A new anthropogenic emission inventory was compiled within the TRANSPHORM project, with substantial updates regarding the EU-wide transport activities. The baseline emission data contains the following substances: NOx, SO2, NMVOC, CH4, NH3, CO, PM10, PM2.5, EC, B[a]P (benzo[a]pyrene), and particle number (Denier van der Gon et al., 2014).

The natural emissions of biogenic VOCs and sea salt were calculated online by each model. The wild-land fire emissions were provided by the Integrated System for wild-land fires IS4FIRES v.1 (Sofiev et al., 2009) and were injected as primary particles to a homogeneous layer up to 1 km above the surface. An exception was the SILAM model that calculates the wildfire emissions online, based on the IS4FIRES v.2 calibration (Soares et al., 2015) and vertical profiles of (Sofiev et al., 2012). Desert dust was included only through the lateral boundary conditions; no wind-blown dust was emitted inside the modelling domain.

The inflow of PM and gases through the lateral boundaries was prescribed according to global simulations by two different models. The aerosol boundary conditions were generated by the EMAC (ECHAM/MESSy Atmospheric Chemistry, Jöckel et al., 2006) global model including the aerosol sub-model MADE (Modal Aerosol Dynamics model for Europe, adapted for global applications, Lauer et al., 2005, 2007). Boundary conditions for gas phase chemical species were provided by the global chemical transport model MATCH-MPIC (Model for Atmospheric CHemistry and Transport, Max Planck Institute for Chemistry version, Lawrence et al., 1999; von Kuhlmann et al., 2003; Butler et al., 2012). A detailed description of the models and the simulation setups can be found in Appendix A.

The setups of the four participating models are summarized in Table 1. The detailed model descriptions are given in Appendix B.

The collected model output consists of hourly concentrations of each PM component, separately for fine (PM2.5) and coarse (PM2.5-10) fractions: SO42-, NO3-, NH4+, EC, OC, SOA, sea salt, mineral dust, wild-land fire originated particulate matter, unspeciated other primary PM, and additionally also total PM2.5 and PM10 fields. While the primary anthropogenic PM, EC, secondary inorganic species and sea salt were computed by all models, other components were not always available (Table 2). For instance, OC was provided as a separate species only by EMEP and CMAQ models that included the secondary organic aerosol formation, while in the case of SILAM and LOTOS-EUROS primary OC was lumped together with the rest of the anthropogenic primary PM. Due to very high uncertainties in the forest fire emission inventory, this component was left out of the total PM output of EMEP and LOTOS-EUROS, but was still provided as a separate field. In CMAQ the fine fraction of fire-emitted PM was included in primary OA and the coarse fraction in unspeciated coarse primary PM.

Models also computed the concentration of benzo[a]pyrene (BaP), which was assumed to be an inert fine aerosol not participating in any chemical transformations and not affecting the total-PM budget due to its very low concentrations.

The ensemble median fields for total PM and each separate chemical component listed in Table 2 were computed from the hourly model data from the CTMs (hereinafter, median model). To reduce the influence of the components omitted in some of the models to the total PM, the median fields of the PM components were added up to form another data set of total PM (hereinafter, medianComp model). When computing the median field for every component only those models are used which provided a valid field for that component, and thus the medianComp PM includes valid fields for all species computed by at least one model.

The PM observations of the EMEP network were used for the model evaluation (Table 3). A detailed description of EMEP observations of PM and its components for 2005 is available in Yttri et al. (2007). Table S1 in the Supplement shows the location and altitude of all stations together with a list of observed species. EC/OC observations were available at four stations, which, along with data for a wide range of other species at these sites (Table 4) allowed a detailed evaluation of the PM composition along a transect from northern to southern Europe formed by these stations (Birkenes in Norway, Melpitz in Germany, Ispra in Italy and Montseny in Spain, Fig. S1 in the Supplement).

In addition to the regular monitoring data, the EMEP 2002–2003 EC/OC campaign data are used for evaluation of the seasonality of the carbonaceous aerosols. The data were collected at 12 stations, 1 day per week from July 2002 to June 2003. One station in Portugal (Braganca, PTR0001R) was excluded from the comparison due extremely high modelled wild-land fire contribution which made that station not representative of average conditions – 2005 and 2003 were both record high wild-land fire years in Portugal, while being closer to average in rest of Europe. However, the 2002–2003 EMEP intensive campaign ends in the beginning of July 2003 and thus does not cover the 2003 Portuguese fires which mostly took place in August.

For the model-measurement comparison, the hourly model results were extracted at the station locations and averaged to the temporal resolution of the observations. The model data were converted to the observed quantities. The observed Na+ was assumed to originate only from sea salt, sea salt consisting 30.8 % of sodium by dry weight. The part of the Ca2+ observations not related to sea salt (nss-Ca2+) was used to evaluate the modelled mineral aerosol. The sea-salt-related calcium was subtracted from the observations proportionally to observed Na+ concentrations, sea salt including 1.2 % of calcium by dry weight. Widely varying calcium contents have been reported for Saharan dust from different origin areas ranging from < 5 % to > 15 % (Avila et al., 1998; Formenti et al., 2011; Marconi et al., 2014; Putaud et al., 2004a). The calcium content of anthropogenic emissions also varies between the sources, ranging from less than a percent for biomass burning (Akagi et al., 2011; Larson and Koenig, 1993) to ∼ 30 % for cement and lime production (Lee and Pacyna, 1999; van Loon et al., 2005). In the current study the modelled dust originating from the boundary conditions was assumed to come mainly from Sahara and was attributed 10 % Ca2+ content (Marconi et al., 2014). In addition, 3.5 % Ca2+ content was attributed to the mineral part of primary anthropogenic emissions. This value was chosen as it maximizes the correlation between the observed nss-Ca2+ and the model results. It stays well within the reported range for the anthropogenic emissions. The simulated nss-Ca concentrations were estimated as the sum of the 10 % of dust concentrations plus 3.5 % of the unspeciated other primary PM concentrations.

