CAMx (Comprehensive Air Quality Model with Extensions) is an Eulerian photochemical dispersion model developed by Ramboll Environ. Version CAMx v6.50 of the model was used in the present study.

For this study, a first domain with a 36 km resolution was defined at the European scale. A second nested domain was defined and named MEDI12 (147×249 points), which covered the center of Europe with a resolution of 12 km. Both meteorological and chemical transport simulations were provided for these domains. WRFv3.9 was run for the simulation of meteorological conditions with 28 vertical layers up to 50 hPa, with FNL data for initial conditions.

For the CAMx simulation, boundary conditions from MOZART-4 were used.

Sea salt emissions are calculated in the SEASALT pre-processor of CAMx. This program generates aerosol emissions of sodium, sulfate and chloride and gaseous emissions of chlorine using CAMx-ready meteorological and land use files. The sea salt emissions program calculates the flux of sea salt over the open ocean using parameterizations developed by Ovadnevaite et al. (2014). The surf zone aerosol flux is calculated by using the Gong (2003) open-ocean approach with an assumed 100 % whitecap coverage. Biogenic emissions were calculated separately with MEGANv2.03 (Model of Emissions of Gases and Aerosols from Nature; Guenther et al., 2006) and then included in the land-based emissions. WBDUST pre-processors deliver dust emissions in CAMx and generate gridded windblown dust emissions. The scheme is based on an updated approach used in the global EMAC (ECHAM/MESSy) atmospheric chemistry–climate model (Klingmüller et al., 2018; Astitha et al., 2012). The mechanism for lightning NOx was not activated in CAMx.

The gas phase chemical mechanism is Carbon Bond 5 (CB05), in which the NMVOC emissions are split into 13 species (TERP, ISOP, XYL, TOL, ETOH, MEOH, IOLE, OLE, ETH, ALD2, PAR, ETHA and FORM) and describe approximately 156 reactions. For semivolatile inorganic species (sulfate, nitrate and ammonium), the equilibrium concentration is calculated using the thermodynamic model ISORROPIA (Nenes et al., 1998). Fourteen vertical levels are simulated with a first layer height of approximately 10 m.

CHIMERE is an offline chemistry transport model developed by LMD-IPSL/CNRS (Menut et al., 2013). The CHIMERE2017r4 version of the model was used in this study.

WRFv3.9 (Weather Research and Forecasting Model) was run for the simulation of meteorological conditions with 28 vertical layers up to 50 hPa, with FNL data for initial conditions.

Concerning the CHIMERE simulation, boundary conditions are monthly mean climatologies taken from the LMDz-INCA model (Laboratoire de Météorologie Dynamique General Circulation Model – INteraction with Chemistry and Aerosols; Schultz et al., 2006) for gaseous species and from the GOCART model (Global zone Chemistry Aerosol Radiation and Transport; Ginoux et al., 2001) for aerosols (desert dust, carbonaceous species and sulfate). Sea salt emissions were calculated as described in Monahan et al. (1986). MEGANv2.04 calculated biogenic emissions (Guenther et al., 2006). MEGAN is run directly by CHIMERE code, and biogenic emissions are just generated before the air quality run. The mineral dust emissions are calculated on-line. The soil is represented by relative percentages of sand, silt and clay with the USGS soil texture (https://www.usgs.gov/, last access: 20 January 2023). The aeolian roughness length used in CHIMERE is the GARLAP (Global Aeolian Roughness Lengths from ASCAT and PARASOL) dataset as in Prigent et al. (2012). There is no treatment of NOx lightning in CHIMERE.

The gas phase chemical mechanism is MELCHIOR2 (Modele Lagrangien de Chimie de l'Ozone a l'echelle Regionale), in which the NMVOC emissions are split into 10 species (C2H6, NC4H10, C2H4, C3H6, C5H8, OXYL, HCHO, CH3CHO, CH3COE and APINEN) and describe approximately 120 reactions. For semivolatile inorganic species (sulfate, nitrate and ammonium), the equilibrium concentration is calculated using the thermodynamic model ISORROPIA (Nenes et al., 1998). Nine vertical levels are selected with a first layer height at 20 to 25 m.

