The physical area (displayed in Fig. ) of 5040km×4960km, centered north of Utqiaġvik is modeled for the time interval of 1 February through 1 May 2009, for which Global Ozone Monitoring Experiment-2 (GOME-2) data with a stratospheric correction for BrO VCDs  as well as surface ozone and ozone sonde data are available for model evaluation.

The software Weather Research and Forecasting model coupled with Chemistry (WRF-Chem) version 3.9 is employed. WRF-Chem  is a state-of-the-art regional numerical weather prediction system with online computation of chemistry. Table  summarizes the configuration of the software. The physics modules are chosen following recommendations of the polar WRF community ; the modules include the meteorology and the emission, transport, mixing, and chemical reactions of trace gases as well as aerosols.

The simulation domain is centered north of Utqiaġvik using the polar stereographic projection at a true latitude of 83∘ with a reference longitude of 156∘ W. A horizontal grid resolution of 20 km for the 5040km×4960km domain is employed, allowing a comparison to GOME-2 BrO satellite data  with a resolution of approximately 40km×30km. In the vertical direction, 64 non-equidistant grid cells with a finer resolution near the ground are used, starting with approximately 25 m at the ground level. Half of the grid cells used in the present study are in the first 2 km of the atmosphere, allowing a detailed representation of the Arctic boundary layer. The vertical grid is provided in the Supplement of this paper.

The meteorological time step of 1 min is chosen to fulfill the Courant criterion. Chemistry is updated between every meteorology time step, and radiative transfer is updated every 10th meteorological time step.

In the present model, the Mellor–Yamada–Janjić (MYJ) planetary boundary layer (PBL) scheme  is employed, which is a 1.5-order local turbulence closure model. Prognostically determined turbulent kinetic energy is used to determine the eddy diffusion coefficients. The MYJ PBL scheme is best suited for stable to slightly unstable conditions .

WRF-Chem offers several implementations of chemical reaction schemes. In the present study, the MOZART–MOSAIC mechanism based on MOZART-4 gas-phase chemistry  is used, which includes 85 gas-phase species, 237 gas-phase reactions, and 49 photolysis reactions. An additional 18 gas-phase species, 73 gas-phase reactions, and 13 photolysis reactions  account for the bromine and chlorine chemistry (termed “full chemistry”, see Table ). Observations of reactive iodine in the Arctic region  suggest only low mixing ratios of iodine. Even though small mixing ratios of iodine can significantly enhance ozone depletion , iodine is neglected due to the uncertainties in the abundance of iodine in the Arctic atmosphere and in snowpacks. The photolysis rates are calculated with the “Updated TUV” (Tropospheric Ultraviolet–Visible) scheme , which already contains the halogen photolysis reaction rates. The added bromine and chlorine chemical reactions are provided in the Supplement.

The MOZART–MOSAIC mechanism employs four-bin MOSAIC aerosols . In WRF-Chem, MOSAIC is implemented using a sectional approach, where size bins are defined by the upper and lower dry particle diameters. In MOSAIC, the mass and number density for each bin are considered, and the processes of nucleation, coagulation, condensation, evaporation, and aerosol chemistry are modeled. The mass transfer rate ki,m for gas species i and aerosol size section m is calculated using the following parameterization : 1ki,m=4πRp,mDg,iNmf(Knm,γi), where Dg,i is the gas diffusivity of species i, Rp,m is the wet mean particle radius of size bin m, Nm the number density of size bin m, and Knm=λ/Rp,m is the Knudsen number of size bin m with the free mean path λ. f(Knm,γi) is the transition regime correction factor and accounts for the interfacial mass transport limitation: 2f(Knm,γi)=0.75γi1+KnmKnm1+Knm+0.283γiKnm+0.75γi, where γi is the accommodation coefficient for gas-phase species i taken from the CAABA/MECCA (Chemistry As A Boxmodel Application/Module Efficiently Calculating the Chemistry of the Atmosphere) model . Aerosol forms of bromine are currently not implemented in the MOSAIC framework and are treated as gas-phase species. The transfer reactions of bromine gas-phase species X to aerosol-size bin m are assumed to produce species Xaq,m as R17HBr(g)→HBrm(aq),R18HOBr(g)→HOBrm(aq),R19BrONO2(g)→HOBrm(aq)+HNO3(g), which may produce gas-phase Br2 : R20HOBrm(aq)+HBrm(aq)→Br2(g). Reactions ()–() may only occur if the aerosol is in a liquid state, and, in addition, Reaction () requires the aerosol to have a pH of 6 or less. The heterogeneous reactions and parameters required to calculate the reaction rates are listed in the Supplement. Heterogeneous BrCl production (Reactions  and ) is not implemented in the model.

