Emissions of biogenic volatile organic compounds (BVOCs) were not prescribed by the EDT experiments. Each participating team used their own emission model to calculate biogenic emissions.

One group of models used the Model of Emissions of Gases and Aerosols from Nature (MEGAN) v2.04 (Guenther et al., 2006) and MEGANv2.1 (Guenther et al., 2012) emission models: CHIMERE and MINNI, respectively. CHIMERE uses highly resolved spatiotemporal data (30 arcsec every 8 d) generated from MODIS for leaf area index (LAI) inputs. The 30 arcsec USGS (US Geophysical Survey) land-use database is used to provide information on the plant functional type (PFT). The PFT is then combined with the emission factors for each functional type of Guenther et al. (2012) to compute the landscape average emission factors. MINNI derived them from the CORINE Land Cover (CLC2006) inventory. The MEGAN model is driven with meteorological variables, such as temperature, wind speed, humidity, solar radiation and soil moisture. The leaf area index retrieved from the Terra MODIS satellite is used to simulate the vegetation growth (8 d and 1-month average LAI data at 0.25∘×0.25∘ resolution for CHIMERE and MINNI, respectively). Common BVOCs species such as isoprene, α-pinene as well as other classes of monoterpenes are generated for each hour and grid cell of the domain. In the CHIMERE model, emissions of sesquiterpenes are also included and used as an input for SOA chemistry. More information on the MEGAN emission algorithms can be found in Guenther et al. (2006, 2012).

The second group of models (LOTOS-EUROS, MATCH and EMEP MSC-W) used a detailed tree inventory of 115 species for 30 European countries based on the work of Koeble and Seufert (2001) and aggregated tree species based on land-cover types. For this group of models, the environmental factors to derive biogenic emissions include the light correction factor (γL) and the temperature correction function (γT), which are applied to three types of emission categories: isoprene, pool-dependent monoterpenes and light-dependent monoterpenes based on Guenther et al. (1993). More information on the EMEP MSC-W BVOC emission algorithm can be found in Simpson et al. (2012).

Finally, sea salt, emitted in water droplets from the sea during high wind speed conditions and as a result of breaking of waves and/or bursting of air bubbles, is included in all the models, based on different schemes, as described in Colette et al. (2017). Windblown dust emission were taken into account by all the models except MATCH, while road traffic dust resuspension was only included in the EMEP MSC-W model (Colette et al., 2017).

Anthropogenic gridded emissions by country and activity, i.e., SNAP (Selected Nomenclature for reporting of Air Pollutants) codes, were estimated using the Greenhouse gas – Air Pollution Interactions and Synergies (GAINS) model (Amann et al., 2011). Emission of SOx, NOx, NH3, non-methane volatile organic compounds (NMVOCs) as well as primary PM2.5, PM10, black carbon and primary organic aerosol were prepared at a 0.25∘×0.40∘ resolution (latitude × longitude). Anthropogenic emissions were calculated for the years 1990, 1995, 2000, 2005 and 2010, and linearly interpolated by country and activity sector for the 5-year periods to obtain the continuous 21-year emission dataset. Data for the different emission sectors were obtained from Eurostat (http://ec.europa.eu/eurostat, last access: 14 November 2019), the International Energy Agency (IEA, 2012) and the UN Food and Agriculture Organization (FAO) (http://www.fao.org/statistics/en/, last access: 14 November 2019). Additionally, data from the International Fertilizer Association (IFA) and the COPERT model (Athanasiadis et al., 2009) were used for the agriculture and transportation sectors, respectively. An error in primary particulate emission matter for Russia, north Africa and maritime areas for the period of 1991–1999 was identified at the end of the exercise. However, the effect of the error was estimated to be very limited (Theobald et al., 2019).

The complete anthropogenic emission dataset accounts for source-specific emission limits as well as for various European air quality directives (e.g., the UNECE Gothenburg Protocol; UNECE, 1999). This emission dataset, referred to as ECLIPSE_V5, was delivered by IIASA as country national totals by activity sector. It was subsequently spatialized by INERIS on the EURODELTA-Trends grid for use in the CTMs using the gridding process described in Terrenoire et al. (2015) and Bessagnet et al. (2016). For the residential heating sector (SNAP2), a proxy based on population density was applied using a bottom-up inventory available for France. More information about the regridding can be found in Colette et al. (2017).