The OC to OM ratios have been reported to range from 1.2 to 1.6 for fresh anthropogenic emissions, while factors around 2 have been found for aged, secondary and oxygenated aerosol and particles originating from biomass burning (Aiken et al., 2008; Turpin and Lim, 2001). Factor 1.6 was used in this study, analogously to Bessagnet et al. (2014); however, this might be an underestimation for the EMEP stations, which are mostly located in rural areas and would thus be largely influenced by aged aerosols.

The aerosols emitted by wild-land fires also consist mainly of carbonaceous compounds. The fire emissions originated from IS4FIRES (Sofiev et al., 2009), which provides unspeciated PM10 and PM2.5 emissions. In SILAM, EMEP and LOTOS-EUROS the emitted PM was transported as a separate field of unspeciated particulates, while in CMAQ the fine fraction was included in primary OA and the coarse fraction in coarse primary PM. The fire OA in CMAQ cannot be distinguished from the anthropogenic OA, and fire EC was not included in that model. In the other models the fire PM has been further speciated as post-processing following Akagi et al. (2011) and Andreae and Merlet (2001). On average these papers suggest roughly 5 % EC and 50 % OC content for fire-emitted aerosol, the rest mainly consisting of non-carbon atoms in the organic compounds and some inorganics (up to 5 %). The fire contribution to EC and OC has been calculated following this composition and added to the modelled EC and OC.

The models provided dry PM concentrations, which exclude the aerosol-bound water. In gravimetric sampling, which is the reference method for PM observations defined by the European Committee of Standardization, the filters are weighted in laboratory conditions of 20 ∘C and 50 % relative humidity. While the deliquescence relative humidity of most pure inorganic salts present in aerosol is higher than 50 % (Martin, 2000), it can be lower for mixed particles (Seinfeld and Pandis, 2006, chapter 10.2). Apart from that, hysteresis exists in the particle deliquescence-crystallization cycle. For some common aerosol components, such as ammonium sulfate and sodium chloride, the efflorescence humidity, at which the particle crystallizes and loses its water content, is below 50 % (Martin, 2000). Therefore, if the particle has been exposed to a more humid outdoor environment, crystallization might not occur in the standard laboratory conditions, leaving some water bound to the particles on the filter. Based on the dry PM mass and speciation provided by the models, the aerosol thermodynamic model ISORROPIA2 (Fountoukis and Nenes, 2007) was applied to estimate the water content of the aerosol at the conditions where the filters were weighted (20 ∘C, 50 % relative humidity). ISORROPIA2 was run in the reverse mode, where the input quantities were the soluble inorganic components (SIA and sea salt, Ca) in the aerosol phase. Both stable and metastable states were computed, corresponding to the lower and upper branches of the deliquescence hysteresis loop, providing the lower and upper limits of the aerosol-bound water amount.

The model results were evaluated in terms of bias, temporal and spatial correlations and the fraction of model values that are within a factor of 2 of the observations (FAC2).

The annual mean PM2.5 and PM10 fields are presented in Figs. 1 and 2, for the individual models and the ensemble median. All models predict generally similar patterns of the near-surface concentrations for both PM2.5 and PM10 although there are significant quantitative differences between the models' predictions. For PM2.5, the highest concentrations are in densely populated areas such as Benelux and Po Valley, which reflects the large contribution of anthropogenic sources. The PM2.5 concentrations are lower over the open sea, whereas all models predict high PM10 concentrations at marine areas due to coarse sea salt contribution. However, large differences are visible in absolute PM10 concentrations over sea, reflecting the differences between the sea salt emission algorithms. For example, the PM10 level predicted by the EMEP model over sea is up to 4 times higher than that of LOTOS-EUROS, whereas SILAM predicts a considerable south to north decrease in the marine PM10 concentrations due to the strong temperature dependence of its sea salt emissions. The LOTOS-EUROS predictions did not include desert dust and wildland fire smoke, which explains the low values of both PM fractions close to the southern border of the domain.

The MedianComp model that sums up the ensemble medians of all the PM components and thus fully includes the wildfire emissions, desert dust and secondary organics, shows higher PM concentrations than the median model in various areas. The difference between the MedianComp and median models in the central Europe is mainly due to SOA. PM10 in the southern part of the domain is influenced by the dust inflow from Sahara, while the fire impact is visible in Portugal. In CMAQ the dust and fire contributions are very low, and LOTOS-EUROS does not have them at all, so the median total PM is based on half of the models with zero or very low dust concentration. MedianComp is based only on the valid dust fields of SILAM and EMEP and thus includes noticeably higher dust contribution.

Figures S2 and S3 show the spatial patterns of model bias for PM2.5 and PM10 with regard to the EMEP network. The individual models and the ensemble median underestimate both PM2.5 and PM10 concentrations quite homogeneously in space. The only station, where the models noticeably overestimate the PM concentrations, is located on the Schauinsland mountain in the Black Forest, with an elevation above 1200 m. About 10 km from the station and about 1 km below it is the city of Freiburg. The overestimation occurs in winter (see the monthly average time series on Fig. S4), when the site is actually in the clear air above the low wintertime boundary layer, while in the models, both the city and the station are covered with one uniformly mixed grid cell. In summer, when the site is located within the boundary layer, the PM concentration there is mostly underestimated.

As seen from Fig. 3, all models report stronger seasonal variations in total PM than is observed. The models report highest concentrations in autumn or winter, while the observations peak in spring. There are also noticeable differences between the models. In SILAM and LOTOS-EUROS the PM2.5 concentration is noticeably lower in summer, while in CMAQ the autumn concentrations are substantially higher than during the other seasons. EMEP predictions show very small seasonal variations for both PM2.5 and PM10. The different anthropogenic emission seasonalities applied in the models (Table 1) explain part of the differences in Fig. 3. However, omitting the secondary organic aerosol (SOA) is probably the main explanation for the exaggerated PM2.5 summer minimums calculated by the LOTOS-EUROS and SILAM models. SOA is present in larger quantities in summer due to biogenic emissions of semivolatile organic compounds.