Emissions of bromine species on snow surfaces are parameterized following . Numerically, bromine emissions are coupled to vertical diffusion. In WRF-Chem, vertical (turbulent) diffusion for each species and horizontal grid cell is solved using a Peaceman–Rachford alternating direction implicit method . The bromine emissions are added as boundary conditions to the tridiagonal diffusion matrix. For the surface emission in Reactions (), (), and (), the boundary flux for instance of Reaction (), Fd(Br2|HOBr) for Br2 due to HOBr is 3Fd(Br2|HOBr)=βρd,0vd(HOBr)HOBr0, where ρd,0 is the dry air density of the lowest grid cell and HOBr0 is the HOBr mixing ratio in the lowest grid cell. The species-dependent deposition velocity vd≈1 cms-1 is calculated using the WRF-Chem Wesely deposition module  under an additional assumption of near-zero surface resistance. Thus, the turbulent transfer resistance dominates the deposition velocity, and the bromine emissions increase with larger wind speeds. β≥1.0 is the reactive surface ratio  of the ice or snow surface, accounting for non-flat surfaces such as ice or snow and frost flowers. For simplicity, β is set as a global value in this study to allow for the investigation of the strength of bromine emissions in a parameter study. For the direct emission of bromine due to ozone oxidation of bromide (see Reaction  above), the factor α is used to control the emission probability: R21O3→αBr2 and 4Fd(Br2|O3)=αβρd,0vd(O3)O30. The value of α is parameterized with a dependence on the solar zenith angle (SZA) : 5α(SZA)=0.1%if SZA>85∘7.5%otherwise. The deposition velocity for ozone is dominated by the surface resistance , leading to vd(O3)≈0.01 cms-1. An emission mechanism relating to the bromide oxidation by the hydroxyl radical (see Reaction ) is currently not implemented in the model. All sea ice is assumed to be snow covered for the simulated time range. On snow covering FY ice, it is assumed that the bromide content is infinite, so that unrestricted gaseous bromine emissions are possible, and emissions of Br2 due to O3 and N2O5 depositions are only active on snow covering FY ice. On snow covering MY ice, no bromide content but infinite chlorine is assumed. HOBr depositions only release Br2 up to the combined depositions of gaseous and aerosol HBr, whereas excess HOBr depositions release BrCl. On snow-covered land, neither bromide nor chloride content is assumed, so that excess HOBr depositions are lost. A list of the depositions and emissions added to the MOZART mechanism can be found in the Supplement. Br2 production from the sunlit condensed phase without any depositions of gas-phase species as found under certain conditions by and as well as possible oceanic emissions of very short-lived brominated species are currently not considered in the model.

ERA-Interim  is used to generate both the initial and boundary meteorological and sea ice cover data. ERA-Interim was found to perform well in polar regions in various studies e.g., and was successfully used in various modeling studies in polar regions e.g.,, which is why it was chosen in the present study. Nudging of temperature, horizontal wind speed, humidity, and surface fields to ERA-Interim data ensures the validity of the simulation meteorology over the simulated 3-month period. The idea of the present work is not to try to make meteorological predictions (which would not be meaningful anyway on the timescale of a few months) but rather to model chemistry under meteorological conditions prevailing over a particular period of time. Nudging is active for the entire duration of the simulation and is inactive inside the boundary layer. The nudging timescale is set to 1 h. MOZART-4 results driven by Goddard Earth Observing System (GEOS-5) meteorological fields are used as initial and boundary data for all non-halogen species . For most halogen species, initial and boundary conditions are set to near-zero values. The initial mixing ratio of HBr and Br2 is set to 0.3 ppt in the lowest 200 m of the atmosphere. The mixing ratio of CHBr3 is fixed to 3.5 ppt . The bromide oxidation of ozone in the dark for an ozone deposition velocity of 0.01 cms-1, a boundary layer height of 200 m, an emission probability of Φ=0.001, and 40 nmolmol-1 ozone will release approximately 2 pmolmol-1 Br2 on FY ice per day. This emission rate is assumed to prevail for all simulations with active halogen chemistry. The chosen initial halogen concentrations and the fixed mixing ratio of CHBr3 are thus irrelevant. The RTG_SST (where RTG stands for real-time global) high-resolution dataset is used for the sea surface temperature (SST). The present model differentiates between FY and MY sea ice in order to estimate bromine emissions. For this purpose, the OSI-403-c (where OSI stands for ocean and sea ice) sea ice type dataset   is used. The original dataset does not provide values for latitudes larger than about 88∘ due to a lack of satellite measurements for these latitudes. In the present study, these values are filled with first-year sea ice. Figure  shows the simulation domain and the locations of FY and MY sea ice. Grid cells with a mixed FY–MY sea ice type are treated as multi-year sea ice in the bromine emission mechanism described above. Sea ice cover, SST, and sea ice type are updated online during the numerical simulations. EDGAR-HTAP and MEGAN are used as anthropogenic emissions and bio-emissions, respectively.

