppb in mean). This tendency is likely also related to the lack of O 3 titration by NO 2 due to the previously described larger underestimation of NO 2 in the morning (e.g. 8 a.m.). The FB of O 3 has a seasonal variation, with a low positive FB in the summer (14.9 %) and the highest overestimations during the autumn at RB (FB = 30.6 %) and UB sites (FB = 36.9 %).Data from the remote Valentia observatory station (51.94 • N; 10.24 • W), located in Ireland at the western lateral boundary of the domain, show that the background concentrations are slightly overestimated throughout the year (2.4 ppb for the median), with a maximum during the autumn (12 ppb for the median). The lateral boundary condi-Geosci. Model Dev., 8, 21-42, 2015 www.geosci-model-dev.net/8/21/2015/  (Chen et al., 2003;Szopa et al., 2009;Van Loon et al., 2007). The overestimation of O 3 during the cold season is therefore attributed to the overestimation of the background concentrations at the boundaries of the domain.

PM 10 and PM 2.5
The model reproduces the temporal variability of PM 10 observations throughout the year (R = 0.62 and 0.56, respectively) at both RB AirBase and EMEP stations. Interestingly, by contrast to the AirBase RB stations (FB = −5.5 %), the PM 10 concentrations are overestimated at EMEP RB stations (FB = 2.90 %). The lowest FBs are observed during autumn for AirBase (FB = 0.30 %) and summer for EMEP (FB = −2.30 %), while the highest FB occurs during the winter for both networks (FB = −20.30 % for the AirBase and −14.40 % for EMEP sites). It should be noted that such differences in model performance point out that RB stations of EMEP and AirBase networks are characterized by a different representativeness, with the latter more influenced by local emissions. This is confirmed by comparing the statistics of the observed PM 10 and PM 2.5 data, with the observed mean and standard deviation at EMEP RB sites always lower than at AirBase RB sites. At UB stations, the performance of the model for PM 10 is good over the year (FB = −20.1 %). The FB is lower during spring (−12.2 %) than in winter (−36.6 %) and R is highest during the autumn (0.56) and lowest during summer and winter (0.47). Overall, using the AirBase network, the model agrees better, in terms of correlation at RB, than at UB sites (0.62 and 0.52 respectively) and more precisely during winter (0.67 and 0.47, respectively).
Throughout the year, the model correctly reproduces the temporal variation of the PM 2.5 concentrations at both RB AirBase (R = 0.71) and EMEP (R = 0.68) sites and the highest correlation coefficient is observed during the winter (0.74 and 0.77, respectively). Similar to PM 10 , at EMEP sites, the yearly mean FBs show that the model overestimates PM 2.5 concentrations (7.50 % at AirBase and 8.6 % at EMEP sites). The highest overestimation is observed during the autumn (FB = +14.2 % at AirBase and FB = +15.7 % at EMEP sites). However, during winter the model underestimates the RB concentrations (−8.7 % at AirBase and −11.6 % at EMEP sites).
At UB sites, the model captures the temporal variation throughout the year (R = 0.65) and the FB is rather low (FB = −6.4 %). The highest FB is observed during the winter season (−24.6 %), while according to the RMSE (5.94 µg m −3 ), the best model performance takes place during the summer period, thus confirming the findings of Hodzic et al. (2005).
The intra-annual variability of model performances shows that CHIMERE has more difficulty in reproducing the PM concentration levels during winter, especially at the UB stations (RMSE = 34.88 µg m −3 for PM 10 and 19.87 µg m −3 for PM 2.5 ), when models generally are not able to correctly simulate the stable meteorological conditions that lead to high PM episodes (Stern et al., 2008). For both networks, CHIMERE performed better in reproducing the low PM 10 concentrations as shown by the low quantiles indicated on the daily box-whisker plots time series (Figs. 5s and 6s). Moreover, the comparison of PM 10 and PM 2.5 model performance shows that the highest yearly mean correlation coefficient is calculated for PM 2.5 at UB (0.65) and RB AirBase sites (0.71). This indicates that CHIMERE reproduces the temporal variability of PM 2.5 across Europe better than the PM 10 on a yearly basis. The better performance of the model for PM 2.5 at UB stations confirms that the underestimation of PM 10 is likely due to an underestimation of PM coarse, as reported in other studies Kim et al., 2011;Matthias et al., 2008). The modelled concentration fields described in Sect. 3.1 show that sea salt and dust (coarse particles) can represent a significant part of the PM 10 mass. Indeed, the dust in summer in south Spain and the sea salt during the winter over the North Sea, can reach, respectively, 90 and 80 % of the PM 10 masses. In these cases, the underestimation of PM coarse is reinforced in these parts of Europe. Conversely, the statistical indicators also show that the overestimation of PM 2.5 is larger than that of PM 10 at RB stations and is likely due to the corresponding overestimation of sulfate, the effect of which is less visible in the PM 10 scores due to the compensating underestimation of the PM coarse. PM 2.5 and PM 10 speciation data are available for several EMEP sites. In winter, as reported by  an important lack of organic compounds in models is responsible for large underestimate of models. As the model performance for PM 10 is reflected by the quality of the reproduction of its different components, we also looked at the capacity of the model to reproduce three main inorganic PM 10 compounds: sulfate, nitrate and ammonium.