The model skill scores for PM10 and PM2.5 in winter and summer are presented in Table 5. The fraction of model values that are within a factor of 2 from the observations is larger in winter than in summer for all models, except EMEP. The temporal correlation of daily concentrations tends to be higher in winter, with the exception of CMAQ that has the lowest wintertime correlations among the models. The models' ability to reproduce the average seasonal concentration patterns differs between finer and coarser particles – spatial correlation of PM10 is higher in summer, while PM2.5 correlates better in winter for most of the models. Low summer-time correlations of LOTOS-EUROS result from large underestimations in Spanish stations due to missing Saharan dust. The worse summertime scores are probably due to the highly uncertain components that dominate the summer aerosol – wind-blown dust, wild-land fires, and biogenic secondary organic aerosols. The only score that is better in summer than in winter is the spatial correlation for PM10. In summer, the PM10 pattern over Europe is formed by the inflow of Saharan dust and wild-land fires in Portugal and Spain, which create a strong south–north gradient. This gradient is reproduced by the models, although with smaller magnitude. As the species contributing to this summertime south–north gradient are desert dust and wild-land fires, which by nature are episodic and hard to model, the temporal correlation and factor-two agreement are still generally lower and bias is larger in summer. In winter, the particulate matter is dominated by the anthropogenic emissions, forming a more complex pattern, and thus the spatial correlation is worse.

As seen from Table 5, while the bias of the ensemble median follows the mean bias of the models, the temporal and spatial correlations exhibit more complicated relations. In winter, the ensemble median shows the overall best temporal correlation for both PM2.5 and PM10, while in summer EMEP performs better. The spatial correlations of SILAM or EMEP models usually slightly exceed that of the median model.

The medianComp fully includes SOA, desert dust, and fire-induced PM. As the contributions of those components are more important in summer, the difference between the median and medianComp is largest in summer, and small in winter (Table 5). MedianComp thus shows a noticeably smaller summer-time bias than the median model for both PM10 and PM2.5. For PM10 the medianComp outperforms the median model in summer in all quality scores, while for PM2.5 its spatial and temporal correlations are worse. This indicates that accounting for desert dust, which is an important component in PM10 and less so in PM2.5, improves significantly the models' ability to reproduce the observed coarse PM patterns. It is worth pointing out that the measurement network includes a large number of Spanish sites, where mineral dust is more important than for the rest of the modelling domain. The worsening of the summer time correlations of PM2.5, on the other hand, indicates that improvements are necessary also for modelling the other components that were included only by few models, such as smoke from the vegetation fires and formation of secondary organic aerosols from the biogenic precursors.

The water contribution estimated with ISORROPIA2 based on the modelled aerosol composition is shown on Fig. 3 with light blue. The solid part indicates the stable water content (lower branch of the hysteresis cycle) and the striped part the metastable phase (the upper branch of the hysteresis cycle). In the stable case, the annual mean PM2.5 water content, average over all EMEP stations, stayed between 4 and 9 % depending on the model, and between 11 and 17 % for PM10. For PM2.5, the models predicted annual average water content above 10 % for only a few stations. For PM10, CMAQ and EMEP predict majority of the stations to have less than 10 % of water content, while LOTOS-EUROS and SILAM predict the majority to be between 10 and 20 %. Annual average water contents of more than 25 % were predicted for some stations. The water content of PM10 computed in the metastable mode was on average about twice as high (∼ 25 %); ∼ 20 % water content was predicted for PM2.5.

As seen from Table 6, adding the aerosol-bound water reduces noticeably the model bias for both PM10 and PM2.5. For PM2.5 the correlation coefficients are not much affected, while for PM10, they are noticeably reduced. The factor-2 agreements improve due to the bias reduction. The worsening correlations could be related to the models overestimating the sea salt concentrations that can lead to overestimation of the water content in PM10, as sea salt is the most hydrophilic of the considered aerosol components.

The ensemble median maps of the PM components are shown in Figs. S5, S6, S7 and S8. In the continental Europe the models predict the highest contribution from the summed secondary inorganic species, nitrate being most important in central Europe and sulfate contributing mostly in southern and eastern regions. Sea salt concentrations are high over the marine areas and shores but decrease rapidly inland. Desert dust and wild-land fires can be the main contributors to aerosol in some areas, but their impact is spatially limited.

The models' performance in comparison to the measurements of the PM chemical components is shown in Table 7 (note different units (µg N, S, C, Na, Ca m-3); the model maps (Figs. S5–S8) are plotted in full modelled species mass). The right columns of Figs. S5 and S6 show the spatial spread of the model bias. PM10 is underestimated slightly more than PM2.5 by all models except EMEP, possibly due to the missing emissions of wind-blown dust, which mainly resides in the coarse fraction. Sodium and NO3- are on average overestimated, whereas NH4+ and SO42- are underestimated but much less than total PM. The overestimation of NO3- is most noticeable in the central and eastern Europe, whereas the western areas are predicted accurately and the northern ones are underestimated (Fig. S5). The carbonaceous aerosols and the mineral dust are underestimated more than the total PM.

Temporal correlation of the daily time series is usually lower for the specific components than for the PM10 and PM2.5, and same is true for FAC2. One possible reason for this is that summing up the non-correlated individual components smooths the gradients and reduces the penalty for slight mislocations of plumes. It cannot be ruled out that the lower correlation can in some cases be also due to higher observation errors. In particular, higher uncertainties are present in observations of mineral dust and carbonaceous species (Putaud et al., 2010, Annex 5; Sillanpää et al., 2006), but observation artefacts also influence the species with dynamic equilibrium partitioning between particulate and gaseous phases, such as NH4NO3 (EMEP, 2001; Putaud et al., 2010, Annex 5). It also has to be noted that different pollutants are observed by different sets of stations in the EMEP network, which might induce some extra variations to the average model scores.

