The Weather Research and Forecasting (WRF) model is a state-of-the-art regional meteorological model. Various gas-phase chemistry and aerosol mechanisms have been implemented into the WRF model, creating the online WRF-Chem model . In this study, we employ WRF-Chem version 3.61 with three widely used gas-phase chemistry and two aerosol mechanisms to simulate air quality over the eastern Mediterranean: CBMZ-MOSAIC (CM)

The lumped CBM-Z chemical mechanism is based on the Carbon Bond Mechanism (CBM-IV) developed by . The Carbon Bond Mechanism includes 73 chemical species and 237 reactions. CBM-Z is coupled with the Model for Simulating Aerosol Interactions and Chemistry (MOSAIC) developed by . MOSAIC uses a sectional bin approach for the representation of the aerosol size distribution. In the WRF-Chem model the user can choose between four and eight size bins which are defined by their lower and upper dry particle diameters. In both cases, only one bin is dedicated to aerosols with diameter between 2.5 and 10 µm. Therefore, when four aerosol bins are used, three bins are dedicated to aerosols less than 2.5 µm in diameter, and when eight aerosol bins are used, seven bins are dedicated to aerosols with diameters within this range. Since this study focuses on the total PM2.5 mass concentrations and not on detailed aerosol microphysics or effects on clouds, it is sufficient to use the four-bin option to reduce computational complexity.

All our simulations are conducted with the same model physics configuration (Table ) to facilitate intercomparison. We modified the WRF-Chem v3.6.1 code to take into account dust particles in the accumulation size mode (0.1 µg m-3 ≤ particle size ≤ 2.5 µg m-3) for the calculation of the total PM mass concentration in the RMS mechanism. Three nested domains are used, as shown in Fig. a. The outermost domain (d1) with a horizontal grid resolution of 80 km extends from 16 to 4∘ N and from 10∘ W to 50∘ E in order to include a large part of Europe and the Black Sea region, which have a significant contribution to the pollution of air masses that reach the EMME region, as well as a large part of the Sahara and Middle Eastern deserts in order to utilize the dust emission schemes included in the WRF-Chem model. The second domain (d2) with a horizontal grid resolution of 16 km is located over the Levantine Basin, including all the surrounding major urban centres. The third innermost domain (d3) is located over the island of Cyprus (Fig. b) with a horizontal grid resolution of 4 km, allowing for a more accurate representation of the state of the atmosphere over the complex terrain of the island close to the surface observation stations. The WRF-Chem model uses a terrain-following hydrostatic-pressure vertical coordinate system. In our study 29 layers are used from the surface up to 50 hPa. The first layer on average extends to a height of 70 m. Control experiments that were conducted during the model set-up showed that increasing the number of vertical layers (in the lowest 70 m or throughout the vertical extent of the atmosphere) does not significantly alter the concentrations of pollutants near the surface, at the station locations, due to mixing within the boundary layer.

Meteorological initial and boundary conditions are provided by the Global Forecast System (GFS) at a horizontal grid resolution of 0.25∘×0.25∘. Time-variant chemical boundary conditions are provided from the global Model for OZone And Related chemical Tracers (MOZART-4; ). The MOZART-4 model output datasets are available at a horizontal grid resolution of 1.9∘×2.5∘ and interpolated in space every 6 h to our model domain and the chemical species of each mechanism. Biogenic emissions are calculated online by the the Model of Emissions of Gases and Aerosols from Nature version 2.1 (MEGAN2.1) by . We use the Air Force Weather Agency (AFWA) dust scheme that was developed based on the dust emission scheme in the Goddard Global Ozone Chemistry Aerosol Radiation and Transport (GOCART) model . The EDGAR-HTAP (Emission Database for Global Atmos. Res. for Hemispheric Transport of Air Pollution) Version 2, compiled by the European Commission, Joint Research Center (JRC)/Netherlands Environmental Assessment Agency is utilized. This dataset includes emissions of gaseous pollutants such as sulfur dioxide (SO2), NOx, CO, non-methane volatile organic compounds (NMVOCs) and ammonia (NH3) and particulate matter with carbonaceous speciation (PM10; PM2.5; black carbon, BC; and organic carbon, OC) from anthropogenic and biomass burning sectors . PM2.5 is a subset of PM10 and includes BC, OC, SO42-, NO3-, crustal material, metal, and other dust particles. The dataset used in this study is available in 0.1∘ × 0.1∘ emission grid maps for the year 2010 and can be downloaded from the EDGAR JRC website per year, per substance, and per sector. Anthropogenic emissions were interpolated in space and time to produce daily emissions using the anthro_emiss utility .