On the basis of emission input data, the CMAQ Model v5.2 with the AERO6 model calculates air concentration as well as deposition fluxes of atmospheric gases and aerosols (Byun and Schere, 2006; Appel et al., 2017). Atmospheric chemistry is used by the chemical Carbon Bond 5 (CB05) mechanism (Sarwar et al., 2008) cb05tucl with updated toluene chemistry (Whitten et al., 2010), including the chlorine chemistry extension (CB05-TUCL; https://www.airqualitymodeling.org/index.php/CMAQv5.0_Chemistry_ Notes, last access: 20 January 2023). The aerosol scheme AERO6 is used for the formation of secondary inorganic aerosols. Sulfuric acid (H2SO4), nitric acid (HNO3), hydrochloric acid (HCl) and ammonia (NH3) gas-phase–aerosol partition equilibria are solved by the ISORROPIA mechanism (Fountoukis and Nenes, 2007; Nenes et al., 1998). Contained within is the formation of secondary organic aerosol (SOA) from isoprene, terpenes, benzene, toluene, xylene and alkanes (Carlton et al., 2010; Pye and Pouliot, 2012).

Sea salt emissions were calculated as described in Kelly et al. (2010). Biogenic emissions (NMVOC from vegetation and soil NO) were calculated previously with MEGANv3 (Guenther et al., 2012) and then included in the land-based emissions. Emissions of windblown dust were not considered. CMAQ models 30 vertical layers, with the lowest layer from 0 to 42 m and the second layer from 42 to 85 m. The NOx lightning treatment in CMAQ was not activated for the present study.

The COSMO model simulated the meteorological data for CMAQ, applying the version COSMO5-CLM16 (Schultze and Rockel, 2018; Petrik et al., 2021). The MCIP (Meteorology-Chemistry Interface Processor) processed meteorological model output into the input format required for CMAQ. The vertical resolution of the meteorological model was 40 terrain-following geometric height levels up to 22 km. The boundary condition driver used was IFS-CAMS cycle45r1 (Integrated Forecasting System – Copernicus Atmosphere Monitoring Service; Inness et al., 2019) with a vertical resolution of 60 sigma levels up to 65 km.

To prevent the effects from initial conditions on the simulated atmospheric concentrations in 2015, the model run started with a spinup run in mid-December 2014. The grid size of the Mediterranean Sea domain was 12 km × 12 km, nested in a 36 km × 36 km domain covering all of Europe.

The EMEP MSC-W (European Monitoring and Evaluation Programme, Meteorological Synthesizing Centre – West, https://www.emep.int/mscw/index.html, last access: 20 January 2023) model is a limited-area, terrain-following hybrid coordinate model designed to calculate air concentrations and deposition fields for major acidifying and eutrophying pollutants, photooxidants and particulate matter (Simpson et al., 2012, 2020).

In this study, a 0.1∘×0.1∘ resolution grid on a long–lat projection and with 20 vertical levels was used. The meteorological input data are based on forecast experiment runs with the Integrated Forecast System (IFS), a global operational forecasting model from the European Centre for Medium-Range Weather Forecasts (ECMWF). The meteorological fields are retrieved on 0.1∘×0.1∘ long–lat coordinates. Vertically, the fields on 60 eta (η) levels from the IFS model are interpolated onto the 20 EMEP eta levels.

The model version used was rv4.34 with chemical mechanism EmChem 19a (Simpson et al., 2012, 2020). The mechanism builds on surrogate VOC species (Simpson et al., 2012; extended with benzene and toluene) and has 171 gas phase and heterogeneous reactions. The model always assumes equilibrium between the gas and aerosol phases using the MARS equilibrium module (Model for an Aerosol Reacting System) of Binkowski and Shankar (1995). For secondary organic aerosol (SOA), a so-called volatility basis set (VBS) approach (Robinson et al., 2007; Donahue et al., 2009; Bergström et al., 2012) is used. All primary organic aerosol (POA) emissions are treated as nonvolatile to keep emission totals of both PM and VOC components the same as in the official emission inventories, while the semivolatile ASOA and BSOA species are assumed to oxidize (age) in the atmosphere by OH reactions (Simpson et al., 2012).

The generation of sea salt aerosol over oceans is driven by the surface wind, and the EMEP model's parameterization scheme for calculating sea salt generation is based on two source functions: those of Monahan et al. (1986) and Mårtensson et al. (2003). The following natural emissions are calculated in the model for each grid cell and at every model time step: biogenic emissions of isoprene and monoterpenes use near-surface air temperature and photosynthetically active radiation. Soil NO emissions from soils of seminatural ecosystems are specified as a function of N deposition and temperature. The key parameter driving dust emissions is wind friction velocity. Additionally, daily emissions from forest and vegetation fires are taken from the “Fire INventory from NCAR version 1.0” (FINNv1; Wiedinmyer et al., 2011). Emissions of NOx from lightning are included as monthly averages of global 3-D fields on a T21 (5.65∘×5.65∘) resolution (Köhler et al., 1997). For this study, the initial and boundary conditions provided with the open-source model distribution for 2015 were used.