The temporal correlations of the modelled carbonaceous compounds with their observed concentrations in PM2.5 are among the highest for the PM components, and substantially lower for the observations of the same compounds in PM10. The correlation coefficients are lowest for dust, but also below average for NO3-. One can also notice a better agreement for the sum of HNO3 and NO3- than for HNO3 or for nitrates. Temporal correlation coefficients and factor-2 agreements are noticeably worse for NH3 and HNO3 compared with NH4 and NO3 aerosols or the gas-aerosol sums. The lower scores reflect the complexity of the gas-particle equilibrium between the NH4NO3 and HNO3 and NH3. Another possible reason for higher scores for the summed gases and aerosols (NH3+ NH4 and NO3+HNO3) than for NH3, HNO3 and nitrate aerosol separately is the higher uncertainties in the observations of the latter (Chang et al., 2002; Schaap et al., 2011; EMEP, 2001; Putaud et al., 2010, Annex 5). Also the pulsed behaviour of the aerosol nitrate production in the PBL, recently described by (Curci et al., 2015), could be a reason for inaccuracies in modelling the nitrate aerosol. Conversely, for NH3+ NH4+ the temporal correlation is lower than for NH4+ only, albeit the bias is smaller and FAC2 is better. Sulfate and NH4+ show very similar correlation values, as large fraction of NH4+ is present in the form of ammonium sulfate. The correlation for sulfates is higher than for SO2, probably mainly due to the smoother features of the sulfate field – SO42- as a secondary pollutant is less affected by the local sources.

The evaluation of the secondary inorganic aerosols (Fig. 4) shows that the models adequately reproduce the observed seasonal variation of SIA and its precursors. Moderate deviations exist: somewhat exaggerated seasonal cycle of SO2 is shown by EMEP; CMAQ overestimates the autumn levels of both NO3+HNO3 and NH3+NH4,and predicts an autumn peak for all three SIA species; high autumn NH3+NH4 and SO42- are also predicted by SILAM and high autumn levels of NO3+HNO3 by EMEP. SILAM manifests strong over-estimation of sulfates in autumn, but no overstatement of SO2. While the models adequately reproduce the summertime drop in the concentrations of NH4 and NO3, they tend to overestimate the autumn concentrations and while the observations show the highest concentrations in spring for all three SIA species, this is not reproduced by the models. This could be one of the reasons for the errors in the seasonal cycle of total PM.

The contribution of the gas phase HNO3 and NH3 to the sums of NO3+HNO3 and NH3+NH4, is shown on Fig. 4 with dark shading. On annual average level the gas phase fraction is in both cases relatively well reproduced by most of the models, only SILAM underestimates HNO3 and overestimates NH3 contributions. However, the models do not adequately reproduce the seasonal variations in HNO3 and NH3 concentrations – EMEP and LOTOS-EUROS overestimate the seasonal variability of HNO3, CMAQ strongly overestimates the autumn NH3 concentrations and so does with smaller magnitude also SILAM. The seasonal variations of NO2 are reproduced well, but all models apart from CMAQ overestimate NO2 all of the seasons Table S3 and Fig. S9. This can be one of the reasons for the overestimation of the sum of NO3- and HNO3 seen on Fig. 4.

For the sea-salt concentrations, EMEP and CMAQ predict higher levels than the other models and are also higher than observations in all seasons (Fig. 5, left-hand column). However, the seasonal cycle is reproduced well. Conversely, LOTOS-EUROS, while being closest to the average annual level, underestimates the seasonal variations. SILAM is also close to the observations but seems to have an exaggerated temperature dependence of the sea salt emission as it overpredicts the summer and autumn concentrations while underestimating in winter. For all models the Na concentration for model-measurement comparison was computed from sea salt in PM10, while the Na observations in the EMEP network are made mostly in whole aerosol without size limits. As the models already overestimate Na concentration, including also the particles larger than 10 µm would lead to overestimation even larger than what is shown. However, comparing Na in PM10 with Na in whole sea salt in SILAM (size range 0.01 to 30 µm), the changes are minor for majority of the EMEP stations that observed Na in 2005: below 5 % for 65 % of the stations, below 10 % for 77 %, and below 20 % for all stations. The concentration changed more than 10 % only in the stations located directly at seaside.

Only SILAM and EMEP modelled the transport of desert dust from the boundaries (mainly Sahara) as a separate tracer. A 10 % Ca2+ content was assumed for it (right panel of Fig. 5, shaded part of the bars) and in addition a 3.5 % Ca2+ content was attributed to the mineral part of primary anthropogenic emissions (non-shaded part of the bars). The modelled contributions from these sources are about equal, except for winter when the models predict almost no dust from Sahara. The nss-Ca2+ concentrations are substantially underestimated by the models for the whole year (the EMEP model underestimated the nss-Ca by 75 % and SILAM by 58 %). Considering that the models omitted the wind-blown dust emissions inside the European modelling domain, this underestimation is not surprising. The seasonal patterns of the models differ from the observations, where the autumn concentrations are noticeably lower than the summer ones and close to the winter levels – the models rather suggest similar dust levels for most of the year, except for winter when the predicted concentrations are lower.

The available observations of the carbonaceous aerosols for 2005 point out a strong underestimation of these components by the models (Fig. 6, upper panels). The models underestimated the EC in PM2.5 by ∼ 20–60 % and OC by 40–80 % (Table 7). The models only provided the fine fraction of these compounds as separate tracers; the anthropogenic coarse mode emissions were included in the coarse unspeciated primary aerosol. The fire PM concentrations modelled by EMEP, LOTOS-EUROS and SILAM have been speciated as post-processing following Akagi et al. (2011) or Andreae and Merlet, (2001). On average these papers suggest roughly 5 % EC and 50 % OC content for fire-emitted aerosol. The fire contribution to EC and OC calculated following this composition is shown on Fig. 6 with darker shading.

