We use the Weather Research and Forecasting model (v4.1.1) coupled to chemistry (Chem) and optimized for Polar regions . Polar-WRF-Chem (hereafter WRF) is a non-hydrostatic mesoscale numerical weather prediction and atmospheric chemistry model used for operational and research purposes. Figure  shows the selected study area including the locations of surface O3 observational sites selected for this study (more information in Sect. ). WRF is set up with a polar projection centered at 90∘ N, 250 × 250 horizontal grid points (30 × 30 km resolution) and 44 vertical levels up to 100 hPa, with a finer vertical grid spacing in the ABL and lower troposphere. The simulation period is 8 August to 7 September 2008 including 3 d of spin-up. This end-of-summer 2008 period is chosen (1) to limit the role of active halogen chemistry during springtime and (2) the additional availability of O3 observations in the High Arctic over sea ice from the Arctic Summer Cloud Ocean Study (ASCOS) campaign . The ECMWF ERA5 meteorology (0.25∘ × 0.25∘) and CAMS reanalysis chemistry (0.75∘ × 0.75∘) products are used for the initial and boundary conditions. Boundary conditions, SSTs and sea ice fractions are updated every 3 h to these reanalysis products to allow for the sea ice retreat during the simulation. Other relevant parameterization schemes and emission datasets have been listed in Table  and are mostly based on .

The first WRF simulation, without any adjustments to O3 deposition, indicated that WRF was misrepresenting the temporal variability in surface O3 observations, most prominently starting from a few days into the simulation. We hypothesize that this misrepresentation is caused by deviations in the synoptic conditions in the free-running WRF simulation. This was confirmed with a comparison of simulated and satellite observed wind speeds above oceans at a spatial resolution of 0.25∘ × 0.25∘ . To overcome the impact of this deficiency on our O3 study, nudging is applied to ensure an optimal model evaluation with observations. Hence, WRF is nudged every 3 h to the ECMWF ERA5 specific humidity, temperature and wind fields in the free troposphere with nudging coefficients of 1 × 10-5, 3 × 10-4 and 3 × 10-4 s-1, respectively.

The Coupled Ocean-Atmosphere Response Experiment (COARE) has been developed to study physical exchange processes (sensible heat, latent heat and momentum) at the ocean–atmosphere interface. Later, COARE has been extended to include the exchange of gaseous species such as O3, dimethyl sulfide (DMS) and carbon dioxide (CO2) . Many studies have used the COARE Gas transfer algorithm (COAREG) in combination with eddy-covariance measurements to study the effects of wind speed and sea state on ocean–atmosphere gas exchange (e.g., ). Furthermore, the COAREG algorithm has also been previously used in global O3 modeling studies . The choice for COAREG is further motivated by the consistent coupling with other species such as DMS.

Here we use COAREG version 3.6, which is extended with a two-layer scheme for surface resistance compared to the previous version described by . The two-layer scheme is similar to building upon a first application of a one-layer version of COAREG by . In that study, chemical enhancement of ocean O3 deposition by its reaction with iodide was considered using a global climatology of ocean surface water concentrations of nitrate serving as a proxy for oceanic iodide concentrations (Iaq-). Besides nitrate, satellite-derived chlorophyll-α concentrations have been used as a proxy for Iaq- . Since then, alternative parameterizations of oceanic Iaq- have been proposed e.g., using SST as a proxy for this reactant. In COAREG, chemical reactivity of O3 with Iaq- is present through the depth of the oceanic mixing layer. O3 loss by waterside turbulent transfer is negligible in the top water layer (few micrometers), but is accounted for in the underlying water column. The waterside turbulent transfer term is especially relevant for relatively cold waters because the chemical enhancement term is then relatively low . The last two important waterside processes that determine the total O3 deposition are molecular diffusion and solubility of O3 in seawater which both depend on the SST. In Appendix  we list the formulation of the air side and waterside resistance terms in the COAREG routine applied in this study and show the sensitivity to the environmental factors wind speed, SST and Iaq- for typical Arctic conditions.

