We used the offline (i.e., the meteorology is prescribed) regional air quality model, CAMx (Comprehensive Air Quality Model with Extensions; http://www.camx.com, last access: 2 July 2018) version 6.30. We modeled the summer season (June, July, August – JJA) in 2010 plus the last 2 weeks of May which were used as spin-up time. The model domain had a horizontal resolution of 0.250∘ × 0.125∘ and covered all of Europe from 15∘ W to 35∘ E and 35∘ N to 70∘ N. The vertical extension was up to 460 hPa using 14 sigma layers. The thickness of the first layer was ∼ 20 m but its modeled values corresponded to ∼ 10 m, as the concentrations are calculated at the midpoint of each layer. We used the CB6r2 (Carbon Bond mechanism, version 6, revision 2; Hildebrandt Ruiz and Yarwood, 2013) gas-phase mechanism, and we simulated the PM concentrations using a static two-mode (fine/coarse) scheme for the aerosol size distribution. For the inorganic thermodynamics and gas–aerosol partitioning calculations, the ISORROPIA scheme (Nenes et al., 1998, 1999) was used, while for the calculations of the organic aerosol concentrations we used the SOAP model (Strader et al., 1999). The dry deposition was calculated according to the scheme of Zhang et al. (2003). The MOZART (Model of Ozone and Related Chemical Tracers) global model data for 2010 (Horowitz et al., 2003) served as initial and boundary conditions for the chemical species. The MOZART data had a time resolution of 6 h and were interpolated to the size and resolution of our grid using the CAMx preprocessor MOZART2CAMx (Ramboll Environ, 2016). The full-science tropospheric ultraviolet and visible (TUV) radiation model (NCAR, 2011) is used as a preprocessor to provide CAMx with clear-sky photolysis rates, where a climatological aerosol profile determined by Elterman (1968) is used. Then, these rates are internally adjusted in CAMx every hour for clouds and aerosols as well as for pressure and temperature using a fast in-line version of TUV (Emery et al., 2010; Ramboll Environ, 2016). The internal adjustment for clouds and aerosols inside CAMx is performed in two steps. First, the clear-sky radiative transfer calculations with in-line TUV are repeated inside CAMx. In the second step, the radiative transfer calculations are repeated including the impact of clouds and aerosols (simulated by CAMx). A ratio of cloudy- (and aerosols) to clear-sky solar radiation is derived by the aforementioned two-step radiative transfer calculations in CAMx. This ratio is then applied to the clear-sky photolysis rates and SSR which were calculated by the full-science TUV preprocessor at the beginning. This internal adjustment (i.e., in-line TUV) is carried out only for a single representative wavelength (350 nm), as tests against the full-science TUV indicated a difference smaller than 1 % in the ratio of cloudy- to clear-sky solar actinic flux for a variety of cloudy conditions (Emery et al., 2010). Inside CAMx, the COD is calculated for each model grid cell based on the approach of Genio et al. (1996) and Voulgarakis et al. (2009), while the dry extinction efficiency of the aerosol species, which is needed for the calculation of the AOD, as well as the single-scattering albedo (SSA) were provided by Takemura et al. (2002) for the wavelength of 350 nm (Table S1 in the Supplement). These values of aerosol optical properties were provided for sulfate, organics, soot, total dust and sea salt, and the sulfate values were extended to nitrate and ammonium (Ramboll Environ, 2016). The asymmetry factor for aerosols was set to have a default value of 0.61 regardless of their composition. For clouds, the default values of the asymmetry factor and SSA were 0.85 and 0.99, respectively. In addition, the eight-stream discrete ordinates scheme was used for the radiative transfer calculations compared to the more common (and computationally faster) two-stream delta-Eddington approximation scheme, as the calculations' accuracy increases with the number of streams (Stamnes et al., 1988; Toon et al., 1989). The choice of eight streams has been suggested to offer high accuracy (1 % or better compared to 32 streams) without having a significantly higher computation cost (Petropavlovskikh, 1995). TOMS (Total Ozone Mapping Spectrometer) data, which were provided by NASA (National Aeronautics and Space Administration; ftp://toms.gsfc.nasa.gov/pub/omi/data/, last access: 2 July 2018), were used as input for total ozone column in both TUV and CAMx. In addition, the radiative transfer algorithms of both full-science TUV and CAMx (i.e., in-line TUV) were modified to extract the modeled AOD and SSR data. In other words, both the SSR (used in the photolysis rate calculation) and the photolysis rates were calculated according to the same parameterization that was described above.