The observations on the upper panels of Fig. 6 are shown for OC and EC in both PM2.5 (shaded part of the bars) and PM10 (whole bars). The observations in PM10 are about 20 % higher than those in PM2.5. The modelled fine EC and OC correlate substantially better with the observations in PM2.5 than with those in PM10 (Table 7). This agrees quite well with the emission estimates of Kuenen et al. (2014), according to which the anthropogenic emissions of coarse EC and OC are about 5 times lower than their fine mode (PM2.5) emissions and also originate mostly from different sectors than the fine mode – coarse EC from large scale combustion and coarse OC from agriculture, while the most contributing sources of fine carbonaceous aerosol are residential combustion and traffic. As large part of OC is secondary and also resides in fine fraction, some extra sources are still necessary to explain the observed coarse OC. The contribution of the coarse mode to the OC concentrations is highest in summer and autumn and lowest in winter, consistent with origin from biological and agricultural sources.

SILAM only shows a small fire contribution to EC in spring and summer, while in EMEP and LOTOS-EUROS the contribution is larger and visible all year round. EMEP also predicts a noticeable fire contribution to OC for all seasons. For EMEP and LOTOS-EUROS, the fire contribution reduces the model bias for the carbonaceous species, while at the same time reducing the correlation with the measurements of EC and OC in PM2.5 (Table S2). The SILAM EC prediction quality does not noticeably change. The correlation with EC and OC observations in PM10 in some cases rises when including the fire-emitted part.

The models reproduce the observed seasonal variation in EC concentration, but all underestimate with varying magnitude. As elemental carbon emission data were the same for all models and no chemical transformations affect its concentrations in the atmosphere, the large differences of the average EC concentrations between the models are rather surprising. SILAM predicted the highest concentrations, being more than twice as high as CMAQ and EMEP in winter, the difference being smaller for the other seasons. A possible explanation is the considerably lower dry deposition of fine aerosols in SILAM (Kouznetsov and Sofiev, 2012). Different treatment of EC hygroscopicity and ageing, affecting the efficiency of its wet scavenging, could also contribute to differences in the model results. The relatively coarse vertical resolution near the surface is a plausible explanation of EMEP's underestimation of EC, especially in winter. Finally, the emissions of carbonaceous particles are likely to be underestimated during the cold seasons due to large uncertainties in the emission factors for the residential wood burning (Denier van der Gon et al., 2015).

For OC only CMAQ and EMEP results are included in the analyses, as OC was not available from LOTOS-EUROS and SILAM (these models did not calculate the secondary OC and lumped together the primary anthropogenic OC with the primary PM emissions). The models did not reproduce the observed seasonal variations in OC concentration, which peak in winter and autumn. Both models show quite flat seasonal profiles and if accounting for the wildfire emissions, EMEP even overestimates the summer concentrations. The large underestimation in winter could be caused by missing emissions of domestic heating (Denier van der Gon et al., 2015), but also the SOA formation from anthropogenic aromatics could be underestimated. A rather large portion of semi-volatile organics is believed to be missing in current anthropogenic emission inventories of PM2.5 and NMVOCs (Denier van der Gon et al., 2015; Donahue et al., 2006; Ots et al., 2016; Robinson et al., 2007). Cooking emissions have been pointed out as another missing source of organic aerosols (Fountoukis et al., 2016; Young et al., 2015).

The above analysis was based on only four stations that measured the carbonaceous compounds during 2005, which makes it uncertain. To better understand the results for carbonaceous compounds, we used OC / EC observations from the EMEP campaign in 2002/2003 (Simpson et al., 2007; Tsyro et al., 2007), when the carbonaceous aerosols in PM10 were observed at 12 stations. Keeping in mind the inter-annual variability, some kind of indication of model biases can still be obtained from comparing the modelled seasonal average concentrations of EC and OC for 2005 with the seasonal averages of these observations, especially as the PM10 concentrations observed during this campaign were underestimated by the models by about the same factor as the PM10 observations of 2005. The comparison supports the previous conclusion: the modelled OC concentrations, and also those of EC at many sites, are substantially lower than the observations (Fig. 6, lower row), and models completely miss the observed OC winter maximum.

All models of this study overestimated the Benzo(a)pyrene concentrations all year round (Fig. 7), whereas the seasonal cycles are qualitatively similar to the observed cycles. This is somewhat unexpected, as the models underestimate the concentrations of black carbon and the sources of these two pollutants significantly overlap. One possible reason for this can be a simplified approach taken by the models to simulate this species: BaP was assumed to be an inert fine aerosol not participating in chemical transformations and not partitioning to gas-phase. In more complex models the heterogeneous oxidation by ozone has been reported to efficiently reduce the BaP concentrations (Friedman and Selin, 2012; Matthias et al., 2009). It is also probable that some part of the over-estimation, especially in wintertime when the oxidation is slower, may be attributed to the emissions.

The PM composition was evaluated at the four stations that provided more complete data on the chemical speciation of the PM concentrations (Fig. S1). All the modelled and observed species in Fig. 8 are converted to total masses of the species in order to add up to total PM2.5 or PM10. OC is converted to total organic aerosol mass by multiplying with 1.6 and nss-Ca2+ to mineral dust by multiplying with 10. Observed sea salt is taken as the sum of Na+ and Cl. However, the modelled and observed species are not always directly comparable, e.g. some models include carbonaceous aerosol in primary anthropogenic PM or wildfire smoke and mineral dust in the primary unspeciated aerosol. OA comparison with observations is based on CMAQ and EMEP results only, since the other models include OC in other PM. Dust comparison is based on EMEP and SILAM.