The COAREG algorithm is coupled such that WRF provides the meteorological and SST input for the COAREG routine. In turn, the COAREG calculated ocean–atmosphere exchange velocities are used in the WRF model to calculate the oceanic O3 deposition flux replacing the default oceanic O3 deposition fluxes calculated by the scheme reflecting use of the default constant rs of 2000 s m-1. For grid boxes with fractional sea ice cover, COAREG replaces the Wesely deposition scheme for the fraction that is ice free. Note that in this study, only O3 ocean–atmosphere exchange is represented by COAREG not having modified simulations of ocean–atmosphere exchange of other compounds (e.g., DMS).

Moreover, we apply the monthly mean Iaq- distribution by (0.125∘ × 0.125∘ resolution) which applies a machine learning approach, namely the random forest regressor algorithm , using various physical and chemical variables such as SST, nitrate, salinity and mixed layer depth. This distribution replaces the previously applied Iaq- estimations only using SST . At high latitudes, these Iaq- distributions are highly uncertain due to the limited number of observations. The choice for is motivated by the most accurate representation of observed Iaq- by the introduction of other predictors besides SST. Furthermore, this product will be further updated with newly available measurements. Figure  shows the spatial distribution of Iaq- used in the calculation of the O3 deposition velocities. Using the distribution for August/September we found Iaq- concentrations ranging between 30 and 80 nM for the open oceans up to 130 nM in coastal waters. In and , Iaq- is solely a function of SST which leads to Iaq- in the order of 5 to 50 nM and thus low reactivity and O3 deposition velocities.

Reported atmosphere–snow gas exchange spans a wide range of observed O3 deposition velocities. Some studies even report episodes of negative deposition fluxes (emissions) over snow or sea ice . recently summarized observed O3 deposition velocities to snow having a range of -3.6 to 1.8 cm s-1 with most of the observations indicating a deposition velocity between 0 and 0.1 cm s-1 for multiple snow-covered surfaces (e.g., grass, forest and sea ice). Generally, O3 concentrations in the interstitial air of the snowpack are lower than in the air above making the snowpack not a direct source of O3 in terms of emissions . However, the emissions of O3 precursors from the snowpack can enhance O3 production in the very stable atmosphere above the snowpack . investigated the sensitivity of a global chemistry and tracer transport model to the prescribed O3 deposition velocity and found the best agreement between modeled and observed O3 concentrations at four Arctic sites by applying deposition velocities in the order of 0.00–0.01 cm s-1. Following we have increased the O3 surface uptake resistance (rs) for snow and ice land use classes to 104 s m-1. This corresponds to total deposition velocities of ≤ 0.01 cm s-1, which is a reduction of ∼ 66 % compared to the Wesely deposition routine that is the default being applied in WRF .

The new modeling setup, including nudging to ECMWF ERA5 and the revised O3 deposition to snow, ice and oceans, is evaluated against observational data of pan-Arctic surface O3 concentrations. We expect that the different representation of O3 deposition mostly affects O3 concentrations in the ABL. Therefore, we evaluate our simulations against hourly averaged surface O3 observations from 25 measurement sites above 60∘ N. These sites are further categorized in three site selections: “High Arctic”, “Terrestrial” and “Remote”. High Arctic refers to sites having latitudes > 70∘ N and for which we expect that the deposition footprint is a combination of ocean and sea ice e.g.,. The Terrestrial sites are located below 70∘ N and show a clear diurnal cycle in observed O3. Sites are characterized as Terrestrial when the average observed minimum nighttime mixing ratio is > 8 ppb smaller than the average observed maximum daytime mixing ratio during the ∼ 1 month of simulation. This criterion is based on a preparatory analysis of the observational data, footprint and site characteristics. The Remote sites have been identified as such based on their location below 70∘ N and showing no clear diurnal cycle in O3 concentrations. The analysis also includes the observations during the Arctic Summer Cloud Ocean Study (ASCOS) campaign, when the icebreaker Oden was located in the Arctic sea ice . In total, 25 surface O3 measurement sites are included (Fig. ), of which 6, 8 and 11 sites are characterized as High Arctic, Remote and Terrestrial sites, respectively. A full list of available measurement sites is available in Table .