The required meteorological input for CAMx was generated by the WRF-ARW (Advanced Research Weather Research and Forecasting model, version 3.7.1; Skamarock et al., 2008). Reanalysis global data, with time resolution of 6 h and horizontal resolution of 0.72∘ × 0.72∘, were provided by ECMWF (European Centre for Medium-Range Weather Forecasts) and served as initial and boundary conditions for WRF. Both CAMx and WRF had the same model domain and horizontal resolution. However, for the WRF runs, 31 vertical layers, up to 100 hPa were used instead of 14, which was the case for the CAMx runs for computational efficiency. More details about the WRF parameterization are provided in Oikonomakis et al. (2018).

For the anthropogenic emissions, we used the TNO-MACC-III emission inventory for 2010. This inventory was provided by the Netherlands Organization for Applied Scientific Research (TNO) and is an extension of the TNO-MACC-II emission inventory (Kuenen et al., 2014). More details about the TNO-MACC-III emission inventory are given in Kuik et al. (2016). The TNO European emission domain is the same as our domain but with a finer horizontal resolution (0.125∘ × 0.0625∘). The mineral dust, sea salt and wildfire emissions are not included in the inventory. However, in the model's initial and boundary conditions, the concentrations of mineral dust and sea salt are included. For the calculation of the biogenic emissions (isoprene, monoterpenes, sesquiterpenes and soil NO), we followed the methods described by Andreani-Aksoyoglu and Keller (1995) using temperature and SSR data from the WRF output (the SSR data from WRF were not used in any calculation in CAMx) as well as land use data from the GlobCover 2005–2006 land use inventory (http://due.esrin.esa.int/page_globcover.php, last access: 2 July 2018) and the United States Geological Survey (USGS). All emissions were injected in the first model layer and were treated as area emissions. A detailed discussion and values of the emissions used in this study are given in Oikonomakis et al. (2018).

The European Air Quality Database v7 (AirBase; Mol and de Leeuw, 2005) provided observational surface data for the air pollutant concentrations (http://acm.eionet.europa.eu/databases/, last access: 2 July 2018) with an hourly time resolution, which were used for chemical model evaluation. For a better comparison between the model and the observations, we used only rural background stations due to our grid resolution. Furthermore, we evaluated the daily mean of the chemical species in order to be able to compare our results with other studies (e.g., Bessagnet et al., 2016). More details about the observational data treatment and the statistical methods are described in the model evaluation part of Oikonomakis et al. (2018). Furthermore, PM10 (particles with an aerodynamic diameter, d<10 µm) and PM2.5 (d<2.5 µm) data from the AirBase database as well as from the Swiss National Air Pollution Monitoring Network (NABEL; Empa, 2010) were used for trend analysis. Switzerland and the Netherlands were chosen for the PM trend analysis as they have PM10 data going back to 1990 and 1992, respectively. For Switzerland, the PM10 data until 1997 are actually corrected total suspended particle (TSP) data (Empa, 2010), but they are suitable for PM10 trend analysis (Barmpadimos et al., 2011). Hourly high-quality SSR data from the Baseline Surface Radiation Network (BSRN; König-Langlo et al., 2013) for seven stations were used for model evaluation. An overview of the seven BSRN stations is given in Table S2. Finally, AOD data were retrieved by the Aerosol Robotic Network (AERONET), which is a network of ground-based Sun photometer measurements of aerosol optical properties (Holben et al., 1998; O'Neill et al., 2003). We used level 2.0 (quality assured) data for the 340 nm wavelength band to compare with the respective modeled AOD values. The calibration error of the AOD measurements is of the order of 0.015 (Holben et al., 1998; Eck et al., 1999). Since the temporal resolution of the AOD measurements is not constant (e.g., at specific hours), the calculated daily mean does not correspond to a 24 h time interval but to intra-day time intervals with available measurements. The daily average of the modeled AOD was calculated using only the times of available AOD measurements for each site, for a more consistent comparison between model and observations.

The description of 12 model runs is shown in Table 1. We used two base case scenarios: one with the default parameterization (BASE) and a second one with increased NOx emissions (BASE_NOx) which produced higher ozone concentrations, in order to incorporate any potential underestimation of the ARI effects on ozone due to underestimated modeled ozone production as suggested by Oikonomakis et al. (2018). All sensitivity tests were performed using both base case scenarios (see Table 1).