The model output is compared against observational data from a dense station network which spans the island of Cyprus and covers a large variety of monitoring sites, including seaside and mountainous areas. Specifically, the modelled meteorology is compared against meteorological hourly observations from eight ground stations operated by the Cyprus Department of Meteorology (Fig. , squares), and meteorological data from the CYPHEX campaign, which took place from 7 to 31 July 2014, near the village of Ineia (Fig. , star). Modelled pollutant concentrations are compared against observational data from five background air quality monitoring ground stations operated by the Cyprus Department of Labour Inspection – DLI (Fig. , circles) and data from the CYPHEX campaign. The Finokalia station in Crete, which is part of the European Monitoring and Evaluation Programme (EMEP) network, is used as a reference station to discuss O3 discrepancies on measurements over Cyprus. The frequency of all air pollutant concentrations measurements is hourly, except for PM2.5 concentrations measurements by the CYPHEX campaign, which are provided every 6 h. The location of the air pollution and meteorology monitoring stations and the measurements carried out at each station are given in detail in Table .

Previous studies have shown that lateral boundary conditions affect the modelled near-surface O3 concentrations. highlighted the importance of time variant chemical boundary conditions on O3 concentrations over Europe. (accepted for publication) showed that chemical boundary conditions from the MOZART-4 global model have an important effect on the modelled concentrations over the region of study. More specifically, in their study, an important O3 overestimation by the WRF-Chem model was attributed to the effect of chemical boundary conditions. When the O3 inflow from the boundaries was reduced by 30 %, model results were closer to observations. Based on these results and the MOZART-4 model evaluation , O3 inflow from the global model was reduced by 30 % in our study. Figure a shows the observed O3 concentrations at the five air quality stations and the CYPHEX campaign and the modelled concentrations from (a) the base run using the RMS mechanism (blue) and (b) a simulation where initial O3 concentrations and O3 inflow from the global model were reduced by 30 % (red). The average NMB decreased from 21 to 9 % when O3 from the global model was reduced. Similar results appear at the Finokalia background station, where NMB was reduced from 18 to 7 % when O3 inflow from the global model was reduced by 30 %. The results for the CM and MM mechanisms are analogous. O3 overestimation due to the effect of boundary conditions from the MOZART-4 model was also reported by . In their study, the Polyphemus chemical transport model was found to highly overestimate O3 concentrations over Lebanon when using boundary conditions from the MOZART-4 model. also reported a significant contribution of boundary conditions over the O3 levels in the area. Therefore, O3 from the global model was reduced by 30 % for the simulations of this study.

Since dust is an important parameter of air quality in the region of study and important dust sources are not included in our outermost domain, dust inflow from the global model was taken into account in our simulations. The effect of dust from the boundaries by the MOZART-4 on the WRF-Chem PM2.5 concentrations is examined. In the CM mechanism much higher PM2.5 concentrations than those observed occur between 11 and 13 July. To investigate this discrepancy we performed two CM simulations (Fig. ) with (continuous blue line) and without (dashed blue line) dust influx from the boundaries.

Incoming dust results in an increase of the order of 19 % in PM2.5 modelled concentrations during the whole study period. This increase is more pronounced from 11 to 13 July (40 %). Dust presence also influences O3 concentrations though aerosol–radiation feedbacks and their impact on photolysis rates. Figure b shows the observed and modelled O3 concentrations with (blue markers) and without (red markers) dust influx from the global model. The inclusion of dust particles results in a decrease of 10 % in modelled O3 concentrations due to changes in solar radiation through aerosol–radiation interactions.

The comparison of model to station data is performed using the first free model layer. The actual altitude of four stations, located in regions with very complex terrain, was found to differ from the model terrain height (CYPHEX campaign, Ineia village, Troodos air quality monitoring station, and the nearby Prodromos meteorological station). As a test, we performed the comparison using modelled concentrations taking into account the actual altitude of the stations. This resulted only in a slightly better agreement in the predicted surface pressure at the CYPHEX campaign and the Prodromos station. Results in all other locations were not influenced because of the mixing within the model boundary layer.