In total, we perform two simulations. The first WRF simulation (NUDGED) is a run with the setup described in Sect.  and nudged with the synoptic conditions to the ECMWF ERA5 product as described in Sect. . The second simulation (COAREG) includes also includes the adjustments to the O3 deposition to oceans as described in Sect.  and the O3 deposition to snow and ice as described in Sect. . Furthermore, we also compare our results with the state-of-the-art CAMS global reanalysis data product . This product has a temporal resolution of 3 h, a spatial resolution of 0.75∘ × 0.75∘ and does not include a mechanistic representation of ocean–atmosphere O3 exchange. CAMS assimilates satellite observations of O3 but it does not assimilate O3 observations from radiosondes or in situ measurement sites such as the 25 sites used in the evaluation presented here. This implies that lower-tropospheric O3 is weakly constrained by observations in this CAMS product making an accurate model representation of the sources and sinks important. We opted to include the CAMS reanalysis data as another tool to study Arctic surface O3 and to address potential limitations in its model setup. Moreover, CAMS is being widely used for air quality forecasts and assessments but also to constrain regional-scale modeling experiments such as presented in this study. Therefore, an analysis of the performance of the CAMS reanalysis data might also benefit future Arctic air quality assessments.

Figure a and b show the mean deposition velocities for the NUDGED and COAREG runs, respectively. As expected, in the NUDGED run (Fig. a) the mean Vd,O3 to oceans is in the order of 0.05 cm s-1. Furthermore, the spatial distribution shows a relatively low heterogeneity and no increase in deposition velocities towards the warmer waters. The COAREG run (Fig. b) provides a mean Vd,O3 in the order of 0.01 cm s-1 for the Arctic Ocean > 70∘ N up to 0.018 cm s-1 for oceans with high Iaq- concentrations (Fig. ). Simulated oceanic O3 deposition is elevated in coastal waters (e.g., Baltic Sea and around the Bering Strait) with Iaq- concentrations reaching up to 130 nM compared to 30–50 nM for the open Arctic Ocean waters (Fig. ). This highlights the sensitivity of the COAREG scheme to chemical enhancement with dissolved iodide.

Figure c shows the temporal variability in Vd,O3 for one of the grid boxes, which is in terms of temporal variability representative of the whole domain. The temporal variability in the NUDGED run is mainly governed by temporal variability in ra. During episodes with high wind speeds (> 10 m s-1), ra becomes so small that it is negligible over the constant surface uptake resistance of 2000 s m-1, corresponding to a maximum Vd,O3 of 0.05 cm s-1. During episodes with low wind speeds (< 5 m s-1), reduced turbulent transport poses some additional restriction on O3 removal with increasing ra, which reduces the Vd,O3 to ∼ 0.04 cm s-1. In the COAREG run, temporal variability in Vd,O3 is also governed by wind speeds that control the waterside turbulent transport of O3 in seawater besides atmospheric turbulent transport. For high wind speeds, the waterside turbulent transport increases (Fig. ) and more O3 is transported through the turbulent layers. For our simulation, we found that the temporal variability in O3 deposition due to waterside turbulent transport can be up to ±20 % around the mean. Only during episodes of very low wind speeds (< 2.5 m s-1) does the ra+rb term pose an additional restriction on O3 deposition in the COAREG run. Overall, the Vd,O3 to oceans in the COAREG run is reduced by ∼ 60 %–80 % compared to the NUDGED run. The mean Vd,O3 to snow and ice is reduced by ∼ 66 %, from ∼ 0.03 cm s-1 in the NUDGED run to ∼ 0.01 cm s-1 in the COAREG run.

The temporal evolution in oceanic O3 deposition velocities simulated by the COAREG run appears to be on the low side of observed Vd,O3 and of that simulated elsewhere e.g.,. showed that Vd,O3 can increase by a factor of 5 with wind speed increasing from 0 to 20 m s-1. (their Fig. 7) shows a wide range of observed and simulated sensitivities to wind speed. Observations from the TexAQS06 summer campaign in the Gulf of Mexico show a large sensitivity to 10 m wind speeds even though the model seems unable to capture these high deposition velocities at high wind speeds . However, also shows that for the GasEx08 campaign in the cold Southern Ocean the sensitivity of observed and simulated Vd,O3 to 10 m wind speeds is very limited. This limited sensitivity is most accurately represented by the modified two-layer reactivity scheme compared to the older one-layer scheme due to a more limited interaction between chemical reactivity and waterside turbulent transport . Furthermore, the variability around the mean presented in Table  (0.012 ± 0.002 cm s-1) seems to correspond to the (0.016 ± 0.0015 cm s-1) 1-month simulation including O3 removal by Iaq-. In this study we show the intra-monthly variability in oceanic O3 deposition, which is expected to be relatively low compared to the seasonal variability which will also be driven by temporal changes in solubility and reactivity due to the seasonal changes in SST and Iaq-.