As seen in Fig. 8, for these stations the sum of measured PM components was up to ∼ 20 % lower than measured total PM2.5 and PM10. The water contribution estimated with ISORROPIA2 can be seen on Fig. 8 in light blue. Adding the aerosol-bound water in metastable state closes the gap between the observed total PM10 and the sum of the individual components in almost all cases (in Montseny the PM estimate based on nearby stations can be inaccurate). In Ispra and Birkenes the observed PM is exceeded, which could indicate that the aerosol on the filters is in crystallized state or be due to inaccuracies in other observed species.

At Melpitz the models are close to the observations for SIA and overestimate the sea-salt contribution. Carbonaceous part is underestimated, though accounting also for the wild-fire emissions (striped orange on Fig. 8) brings EMEP very close to the OC observations in PM2.5. The mineral dust transported from the boundaries (separately only in EMEP and SILAM) shows lower values than the observed dust concentration. EMEP is the only model, where the unspeciated part of the primary PM (PPMr) consists solely of mineral components, while in the other models it is mixed with either the primary organic aerosol or wild fire smoke. The sum of EMEP PPMr and desert dust is very close to the observation. However, here the observed total mineral dust concentration is estimated assuming 10 % Ca2+ fraction, which is an overestimation for majority of the anthropogenic emissions.

At Montseny all models overestimate NO3-, whereas NH4+ is overestimated by EMEP and LOTOS-EUROS and SO42- by SILAM. Considering that forest fire emissions also have substantial organic aerosol content, EMEP model even overestimates the observed OA, while EC is overestimated by all models. Due to over-predicted NO3-, PM2.5 is overestimated by EMEP and LOTOS-EUROS at this station. The modelled desert dust values are again substantially lower than the observed dust, while adding the PPMr concentration brings EMEP very close to the observation in PM2.5, although still underestimating the mineral part of PM10.

At Ispra, the major contributor to the observed PM is organic aerosol, while the models show a few times lower values. Elemental carbon is also somewhat underestimated. However, Yttri et al. (2007) warn against possible errors in the observations of carbonaceous aerosols at that site for 2005, especially in the case of PM10. CMAQ also underestimates all SIA in Ispra and all models miss some SO42-, while fine NO3- is overestimated by LOTOS-EUROS and SILAM. Sea salt and dust cannot be evaluated in Ispra, as no Na+ or Ca2+ observations were available in 2005. This also makes the estimation of water content in the observed PM inaccurate.

At Birkenes all models but LOTOS-EUROS overestimate the measured PM10. PM2.5 is not shown, as the SIA, Ca2+ and Na+ observations were not available separately for fine and coarse aerosol. As these species were measured in total aerosol, they might partly also originate from larger particles than PM10. Elemental carbon concentrations are somewhat overestimated by CMAQ and SILAM. EMEP overestimates the organic aerosol. All models overestimate the sea salt contribution in PM10 by factor of 2–3, leading to very high water uptake of the aerosol. Modelled desert dust alone is lower than the nss-Ca2+-based observation, while its sum with PPMr brings EMEP again very close to the observation.

All-in-all, overestimations of some components can bring the models very close or even over the observed PM levels, while still underestimating other components. The sea-salt concentrations are usually overestimated by all models – up to a factor of 2–4 – and this becomes important at the sites with a significant sea salt fraction in the mass budget. Sulfates are reproduced comparatively well with limited regional differences, probably driven by emission data quality. NH4+ is quite well reproduced by all models, except for CMAQ, which underestimates it. For nitrates, the models showed varying degree of agreement. OA is mostly underestimated, while EMEP can also sometimes overpredict its concentration. Models underestimate high observed EC observations, while low concentrations can be overestimated. Mineral dust, which was taken only from global boundary conditions, is not enough to explain the observed nss-Ca2+ concentrations. Adding it up with the mineral part of the anthropogenic PM brings EMEP model close to observations, at least for PM2.5. However, EMEP still underestimates the mineral contribution to PM10 in Montseny, which is the station most influenced by Saharan dust. The underestimation of nss-Ca2+ is smaller in the north, further away from Sahara (Fig. S6, lowest right panel).

The individual PM components are reproduced with about the same or lower quality as the total PM. The temporal correlation of the daily time series is usually lower for the specific components than for the total PM, and same is true for the FAC2 agreement. This could indicate compensating errors in the model parameterizations, but even without that the comparison for the sum of the non-correlating components would benefit from the averaging of the errors in the components.

The considered models are found to underestimate the observed total PM2.5 and PM10 concentrations. However, not all individual PM components are equally underestimated: secondary inorganic species are reproduced quite accurately and sea salt is usually overestimated. This suggests large underestimations for carbonaceous and mineral aerosols, which is supported by the few available observations. However, the mismatch between the modelled and observed quantities leaves large uncertainties in evaluating how much exactly these aerosol components are underestimated in this study.

Wind-blown crustal aerosols have been pointed out as a potentially underestimated fraction of PM (Im et al., 2015) and substantial underestimation is found also strongly indicated by this study. The fraction of calcium observations not related to sea salt was used to evaluate the mineral dust concentration in this study. However, the evaluation of the wind-blown dust against non-sea-salt calcium observations is highly uncertain. Various options exist for deriving the total mineral dust concentration from observations of e.g. aluminium or non-sea-salt fraction of calcium (nss-Ca2+) (Marconi et al., 2014; Putaud et al., 2004b), but fractions of these vary among different minerals and dust source areas (Avila et al., 1998; Formenti et al., 2011). Putaud et al. (2010) provided various formulas for estimating the mineral dust concentration from several related tracers, such as Si, Al and Fe and nss-Ca2+. They estimated that the uncertainty of deriving mineral dust concentration from observations can reach ±150 %. Observations of Si, Al and Fe were available only in Montseny station. The location of Montseny station about 30 km from the Mediterranean coast, at 700 m height from sea level exposes it to Saharan dust episodes (the high dust contribution there is visible on Fig. 8) and thus allows for evaluating the nss-Ca2+ as a desert dust tracer. The nss-Ca2+ concentrations there correlate well (correlation coefficient above 0.9) with the observations of the other mineral dust tracers, and the dust concentration obtained by assuming 10 % Ca2+ content is not far from the estimates provided by the most detailed formulas presented in the Annex 5 of Putaud et al. (2010).