LOTOS-EUROS is an Eulerian chemistry transport model (Manders et al., 2017). The model simulates air pollution in the lower troposphere and is of intermediate complexity, allowing ensemble-based simulations and assimilation studies. LOTOS-EUROS performs hourly model output using ECMWF (European Centre for Medium-Range Weather Forecasts) meteorological data. The gas phase chemistry follows the TNO CBM-IV scheme (Schaap et al., 2008).

For sea salt two parametrizations are used for online calculation of emissions: Mårtensson et al. (2003) for fine particles and Monahan et al. (1986) for coarse particles. Biogenic emissions are calculated online during the CTM run. For isoprene, a tree-species-dependent emission factor was used (Schaap et al., 2008; Beltman et al., 2013). NO emissions from soil were calculated as in Novak and Pierce (1993). Dust emissions are also calculated online for three sources of dust. Desert dust follows Mokhtari et al. (2012), and road resuspension and dust from agricultural processes follow a module developed by Schaap et al. (2008). There is no treatment of NOx lightning in LOTOS-EUROS.

LOTOS-EUROS has a dynamical vertical layer structure with five layers in total. The first layer is at 25 m, while the second layer follows the meteorological boundary layer. On top of that, two evenly distributed reservoir layers are defined: one up to 3500 m and one top layer up to 5000 m above sea level. The model has participated in multiple model intercomparison studies (Bessagnet et al., 2016; Colette et al., 2017), showing overall good performance.

Annual anthropogenic land-based gridded emissions for 2015 obtained from the CAMS-REG v2.2 emission inventory were used as input by all five compared models. Gridded emission files contain GNFR (Gridded Nomenclature for Reporting) emission sectors for each country for the air pollutants NOx, SO2, NMVOC, NH3, CO, PM10, PM2.5 and CH4. The emissions are provided at a spatial resolution of 1/10∘×1/20∘ in longitude and latitude (i.e., ∼6 km × 6 km over central Europe).

The height distribution of emissions per GNFR sector was determined as described in Bieser et al. (2011b). The temporal distribution was determined by separating the annual emissions of each sector into hourly emission data with data splitting as described in Granier et al. (2019). PM was split as described in Bieser et al. (2011a); NOx was split according to Manders-Groot et al. (2016). NMVOC emissions were given for different sectors, using the GNFR, and were separated country-wise. This split was used as provided in the CAMS-REG v2.2 emission inventory (Granier et al., 2019). The species were afterwards split within each CTM according to their chemical mechanism. Information on biogenic emission totals for the whole model domain can be found in Table S20.

The shipping emission dataset produced with the STEAM model has a spatial resolution of 12 km × 12 km and a temporal resolution of 1 h. The STEAM emissions are divided into two vertical layers (0 to 36 m; 36 to 1000 m) and are provided for mineral ash, carbon monoxide (CO), carbon dioxide (CO2), elemental carbon (EC), NOx, organic carbon (OC), PM2.5, particle number count (PNC), sulfate (SO4), SOx (containing SO2 and SO3) and VOC. To reduce the number of generated emission maps and the computational resources needed to run the STEAM model, VOC emissions were divided into four categories based on how their emission factors change as a function of the engine load. Emissions of individual VOC species were calculated afterwards based on their mass fractions of the total emissions in the VOC group. Emission factors for VOC are based on the average values taken from various publications (Agrawal et al., 2008, 2010; Sippula et al., 2014; Reichle et al., 2015).

In CAMx, all shipping emissions are put in the first layer. For CHIMERE, all shipping emissions above 36 m and 88 % of the emissions below 36 m have been added to the second layer. Only 12 % of the emissions below 36 m were emitted in the first layer of the model. This was calculated based on the STEAM emission dataset and the stack heights contained therein. Additionally, in CMAQ, shipping emissions were distributed in the two lowest layers, emissions below 36 m were attributed to the lowest layer, and emissions above 36 m were in the second layer. For EMEP simulations, the STEAM emissions were summed from hourly to daily emissions and attributed to the lowest layer (up to 90 m). In LOTOS-EUROS, ∼70 % of emissions below 36 m are assigned to the first layer, which is 25 m thick, and ∼30 % to the second layer. Emissions above 36 m are divided over different height classes: 30 % between 36 and 90 m, 30 % between 90 and 170 m, 30 % between 170 to 310 m, and 10 % between 310 cm and 470 m. Due to the dynamic second model layer (following the meteorological boundary layer), those emissions are put in the second and/or third model layer. In the case of a well-mixed and vertically extended meteorological boundary layer (above 470 m), all emissions are in this second layer, whereas when the boundary layer is shallow, some emissions are put in the third layer.