To provide meteorological inputs to the modeling teams, dynamically downscaled regional climate model simulations were used in combination with ERA-Interim global reanalysis data (Dee et al., 2011). The Weather Research and Forecasting model (WRF version 3.3.1; Skamarock et al., 2008) was used at a resolution of 0.44∘ to generate the meteorological parameters. To reduce the uncertainty of the meteorological data, WRF was re-run with ERA-Interim reanalysis data in grid-nudging mode as described in Stegehuis et al. (2015) and subsequently interpolated at a 25 km resolution to match the EDT grid, although there were a few differences between the procedures of the modeling team. LOTOS-EUROS used RACMO2-downscaled data and MATCH used HIRLAM-downscaled data. More information on the meteorological inputs can be found in Colette et al. (2017).

The observations are reported to EMEP, and the original time series are available in EBAS (http://ebas.nilu.no, last access: 14 November 2019). The annual observational datasets chosen for the trend assessment have passed the completeness criteria of 75 % of data available over the full 1990–2010 period and have undergone visual screening tests. The secondary dataset with annual and seasonal average concentrations is available from the webpage set up by the Task Force on Measurements and Modeling (TFMM) for this study (https://wiki.met.no/emep/emep-experts/tfmmtrendstations, last access: 14 November 2019). The datasets include yearly measurements of long-term air concentrations of sulfur dioxide (SO2), particle sulfate (SO42-), nitrogen dioxide (NO2) total nitrate (TNO3=HNO3(g)+NO3-(p)) and total ammonium (TNHx=NH3(g)+NH4+(p)) performed in Europe between 1990 and 2010. Overall, the numbers of observational sites available for each of the species are 30, 20, 25, 13 and 16 for SO2, SO42-, NO2, TNO3 and TNHx, respectively. Figure  illustrates the geographical distribution of the observational sites for each of the species, all classified as rural background stations. It can be noted that most of the stations are located over the northern and central parts of the domain, therefore limiting the evaluation of the model results to these specific sites. The complete list of the observational sites is reported in Table S1.

NO2 is mainly sampled with the manual method where NO2 is selectively absorbed on impregnated glass sinters. Some sites do, however, use a chemiluminescence monitor with a molybdenum converter, which is not selective for NO2; thus, these measurements might be biased, and this is especially important in areas with low concentrations (Reed et al., 2016), but it is not assumed that the trends will be largely affected when same method is used during the whole period. The other components are mostly measured using a filterpack sampler with no size cutoff in the inlet. The three-stage filterpack separates gas and aerosol species, but for nitrogen compounds this separation might be biased due to the volatile nature of NH4NO3. Therefore, TNO3 and TNHx are usually used for robust estimate of the atmospheric nitrogen loading (Tørseth et al., 2012). However, it is recommended to report the measurements of all the species since it may give valuable insight into the gas–particle ratio despite possible biases. Details of the method used are found in the annual data report (i.e., EMEP, 2012, for the 2010 data).

The linear trends for each species and observational site were calculated with the Theil–Sen method (Sen, 1968) and their significance was evaluated at 95 % confidence level (p<0.05) using the non-parametric Mann–Kendall test (Kendall, 1948; Mann, 1945). Trends in observational data were compared with trends in modeled data calculated with the same methodology. Since anthropogenic emissions did not decline linearly during the full period covered by the experiment (1990–2010, referred to as PT), and larger emissions reductions are expected during the early 1990s, the trend analysis was performed for two subperiods: the first period between 1990 and 2000, referred to as P1, and the second period between 2000 and 2010, referred as to P2. The linear trends are presented as relative changes with respect to the years 1990 and 2000 for the two 11-year periods and as relative changes with respect to the year 1990 for the full 21-year period. In addition, to provide a more comprehensive picture of the trends in the air pollutant concentrations, the trend analysis was also performed for several subregions adapted from the commonly used the Prediction of Regional scenarios and Uncertainties for Defining EuropeaN Climate change risks and Effects (PRUDENCE) climatic zone classification (http://ensemblesrt3.dmi.dk/quicklook/regions.html, last access: 14 November 2019). The extension of the subregions used in the study is reported in Fig. .