However, the wind-blown crustal emissions are not the only source of mineral aerosols. Generally, about half of primary fine anthropogenic aerosol emission consists of carbonaceous components (Kuenen et al., 2014), while the rest is mainly associated with mineral compounds. For coarse fractions, the carbon content is low; hence the bulk of mass consists of mineral components. Therefore, the unspeciated primary PM in the models has to also be included to the comparison with the nss-Ca observations. However, the variations of the calcium content are even wider there, ranging from less than a percent for biomass burning (Akagi et al., 2011; Larson and Koenig, 1993) to ∼ 30 % for cement and lime production (Lee and Pacyna, 1999; van Loon et al., 2005). According to Lee and Pacyna (1999), the emissions from coal combustion include 2 % of Ca2+ and steel and iron production emissions 0.7–3.6 %. The Ca2+ content in the top soil layer, influencing the dust emissions form agricultural activities, but also the dust suspended by wind and traffic, stays in Europe below 5 % and below 1 % in the northern areas (van Loon et al., 2005). Although the 3.5 % Ca2+ content used in this study for the anthropogenic mineral aerosol is well within these limits, good model-measurement agreement cannot be expected due to these large variations. The uncertainty in aerosol Ca content can be expected to be a few times. However, with the 10 and 3.5 % Ca content, the EMEP model underestimated the nss-Ca by 75 % and SILAM by 58 %, so even assuming twice the calcium content, the nss-calcium concentrations would still be underestimated by the models.

In 2005, the wild-land fires took place in a comparatively small part of the domain and affected noticeably only a few stations in Spain and Portugal. However, the very strong emission within short time had a significant impact on PM concentrations even at annual level. Therefore, exclusion of this component from the computations results in strong underestimation and poor correlation, both in space and in time. On the other hand, fire emission is arguably among the most uncertain input data sets (Soares et al., 2015) and requires careful treatment, accounting for the strong diurnal variation of the fluxes, as well as the vertical injection profile. The fires emit wide spectrum of pollutants and the observations rarely distinguish the fire-originated aerosol from the rest of atmospheric PM. Specific tracers of combustion of organic materials, such as levoglucosan, are occasionally measured, but their relation to the total emitted PM is not fixed. Also, wood burning is common in many other sources, such as domestic heating, which cannot be told apart from large scale fires. As a result, evaluating the modelled fire smoke becomes possible only for episodes with strong domination of fire-induced pollution. On the other hand, inaccurate representation of the fire emissions and their temporal and vertical profiles can result in a very poor correlation with the measured concentrations. In EMEP and LOTOS-EUROS, using the IS4FIRES v1 emissions resulted in degradation of model scores for PM2.5 and PM10, and thus these models excluded the fire PM from their total PM2.5 and PM10 fields, the correlations for EC and OC also reduced when fire contribution was added (Table S2). In SILAM, a newer version of the emission data was used (IS4FIRES v2, Soares et al., 2015), together with dynamic emission vertical profiles (Sofiev et al., 2012), while in other models the IS4FIRES v1 emission data was spread evenly to the first 1000 m. Mainly due to the vertical profiles that release most of the smoke high aloft, the ground level concentrations of fire PM were substantially lower in SILAM and the fire PM did not negatively affect the model performance, demonstrating that the quality of the fire emission data is essential for simulating the particulate matter concentrations.

The spatial features of the compared data can also lead to uncertainties in model-measurement comparison. Regional models with grid-cell sizes of a few tens of kilometres are not designed to reproduce the concentration patterns with smaller spatial scales, e.g. in the vicinity of strong sources, in urban conditions or mountainous areas. For instance, the study by (Im et al., 2015) found a stronger underestimation of PM in urban stations than in rural ones, which, apart from emission underestimation, could also be explained by the limited representative area of these stations. Also Vautard et al. (2007) found larger PM underestimation in the urban stations by the large scale models than by those with higher resolution. Even for stations of the EMEP network, whose locations have been carefully selected to represent the regional background (EMEP, 2001), the effects of local topography and sources may still be noticeable. The models' performance was found to degrade in the higher stations; there is a strong negative correlation between the station altitude and the models' temporal correlation coefficients for both PM10 and PM2.5. The bad model performance is caused not only by the altitude difference between the station and the model grid cell average, but also other inhomogeneities, such as strong emission sources in the area. Indications of wintertime overestimation are visible for some of the high stations, but not all the high stations are located at extreme points of the terrain, such as mountain summits, and not all of them have strong emission sources in the immediate vicinity. Opposite problems arise for sites located in narrow valleys, where the models' cell-mean altitude is higher than the station and the models overspread the pollution that in reality can be trapped in the valley.

Wang et al. (2014) and Samset et al. (2014) demonstrate that shorter EC lifetimes are necessary for reproducing the EC vertical profiles and low concentrations in remote regions. This result contradicts with the current model intercomparison, where SILAM was found to best reproduce the observed EC concentrations, and longer EC lifetime due to slower deposition in that model was assumed as the main reason for the model-to-model differences. Also the temporal correlation with 2005 observations and spatial correlation between the models 2005 average EC and EC observed during the 2002–2003 EMEP campaign is no worse for SILAM than it is for the other models, and hence there is no clear indication that the slower deposition would not be consistent with the surface EC observations in European scale. However, as indications of strong underestimation of EC emission were found, the slower deposition in SILAM is likely to be compensating for the missing emissions. Observations of vertical profiles and concentrations in more remote locations would be necessary for investigating this issue; unfortunately such were not available in Europe in 2005.