Table 2 contains R, NMB and RMSE based on the annual time series for NO2 at all stations. The highest correlation across all 62 stations showed LOTOS-EUROS followed by CMAQ with a slightly lower correlation (LOTOS-EUROS: R=0.45; CMAQ: R=0.42), whereas for CHIMERE, EMEP and CAMx, non-existent to weak correlation was found (R=0.08 to R=0.10). The NMB suggests that all five CTMs underestimate the annual mean concentrations at most measurement sites; the NMB for all stations is negative for all models. The RMSE is within the same range for all models (RMSE =15.6 to 19.5 µgm-3; Table 3).

Time series for three example stations show the temporal variations between measured and modeled data (Appendix C). The supplements provide an overview of the mean values of stations in each map parcel (west, north, south, east; Fig. S2). Figure C1 displays a time series at an urban background station in France (fr08614, “Gauzy”; lat 43.8344, long 4.374219), which was chosen because southern France will be investigated in greater detail as part of this study. Figure C2 shows a rural background station in Italy (it1773a, “Genga – Parco Gola della Rossa”; lat 43.46806, long 12.95222), which was chosen due to its central location in the domain and the high number of stations in Italy. Figure C3 displays the time series at a station in Greece (gr0035a, “Lykovrysi”; lat 38.06963, long 23.77689) to include a station in the eastern part of the domain.

Measurements at the French station show the highest NO2 values in winter, with peaks between 40 and 55 µgm-3 (Fig. C1). LOTOS-EUROS and EMEP underestimate the values throughout the year. Moderate correlation was calculated for CMAQ (R=0.6) and LOTOS-EUROS (R=0.65) at this station. The simulated ship impact has annual mean values from 0.2 µgm-3 (EMEP, CAMx) to 0.6 µgm-3 (CMAQ) at station fr08614. Shipping emissions have a potential relative impact between 1.8 % (EMEP) and 6.7 % (CMAQ) on the total concentration in the annual mean. The highest potential ship impact at this station was modeled by CMAQ. At the Italian station, it1773a lower NO2 concentrations were measured compared to the station in France. The highest peaks are approximately 20 µgm-3 in winter. At station it1773a, the potential ship impact on the total NO2 concentration has annual mean values between 0.07 µgm-3 (LOTOS-EUROS) and 0.5 µgm-3 (CAMx). The highest relative potential ship impact was 7.9 % and was modeled by CAMx. At station gr0035a, the lowest simulated values are shown by CMAQ and LOTOS-EUROS. The highest values display EMEP at this station, also with the highest correlation between measured and simulated data (R=0.55). The potential ship impact at the Greek station is between 5.0 % (EMEP) and 15.3 % (CAMx), which is higher than the potential ship impact at the other two stations.

All CTMs underestimate the actual measured total NO2 values at both stations, except for LOTOS-EUROS in Italy. None of the models is able to model matching peak values. Neither at the station in France, Italy nor Greece did models show seasonal variation in concentrations, whereas NO2 usually has higher values in winter and lower values in summer, mainly because of lower photolytical degradation and suppressed vertical mixing, as described, e.g., in Ordóñez (2006).

Differences in potential ship impacts between the stations are caused by the location and station type (fr08614: urban background; it1773a: rural background; gr0035a: suburban background). At the French station, the traffic-related NO2 concentration might supersede the ship-related NO2. The station in Italy is not located in a city, so the NO2 concentration caused by ships comes to the fore. The highest potential ship impact was simulated at the station in Greece because it is suburban but close to the Port of Piraeus, which is one of the largest ports in the Mediterranean Sea. As expected, the average potential ship impact is low at stations that are not directly located on the coast or near a harbor.

To compare the correlation R, NMB and RMSE at all measurement stations for all models, the results of the comparison are divided by country and displayed in boxplots (Fig. 2). Each dot displays one measurement station. The correlation measured against the simulated annual mean NO2 is highest for LOTOS-EUROS and CMAQ in all countries, reflecting the results shown in Table 3 for correlation. Nevertheless, boxplots for NMB and, in particular, for RMSE show that differences among countries are larger than differences among the models (Fig. 2b, c). This means that all models show good or bad performance at some stations, which was not found to be statistically relevant.