The impact of solar radiation changes due to the ARI on ozone chemistry was investigated via two pathways: (i) via impact on photolysis rates and (ii) via impact on biogenic volatile organic compound (BVOC) emissions. In order to quantify these impacts, we first simulated the summer of 2010, then applied sensitivity tests that would represent the radiation conditions in the summer of 1990 (i.e., different solar radiation due to ARI) and finally compared the two cases. In other words, we used the same meteorology and emissions for both cases (except for the BVOC emission sensitivity tests where we used different BVOC emissions) and we designed special sensitivity tests to isolate and quantify the effect of changes in the ARI between those years on ozone concentrations. Finally, it is noted that the chemistry simulated by CAMx (for any scenario) does not affect the meteorology, as it is prescribed (see Sect. 2.1), and hence the impact of ARI on atmospheric dynamics and other meteorological related effects (e.g., vertical mixing, dry deposition, Xing et al., 2017) are excluded in this study.

The model performance evaluation for both WRF and CAMx models was carried out and discussed in detail in Oikonomakis et al. (2018). A summary of the statistical metrics and model performance evaluation is given in Tables 3 and 4, respectively, for the daily mean O3, PM2.5 and PM10 (see also Fig. S1). The model performance for O3 and PM2.5 was satisfactory, as discussed in detail by Oikonomakis et al. (2018). On the other hand, there was a consistent underestimation of PM10 with a mean bias (MB) of -6 µg m-3 and normalized mean bias (NMB) of -33 %. However, the correlation coefficient for the PM10 is 0.5, suggesting that the model can capture the observed PM10 temporal evolution (Fig. S1). Also, since the model performance for PM2.5 is better, this implies that the discrepancy in the PM10 is more likely due to missing emissions in the coarse mode such as sea salt, mineral dust and wildfires (see Sect. 2.1). Even with the inclusion of such emissions, models still have difficulties simulating the PM10 concentrations accurately, as the uncertainties related to these emissions are large and meteorological uncertainties (e.g., in wind speed, vertical mixing) also play an important role (Karamchandani et al., 2017; Solazzo et al., 2017).

The systematic model underestimation of the PM10 concentrations is also evident in the AOD (Table 4), where the model consistently underestimates the AERONET observations (MB of -0.15, MGE of 0.16). Despite this systematic negative bias, the model is able to represent quite accurately the spatial and temporal variability of the observed AOD, indicated by the relatively high correlation (r=0.6) between the model and the observations which is shown in more detail in Fig. S1. Other possible error sources for the modeled AOD could be (i) the simplified treatment of the aerosol size distribution, the optical properties and the mixing state (Curci et al., 2015); (ii) the use of the constant climatological aerosol Elterman (1968) profile for the upper troposphere and stratosphere; (iii) uncertainties in RH and f(RH) for inorganic aerosols; and/or (iv) uncertainties due to grid resolution (horizontal or vertical). Overall, our model AOD discrepancies are within range with other modeling studies (Cesnulyte et al., 2014; Im et al., 2015), where they underline the importance of dust and sea salt treatment in the models.

In the case of SSR, the model performance is better, with a slight overestimation (NMB of 6 %; Table 4). The diurnal and inter-daily variability was captured as well (Figs. 2 and S1; r=0.8). In general, the overestimation of the downward shortwave radiation is a long-standing issue in the models (Wild, 2008; Wild et al., 2013), which indicates that it might be related not only to aerosols but also to other important sources of uncertainty such as parameters related to clouds and water vapor. Since the modeling framework of this study is based on the PM2.5, we believe that the systematic PM10 model bias would not affect the results and conclusions significantly.