Seasonal variations of secondary aerosols result from a wide range of processes. Firstly, the emissions of precursors vary seasonally and some of these depend on meteorology. For instance, NH3 emission depends strongly on the seasonality and type of agricultural activities, as well as on the temperature. Secondly, formation of secondary pollutants from precursor gases is controlled by multiple factors with strong seasonal cycles: the abundance of oxidants and water, ambient temperature and solar radiation, etc. Thirdly, gas-particle partitioning of semi-volatile species depends on temperature and relative humidity. There are significant differences in the treatment of these processes in the models, leading to substantial variations between the modelled seasonal cycles of the secondary aerosol concentrations. Resulting from these variations, the ability of the models to represent the observed PM2.5 and PM10 concentrations also varies seasonally and largely depends on the completeness of PM chemical composition in each specific model. For instance, the models that do not include SOA have larger bias in summer. Missing the contribution of the desert dust and wild-land fires also leads to negative bias and strongly reduces spatial correlation during summer time.

Especially for NH3 the timing of the emissions as used in the models (fixed temporal profiles) can deviate substantially from real world emission timing which is largely controlled by meteorology (Backes et al., 2016; Hamaoui-Laguel et al., 2014; Hendriks et al., 2016). Meteorology also influences the total amount of emitted NH3 but the strongest influence is on the timing. The timing of agricultural activities, such as manure spreading has also a strong impact on the emissions (Hendriks et al., 2016). The inaccurate temporal emission profiles lead to models not reproducing the seasonal cycle of SIA (Fig. 4) and PM.

The observed nss-Ca concentrations peak in spring and so do the modelled Saharan dust concentrations. Previous studies about Saharan dust confirm the emissions peaking at spring (Fiedler et al., 2013; Laurent et al., 2008). In addition to Saharan emissions, there are other reasons for elevated crustal aerosol concentration in spring, such as agricultural activities and vehicle-caused erosion of roads – in colder regions where winter tires are used and the roads are sanded against slipperiness, high dust emissions occur when the road conditions get dry in spring. These emissions were not included in the model runs, which could be another reason why the models miss the spring peak in PM.

Another source of OC that has received very little attention is the primary biogenic particles, such as plant debris, fungal spores and pollen. While majority of these particles are larger than 10 µm, the aerodynamic diameter of some common fungal spores is below 10 µm and in some cases even below 2.5 µm (Reponen et al., 2001), making them relevant to even PM2.5. According to Hummel et al. (2015) and Winiwarter et al. (2009) the fungal spores could contribute noticeably to aerosol concentration in summer and autumn (up to a microgram m-3 in long-term average and even more during specific episodes).

The PM components mentioned above as the most uncertain and sometimes omitted in the models (wind-blown dust, wild-land fire smoke, biogenic primary and secondary particles), are all more common in summer time. The models mostly do underestimate PM by a larger fraction in summer. On the contrary, organic aerosol is underestimated by a larger fraction in winter. As noted by Denier van der Gon et al. (2015), Lefebvre et al. (2016), the residential wood combustion emissions are severely underestimated in the current emission inventories and that would cause underestimation in carbonaceous particles during the cold seasons. According to Fountoukis et al. (2015) underestimation of the SOA formation rate in low light conditions could be another reason for the wintertime OA underprediction.

Previous publications (Putaud et al., 2010; Sillanpää et al., 2006; Tsyro, 2005) have pointed out that a gap exists between the gravimetric total-PM observations and the sum of individual PM components (also seen in Fig. 8). The main reason for this has been found to be aerosol-bound water contribution to the gravimetric observations, which can contribute ∼ 20 % of mass to annual average observations. Based on the modelled aerosol composition, the average water content at laboratory conditions was estimated roughly between 5 and 20 % for PM2.5 and between 10 and 25 % for PM10, depending on whether the aerosol was assumed to be in stable or metastable state, the latter corresponding to situation, when the aerosol has been exposed to more humid conditions and crystallization has not occurred. Adding this contribution to the modelled PM reduced the model bias 25–70 %, but also reduced both spatial and temporal correlations with the observations.

There are several uncertainties in estimating the PM water content. Firstly, the water content depends on the outdoor humidity at the measurement location as well as the filter transportation and storage conditions, so it cannot be determined, whether the aerosol is in stable or metastable branch of the hysteresis cycle. Secondly, ISORROPIA2 computes the water content based on the inorganic part of aerosol – SIA, sea salt, calcium; it does not take into account the water related to the hydrophilic part of the organic aerosol, which could also influence the water uptake of the inorganic species (Jing et al., 2016). Thirdly, the aerosols were assumed fully internally mixed, which lowers the deliquescence humidity compared to external mixtures and might lead to overestimation of water uptake. Overestimating hydrophilic compounds, such as sea salt can also lead to overestimation of the water content in PM. Also, in addition to the particle-bound water, the filters themselves can accumulate humidity and influence the measurement results (Brown et al., 2006). Taking into account all these uncertainties, the water content estimated based on the observed PM composition (Fig. 8) assuming metastable state closes the budget for several stations (e.g. PM10 and PM2.5 in Melpitz and PM2.5 in Montesny) surprisingly well.

Even when non-gravimetric measurement methods are used, they often include processing steps to obtain similar values to the gravimetric method, which is defined as the reference for PM measurements by the European Committee of Standardization. The reason for these corrections is that a substantial fraction of secondary aerosols consists of components, such as ammonium nitrate and semivolatile organic species, whose partitioning between gaseous and particulate phase depends on the atmospheric conditions and concentrations of the compounds. Apart from water, also the semivolatile compounds can condense or evaporate during the measurement process. Loss of semivolatiles is an especially important issue for observation techniques that involve heated inlets, and dedicated methodologies have been developed to compensate for such losses and bring the results closer to the standard gravimetric observations (Alastuey et al., 2012; Charron et al., 2004; Hauck et al., 2004). However, such corrections implicitly introduce the particle-water-related offset also to observations that should by their design avoid it. As various applications using the PM10 and PM2.5 concentrations as an input (e.g. health impact assessment) are often calibrated using the total PM observations, using the model-produced dry PM masses will introduce a bias to the impact analysis.