Underestimations by models of NO2 at urban sites were found in other studies (Karl et al., 2019b; Giordano et al., 2015), despite differences in grid size. Karl et al. (2019b) used a grid resolution of 4 km, and Giordano et al. (2015) used a grid resolution of ∼0.25∘ (27 to 28 km). The underestimation might be due to too low emissions in the inventory used by the models and the heterogeneity of emissions. Regional CTMs cannot display small-scale spatial heterogeneity; coarse grid cells are not representative of the measurement location. Giordano et al. (2015) suggested in their study that the underestimation of NO2 could be caused by either an underestimation of the chemical lifetime of NOx, excessively high dry deposition, an underestimation of natural emissions at rural and remote stations, or a combination of these factors. Differences in radical concentrations and reactive nitrogen might be additional reasons for underestimation (Knote et al., 2015).

The model performance of NO2 has shown that differences in time series between the models occur, caused by the differences in meteorology and large grid size. Large grid sizes can cause errors insofar as, in simulations, the land areas are not seen as such but as water areas. This is especially problematic when having measurement stations located close to the sea.

The simulated annual mean NO2 concentrations considering all emission sectors are similar for all CTMs, with a median ensemble mean of ∼2.5 µgm-3 (Fig. 3a). Regarding the spatial distribution CAMx and CHIMERE have the largest areas, with values exceeding 5.0 µgm-3, especially along the main shipping routes and in urban areas (Fig. 4). The CMAQ, EMEP and LOTOS-EUROS figures look similar, which is in good agreement with the displayed time series in Sect. 3.1, where the results are within the same range.

Over land area, all model simulations display a concentration pattern ranging within 1 order of magnitude. Nevertheless, the frequency distributions of the CMAQ, EMEP and LOTOS-EUROS simulations show the highest frequency between 1.0 and 2.0 µgm-3, whereas for CAMx and CHIMERE, they are more equally distributed. Higher values of NO2 concentrations simulated by CAMx and CHIMERE might indicate a longer lifetime of NO2 in the atmosphere. NO2 reacts quickly with hydroxyl radicals (OH) and forms HNO3, or NO2 photolysis creates O3 during the daytime. The annual mean HNO3 concentrations are between 2.0 to 5.0 µgm-3 for CAMx and CHIMERE over water areas and are 0.8 to 2.0 µgm-3 over water areas for CMAQ, EMEP and LOTOS-EUROS (Fig. S11). Over land areas, the HNO3 concentrations are within one range for all models. A lower HNO3 concentration is expected for CTMs with a longer lifetime of atmospheric NO2. Nevertheless, there can be a misinterpretation when both concentrations are high. Therefore the data were normalized by using the HNO3:NO2 ratio (Fig. S12). Differences are displayed especially along the main shipping routes. There, values are lower in CAMx and EMEP compared to the other models. This can be explained by the lower HNO3 formation by these models along the shipping routes.

Also the meteorology might influence the vertical mixing of NO2. This leads to differences between the models or explains the similarity between CAMx and CHIMERE due to the use of the same meteorology. Nevertheless, this point will not be discussed here in detail since in the present study only the lowest layer was considered and the vertical mixing processes were not evaluated.

The correlation between the models for total NO2 concentration was calculated based on hourly data (Table 4). The highest correlation was found between CAMx and CHIMERE (R=0.80). Weak correlations were found between LOTOS-EUROS and CAMx (R=0.31) and LOTOS-EUROS and CHIMERE (R=0.36). This weak correlation is due to the differences in frequency distribution, with LOTOS-EUROS showing most values below 1.0 µgm-3, whereas for CAMx and CHIMERE, more values are located in the higher value ranges. Overall, the models can give a robust estimate regarding the base run of the annual mean of NO2.

The highest potential impact of ships on total NO2 concentrations was found on the main shipping routes, with values >85 % (Fig. 5). Similar values were found for the Baltic Sea (Karl et al., 2019b) and for the Iberian Peninsula (Nunes et al., 2020). CHIMERE and CAMx model the highest values over the sea region, with a potential ship impact on NO2 between 60 % and 85 %. CMAQ, LOTOS-EUROS and EMEP have similar patterns for ship impacts over the sea.

On the Mediterranean coastline, CMAQ, CHIMERE, LOTOS-EUROS and EMEP simulate a similar potential impact, with 25 % to 45 % potential ship impacts on total NO2. Merico et al. (2017) found similar results in a study with an NO2 shipping impact of up to 32.5 % regarding four port cities in the Adriatic–Ionian Sea. CAMx reveals a higher impact with >85 % at the coastline. The potential ship impact displayed in the time series in Sect. 3.1 was lower, although the measurement stations were not far from the coast. This shows that although the potential impact from ships reaches regions far from the coast, the highest impact is over the sea area. The frequency distribution for the relative ship impact shows that all models simulate the most values between 0 % and 5.0 % of the potential ship impact. Interestingly, the distribution is lowest at values between 20 % and 40 % (CMAQ, EMEP, LOTOS-EUROS) and 60 % (CAMx, CHIMERE) and then increases again at higher values, showing a bimodal distribution. This is due to large areas with high potential impacts over water and large areas with low potential impacts over land or near harbors.