By estimating the total deposition flux for the water, snow/ice and land surfaces we can quantify the total simulated O3 deposition budget (Table ) for the Arctic modeling domain. Land, not covered with snow or ice, is the dominant surface type for this specific domain setup in summer with 48 %. Combined with a relatively high simulated Vd,O3 of ∼ 0.45 cm s-1, this is the most important sink, in terms of deposition, of simulated O3 with ∼ 135 Tg O3 yr-1. The simulated O3 deposition budget to water bodies, covering 37 % of the total surface area, contributes ∼ 10 % in the NUDGED run (15.4 Tg O3 yr-1) to the total O3 deposition sink. In the COAREG run, this reduces to only ∼ 3 % (4.6 Tg O3 yr-1) of the total O3 deposition sink. Simulated O3 deposition to snow and ice, covering 15 % of the total surface area, is the least important deposition sink removing 4.1 and 1.7 Tg O3 yr-1 in the NUDGED and COAREG runs, respectively.

Figure  shows the spatial distribution in the simulated mean surface O3 mixing ratios overlain with the observed mean surface O3 mixing ratios. In the NUDGED and COAREG runs (Fig. a and b, respectively) we find similar surface O3 mixing ratios of ∼ 15–20 ppb over the Russian, Canadian and Alaskan landmasses. Over Scandinavia, slightly higher surface O3 mixing ratios of ∼ 20–25 ppb are simulated due to more anthropogenic emissions of precursors in the EDGAR emission inventory and advection of O3 and its precursors from outside the domain. As expected, we find a limited effect of reduced deposition to water and snow/ice to the simulated mean O3 mixing ratios over land. In general, the model appears to simulate the mean observed surface O3 mixing ratios for the Remote and Terrestrial sites (all sites < 70∘ N) generally well without clear positive or negative bias. Due to the altitude effect, relatively high surface O3 concentrations are simulated over Greenland even though the deposition velocity to snow and the surrounding oceans is of similar magnitude (∼ 0.01 cm s-1).

The reduced O3 deposition to water and snow/ice surfaces, comparing the NUDGED and COAREG simulation results (Sect. , Table ), appears to be limited in terms of relative changes in Vd,O3 and the total simulated O3 deposition budget. However, these relatively small changes do substantially affect the simulated spatial distribution of surface O3 mixing ratios over oceans and sea ice as indicated in Fig. . We find that the NUDGED run (Fig. a) systematically underestimates the mean observed surface O3 mixing ratios for the High Arctic sites (all sites > 70∘ N) by ∼ 5–10 ppb, which appears to be caused by an overestimated deposition to ocean, snow and ice surfaces, also further substantiated by the following analysis of temporal variability in O3 concentrations (Sect. ). Over the Arctic sea ice and oceans the ABL is typically very shallow and atmospheric turbulence is relatively weak. This suppresses vertical mixing and entrainment of O3-rich air from the free troposphere. Dry deposition of O3 to the ocean or snow/ice surfaces appears to be an important removal mechanism that has a large impact on O3 concentrations in these shallow ABLs both in terms of magnitude but also temporal variability (see Sect. ). In the COAREG run, surface O3 mixing ratios over oceans and Arctic sea ice have increased by up to 50 %. Furthermore, the reduced deposition to snow/ice has also clearly affected simulated surface O3 mixing ratios over Greenland. Most importantly, the negative bias in simulated surface O3 mixing ratios is reduced in the COAREG run with respect to the NUDGED run (see Sect. ).