The conducted simulations are summarized in Table . Five different observational datasets are used for comparison to the simulation results:

ground-based in situ ozone measurements at Utqiaġvik, Alaska, and Summit, Greenland ;

For comparison of the observations and the simulations, three different statistical parameters are used. For model variable M and the corresponding observation variable O, the Pearson correlation R, the mean bias MB, and the root mean square error RMSE are calculated by 6R=M-MO-OσMσO,7MB=M-O,8RMSE=M-O2, where is the mean and σM and σO denote the standard deviations of M and O, respectively.

The NOAA and ESRL (Earth System Research Laboratories) Global Monitoring Division Surface Ozone measurements near Utqiaġvik and Summit are compared to the simulation results for the numerical grid cell closest to the observation site under consideration where the numerical results in the lowest grid cell are used. The temperature at 2 m, wind speed, and wind directions at 10 m of the Barrow Atmospheric Baseline Observatory  are compared to the corresponding simulated surface fields.

Figure  shows simulated and observed temperatures, T, at 2 m height and wind speeds u at 10 m height at Utqiaġvik. Simulations 1–5 share the meteorology shown in the left of Fig. , whereas results of simulation 6 with deactivated meteorological nudging are shown in the right of Fig. . The first 11 d in February are very cold, reaching temperatures as low as -40 ∘C, and the wind speed is very low during this period of time, which is likely to inhibit BrO emission due to the wind dependence of the emission. Both the wind speed and the temperature increase during the following 3 weeks: wind speeds increase to values up to 16 ms-1 and temperature reaches up to -5 ∘C. On 21 and 23 February and 1 March, wind speed is notably underpredicted by the model with nudging. Both temperature and wind speed vary strongly during that time. From mid March onwards, temperature increases gradually with fewer day-to-day variations compared to the previous weeks. Simulations 1–5 predict both temperature and wind speed very well during this time period with the exception of underpredictions of wind speed occurring on 16–17 March and at the end of April. Simulation 6 produces higher errors in the second half of the simulation where temperature is consistently too large by several degrees in April and overpredictions of wind speed on 18–22 March and 22 and 29 April. The results of simulation 6 appear not to be very realistic.

Figure  shows the correlation of the observed (vertical axis) and the modeled (horizontal axis) temperatures, where a correlation of unity applies if the data lie on the diagonal marked in the figure. Shown in blue is the regression line, for which the observed and measured variables are assumed to be the independent and dependent variables, respectively. The results of the entire simulation period are displayed, where the first week should be regarded as spin-up period. For simulations 1–5, there is an overestimation of the temperature when it is cold, which is likely due to the lowest temperatures occurring during the spin-up time during which the modeling errors are larger compared to other times. ERA-Interim is known to have a warm bias for temperatures below -25 ∘C , which may also explain the deviations. Simulations 1–5 perform well throughout the simulation in contrast to simulation 6 with no nudging. In simulations 1–5, a maximum deviation in temperature of about 8 ∘C occurs, and in simulation 6, a stronger temperature difference of up to 20 ∘C is observed. The statistical parameters (see Eq. ), at Utqiaġvik for the entire time range are shown in Table . The simulations with nudging perform better in all regards, emphasizing the necessity of data assimilation. Temperature is predicted best with almost perfect correlation and relatively small mean bias and RMSE. Temperature is overpredicted in all simulations by approximately 0.55 and 1.71 ∘C for simulations 1–5 and 6, respectively. Colder temperatures are generally favorable for ODEs, both by changing the boundary layer configuration and affecting chemical reaction constants, which could result in an underestimation of ODEs. Both wind speed and direction are predicted less accurately, which might result in wrong source locations or times of the occurrence of ODEs; this is likely to explain some of the differences between simulations and observations. Wind speed is underestimated on average by about 0.52 and 0.66 ms-1 for simulations 1–5 and 6, respectively, which may contribute to a slight underestimation of bromine emissions due to the dependence of the deposition velocity on wind speed. The Barrow Meteorological Station (BMET) Handbook mentions an instrument accuracy of 0.17 ms-1 for wind speeds between 0.4 and 75 ms-1, a 5.6∘ wind direction resolution, and 0.25 ∘C instrument accuracy for temperatures between -65 to -20 ∘C. The RMSE is at least 1 order of magnitude higher than the mentioned instrument accuracies and resolutions for all simulations, so that the errors of the observations can be neglected in comparison to the model errors.