Over land in the northeast area of the domain, slightly negative potential ship impacts are derived from the CMAQ, CAMx, LOTOS-EUROS and EMEP results. CHIMERE shows only very few negative values but in the same region. Negative potential ship impacts on NO2 concentrations may arise when the zero-out method is applied. They are a consequence of the nonlinear NOx gas phase chemistry. Especially in areas where the impact of NOx emissions from shipping is very low, less NO oxidation takes place because the additional NO from shipping in other areas has already consumed the oxidants (e.g., O3).

The boxplots in Fig. 3 display the annual mean values for the whole model domain of NO2. Model results vary for the base run but also for the potential ship impact. This variability needs to be taken into account when the predictive power of CTMs is considered. The “all_mean” boxplot displays the mean of all models and shows that in comparison with other models, CAMx has high values. It further helps to show which CTM tends to simulate higher or lower values compared to others. The all_mean boxplots show similar ranges as boxplots for CMAQ and EMEP, particularly regarding absolute and relative potential ship impacts. Additionally, models simulating a higher overall concentration of pollutants also tend to simulate a higher potential ship impact. The relative potential ship impact is highest for CAMx and CHIMERE and lowest for LOTOS-EUROS.

The tropospheric O3 concentrations are strongly connected to the NO2 concentration and to the oxidized nitrogen chemistry in the atmosphere. O3 can be both an initiator and a product of photochemistry; thus, it is crucial in tropospheric chemistry.

Simulated versus measured data of 1-year daily mean O3 time series show a weak (EMEP: R=0.38) to moderate correlation (CAMx: R=0.40; CHIMERE: R=0.47; CMAQ: R=0.60; LOTOS-EUROS: R=0.69; Table 5).

Selected time series represent these differences in correlation (Appendix D). Nevertheless, for the first months of the year CHIMERE, CAMx and CMAQ overestimate the actual measured O3 values (Fig. D1: station fr08614; Fig. D2: station it1773a; Fig. D3: gr0035a).

During summer months, O3 shows the highest values due to increased photochemical activity. The simulated potential ship impact is between 1.1 µgm-3 (CAMx) and 2.8 µgm-3 (LOTOS-EUROS) at station fr08614 and has a relative potential impact between 1.3 % (CAMx) and 4.0 % (CHIMERE) on the total concentration. At station it1773a, the mean O3 potential ship impact is between 1.0 µgm-3 (CAMx) and 3.0 µgm-3 (CHIMERE), and the relative potential impact ranges from 1.1 % (CAMx) to 3.5 % (LOTOS-EUROS). The potential ship impact of station gr0035s ranges from -0.1 µgm-3 (CAMx) to 3.7 µgm-3 (CMAQ; LOTOS-EUROS), which is a relative potential impact of -0.1 % (CAMx) and 3.7 % (CMAQ).

The O3 potential ship impact is within the same range at both stations and for all five CTMs. Figure 6 shows that CMAQ has the smallest bias compared to the other models (NMB =0.28), followed by LOTOS-EUROS (NMB =0.36). The RMSE is lowest for CMAQ (RMSE =31.2 µgm-3) and LOTOS-EUROS (RMSE =32.6 µgm-3), along with the lower NMB compared to the other models. The performance analysis revealed that all five models predict higher O3 concentrations than those measured at almost all stations (NMB >0). The overestimation of actual measured O3 by the models is in line with results from previous studies (Karl et al., 2019b; Appel et al., 2017; Im et al., 2015a, b). Im et al. (2015a) showed that O3 concentrations above 140 µgm-3 are underestimated, while concentrations below 50 µgm-3 are overestimated by 40 % to 80 % in all considered models. This overestimation of O3 by the models is likely linked to the chemical boundary conditions used in the regional CTMs. Analyses of the boundary conditions revealed that, especially in winter, O3 levels are mostly driven by transport instead of local production due to limited photochemistry (Giordano et al., 2015).

CHIMERE uses boundary conditions from monthly mean climatologies simulated with the LMDz-INCA model, CAMx uses MOZART-4 output, LOTOS-EUROS and CMAQ use IFS-CAMS reanalysis data, and the EMEP model uses ozone boundary conditions provided with the open-source model distribution for 2015. These differences in input for the boundary conditions can be seen as the reason for the varying results in O3 (Figs. S13 to S16).

All CTMs performed relatively well and are able to represent the course of the year, with higher values in summer and lower values in winter. Nevertheless, in some cases, the values in spring are overestimated.