The evaluation of modeled SOA was performed using an extensive dataset of secondary organic aerosol concentrations retrieved with positive matrix factorization (PMF) analysis (Paatero, 1999) and recently compiled by Tsimpidi et al. (2016). This dataset includes SOA average concentrations at various sites in Europe during the P2 period. In order to remove local pollution events, likely not included in emission inventories, stations with average SOA concentrations higher than 7 µg m-3 during the measurement period were excluded from the analysis (three sites excluded – in total, 28 sites were kept: six urban sites, eight urban downwind sites and 14 rural/remote sites). Most of the measurements were performed during short campaigns using aerosol mass spectrometers (AMSs) in different periods of the years, lasting from about 2 weeks to 1 month. The spatial distribution of the stations is presented in Fig. S1. The complete list of stations used is reported in Table S2 along with information regarding the year and the seasons during which measurements were made.

Figure  and Table 3 report the mean relative trends of all the sites included in the analysis (Table S1).

For SO2, the observed relative reductions were 82 %, 47 % and 97 % for the P1, P2 and full periods, respectively (Fig.  and Table 3). The models indicate very similar ranges of SO2 reductions, i.e., 64 %–76 % for the P1 periods, 34 %–54 % for the P2 period and between 84 % and 97 % for the full period, depending on the model (Table 3). This is in line with the emission reduction trends presented in Sect.  and with previous trend studies for Europe (Tørseth et al., 2012). Table 3 also reports the fraction of model estimates within a factor of 2 of the observed trends. Most of the models were able to reproduce the observed SO2 trends within a factor of 2 at most of the individual sites (Fig.  and Table 3) and with model performance being better during the P1 period compared with P2.

Overall, the observations indicate relative reductions of 25 %, 12 % and 36 % in NO2 concentrations for the P1, P2 and PT, respectively, with the models estimating similar ranges of relative reductions, i.e., 19 %–23 % for the P1 period, 22 %–26 % for the P2 period and 44 %–47 % for the full period, depending on the model (Table 3). Only about half of the individual observed trends were reproduced within a factor of 2 by individual models and all models performed worse in the second period (P2), overestimating the observed trends (Fig.  and Table 3). Such behavior could indicate possible difficulties for CTMs in capturing long-term trends at relatively low concentrations with small annual changes, typical of the P2 period.

Figure  shows the percentage of statistically significant/non-significant increasing/decreasing observed and modeled SO2 and NO2 trends.

For SO2, most of the stations had significant decreasing trends in concentrations during the P1 period, with only a small fraction of the stations with non-significant decreasing trends. All the models were able to reproduce this pattern, albeit with a slight overestimation of the significant decreasing fraction. During the P2 periods, most of the models tend to overestimate the number of significant trends, with CHIMERE and LOTOS-EUROS being closer to the fraction of significant/non-significant decreasing trends indicated by the observations. For the full period (PT), the agreement between the modeled and observed fractions of significant/non-significant increasing/decreasing appears very good, mainly because of the larger number of data points in the time series, and with all the sites indicating significant observed and modeled decreasing trends.

For NO2, the models were able to reproduce the observed fraction of significant/non-significant increasing/decreasing trends in the P1 period, with most of the models indicating a significant decrease in NO2 concentrations at most of the stations (slightly lower for EMEP MSC-W). The analysis for the P2 period shows that the reduced fraction of observed significant decreasing trends compared with the P1 period was only partially reproduced by the models, all of them tending to overestimate the fraction of significantly decreasing trends (Fig. ). Again, a possible explanation for the degraded model performance during the P2 period could be related to the relatively low pollutant concentrations, which might be challenging to model at such coarse resolution, as well as to uncertainties in the measurement data (see Sect. ). One site in Ireland (IE0001R) was the only site with significant increasing observed trends during the P1 and PT periods, a result which was not reproduced by any of the models (the significant increase in the LOTOS-EUROS model is for the SE0014R station located in Sweden).

An additional trend analysis was performed using a generalized least squares (GLS) fit model that accounts for the temporal autocorrelation of the data. The results from the GLS model were in line with the one predicted by the Theil–Sen and the Mann–Kendall methods (Fig. S2 and Table S3) and with the GLS model showing a slightly higher fraction of non-significant decreasing trends during the P1, P2 and PT periods.