Figure  shows modeled and observed surface ozone and BrO at Utqiaġvik and at Summit. Only results of simulations 1 and 3 are shown for visual clarity. Figure S1 in the Supplement displays ozone mixing ratios modeled by simulations 1–4 and 6. The correlations of modeled and observed ozone can be seen in Fig. . Statistics are summarized in Table . Simulations 2–5 perform considerably better than simulation 1 for which halogen chemistry is turned off. Simulation 3 with enhanced emission performs best, with the correlation increasing from -0.31 to 0.644 compared to simulation 1. Quite a few ODEs are not captured by simulation 4, for which the emission probability for bromine emissions due to ozone under sunlight is reduced from 7.5 % to 0.1 %. Thus, direct emissions of bromine due to ozone are nearly completely turned off in simulation 4. This suggests a strong underestimation of bromine emissions without a direct emission of bromine due to ozone. A possible conclusion is that the bromine explosion mechanism is insufficient to explain ODEs in the Arctic, or the present bromine explosion scheme is incomplete, for instance with respect to emissions of bromide containing aerosols due to blowing snow and/or regions of increased β such as frost flowers. On 4 March an ODE is predicted by simulation 4 which, however, is not seen in the observations. The model predicts too large wind speeds for the preceding days, causing larger BrO emissions that ultimately result in a predicted ODE being advected to Utqiaġvik. For the first 3 weeks of February, the observations and results of simulations 2–6 are similar to those of simulation 1, in which halogen reactions are turned off, but afterwards, they differ increasingly. This suggests a weak initial influence of halogen chemistry during the first 3 weeks of February, which might be due to the low wind speeds during this time or due to the weak solar irradiation. Partial ODEs occur on 14, 17, 19, and 22 February 2009. The first full ODE in the observations occurred on 13 February, which is predicted by the model only as a partial ODE with 1 d of delay. The partial ODE observed on 17 February is found in simulations 2–5 with a delay of a few hours; simulations 3 and 4 find a stronger ozone depletion more consistent with the observations. On 21 February 2009, simulations 2 and 3 and simulations 4 and 5 predict partial and full ODEs, respectively, which are not seen in the observations. The strength of the ODEs in February is underestimated by the model. A possible cause for this is an overestimation of halogen deposition over land, which can be seen in the comparison to satellite data and is discussed in Sect. . Most of the model BrO capable of reaching Utqiaġvik can only be produced in the Bering Sea during February due to a lack of sunlight in the northern regions. Since BrO over land is removed too quickly in the model, BrO can only be sustained through heterogeneous reactions while being transported from Bering Sea to Utqiaġvik by trajectories that go mostly over the sea ice.

In March, both simulations and observations agree in the occurrence of at least partial ODEs during most of the month, whereas times without any ozone depletion at all are rare. Around 4 March, the model predicts a partial ODE in simulations 2–4, whereas simulation 6 predicts a full ODE, neither of which is found in the observations. Four days later, all simulations predict a partial ODE even though a full ODE is seen in the observations. The following ODE-free time period until 13 March is predicted in agreement with the observations; however, the full ODE on 15 March appears as partial ODE in all simulations, and the simulations with enhanced emission find the partial ODE to continue for 3 more days. The ODE on March 19 is found in simulations 2–6. The simulations predict a near-full recovery of ozone levels over 3 d, which, however, is interrupted in the observations on 21 March. The following ODE episodes are captured quite well by the simulations with an overprediction of ozone levels on 25 and 28 March. ODEs around 1, 14, and 18 April are underestimated in the simulations, whereas all other ODEs and ozone regeneration episodes are predicted quite well. At the end of April, the observations find enhanced ozone levels which are not captured by the model, not even by the simulation without the halogens. The enhanced ozone levels in the observations might be due to Arctic haze, i.e., enhanced photochemical ozone formation due to air pollution originating from lower latitudes. found that the decomposition of peroxyacyl nitrates (PAN), transported from lower latitudes or the upper troposphere to the Arctic boundary layer, can account for up to 93 % of the ozone production in the Arctic. The domain modeled in this work (see Fig. ) does not consider the lower latitudes, so that the simulation itself cannot predict the production and transport of Arctic haze. However, pollution from the lower latitudes might be correctly modeled by the MOZART-4 model and thus be present in the lateral boundary conditions. The model does not find these enhanced ozone levels, which suggests inaccuracies in the MOZART-4 boundary conditions. Simulation 3 finds a partial ODE on 29 April, which is not present in the observations. The other simulations also find a slight decrease in the ozone mixing ratio; however, for these simulations, the BrO levels are not predicted to be large enough for an ODE to happen. Summarizing the entire period of 3 months, simulation 1 shows two ODEs where none were observed. Twenty-two ODEs are identified in the observations, half of which are found by simulation 2. Simulation 3, however, identifies four additional ODEs compared to simulation 2 which were not found in the observations. Simulation 3 misses only 6 of the 22 observed ODEs.