In this section we show how the application of the revised deposition scheme improves the model prediction scores of surface O3 concentrations reflected in a comparison of the simulated and observed hourly surface O3 mixing ratios at the three site selections (High Arctic, Remote and Terrestrial). To our knowledge, this is the first time such an oceanic O3 deposition scheme coupled to a meteorology–chemistry model has been evaluated against a large dataset of hourly surface O3 observations. Figure  shows a comparison between observed and simulated hourly surface O3 mixing ratios subdivided into the three site selections: High Arctic, Remote and Terrestrial. As expected, for the High Arctic sites (Fig. , top row) we find that the NUDGED run is underestimating the observed surface O3 mixing ratios with a mean bias of -3.8 ppb, which is also consistent with the findings in Fig. , where the NUDGED run appears to underestimate surface O3 mixing ratios in the High Arctic region. The COAREG run, having a reduced O3 deposition sink to oceans and snow/ice appears to better represent the surface O3 observations with a slight positive bias of 0.3 ppb. The mean absolute error (MAE) in the COAREG run is reduced to 4.7 ppb from 6.4 ppb for the NUDGED run. Furthermore, we find that the CAMS reanalysis data also underestimate surface O3 in the High Arctic with a bias of -5.0 ppb and an MAE of 6.8 ppb. Note that the performance for the WRF runs and CAMS reanalysis product varies for each observational site, which is further examined in Sect. .

For the Remote sites (Fig. , middle row), having no clear diurnal cycle in surface O3, we again find an improvement by including the mechanistic ocean deposition routine and reduced snow/ice deposition. This improvement appears to be most pronounced for coastal sites like Stórhöfði (63.4∘ N, 20.3∘ W) and Inuvik (68.4∘ N, 133.7∘ W) with a reduction in the MAE of 32 % and 19 %, respectively (not shown here). Overall, the improvement for the COAREG compared to the NUDGED run in the Remote site selection is not as significant compared to the High Arctic sites, also because of the larger role of O3 deposition to land and vegetation, which remained unchanged in this study. We find that the CAMS data show the best performance for the Remote sites with no bias and with an MAE of 5.6 ppb.

For the Terrestrial sites (Fig. , bottom row), having a clear diurnal cycle in surface O3, the WRF runs slightly overestimate the observed surface O3 mixing ratios with mean biases of 0.1 and 1.0 ppb for the NUDGED and COAREG runs, respectively. Reducing the O3 deposition to oceans and snow/ice increases the bias, but the MAE of 6.0 ppb remains unchanged. The CAMS reanalysis data appear to perform worst for the Terrestrial sites with a bias of 6.4 ppb and an MAE of 8.0 ppb. This might be explained by the lower spatial and temporal resolution of CAMS specifically at these sites having a relatively strong diurnal cycle in ABL dynamics, O3 deposition to vegetation and O3 concentrations. Also a misrepresentation of emissions of precursor emissions and concentrations and the O3 deposition to vegetation might explain some of the differences.

In Sect.  we have shown how revising the O3 deposition scheme to oceans and snow/ice can improve the model's capability to represent the observed hourly surface O3 mixing ratios, especially for the High Arctic sites. In this section we show how the NUDGED and COAREG runs and CAMS represent the temporal variation in High Arctic surface O3 observations, focusing on 6 out of the 25 measurement sites. These six High Arctic sites have been selected due to their deposition footprint being dominated by transport over, and deposition to, ocean and sea-ice-covered surfaces. Figure  shows the observed and simulated surface O3 time series for ASCOS, Summit, Villum, Zeppelin, Barrow and Alert. Furthermore, Table  shows the model skill indicators for the High Arctic sites. These skill indicators include the mean absolute error (MAE) that represents the systematic error, the standard deviation of observation minus model prediction σo-p that represents the random error, and the Pearson R correlation coefficient (R) that represents the degree of correlation.

The observations at ASCOS (Fig. a) show a sudden increase in surface O3 mixing ratios from 20 to over 30 ppb around the 17 August due to advection of relatively O3-rich air during a synoptically active period . Only the COAREG run appears to be able to simulate a similar increase in surface O3, while NUDGED and CAMS show a minor increase in simulated surface O3. From the 17 August onwards, the observations show mixing ratios between 25 and 35 ppb. The WRF simulations indicate advection of air over ocean and ice surfaces during this time period (not shown here). In the COAREG simulation, with less deposition to these surfaces, surface O3 mixing ratios are less depleted. Only the COAREG run is able to represent these observed mixing ratios with a bias of -2.0 ppb, whereas NUDGED and CAMS are clearly biased towards lower mixing ratios.