Model performance for SO2 and NO2 was additionally evaluated by calculating the mean fractional bias (MFB) and mean fractional error (MFE) for both the P1 and P2 periods separately (Appendix A). Recommended model performance criteria (MFB ≤ ±60 %, MFE ≤ +75 %) as well as the performance goal (MFB ≤ ±30 %, MFE ≤ +50 %) proposed by Boylan and Russell (2006) were achieved in both periods by most of the models apart for SO2 for MINNI during the P1 and P2 periods and for MATCH during the P2 period, where only the model performance criteria were achieved (Fig. S4). In addition, the evolution of the MFB over the full 1990–2010 period does not indicate any substantial change in 2010 compared to the first year of the exercise (i.e., 1990) with the exception of SO2, which tends to be slightly more positively biased in the latter part of the period compared to 1990 (apart from EMEP MSC-W; Fig. S5).

Figure  shows the mean modeled and observed relative trends in SO42-, TNHx and TNO3 for all the sites included in the analysis (Table S1). Consistent with the gas-phase analysis, the trends are reported for the two subperiods, i.e., 1990–2000 (P1) and 2000–2010 (P2), as well as for the full period (PT).

Overall, the observations indicated that concentrations of SO42- declined by 57 %, 14 % and 66 % for the P1, P2 and PT periods (mean of all the stations; Fig.  and Table 4), with the models indicating relative reductions of 42 %–54 % for the P1 periods, 23 %–35 % for the P2 period and 61 %–78 % for the full period, depending on the models. The reductions in SO42- concentrations were larger during P1 than during the P2 period and most of the model estimates were within a factor of 2 of the observed values for all the periods (Fig.  and Table 4). Two sites, one in Ireland (IE0001R) and one in Poland (PL0003R), showed an increase in SO42- concentrations (Fig. ), which none of the models were able to reproduce.

The percentage of statistically significant/non-significant increasing/decreasing trends in the observed and modeled SO42- trends is reported in Fig. , showing good agreement between the observed and modeled significance (and their direction) for the P1 and PT periods, whereas all the models tend to overpredict the number of statistically significant increasing trends during the P2 period. Statistically significant increasing trends in SO42- concentrations were only observed at the PL0003R site in Poland during the P2 period (Fig. ), a result which none of the models were able to reproduce. Interestingly, observed SO42- concentrations declined less than those of SO2 (Table 3), a behavior also reproduced by all the models. The non-linear dependencies between the reduction in SO2 and SO42- concentrations are influenced by different factors. First, the strong reduction in SOx emissions will increase the availability of OH radicals, which will directly enhance the homogeneous reaction rate of SO2. Second, all the models account for the dependence of the aqueous chemistry of SO2 on pH levels. Thus, heterogeneous reactions of SO2 are also expected to proceed more efficiently due to the increase of pH levels over time.

Observations of TNHx reveal that concentrations declined by 28 %, 22 % and 46 % for P1, P2 and PT, respectively (Table 4). In general, most of the models underpredict the relative changes; the modeled relative reductions for the P1, P2 and PT periods were 15 %–26 %, 14 %–21 % and 27 %–38 %, respectively, with the P1 period showing only a minor fraction of the data points within a factor of 2 (Fig.  and Table 4). Indeed, large uncertainties remain in terms of ammonia emissions, which might affect model performance for TNHx. Moreover, we would like to underline that none of the participating models accounted for the influence of meteorology (e.g., temperature) on ammonia emissions and relied on static emission profiles provided by the EURODELTA exercise. Recent studies, however, have shown that better agreement in terms of the modeled ammonia concentrations can be achieved when ammonia emissions are modulated with local meteorological conditions (Backes et al., 2016; Hendriks et al., 2016). Compared with the other investigated species, a larger variation in terms of the significance of the trends can be seen in Fig. , with most of the models tending to overestimate the fraction of significant decreasing trends during the P2 period. Statistically significant increasing trends in observed TNHx concentrations were found at one station in Norway (NO0039R) for the full period (PT), with none of the models being able to reproduce this feature.