Sulfate
Sulfuric acid is produced from the oxidation of sulfur oxides, and in turn forms sulfate particles. Secondary sulfate aerosol occurs predominantly in the accumulation mode (Altshuller, 1982) (diameter between 0.1 and 1.0 µm). Oxidant and SO 2 availability are the limiting factors for sulfate formation. In 2009, the 37 stations available over Europe indicate that the highest concentrations of sulfate are measured during winter and spring. This tendency is reproduced by the model (R = 0.57 during spring and 0.52 during the winter) but some maxima are overestimated, especially during spring, autumn and winter. Consequently, the FB is rather low during summer (FB = 20.1 %) but indicates a large overestimation during spring (FB = 53.7 %). CHIMERE results are in contrast with the findings of CALIOPE (Pay et al., 2012a) and CMAQ (Matthias et al., 2008), which tend to underestimate the sulfate surface concentrations over Europe throughout the year.
The CHIMERE seasonal trend is in agreement with the study of Baker and Scheff (2007) over North America, but again in opposition with the results obtained by the CMAQ model over Spain (Pay et al., 2012a). In this case, the highest sulfate concentrations occur in summer due to high oxidation of SO 2 during this period. A possible explanation of the CHIMERE overestimation can be inferred by looking at the remote station of Valentia (Ireland), where CHIMERE overestimates the observed yearly mean concentration by 0.40 µg m −3 . The discrepancy detected at this station is comparable to yearly mean bias calculated at RB stations over the whole domain (0.34 µg m −3 ), thus suggesting that the general overestimation of sulfate can be related to the corresponding overestimation of the sulfate at the boundaries of the domain.

Particulate and total nitrate
Ammonia (NH 3 ) and nitric acid (HNO 3 ) are the two main gaseous precursors than can react together to form ammonium nitrate (NH 4 NO 3 ), depending on the temperature and the relative humidity (RH) (Ansari and Pandis, 1998). HNO 3 can be produced through homogeneous reaction of NO 2 with OH radical (daytime), reaction of NO 3 with aldehydes or hydrocarbons (daytime) or hydrolysis of N 2 O 5 in the troposphere (night time) (Richards, 1983;Russell et al., 1986). During the cold seasons (spring, autumn and winter), the equilibrium of the NH 4 NO 3 system shifts towards the aerosol phase. At low RH, NH 4 NO 3 is solid but if RH overcomes the deliquescence threshold it turns to the aqueous phase (NH + 4 + NO − 3 ) (Bauer et al., 2011). Sulfuric acid plays a crucial role in the formation of nitrate and ammonium. Sulfate tends to react preferentially with NH 3 to form (NH 4 ) 2 SO 4 . Two regimes can be identified: the ammonia poor and the ammonia rich regimes (Bauer et al., 2011). In the first case, there is not enough NH 3 to neutralize the available sulfate. In the second case, sufficient ammonia is present to neutralize the sulfate and the remaining ammonia is available to react with nitrate to produce NH 4 NO 3 .
During cold periods, the formation of NH 4 NO 3 is favoured and the associated low dispersive conditions enhance nitrate during these periods. Hence, the highest measured and modelled concentrations are observed during the winter period. The smallest FB is observed during this season (−68.4 %) and a rather high R value (0.67) is calculated, indicating that the temporal variability of nitrate concentrations is well reproduced by CHIMERE during this period. However, it is also shown that the nitrate is largely underestimated throughout the year (FB = −103.5 %). Several explanations concerning the general underestimation of nitrate can be considered. First, the previously described overestimation of sulfate in poor ammonia regimes could contribute to the underestimation. Second, coarse nitrate chemistry is not repre-sented in the CHIMERE version used for the study, leading to an underestimation of the coarse mode nitrate aerosol. Typical reactions involved in the coarse nitrate chemistry include the neutralization of acidic aerosol particles (NO − 3 ) by different basic positive ions such as Ca 2+ and Mg 2+ . Na + and Cl − are also involved along coastal areas, where high sea salt (NaCl) concentrations are observed (Zhuang et al., 1999 andKouyoumdjian andSaliba, 2006). A coarse nitrate formation scheme was implemented in CHIMERE as part of a research project by Hodzic et al. (2006), which showed that it can increase the nitrate model concentrations up to 3 µg m −3 , especially in southern Europe, where coarse nitrates can represent the major part of the nitrate total mass. Hence, the introduction of a coarse nitrate formation scheme into CHIMERE could help reduce the bias between observed and modelled nitrate. Throughout the year, the total nitrate concentrations (R = 0.56 and FB = −55.1 %) are much better reproduced than the nitrate alone (R = 0.28 and FB = −103.5 %). This result is consistent with the NO 2 underestimation previously discussed, thus confirming a possible lack in NO x emissions. Finally, nitrate and total nitrate observed mean values are more similar than the corresponding modelled concentrations. This means that in the observations the equilibrium between HNO 3 and nitrate is shifted more towards the aerosol phase than in the model. This is probably related to the higher availability of modelled sulfate (overestimated by CHIMERE), limiting the conversion of HNO 3 into the aerosol phase, hence explaining the worsening in model performance when aerosol nitrate alone is considered.