The results of simulation 6 differ strongly from the other simulation results starting mid March and the correlation coefficient R of 0.435 compared to simulation 2 with R=0.62. The RMSE is 14.1 compared to 12.1 nmolmol-1. The mean bias is improved, but this is simply due to the enhanced emissions, resulting in more ODEs, and not due to actually predicting the ODEs better. All statistics are worse compared to simulation 3. As discussed previously in this section, simulation 6 predicts meteorology much worse due to the lack of nudging, which also leads to wrong predictions in the ozone mixing ratio. As can be seen in the correlation plots, simulations 2 and 4 rarely find ODEs were there are none in the observations. There is a notable accumulation of points in all four simulations at ozone mixing ratios of about 30–40 nmolmol-1 for both the observations and the model. In this range of ozone mixing ratios, both the model and observations do not show any ODEs. Halogen chemistry, which has large uncertainties regarding the chemical reactions and the source of bromine, is less important in this case, which explains the high density of points in this regime. This accumulation is denser for simulations with weaker bromine emissions, since those simulations less often predict ODEs which do not exist in the observations. There is an additional accumulation of points around an ozone mixing ratio of 0 in both the model and the observations for simulations 2–4, which are ODEs found by both model and simulation. This accumulation is less dense for simulation 4 compared to simulations 2 and 3. Simulation 4 performed worst regarding both mean bias and RMSE. In simulation 4, there is an accumulation of points at around modeled ozone values of 30 nmolmol-1 and observed ozone values of 0, which are the ODEs missed by the simulation, which suggests an underestimation of the occurrences of ODEs. Simulation 4 with a strongly enhanced β=2.0 but a reduced bromine emission due to direct bromide oxidation by ozone during the daytime (Φ=0.1) suggests that the bromine explosion mechanism alone is insufficient to properly predict the bromine production.

Simulations 2–4 and 6 reproduce ozone levels and ODEs much better than simulation 1, where the mean bias is smaller by at least 9 nmolmol-1. For simulation 3, all statistics are improved compared to the base simulation 2, with both the correlation and RMSE being only slightly better and the mean bias being about 80 % smaller (1.1 vs. 5.1 nmolmol-1) than in simulation 2. Figure  shows a strong increase in the number of ODEs that occur in the model but not in the observations, which explains the strongly improved mean bias while the other statistics only improved slightly.

At Summit, ODEs were found by none of the simulations and not in the observations which lack data for 29 April as can be seen in Fig. . The differences between a simulation without halogens and with halogens are negligible. Ozone mixing ratios are underpredicted with a mean bias of -4.3 nmolmol-1 for simulation 2. This is in contrast to Utqiaġvik, where ozone was generally overpredicted. In April, ozone levels at Summit are found to exceed 60 nmolmol-1 for several time periods in the observations. This is probably due to the high elevation of 3200 m a.s.l. of Summit in contrast to Utqiaġvik. At Summit, the time with the highest ozone level, which occurs on 18 April, is found by the model. The high ozone mixing ratio in the model is due to stratospheric ozone, reaching the troposphere due to a tropopause fold event as shown in Fig. . The other time periods of enhanced ozone levels found by the observations may also be due to a tropopause fold or possibly Arctic haze events.

Figure  shows modeled BrNO2 and BrO of simulation 3 and in situ observations of BrO  at Utqiaġvik. In order to improve the comparability of the observed data with a 10 min resolution and the model results, which were saved every 2 h, a seven-point moving average is applied to the observations, taking the average of the time point under consideration and three time points prior to and after that time point. Modeled BrO is underpredicted with a mean bias of -1.65 pmolmol-1, and a correlation of 0.472 is found. In early to mid March, BrO is less underpredicted with an overprediction of BrO for some days. For most of these days, enhanced BrO levels are due to NOx-catalyzed release of reactive bromine. NOx is emitted at Prudhoe Bay and can then produce N2O5, which further releases BrNO2 on FY ice via the heterogeneous reaction R22N2O5(g)+Br-(aq)+H+(aq)→BrNO2(g)+HNO3(g). BrNO2 can then photolyze to Br, which may further release bromine on FY ice through the bromine explosion mechanism. In the current model, the above heterogeneous reaction is the only source of BrNO2, so that any enhanced mixing ratios of BrNO2 at Utqiaġvik can be attributed to polluted air from Prudhoe Bay producing bromine on FY ice through the heterogeneous reaction with N2O5. As can be seen, for many of the days in early March, there are enhanced BrNO2 mixing ratios preceding large BrO levels. Enhanced modeled BrO on 14, 17, and 20 February (see Fig. ) is coincident with large BrNO2 mixing ratios caused by polluted air from Prudhoe Bay, which is transported over sea ice. A similar phenomenon was found by , who discovered large BrO concentrations in February 2017, which are attributed to nighttime photolabile bromine production, possibly by N2O5, over sea ice. These photolabile species may be transported to lower latitudes where they might be photolyzed. A further discussion of modeled N2O5 can be found in Sect. S6 in the Supplement. studied the role of NOx in bromine chemistry from 24 March 2009 to 3 April 2009 at Utqiaġvik using a box model. They found a suppression of ozone destruction for a high-NOx case (concentrations in the range of 800 to 1600 pmolmol-1). During this time frame, the simulation with WRF-Chem predicts negligible production of reactive bromine due to N2O5. In Fig. S6 in the Supplement, modeled NOx, BrONO2, HOBr, and BrO are shown for the time range considered by . Modeled NOx is elevated from 24 to 26 March and again on 2 April, similar to the measurements of NOx shown in Fig. 2 of the paper of . However, the present model does not find NOx mixing ratios on the order of 10 000 pmolmol-1 as identified on 24–27 March in the measurements. The typical modeled NOx concentrations are in the range of 50 to 1000 pmolmol-1, i.e., between the high- and low-NOx scenarios of . The predicted values of BrONO2 shown in Fig. S5 in the Supplement of this paper compare quite well with those of (see Fig. 7c of that work), with peak values around 50 pmolmol-1. From the end of March to 15 April, however, the mixing ratio of modeled BrO is smaller, whereas the BrNO2 mixing ratio drops to almost 0. Due to the higher temperature and stronger sunlight, N2O5 becomes less stable and its mixing ratio drops, suppressing bromine production due to N2O5. At the same time, observed BrO mixing ratios strongly increase. The underprediction of modeled BrO for these later dates is likely due to a general underprediction of bromine near coastal regions and on land, which will be further discussed in the following sections.