The modeled daytime (10:00–18:00 LMT) concentrations of the fine PM species to be adjusted for AOD and SSR calculations are shown in Fig. 3. Sulfate concentrations were predicted to be the highest in summer among all seven species (Fig. 3a) especially over the Mediterranean Sea and southeastern Europe. Although ship emissions are considered to be the main source of elevated sulfate concentrations over the sea, their contribution to the land areas in Southeastern Europe (e.g., Greece and Turkey) is much smaller compared to other emission sources, such as power generation, industries and road transport (Tagaris et al., 2015; Aksoyoglu et al., 2016). Particulate nitrate concentrations, on the other hand, are higher in regions with high NOx and NH3 emissions (around the English Channel, Benelux region, northern Italy). The concentrations of anthropogenic SOA (Fig. 3d) are very low, and the spatial distribution of primary species POA, EC and FPRM is similar to their emission patterns (Fig. 3e–g). The high POA concentrations on the eastern boundary of the model domain are consistent with the summer 2010 Russian wildfires, which influenced mainly the areas around Moscow and to a lesser extent the eastern part of Europe (Mei et al., 2011; Portin et al., 2012; Péré et al., 2015). It is noted that, although wildfire emissions are not included in the model (see Sect. 2.1), they enter the model domain from the model boundaries.

The simulated (in the base case) ground-level photolysis rate of NO2, J(NO2), consistently follows the south-to-north latitudinal gradient of SSR and temperature, as shown in Fig. 6a. The modeled continental mean absolute (relative) differences in ground-level J(NO2) between the base case (BASE) and PHOT1, PHOT2 and PHOT3 tests are 0.7 (3 %), 1.3 (6 %) and 0.9 (4 %) h-1, respectively (Fig. 6b–d). The spatial distribution and relative changes are the same for the ground-level photolysis rate of O3, J(O3 → O1D), with changes in absolute terms being 0.0015, 0.0029 and 0.0020 h-1, respectively, for PHOT1, PHOT2 and PHOT3 tests (Fig. S3). As discussed in Sect. 1, changes in the photolysis rates will affect the chemical production and destruction of ozone as well as other chemical processes in the troposphere such as the secondary aerosol (SA) formation, which in turn can affect the photolysis rates. This implication, however, is rather small with the change in SA concentration (continental grid mean) between base case (BASE) and PHOT1, PHOT2 scenarios being 0.01 and 0.02 µg m-3, respectively (Figs. S4–S5) or in relative terms 0.3 and 0.6 %, respectively (the respective results for the PHOT3 test are very similar to the ones of the PHOT1 test and are therefore not shown). We conclude that these changes in SA have negligible impact on the photolysis rates. However, the changes in SA might not be negligible if the impact of ARI on meteorology and subsequent effects on chemistry are also taken into account (which is not the case for this study).

The enhancement of the photolysis rates leads to higher ozone formation especially in regions where there are significant ozone precursor emissions (i.e., central Europe, northern Italy; Fig. 7). The magnitude of the effect of enhanced photolysis rates is, however, rather small on average. The difference in the surface ozone between BASE and PHOT2 scenarios during the daytime (10:00–18:00 LMT) varies between 0.4 and 0.7 ppb (1–1.5 %) over central Europe and up to 1.4 ppb (2.5 %) in the Po Valley, while it is smaller for the other two scenarios (up to 0.7–0.8 ppb, 0.7–1.4 %). However, on an hourly resolution, the largest difference in surface ozone between BASE and PHOT2 scenarios can go up to 8 ppb (16 %) and up to 4 ppb (10 %) for the PHOT1 and PHOT3 scenarios in high-NOx areas on land (Fig. S6).

We repeated similar tests based on a second base case (BASE_NOx) with increased NOx emissions which improved the model performance for ozone production as discussed in Oikonomakis et al. (2018). In the BASE_NOx case, ozone production was higher over a larger area compared to the BASE (see Figs. 7a and 8a). Consequently, the difference in ozone between BASE_NOx and the PHOT1_NOx, PHOT2_NOx and PHOT3_NOx scenarios was more pronounced over a larger area; the magnitude of the impact, however, only slightly increased (Figs. 8 and S6).

We also investigated the impact of ARI changes on daily maximum ozone, but it was higher only by up to ∼ 0.1 ppb (not shown) compared to the daytime (10:00-18:00 LMT) average. Therefore, the ARI did not have a significantly higher impact on daily maximum ozone. The reason is that the daily maximum ozone occurs at a different time (mid-afternoon) than the times the maximum ARI occurs (morning and evening), as also shown in other studies (Xing et al., 2015a, 2017).