We evaluate the model performance regarding basic meteorological parameters. Statistical metrics are derived by comparing the output of the three model simulations to hourly measurements at ground stations. Pearson's correlation coefficient (R), mean bias (MB), normalized mean bias (NMB), and root mean squared error (RMSE) for temperature at 2 m (T2m), wind speed at 10 m (WS10m), and surface pressure (Psurf) averaged over all stations are shown on Table . Meteorology statistical metrics for individual stations can be found in the Supplement (Table S1). Modelled T2m is in good agreement with observations (NMB = 2 to 3 %) with similar RSME values (2.72 to 2.78 ∘C) for all three mechanisms. The diurnal cycle of T2m is reproduced at the majority of the stations (R∼0.66). The model though does not capture the T2m diurnal variability at the Larnaca meteorological station (R<0.20). The station is very close to the sea and the Larnaca Salt Lake, which might influence the thermal circulation in the area.

The model tends to overestimate WS10m at the majority of the stations by an average of 1.71 to 1.83 m s-1 (R ∼ 0.46) for all three mechanisms. and also reported WS10m overpredictions by the WRF model over the Mediterranean. The latter study attributed this model behaviour to the poor representation of surface drag exerted by the unresolved topography (mountains, hills and valleys) and other smaller scale terrain features.

Local circulation is successfully predicted by the model. Figure a shows the average 10 m wind direction from 12:00 to 17:00 LST in green and from 00:00 to 05:00 LST in red. The model simulates sea breezes during daytime and katabatic winds during the night in agreement with observations. The wind roses at the Athalassa station from the observational data (Fig. a top left panel) and the three simulations show that wind direction is reproduced quite well by the model (Fig. ). The inland dominating wind direction is mainly westerly and north-westerly with frequency of occurrence 60 and 20 % respectively. Similar results appear for the majority of the stations (not shown here). Some discrepancies between model and observations at the Prodromos station are attributed to the complex mountainous topography of the Troodos area. Both model simulations and observational data reveal predominant south-westerly winds at the southern coastline of the island during day and night. The summertime general circulation pattern over the eastern Mediterranean with predominant northerly and westerly winds, as well as the anticyclonic flow over western Africa (Fig. a), is also resembled by the model.

There is very good agreement between observed and modelled Psurf with high hourly correlation coefficients (R≥0.87) and normalized mean bias of 1 %. Some negligible discrepancies exist between the three mechanisms. The differences in the meteorological components are attributed to the inclusion of the aerosol–radiation feedbacks in the simulations. The model performance regarding aerosol concentrations is discussed later in the paper.

Observed average monthly O3 concentrations for July 2014 fall within the climatological averaged summer values given by . In their study, July monthly means of O3 concentrations over a period of 15 years were found to be 54.3 ± 4.7 ppbV over all stations. The mean observed value during our simulation period at the DLI stations is 52 ppbV. The mean modelled values vary from 56.2 ppbV (NMB = 9 %) in the RMS mechanism to 63 ppbV (NMB = 22 %) and 65.2 ppbV (NMB = 23 %) in the CM and MM mechanism respectively, showing a strong overestimation of the latter two. Figure  shows the average O3 ground-level modelled concentrations for the three mechanisms for the outermost domain (Europe – Mediterranean and North Africa). Differences between the three mechanisms are more pronounced over southern Europe and the Mediterranean. Over these regions O3 concentrations predicted by the MM mechanisms are up to 10 and 20 ppbV higher compared to the CM and RMS mechanisms respectively.

The CYPHEX campaign station has been excluded from the analysis of O3. This station gives significantly higher average monthly O3 concentration (71.40 ppbV) that deviates from the climatological and observed mean, even though the station of the campaign was located only a few hundred metres away from the Ineia site of DLI (51.93 ppbV). A direct comparison between these two stations is shown on Fig. S1 in the Supplement. We further investigated this deviation by comparing with the mean monthly O3 for July 2014 at the Finokalia station (from the European Monitoring and Evaluation Programme – EMEP), which reaches 52.43 ppbV. We choose Finokalia because it is located on the island of Crete (approximately 600 km west of Ineia) with no pollution sources in between (Fig. a, red star). Thus Ineia and Finokalia have similar pollution features, being subject to air mass transport from eastern Europe.

Table  shows the statistical performance of the three gas-phase and aerosol mechanisms against hourly observations from six ground stations. The CM and MM mechanisms significantly overestimate O3 concentrations with normalized mean biases of 22 and 23 % respectively. A normalized mean bias of 9 % appears for the RMS mechanism, which corresponds to 4.25 ppbV. This mechanism shows the lowest RMSE (10.77 ppbV) compared to CM and MM (14.79 and 15.30 ppbV respectively), which is a significant improvement compared to the global MOZART-4 model (NMB = 37 %, RMSE = 20.96 ppbV). A further improvement of the order of 3 % is shown when moving from the coarse to the finer WRF-Chem domain on O3 concentrations (Supplement Table S2, Fig. S4).