The annual mean concentration of O3 considering all emission sectors is between 60 and 120 µgm-3 for all models (Fig. 8). This is consistent with the measurements displayed in the time series in Sect. 3.2.1. CHIMERE, CAMx and LOTOS-EUROS show particularly high O3 concentrations over the sea. Interestingly, EMEP results are similarly high over the sea area, but in comparison with other CTMs, concentrations are lower over land, and even values below 60 µgm-3 can be seen in the Po valley (Fig. 8d). Regarding the correlation between the models for total concentration over the whole domain, it is highest between CMAQ and EMEP (R=0.71) and lowest for CAMx and LOTOS-EUROS (R=0.42), but predominantly moderate correlations were found among the models (Table 6).

In general, all CTMs show high annual mean concentrations over the sea areas and low annual mean concentrations over land areas. This is due to lower dry deposition over sea and the overall higher emissions over land. Furthermore, high values of O3 are expected to enter the domain from the eastern part of the Mediterranean Sea. This point will be discussed in Sect. 4. The frequency distribution of the annual mean total concentration of O3 has a bimodal distribution for CHIMERE, CMAQ and EMEP. This reflects photochemical O3 depletion or production, with high values over water areas and lower values over land. Over water, low O3 depletion is expected during the night. A comparison of diurnal cycles of O3 over water and over land shows that this presumption is reflected by CMAQ and EMEP results, showing more pronounced cycles of O3 in grid cells over land (Fig. S17). However, the diurnal cycles of CAMx, CHIMERE and LOTOS-EUROS do not show differences in amplitude over land and water. Despite this, over water, all models show a higher spread of values within diurnal cycles, displaying that there is more variability in the course of the year over water than over land.

The potential relative impact of ships on total O3 concentrations is lowest in areas with a high potential impact of shipping on total NO2 (Fig. 9). It decreases to -20 % in areas with high NO2 concentrations in all model results, displaying a local-scale titration of O3 by NO, which is emitted by ships. This reverse relationship between NO2 and O3 was already shown in other studies (e.g., Karl et al., 2019a). Measurement studies also indicate that emissions of NO lead to local reduction in O3 concentration and showed that there could be an increase at larger distances (Merico et al., 2016). Consequently, the largest areas with O3 destruction for the CAMx and CHIMERE coincide with areas where the models show the highest potential impact of shipping on NO2. The comparison with the time series shows the highest potential ship impact on the total O3 concentration in summer. Likewise, in Sect. 3.1.4 the lowest potential ship impact was found for CAMx.

Figure 7 shows boxplots with annual mean values of the models for the whole domain. It shows that CAMx, CHIMERE and LOTOS-EUROS are within one range regarding the annual mean total concentration. The CMAQ and EMEP simulations are lowest for the annual mean O3 total concentration. Regarding potential ship impact, all CTMs except CAMx are within one range.

The present study does not contain the parts of the Mediterranean Sea furthest east due to the focus of the project on the western Mediterranean Sea with its harbor cities as well as due to the limited extent of the WRF domain. A more detailed investigation of the boundary conditions of CMAQ has shown high O3 values in the eastern part of the domain. A high O3 production over the eastern Mediterranean Sea and a steep west–east gradient of O3 were described in previous studies (i.e., Doche et al., 2014; Safieddine et al., 2014; Liu et al., 2009). This production influences the amount of O3 in the western part of the Mediterranean Sea. Safieddine et al. (2014) found an increase of up to 22 % in O3 in the eastern part of the Mediterranean Basin compared to the middle of the basin. Doche et al. (2014) described a steep west–east O3 gradient with the highest concentrations over the eastern part of the Mediterranean Basin.

Overall, all models showed a relatively good performance for O3 but differed in simulating spatial distribution and potential ship impact mainly over water. Although boxplots for annual mean values of O3 differ, for relative potential ship impact they show that CHIMERE, CMAQ, EMEP and LOTOS-EUROS are within one range. Diurnal cycles did not reveal differences in O3 depletion over water and land among the models.

The annual mean NO2 dry deposition of all four compared CTMs displays similar values over land areas (Fig. 12). In cities and densely populated regions, all models show high NO2 dry deposition, with values over 300 mg m-2 yr-1. Nevertheless, the frequency distribution of all values shows that this is mainly the case for CAMx and LOTOS-EUROS. Additionally, over the sea, the pattern of annual mean dry deposition of NO2 is also similar for CAMx and LOTOS-EUROS.

Table 7 shows that the correlation was strongest between CHIMERE and CAMx (R=0.72). Similarities and strong correlations in the output of both models were also found for the NO2 concentration in Sect. 3.1.2. This can be traced back to the same meteorology data that were used by both CTMs.