Particulate and total ammonia
Along with sulfate, ammonium is the best SIA compound reproduced by CHIMERE. The FB shows a rather low overestimation throughout the year (27.1 %). The lowest FB is observed during the warm season (7.5 %). This overestimation is very likely driven by the corresponding overestimation of the sulfate. The highest R index is calculated during winter (0.77), indicating that the temporal variability of the ammonium concentrations are better reproduced in this season than during the rest of the year. A similar tendency is seen when using the CMAQ model over Spain and the UK (Pay et al., 2012b;Chemel et al., 2010).
Similarly, the total ammonia is also reproduced well by CHIMERE, with a very low bias observed during winter (FB = −1.8 %). The performance degrades during summer, when the model underestimates observations (FB = −10.7 %). In contrast to the total nitrate, total ammonia is rather well reproduced throughout the year (FB = 6.0 %), suggesting that the yearly NH 3 emissions are well estimated. However, the temporal profile of NH 3 still needs to be improved. In that sense, recent work concerning the improvement of the temporal variability, as well as the magnitude and the spatial distribution of NH 3 emissions from the agricultural sector, has been done for France (Hamaoui-Laguel et al., 2014). Unfortunately, a robust monthly time-profile for the NH 3 emission from fertilizer is yet to be finalized for Europe  before its implementation in the model.

Summary and conclusions
A high-resolution air quality CHIMERE simulation (8 km × 8 km) over most of Europe was performed and evaluated for SO 2 , NO 2 , O 3 , PM 10 , PM 2.5 and SIA using both RB and UB available stations for the year 2009. The model reproduces the temporal variability of NO 2 , O 3 , PM 10 , PM 2.5 better at RB than at UB stations, with yearly correlation values for the different pollutants ranging between 0.62 and 0.77 at RB sites and between 0.52 and 0.73 at UB sites. Similarly, FBs show that the model performs slightly better at RB sites than at UB sites for NO 2 (RB = −33.9 %, UB = −53.6 %), O 3 (RB = 20.10 %, UB = 25.2 %) and PM 10 (RB = −5.50 %, UB = −20.1).
The difficulty for the model in reproducing NO 2 concentration is likely to be due to the general underestimation of NO x emissions, especially during the traffic daily peaks, as well as a horizontal resolution that is not high enough to represent correctly the spatial gradients of the emissions over medium and small cities. Moreover, O 3 is overestimated, implying that NO x titration is not a limiting effect. The NO 2 bias at UB sites is larger than at RB sites, so it can reasonably be assumed that NO x emissions are underestimated. Finally, the total nitrate bias is comparable to the NO 2 bias, at least at UB stations, which represent about 70 % of the available sites. This indicates that the chemical pathway of oxidized nitrogen from NO to HNO 3 is correctly balanced with respect to observed values, suggesting that the limiting factor in nitrate production is the availability of NO 2 and in turn of NO x emissions.
The overestimation of O 3 by the model is related to the NO 2 underestimation, as well as to the high O 3 lateral boundary conditions concentrations, especially during the autumn season. Also, PM 10 and PM 2.5 show a less relevant underestimation than other pollutants, indicating that meteorology cannot be considered the only reason for model discrepancies.
At UB sites, CHIMERE performs better at reproducing the PM 2.5 compared to the PM 10 . The degradation of model performance when moving from PM 2.5 to PM 10 , which was also found in Pirovano et al. (2012), can surely be partially explained by uncertainties on SOA chemistry and their precursor emissions (Po valley and Mediterranean basin) as well as the underestimation of dust (south of Spain) and sea salt (western Europe) concentrations. However, the main reasons are likely to lie also in some missing or underestimated processes such as road dust resuspension (Amato et al., 2009), windblown dust (Yin et al., 2005;Park et al., 2010), especially over the Mediterranean area (Putaud et al., 2004), and finally PM coarse chemistry (e.g. nitrate), as discussed in Sect. 3.2.4. Therefore, different areas of work have been identified and suggested as next steps to improve the model performance in the future: -improvement of the CHIMERE urban parameterizations to better account for the urban effect on meteorology over medium and small cities, -introduction of coarse nitrate chemistry and an advanced parameterization accounting for windblown dust emissions, -continued development of national bottom-up emission inventories, as has been done for France (INS) and Spain , to merge them into the existing European emission inventory.
-continuous improvement on emission inventories to better account for semi-volatile organic compounds and their conversion to SOA, particularly for residential and traffic emissions.
Supplementary data associated with this paper can be found in the online version.
The Supplement related to this article is available online at doi:10.5194/gmd-8-21-2015-supplement.