Air quality modelling studies over the eastern Mediterranean in the literature focus on O3. During the second phase of the Air Quality Model Evaluation International Initiative (AQMEII), the majority of the modelling groups using the RMS and CM mechanisms with the WRF-Chem model also reported O3 concentrations overestimation over the eastern Mediterranean . However, reported an underestimation of about 5 ppbV on summertime O3 concentrations WRF-Chem model using the RMS mechanism.

Our sensitivity tests showed that high PM2.5 concentrations affect the O3 concentrations through the aerosol–radiation feedbacks by altering the radiation budget and as a consequence, the photochemical activity and the concentration of secondary pollutants. Specifically, when the dust inflow from the boundaries for the CM mechanism was not taken into account, O3 concentrations at the stations locations increased by 10 %. Since the CM mechanism shows the higher PM2.5 concentrations and relatively high O3 concentrations, we conclude that we can rule out the aerosol concentrations and therefore the different aerosol mechanisms, as the responsible factor for the differences in O3 concentrations between the three simulations. attributed the differences in pollutants concentrations between these mechanisms to the differences in the treatment of VOCs, since the rate constants for the basic O3 production and loss reactions are similar. Hourly correlation coefficients for O3 are low (less than 0.30) for all three mechanisms which is comparable to the findings of . In their study, summertime O3 hourly correlation coefficients at the Ayia Marina station (AQ01) were close to 0.2. The low correlation coefficients for O3 in combination with underestimated NOx concentrations (NMB varies from -54 to -44 %) and low hourly correlation coefficients for NOx as well suggest that nearby anthropogenic emission sources are not represented in the emission inventory used in the simulations. The monthly average diurnal cycles for O3 and NOx (Figs. S2 and S3 respectively) show a weak diurnal cycle from observational data at the majority of the stations, indicating that long-range transport is an important aspect of air quality over Cyprus. A more pronounced diurnal cycle for NOx appears at the Stavrovouni station due to the fact that the station is located close to the highway.

The hourly modelled and observed O3 concentrations at six background stations over Cyprus are depicted in Fig. . The three mechanisms show similar behaviour for 1 to 13 July. However, from 13 July until the end of the month O3 concentrations from CM and MM appear slightly higher than RMS. The fact that such differences do not appear for O3 precursors NOx (Fig. ) indicates that the different behaviour during this period is possibly due to aerosol–radiation interactions and changes in photolysis rates. Similar patterns appear for CO (Fig. ). There is a good agreement between the three mechanisms from 1 to 13 July. The RMS mechanism gives higher CO concentrations for the rest of the simulation period. Due to the long CO atmospheric residence time, these differences can be attributed to long-range transport, and therefore the effect the three different gas-phase chemistry and aerosol mechanisms have on the predicted meteorology.

An abrupt decrease in O3 concentrations in observations from 11 to 13 July is also captured by all three mechanisms at all stations except Cavo Greco, which is located in the eastern part of the island (Fig. ). This decrease in O3 concentrations is accompanied by a decrease in CO concentrations as demonstrated from the observational data from the CYPHEX campaign and the WRF-Chem model. No abrupt changes are shown in NOx concentrations either by the observational data or the model. During this period model results reveal that westerlies account for more than 70 % at the stations where decreases in O3 and CO concentrations occur, indicating the transfer of cleaner maritime air from the Mediterranean. In general, wind direction appears to have an important impact on pollutant concentrations over Cyprus. showed that at the Ayia Marina station northerlies are associated with 3–5 % higher O3 concentrations compared to westerlies and southerlies during all seasons. Similar results appear for modelled O3 concentrations at this station. More specifically, northerlies are associated with 4–12 % higher O3 concentrations compared to westerlies and southerlies for July 2014.

All three mechanisms tend to underestimate NOx concentrations at the majority of the stations (NMB varies from -53 to -44 %). The Ayia Marina and the Troodos stations are located in a mountainous region which is characterized by steep changes in altitude within short distances. The complexity of the terrain results in inaccuracies in the representation of the local wind circulation by the model, which affects the transport of pollutants. This is also supported by the CO model underestimation at the Ayia Marina station. Modelled NOx concentrations are significantly higher, and in better agreement with observations at the Cavo Greco and Stavrovouni stations. The latter is located a few kilometres to the east of an industrial area, which is represented in the anthropogenic emission inventory used in the simulations. On the other hand, a large nearby highway is not captured by the anthropogenic emission inventory, resulting in peaks in NOx concentrations from traffic, which are not captured by the model. The Cavo Greco station is located in the eastern part of the island. Model results showed that when westerlies occur at this location, NOx concentrations are significantly higher compared to southerlies for all three mechanisms, indicating that the eastern part of Cyprus is affected by transported pollutants, which are emitted within the island. Average NOx concentrations during northerlies are also higher compared to southerlies since the station is located south of the main power station of the island (Dekeleia).