The response of isoprene emissions (2.5–3 % changes) to SSR changes (3 %) is nearly linear (Fig. 9), in line with the literature (Guenther et al., 2006). On the contrary, terpene (monoterpene and sesquiterpene) emissions are less sensitive to SSR with changes up to 0.7 % (Fig. 9). Nevertheless, the BVOC emissions' sensitivity to solar radiation can vary depending on the model parameterization of physical processes such as the emission dependence on light and the canopy calculations of diffuse and direct radiation as well as the relative contribution between shaded and sunlit leaves over multiple leaf area index (LAI) layers (Messina et al., 2016). In general, BVOC emission estimates have high uncertainties (a factor of 2–3) due to uncertainties in the land use, LAI and parameterization of physical processes, the large number of compounds and biological sources, and the lack of observations (Guenther et al., 2006; Karl et al., 2009; Guenther, 2013; Oderbolz et al., 2013). Despite these uncertainties, Stavrakou et al. (2014) also reported a linear response of isoprene emissions with the respective SSR changes in Asia between 1979 and 2012, using a different biogenic emission model. On the other hand, other studies suggested that the photosynthetically active radiation (PAR), which depends more on the diffuse component of solar radiation, did not have a significant impact on the increasing BVOC trends in Europe during the solar brightening (after 1980) probably due to the diffuse to direct radiation ratio decrease, compensating for the total increase in SSR (Mercado et al., 2009; Yue et al., 2015). In fact, during the solar dimming (i.e., when the total SSR decreased) between 1960 and 1980, both the diffuse fraction of PAR and the photosynthesis were enhanced (Mercado et al., 2009). It is further suggested that the BVOC emissions are less sensitive to the SSR compared to the temperature, which is identified as a more important driver for the BVOC emission trends (Guenther et al., 2006; Lathière et al., 2006; Yue et al., 2015; Gustafson et al., 2017).

The impact of a 2.5–3 and 0.7 % increase in isoprene and terpene emissions (BIO scenario), respectively, on daytime (10:00–18:00 LMT) average surface ozone is rather small (up to 0.08 ppb, ∼ 0.2 %; Fig. 10a) and an order of magnitude smaller than the respective ozone impact via photolysis rates (see Fig. 8). Both the daytime average and largest hourly (∼ 1 ppb) impacts are higher in central Europe where both BVOC and NOx emissions are ample (Figs. 10a and S7). The effects of increased BVOC emissions are higher in magnitude (up to 0.11 ppb, ∼ 0.3 %) and spatial coverage when applied to the base case with higher NOx emissions (i.e., BASE_NOx–BIO_NOx), as shown in Figs. 10b and S7. The combined effects via BVOC emissions and photolysis rates (COMBO and COMBO_NOx scenarios) on surface ozone appear to be roughly additive, with the photolysis rates effects dominating the overall impact (daytime average difference was up to 0.8 ppb, 1.5 %; Figs. 10c–d and S7). Overall, the direct effects of SSR changes on the BVOC emissions (with the assumptions and parameterizations of this study) were small, and as a result this was also the case for the consequent impact on surface ozone. However, SSR trend implications related to temperature and CO2 changes (Wild et al., 2007; Storelvmo et al., 2016) might have a more significant impact on BVOC emissions and thus on surface ozone, but this was beyond the scope of this study.

Although the effects of ARI changes via photolysis rates and BVOC emissions on surface ozone seem to be small compared to the total ozone concentrations, it might be more meaningful to compare with the magnitude of the observed ozone concentration trends. Wilson et al. (2012) reported an annual (summer) increasing trend of 0.16 ± 0.02 (0.12 ± 0.06) ppb yr-1 in the European ground-level ozone (stations' average) for the period 1996–2005. The total ozone difference (0.2–0.8 ppb) via both the effects on photolysis rates and BVOC emissions (COMBO scenario) would translate (considering the full 20-year time period) to a summer trend of 0.01–0.04 ppb yr-1. These values should not be considered for a direct comparison with the absolute values of the aforementioned observed ozone trends, not only due to differences in the data analysis like time averaging and spatial coverage but most importantly due to the exclusion of other physical and chemical processes influencing the ozone trends. Nevertheless, the comparison of the order of magnitude between the aforementioned values and the reported ozone trends suggests a higher importance of the impact of ARI (only via photolysis rates and BVOC emissions) on surface ozone than when just comparing to the total ozone concentrations. Therefore, this comparison indicates that the ARI (as investigated in this study) might have had an accountable impact on the European surface ozone trends since the 1990s and could have partially dampened the effects of ozone precursor emissions' reduction along with other more influential physical processes like intercontinental transport and stratosphere–troposphere exchange (Ordóñez et al., 2007; Derwent et al., 2008, 2015).