Ozone sonde sounding data  produced near Utqiaġvik are used to validate vertical ozone profiles. Measured ozone and potential temperature for the upward flight of the sonde in the first 2 km are shown in Figs.  and together with the simulation result of the column of the nearest grid cell. The simulation result is interpolated linearly in time to the starting time of the sonde flight.

Figures  and show vertical profiles at Utqiaġvik for various dates. For 14 March, the model fails to find the shallow surface inversion (boundary layer height smaller than 50 m) possibly due to a lack of vertical resolution. The boundary layer height of about 350 m in the observation is overpredicted by approximately 200 m by the model, which might also partially explain the finding of a partial ODE by the model instead of a full ODE as seen in the observations. For this day, simulation 3 performs slightly better than simulation 2. Two days later, both the observations and the simulations show partial ODEs. Simulation 2 predicts the ozone profile very well. The temperature profiles are quite different; however, both model and observations show an inversion at a similar, low height. For 22 March, the enhanced emission case correctly predicts a full ODE, capturing both ozone and temperature profile quite well. The model is however unable to capture a surface inversion. On 15 April, a surface inversion with a second inversion at approximately 500 m is found in the observations. The MYJ PBL scheme also predicts a surface inversion; however it fails to predict the second inversion properly, as can be seen by the lack of a second ozone plateau. While the model is unable to capture the complex boundary layers perfectly, the ozone profiles shows many similarities to the observed profile. For a better prediction, more grid levels closer to the surface and improvements to the PBL schemes might be needed. Even that, however, might not be sufficient, since PBLs in the Arctic can be influenced by very small-scale structures such as open leads, which were found to play an important role in the ozone recovery after an ODE due to down-mixing of ozone-rich air from the free troposphere  and which would require high-resolution sea ice data. Additionally, an accurate modeling of surface inversions might require very high vertical resolutions, which are difficult to obtain in a synoptic-scale simulation.

Figure  shows modeled vertical BrO profiles convoluted with the MAX-DOAS averaging kernel from 28 March 2009 to 16 April 2009 at Utqiaġvik in comparison to BrO measured with a MAX-DOAS instrument . The time range from 26 February 2009 to 27 March is illustrated in Sect. S5, Fig. S4, in the Supplement. BrO from the same observation dataset is shown in Figs.  and . On days with good visibility, the observed data are sensitive for the first 1–2 km. As can be seen, model and observations agree on most dates on the presence of BrO. However, modeled BrO tends to be elevated in comparison to the observations, which can be seen for all days shown in Figs.  and and on 31 March and on 1 and 10 April in Fig. . This is likely due to an underestimation of bromine emissions over snow-covered land, which is also discussed in the next section. Since the model assumptions only allow for partial recycling of bromine over land but not for new emissions, in the lowest grid cells, bromine is lost due to depositions, which results in the elevated modeled BrO profiles. On 9 and 13 March, the model overpredicts BrO. The high BrO mixing ratio on those two dates is due to a heterogeneous reaction involving N2O5; see Fig. . found correlations of the aerosol extinction and BrO, which led to the hypothesis that BrO is released in situ during snowstorms. Currently, there is no model with blowing snow included, which may explain the underprediction of modeled BrO at some days.

Figure  shows vertically integrated modeled (simulation 3) and measured BrO  over the first 2 km. As can be seen, the BrO column is generally underpredicted by the model with a mean bias of -0.98×1013 molec.cm-2. This may partly be attributable to the underprediction of BrO over land in the model; however, there seems to be an offset of around 5.0×1012 molec.cm-2 in the measurements. A correlation of 0.427 is found.