As shown previously in Fig. , all three mechanisms overestimate PM2.5 concentrations at the Ayia Marina station and the CYPHEX campaign site. The lowest MB appears for MADE/SORGAM (MB = 8.96 µg m-3), whereas the MB for MOSAIC is 22.37 µg m-3 when coupled with CBMZ and 16.98 µg m-3 when coupled with MOZART.

Differences in PM2.5 concentrations between the three simulations are pronounced from 11 to 13 July. During this period, the average PM2.5 concentrations at the Ayia Marina station were 93.95, 53.81, 7.17 µg m-3 for CM, MM, and RMS respectively, whereas the average PM2.5 concentration from observational data was 13.72 µg m-3.

Investigating the inorganic mass in the MOSAIC aerosol mechanism, which represents fine dust particles , we find the overestimation is drve by the dust component. Since observations do not show elevated aerosol levels near the surface from 11 to 13 July, the dust component has been subtracted from the simulated total PM2.5 concentrations (Fig. ). PM2.5 concentrations were significantly reduced and are in better agreement with observations for CM and MM simulations, especially from 11 to 13 July. This indicates that in the MOSAIC aerosol mechanism dust has a great contribution to PM2.5 concentrations during this period. On the other hand, smaller differences are shown for the MADE/SORGAM aerosol mechanism. The large difference in PM2.5 from mid-month coincides with the time frame where large discrepancies in gaseous pollutants occur between the three mechanisms, as discussed in Sect. 3.3. This underlines the importance of the interactions between aerosols and radiation and consequently photolytic reactions and air quality.

In order to better understand individual components of PM2.5, and examine differences in behaviour by the aerosol mechanisms, we analyse the aerosol species that dominate the PM2.5 concentrations separately. Figure  presents the components from observed and modelled sulfate (SO42-), ammonium (NH3+), and nitrate (NO3-) aerosol mass concentrations and elemental carbon concentrations at the Ayia Marina station during the study period. Monthly mean sulfate aerosol concentrations for the three mechanisms vary from 5.14 to 7.02 µg m-3, which is close to observed values (5.05 µg m-3). The lowest monthly mean concentrations are produced by the CM mechanism. This mechanism shows the highest sulfate dioxide (SO2) concentrations during the whole study period (Fig. ). Since this simulation uses the same aerosol mechanism with the MM simulation, and therefore the same heterogeneous nucleation rates from sulfuric acid (H2SO4) to sulfate aerosols, the differences between the CM and MM simulations are attributed to the chemical processes that act as sources/sinks of H2SO4. The RMS mechanism includes the heterogeneous SO2 cloud oxidation , which results in higher sulfate aerosol concentrations compared to CM and MM.

Elemental carbon is underestimated by all three simulations. The lowest NMB appears for RMS (-34 %) followed by CM (-35 %). Since the three simulations use the same anthropogenic emission inventory, these differences between RMS and MM can be partially attributed to the different treatment of aerosols by the modal and sectional bin approaches. The MM mechanism highly underestimates EC concentrations due to the absent of anthropogenic emissions in this mechanism.

Ammonium aerosol mean monthly values are close to observed 1.43 µg m-3 for the CM and MM simulations (1.74 and 1.87 µg m-3 respectively), while a higher value is shown for RMS (3.24 µg m-3). Nitrate aerosols are highly overestimated by the RMS mechanism with maximum values and outliers well above the period average. These outliers are due to nitrate aerosol transport from the north, when favoured by wind speed and direction. In contrast, nitrate aerosols are slightly underestimated by CM and MM. These differences lie in the differrent treatment of the gas-to-particle partitioning from the nitric acid to ammonium nitrate as a function of humidity from the two aerosol mechanisms used . MADE uses the approach and MOSAIC uses the method. The diurnal cycle of compounds that are crucial to night-time chemistry (NO3, N2O5) vary significantly between the three mechanisms (not shown). The RMS mechanism exhibits 3 times higher night-time NO3 concentrations than the CM mechanism. This indicates considerable uncertainty in the representation of this important part of tropospheric chemistry that also affects aerosol formation. These results are supported by the findings of that also report 3 times higher pan-European averaged NO3 in the RADM2 mechanism compared to CBMZ in the middle of the night-time chemistry cycle.