The relative potential ship impact on the annual dry deposition of NO2 is displayed in Fig. 13. The lowest potential ship impact on NO2 dry deposition is simulated by CMAQ and LOTOS-EUROS. In particular, CMAQ shows large areas with negative (-2.5 %) potential ship impacts over land. The CHIMERE simulations looks similar to the CAMx simulations over land. Along the coastline, CMAQ and LOTOS-EUROS show a potential impact of ships between 10 % and 25 %; CAMx and CHIMERE expect a potential ship impact on the total annual deposition of 25 % to 75 %. The highest potential impact is displayed by CAMx.

Differences in NO2 dry-deposition model results can be due to the dry-deposition velocities but also due to the different meteorology data used by the models (Wichink Kruit et al., 2014). Dry-deposition velocities of NO2 (Fig. S18) show that deposition velocities of CHMIERE and CMAQ are within one range and are lower compared to CAMx and LOTOS-EUROS deposition velocities. Velocities of the latter two are within one range. High velocities might lead to higher deposition rates, leading to high annual mean deposition. This is reflected in the annual dry deposition of NO2, where CAMx and LOTOS-EUROS simulate the highest values. Overall, the models have more differences in NO2 dry deposition than in air concentration. As was the case for NO2 concentration, CAMx simulated the highest values in dry deposition. The lowest values in NO2 dry deposition are displayed by CMAQ. In addition, the correlation between CMAQ and the other models was lowest.

High NO2 deposition over water areas caused by ships contributes to eutrophication (Vivanco et al., 2018). A study by Im et al. (2013) showed values of approximately 500 kg (N) m-2 yr-1 (=∧50000 mg m-2 yr-1) over the Mediterranean Sea, which means an exceedance of the critical load of 2 g to 3 g (N) m-2 yr-1 (=∧ 2000 to 3000 mg m-2 yr-1) to marine and coastal habitats (Bobbink and Hettelingh, 2011). The present study focused on NO2 dry deposition; thus, a direct comparison with critical load levels or with other studies regarding total N deposition would not be possible. A subsequent calculation of N showed that the simulated values in the present study do not exceed the critical loads (Appendix F). Nevertheless, NO2 dry deposition from ships contributes to the total N deposition budget, thus increasing with ship traffic and affecting the ecosystems in the Mediterranean Sea.

Dry deposition is a major sink for O3 in the lowest model layer. O3 has high destruction rates on vegetated surfaces through plant stomata and lower rates on surfaces such as water or snow (Clifton et al., 2020). Spatial patterns of annual total O3 dry deposition confirm this distribution. Over sea annual totals are lower (250 to 1000 mg m-2 yr-1) compared to values over land (2500 to 10 000 mg m-2 yr-1; Fig. 14). The correlation for the annual total concentration of O3 dry deposition is highest between CHIMERE and CAMx, showing a moderate correlation (R=0.57; Table 8).

Figure 15 shows the potential ship impact on the total dry deposition of O3. CMAQ and LOTOS-EUROS are within a similar range, with potential impacts of ships of 5 % to 10 % over water surfaces. The lowest potential impact of -5 % in the main shipping lanes is simulated by CAMx, showing a similar pattern as for the O3 potential ship impact. Over land areas, ships contribute to dry O3 deposition from 0.25 % to 2.5 %.

In addition to the impact of O3 dry deposition on plant stomata, it is important to explain differences in surface O3 concentration results. The O3 concentration is sensitive to the deposition velocity (Clifton et al., 2020), which differs among the four CTMs. This can be confirmed by studies comparing deposition schemes, where differences in O3 concentration between models are caused by the variety of processes (Clifton et al., 2020). In particular, the variability in deposition velocities across models, as discussed in Sect. 3.3.1, is seen as an originator leading to uncertainties in tropospheric O3 (Wild, 2007). Deposition velocities for the models in the present study (Fig. S19) show the lowest velocities for CMAQ. Highest velocities were found for CAMx over land areas. The deposition velocities go along with the annual dry deposition, with high velocities in areas with high dry deposition.

A model comparison study with 15 models by Hardacre et al. (2015) found the greatest differences in total O3 dry deposition occurring in areas where deposition velocities and O3 concentrations are highest. Additionally, soil moisture has an important impact on O3 deposition and concentration. An evaluation study within the CHIMERE model found that especially in southern Europe, where soil is close to the wilting point during summer and affects stomatal opening, O3 dry deposition declines (Anav et al., 2018). This in turn affects the concentration of gases in the lower atmosphere and thus has an impact on O3 concentrations.