GOME-2 satellite tropospheric BrO VCDs  described in Sect.  are compared with BrO VCDs evaluated from the numerical simulations. All satellite BrO orbits of the same day are averaged and plotted in one figure, where missing satellite data are neglected. Since stratospheric BrO is not generated in the present model, all BrO predicted by the model is of tropospheric origin. Thus, model BrO VCDs are calculated by integrating BrO concentrations vertically from the bottom to the top of the calculation domain. Simulation results are stored every 2 h starting at 00:00 UTC. Each output is assigned a 60∘ segment of a circle with its origin at the North Pole. The segment is centered on a longitude, conforming to GOME-2 orbits for that time. The BrO VCDs are averaged with their neighboring segments with a weight of unity at the center of the segment, and decreasing linearly to 0 at the edge of the segment. This procedure is a linear time interpolation and smoothes the resulting model BrO VCDs. Figure  displays the simulated instantaneous BrO VCDs on 8 March 2009, 16:00 and 18:00 UTC. On the left there are 2 of the 12 full BrO VCDs saved for each day and in the middle the corresponding 60∘ segment multiplied by a weight of unity at the center, which linearly decreases to 0 at the edges of the segment. On the right, the added segments are shown. This procedure is done each day for all 12 time points. Thus 12 segments, not just the two segments shown in Fig. , are added for the average of 1 d, covering the whole domain.

Figure  shows daily averages for the satellite data and simulations 2 and 3 on selected days. On 8 March, both the model simulation and the observations show a high BrO VCD in Nunavut, including King William Island. However, the models predict BrO VCDs to be strongly concentrated in a small area, whereas the satellite BrO cloud is spread out more and reaches deeper into the Canadian mainland. On 15 March, both model simulations and satellite observations find a bromine cloud over the Laptev Sea, reaching to the Siberian land mass. The modeled BrO VCDs are more pronounced, with simulation 3 having a different distribution of BrO being less consistent with the observations than simulation 2. The enhanced emissions in simulation 3 cause a stronger ODE in that region, which in turn depletes BrO in the ozone-depleted area. Ozone mixes back into the ozone-depleted area from the edges of an ODE, which allows BrO to form there which is the reason for the elevated BrO levels seen at the edges of the ODE. The bromine cloud is predicted by the model to extend to the Chukchi Sea in a thin stripe, which is barely seen in the observation. In both model results, a small BrO cloud in Hudson Bay is found, which is more pronounced and less consistent with the observations for simulation 3. On 13 April, a ring-like BrO structure can be seen north of the Kara Sea. The BrO-free center of the ring is due to ozone depletion. Both simulations correctly find a BrO-free area near the North Pole. An enlarged ODE is predicted, resulting in a thinner ring more consistent with the observations. The model, however, underpredicts BrO clouds near the Alaskan coast and finds enhanced BrO VCDs in Greenland in contrast to the observations.

In summary, both simulations 2 and 3 appear to be successful in capturing the general structures. Some of the differences might be explained by a higher model resolution (20km×20km) compared to the satellite data with a resolution of 40km×30km, resulting in more detailed structures in the model. Other differences might be explained by the already discussed errors in the meteorology and underprediction of BrO over land discussed below.

The uncertainties in the satellite data contribute to the differences between model and observations. According to , they are typically below 50 %. Accordingly, differences in absolute values between model and satellite measurement might to a substantial part be caused by measurement uncertainties. However, the spatial patterns found in the satellite data are hardly affected because measurements which are strongly influenced by clouds (cloud shielding) are filtered out using the sensitivity filter of 0.5 for the air mass factor of the lowest 500 m .

Figure  shows monthly averages for the satellite data and results of simulations 2 and 3. and reported BrO observations using MAX-DOAS over the tundra snowpack, which show elevated BrO levels up to more than 100 km inland. found higher BrO concentrations over the tundra than over FY ice. In contrast to that, the simulations conducted in this work underpredict BrO over land and near coasts, which is most likely due to the assumptions in the emission scheme. In the model, it is assumed that snow surfaces have no salt content, which makes depositions of bromine species (excess HOBr is lost) over land a sink, as opposed to depositions over MY ice, which are neutral (excess HOBr is released as BrCl), and over FY (HOBr always releases Br2), which are a source of bromine in most cases. With a deposition velocity of 1 cms-1 and a boundary layer height of 200 m, bromine is removed at a timescale of approximately 5 h over land by surface depositions and possibly even faster by depositions to aerosols. The assumption of zero bromide content of snow covering land or MY ice is of course an idealization and not always correct in reality , contributing to the underprediction of BrO over land mentioned in this paragraph. Future simulations should aim to find ways to incorporate the salinity, pH, and the Br-/Cl- ratio of the snowpack, which where found to be important parameters for the production of Br2 . BrO VCDs are also underpredicted near the boundaries, which is due to the value of 0 of halogens at the boundary. The model overpredicts BrO VCDs at Baffin Bay and at most locations featuring FY sea ice with the exception of the Bering Sea, probably due to its proximity to a domain boundary. The overprediction over FY sea ice is not surprising with the assumption of unlimited BrO in FY sea ice. A relaxation of this assumption, e.g., by allowing finite salt content could solve the issues both over snow covering FY ice, by limiting the bromine emissions, and over land, by allowing salt content of more than 0 and storage instead of loss of deposited bromine. The model prediction for BrO in February is generally too small, which is probably due to a lack of sunlight at higher latitudes and the underprediction of BrO over land. It should be noted that the satellite data are quite incomplete during February and biased towards the end of February, also due to a lack of sunlight necessary for satellite measurements in early February, whereas the model VCDs weights all of February equally.