TNO3 concentrations, on the other hand, declined to a lesser extent than those of SO42- and TNHx. For all periods, the observed relative changes in TNO3 concentration were 16 %, 19 % and 29 %, for the P1, P2 and PT periods, respectively, with the models estimating similar ranges for the P1 and P2 periods, i.e., 16 %–19 %, 8 %–27 % and 28 %–39 %, for the P1, P2 and PT periods, respectively (Fig.  and Table 4). Most of the model estimates were more than a factor of 2 larger than the observed values for the P1 and PT periods (Fig.  and Table 4). The percentage of statistically significant/non-significant increasing/decreasing observed and modeled TNO3 trends revealed that most of the models were able to reproduce the large fraction of non-significant decreasing observed trends. The EMEP and MATCH models estimate a larger fraction of significant decreasing trends than the other models in both the P1 and P2 periods, where CHIMERE and MINNI show the largest fraction of non-significant decreases. CHIMERE also shows the largest fraction of non-significant increasing trends during the P2 period (Fig. ).

As for the SO2 and NO2 gas-phase species, an additional trend analysis was performed using a GLS fit model. For SO42, TNO3 and TNHx results were also in line with the one predicted by the Theil–Sen and Mann–Kendall methods (Fig. S3 and Table S3). For SO42- and TNO3, the GLS model showed higher fraction of non-significant increasing trends during the P2 period in the observation data, whereas for TNHx an higher fraction of non-significant decreasing trends is predicted compared to the Theil–Sen and Mann–Kendall methods during the P1 period.

Model performance for SO42-, TNO3 and TNHx was also satisfactory; the recommended model performance criteria (MFB ≤ ±60 %, MFE ≤ +75 %) as well as the performance goal (MFB ≤ ±30 %, MFE ≤ +50 %) proposed by Boylan and Russell (2006) were achieved in both the P1 and P2 periods by most of the models apart from TNO3 in CHIMERE during the P1 and P2 periods and in EMEP MSC-W during the P1 period, where only the model performance criteria were achieved (Fig. S4). In addition, the evolution of the MFB over the full 1990–2010 period does not indicate any substantial change in 2010 compared to the first year of the exercise (i.e., 1990) apart from TNHx which tends to be slightly more positively biased in the latter part of the period compared to 1990 (Fig. S5).

In order to further investigate the trends in TNO3 concentrations described in the previous paragraph, we also investigated the modeled trends in HNO3 and NO3- concentrations (for the different subregions in Fig. ).

Figure  illustrates the relative trends in HNO3 and NO3- (sum of the coarse and fine particle fractions) for the P1 (first two columns in Fig. ) and P2 (last two columns in Fig. ) periods, for all the models that participated in the experiment. In general, during the P1 period, the models indicate larger significant decreases in HNO3 compared with NO3-, especially over the Fennoscandia and central European regions. A few differences in the spatial distribution of the modeled HNO3 and NO3- trends can be seen in Fig. . For instance, for the LOTOS-EUROS model, the significant relative trends in both HNO3 and NO3- concentrations are more comparable during the P1 period, whereas MINNI and CHIMERE estimate larger areas of non-significant trends in NO3- concentrations, with some significant increases in a few parts of the domain. Figure  and Table 5 show the models' average relative changes and standard deviation (over land) of HNO3 and NO3- for all the PRUDENCE zones for the P1 and P2 periods. During the P1 period, the regions classified as Alps (AL), British Isles (BI), Benelux area (BX), France (FR) and mid-Europe (ME) had the largest decrease in HNO3 concentrations, with average decreases between 36 % and 44 % (Table 5). A comparison of HNO3 and NO3- relative trends for the same regions shows that NO3- concentrations declined to a lesser extent, i.e., around 20 %, which is roughly half of the modeled relative reduction in HNO3 concentrations. The largest difference between the relative reduction in HNO3 and NO3- concentrations occurred over the Scandinavian regions for the P1 period, of 24 % and 5 %, respectively. On the other hand, the reduction of HNO3 and NO3- concentrations were comparable in the eastern European region and over the Iberian Peninsula (Table 5), as well as during the P2 periods.