At Summit (Fig. b), we find a large temporal variability in observed surface O3 between 30 and 55 ppb. From the 11 August onwards we find a decreasing trend in observed surface O3 down to 30 ppb before increasing to 40 ppb around the 17 August. All models capture this specific event in terms of temporal variability even though NUDGED and COAREG are still biased at the observed minimum of 30 ppb. Furthermore, we find that the CAMS reanalysis data represent this specific period very well, also in terms of magnitude. At Summit, the increase in surface O3 in the COAREG run relative to the NUDGED run mostly reflects the reduction in deposition to snow and ice due to the prevailing katabatic wind flow . During episodes with low wind speeds the ABL becomes very stable and shallow during which deposition to snow and ice becomes an important process in removing O3 in the ABL. In the period between the 14 and 26 August this reduction in deposition can increase the surface O3 mixing ratios of up to 10 ppb (e.g., 23 August). In contrast, during episodes with higher wind speeds and deeper ABLs the reduced O3 deposition to snow hardly affects the simulated surface O3 concentrations. Interestingly, we find that the NUDGED and COAREG simulations show a larger negative bias (∼ 5–10 ppb) during the period with low wind speeds and shallow ABLs. Over the entire simulated period, CAMS performs best at Summit, with an MAE of 3.9 ppb, followed by COAREG, with an MAE of 6.1 ppb.

Villum (Fig. c) is the only site for which the NUDGED and COAREG runs as well as the CAMS reanalysis data all systematically overestimate the observed mixing ratios, especially later into the simulation. The observations show an increase in O3 mixing ratios from 10 to 20 ppb in the first 3 d of the simulation, whereafter it remains between 20 and 30 ppb with relatively low temporal variability compared to some of the other sites (e.g., Summit, Barrow). Both the NUDGED and COAREG runs simulate mixing ratios of up to 40 ppb, and CAMS simulates maximum surface O3 mixing ratios of 35 ppb. In terms of representing the magnitude of surface O3 mixing ratios CAMS performs best with an MAE of 4.5.

Zeppelin (Fig. d) and Barrow (Fig. e) show similar behavior in terms of observation–model comparison. For both locations the CAMS reanalysis data systematically underestimate observed O3 mixing ratios with a biases > 10 ppb. In the NUDGED run the bias equals -6.9 and -4.6 ppb for Zeppelin and Barrow, respectively. In the COAREG run the bias is reduced to -1.0 and -0.2 ppb for Zeppelin and Barrow, respectively. This reduction in bias is, together with ASCOS, the largest among the six High Arctic sites and shows the large sensitivity to the representation of O3 deposition. At Barrow, the dominant wind directions during the simulation period are NW–NE reflecting a footprint mostly from the Arctic sea ice and ocean. Especially in the period from the 23 August onward, the COAREG run is very accurate in representing the magnitude as well as the temporal variability in observed surface O3. During this period, the NUDGED run simulates surface O3 mixing ratios of up to 5 ppb lower due to the overestimated deposition to oceans and sea ice. At both sites, the model performance of COAREG is in the same order of magnitude, with an MAE, σo-p and R of 3.5 ppb, 4.2 ppb and 0.65, respectively.

At Alert (Fig. f), we find a relatively steady increase in observed surface O3 from 20 ppb at the start of the simulation to 30 ppb at the end of the simulation. The temporal variability, both in observed and simulated surface O3, appears to be lower compared to some of the other High Arctic sites. Again, the statistical parameters such as MAE, σo-p and R improve in the COAREG run with respect to the NUDGED run. At Alert, we find that CAMS has the lowest MAE and σo-p of 3.0 and 3.4 ppb, respectively.

The model performance in terms of temporal variability in surface O3 observations is diagnosed by using the Pearson R correlation coefficient. The model performance improved for all six sites in the COAREG run with respect to the NUDGED run. The COAREG simulation performs best for five out of the six observational sites in terms of Pearson R correlation coefficient and is only outperformed by CAMS at Summit. Overall, we find that coupling the WRF model to the mechanistic COAREG ocean–atmosphere exchange representation decreases the MAE and σo-p for all High Arctic sites except for Villum by better representing the magnitude of, but also temporal variability in observed surface O3. The CAMS reanalysis data perform well for some locations (e.g., Summit, Alert), while for Zeppelin and Barrow the discrepancy is among the largest we found in the observation–model comparison.