The emission rate of Br2 due to HOBr+BrONO2 and due to bromide oxidation by ozone from the snow surface averaged over the entire simulation period is shown in Fig.  for simulation 3. In Fig. , the production of Br2 from the snow is shown at coordinates 78∘ N, 178∘ W plotted against time. The location has been chosen because it is over FY sea ice and is a strong production site for the bromine that may affect ODEs at Utqiaġvik. As can be seen in these figures, most of the bromine is produced by HOBr, i.e., the bromine explosion mechanism, whereas the oxidation of bromide by ozone provides an initial seed of the bromine formation which then is enhanced by bromine explosion where BrONO2 plays a smaller role than HOBr. Due to a lack of sunlight, bromine is produced only during the second half of February by the bromine explosion and after 1 March 2009 by the bromide oxidation due to ozone.

In the present parameterization, the latter strictly requires an SZA of less than 85∘ for a fast release, whereas the bromine explosion mechanism has a more continuous dependence on SZA. The Br2 photolysis needed by both emission mechanisms requires relatively longwave light and may thus occur even at SZAs slightly above 90∘. The bromine explosion additionally requires HO2 in order to produce HOBr. HO2 is mostly formed by a photolysis of various organic species with shortwave UV and thus occurs generally at smaller SZA; however, it can also be supplied by reactions involving organic compounds, NOx, and/or OH or by their transportation from lower latitudes. Thus, in the present parameterization, the bromine explosion may occur locally at higher SZAs than the bromide oxidation due to ozone.

Emission rates of Br2 from other studies are as follows. In February 2014, measured Br2 fluxes of 0.07–1.2×109 molec.cm-2s-1 above the snow surface near Utqiaġvik with a maximum around noon. found snowpack Br2 emissions of 2.1×108 molec.cm-2s-1 on 15 March 2012 and 3.5×106 molec.cm-2s-1 on 24 March 2012 in a modeling study. Emission fluxes due to the bromine explosion (HOBr+BrONO2) are typically between 2–3×109 (simulation 2) or 4–5×109 molec.cm-2s-1 (simulation 3) around noon and thus are at the higher end of the mentioned values. Bromide oxidation due to ozone, which plays the role of direct snowpack emissions in the present model, is rarely larger than 1×109 molec.cm-2s-1 with an average of around 2×108 molec.cm-2s-1 near Utqiaġvik, which compares quite well to the range found by while being larger than the values calculated by .

For a simulation of 3 months, it should be expected that errors in the simulation pile up, especially considering the nonlinear stochastic nature of ODEs. The meteorological state should be consistent due to the data assimilation via nudging; however, the errors in the chemistry model could grow large over time. As an example, wrongly predicting an ODE probably causes a delay of an ODE at a later date due to the lack of O3, reducing bromine emissions. A test for this is performing a new start of a simulation at a later date, where no ODEs occurred and in which the atmosphere is clean of bromine. For this purpose, simulation 5 was conducted, which is identical to simulation 3 except that the simulation starts on 16 March using ERA-Interim and MOZART-4 data as well as a near-zero bromine concentration, as described in Sect. . These new simulation results are then compared to simulation 3 which started in February.

It is found that these two simulations become very similar after approximately 5 d; see Fig.  which shows the BrO VCDs. After approximately 8 d, the BrO VCDs become nearly indistinguishable. Average BrO concentrations in April are not shown here but are also nearly identical for both simulations. Reasons for the two simulations with different starting times to show such similar results after a few days is due to a combination of several factors. While there is no chemical nudging, the chemical boundary conditions strongly affect the simulation and act similarly to chemical nudging. Assuming a constant wind speed of 20 kmh-1 (corresponding to approximately 5.5 ms-1), a chemical species can be transported from a 2000 km distant boundary to the center of the domain on a timescale of as low as 4 d. Due to the meteorological nudging, chemical boundary conditions are transported in the same way in both simulations. Chemistry boundary conditions transported over land or in the free troposphere behave similarly in simulations 3 and 6, since several aspects of chemistry over land and in the free troposphere are nearly unaffected by the addition of halogen chemistry. Thus, chemical species coming from the lateral boundary condition will only be affected by the halogen chemistry once they reach the sea ice or are mixed into the boundary layer from aloft.

The emission of bromine due to bromide oxidation by O3 is independent of reactive bromine mixing ratios and not of autocatalytic nature as in the bromine explosion mechanism. While it is only responsible for a small fraction of emitted bromine, it produces the initial bromine needed for a bromine explosion. The present emission scheme can be very fast, producing full ODEs in less than a day. All of these effects allow ozone coming from the lateral boundary condition to be depleted in a similar way in simulations 3 and 6 even with leftover bromine from a previous ODE.