The non-linear response of HNO3 concentrations and NO3- concentrations, i.e., larger relative reduction in HNO3 compared to NO3-, could be attributed to the shift in the thermodynamic equilibrium of HNO3 versus particle nitrate NO3-. In fact, for specific regions and especially during the P1 period, the large reduction in SO2 emissions increased the availability of “free ammonia” and thus the transfer of more HNO3 into the particle phase, favoring the formation of ammonium nitrate. This also increases the TNO3 lifetime as dry deposition is much more rapid for HNO3 than for NO3-. This effect could contribute to the reduced TNO3 decreases with respect to the NO2 decrease. Figure  shows the modeled HNO3/NO3- and NH3/NH4+ molar ratio for the P1 and P2 periods, as predicted by all the models. In general, the models indicate significant decreasing trends in the HNO3/NO3- ratio especially over Scandinavian regions (to a lesser extent in the LOTOS-EUROS model) and a strong increase in the NH3/NH4+ ratio over the whole domain except for some eastern European areas.

In this section, the trends in modeled SO42- and SO2 for the subregions reported in Fig. 1 are discussed. Figure  illustrates the relative trends in SO2 and SO42- concentrations for the P1 (first two columns) and P2 (last two columns) periods for all the models that participated in the experiment. Trends were predicted to be statistically significant over the whole domain during the P1 period and to a lesser extent over eastern and northern European regions during the P2 period. A larger decline in SO2 concentrations during the P1 period was predicted in the mid-Europe areas, around 85 % relative reductions, compared to the Iberian Peninsula, Mediterranean and Fennoscandia areas (average reduction 24 %, 48 % and 33 %, respectively) (Fig.  and Table 6). During the P2 period, the modeled relative reductions of SO2 were in the range 33 %–65 %, with the Iberian Peninsula showing a larger reduction compared to the P1 period. As already discussed in the evaluation section, SO42- declined to a lesser degree than SO2, by 34 % and 59 % in the P1 period and 30 %–49 % in the P2 period, likely because of the increased availability of oxidant species and pH-dependent cloud chemistry.

During the P2 period, various field campaigns were performed in Europe using AMS instruments to measure ambient OA concentrations at various sites. Using factor-analysis techniques, i.e., PMF, it was possible to apportion the measured OA to a direct emitted organic factor, i.e., POA, and a more oxidized secondary factor, referred to as SOA. Even though this methodology is affected by various sources of uncertainties, we used this dataset to provide a general benchmark for modeled SOA performance during the P2 period.

Figure  shows the average modeled and observed SOA concentrations (retrieved from PMF analysis) for all the sites included in the analysis (Table S2 and Fig. S1), as well as for campaigns carried out during winter and summer periods (astronomical seasons). All the models underestimate observed SOA concentrations, in general, with larger variabilities between models than for the secondary inorganic species (Sect. ). On average, the models underestimated the SOA concentrations by about a factor of 3 to 11 (Table 7, Appendix A), depending on the specific model (MATCH underestimating the most and EMEP MSC-W the least) and with a larger underestimation during winter periods. The EMEP MSC-W model, which accounts for aging of SOA, was closer to the observations, with average SOA concentrations of about 0.7 µg m-3. The higher SOA mass modeled by the EMEP MSC-W model could be explained by the shift of relatively high-volatility organic compounds toward lower-volatility ranges when aging processes are accounted for (with a reaction rate toward OH of 4.0×10-12 cm3 molecule-1 s-1 in the case of the EMEP MSC-W model for both anthropogenic secondary organic aerosol (ASOA) and biogenic secondary organic aerosol (BSOA)). Such processes will increase the SOA mass since low-volatility oxidation products will rapidly condense into the particle phase. Interestingly, the MATCH model which used the same VBS scheme as the EMEP MSC-W model but without considering SOA aging processes tends to underestimate SOA concentrations substantially (Fig. ). This indicates the importance of these chemical mechanisms in CTMs and their impact on SOA formation. On the other hand, the models based on the two-product scheme and molecular surrogate approach scheme, i.e., MINNI and CHIMERE, respectively, yielded very similar results for the total SOA mass, with SOA concentrations ranging in between the two VBS models (i.e., around 0.4 and 0.6 µg m-3, averaged over all sites).

Figures S6 and S7 illustrate the modeled and observed (retrieved with PMF analysis) SOA concentrations at the individual sites for winter and summer periods. In general, the models had difficulties in reproducing the SOA concentrations at specific urban sites, such as Paris and Manchester (in both summer and winter periods), and in reproducing high levels of SOA concentrations at a few specific sites, e.g., Payerne, where emissions from biomass burning are high. This could be due to missing aerosol precursors (SVOC emissions) in the resident-heating sectors, which have been shown to have high uncertainties (Denier van der Gon et al., 2015).

Figure  shows the modeled relative and absolute contributions of anthropogenic and biogenic secondary organic aerosols to SOA concentrations. For most of the models, larger contributions of ASOA to SOA were estimated during winter periods and/or in urban areas (e.g., in Paris, Manchester and Payerne), whereas the BSOA contribution to SOA was largest during warmer periods. Especially in summer, large emissions of biogenic volatile organic compounds can act as an important source of SOA. The CHIMERE model simulated the largest contribution of BSOA, with only minor variations between the stations and periods.

In this section, the trends in BVOCs emissions, i.e., isoprene and terpenes, are presented together with the trends in BSOA and ASOA concentrations. The trend analysis for BSOA and ASOA is reported for the full 1990–2010 period, for the different PRUDENCE zones and with the methodology as defined in Sect. 2.3.4. Note that not all the participants provided biogenic emissions for the full 21-year period, and only the EMEP MSC-W, CHIMERE and MATCH models provided year-by-year emissions of isoprene and monoterpenes for the EDT experiments. Moreover, for the EDT setup, CHIMERE does not include biogenic emissions for latitudes north of 65∘ N.

Figures 16 and 17 illustrate the trends in isoprene and monoterpenes (first two columns) for the full 1990–2010 period. An increase in both isoprene and monoterpene species was found especially over eastern Europe (EE), with relative increases of a 15 %–27 % for isoprene and 14 %–18 % for monoterpenes, and over the Fennoscandia (SC) regions, with relative increases of 12 %–24 % in isoprene emissions and 7 %–17 % in monoterpene emissions, depending on the model (Figs. 16 and 17). Interestingly, the increase in biogenic emissions was predicted by all biogenic models (i.e., MEGANv2.1 and the one using vegetation data from Koeble and Seufert, 2001). These increases were mainly attributed to the increase in surface temperature used to drive the different biogenic models (Fig. S8), which however were found to be not significant. The increase in surface temperature was found to be larger over Fennoscandia, Mediterranean and eastern European areas compared to the remaining zones (i.e., increases around 0.02, 0.02 and 0.03 K yr-1, respectively). We want to underline that the surface solar radiation (SSR) could also play an important role for the emission of biogenic precursors (especially for isoprene). The strong reduction in SO42- concentrations described in Sect. , especially in the eastern regions of the domain, might have induced an increase in the incoming solar radiation, referred to as brightening periods (Wild, 2009), which could affect the emission of biogenic species. However, none of the models that participated in the exercise have explicitly accounted for such an interaction and the ERA-Interim forcing data rely mainly on climatological aerosol profiles. On the other hand, recent sensitivity studies performed in Europe showed that such effects might be relatively small (Oikonomakis et al., 2018). Figures 16 and 17 also show the relative trends in BSOA and ASOA concentrations for all the models that were able to provide 21 years of data. Even though some models indicated few increases in the biogenic SOA concentrations over the Fennoscandia regions, these increases were found to be smaller than the increase in biogenic emissions, and in some cases, biogenic SOA concentrations were also estimated to have declined (Figs. 16 and 17). This might sound counterintuitive, since one would expect more biogenic SOA to be produced as more biogenic precursors are available and, in general, the increased availability of OH radicals due to the reduction in NOx and SOx emissions, thus increasing the oxidation efficiency of biogenic SOA precursors, especially for isoprene. A possible explanation for this non-linear relation between the trends in biogenic emissions and the trends in BSOA concentrations could be due to the trends in the anthropogenic OA concentrations. As shown in Figs. 16 and 17, a strong decrease in the ASOA concentrations was found for the entire 1990–2010 period, in line with the reduction in the NMVOC precursors described in Sect.  (mainly from transportation sectors). Modeled ASOA concentrations indicate a decline of about 60 %–70 % over the whole domain (considering only land areas). This might have had on effect on the formation of the BSOA fraction; in fact, the strong reduction in ASOA concentrations, and other aerosol organic and inorganic components, will reduce the availability of organic and inorganic material onto which the low-volatility oxidized compounds can condense, directly affecting the formation of the BSOA fraction. Additionally, oxidant levels and thus oxidation pathways could have changed over time, affecting as said before OH but also NO3 concentrations.