Effects of model resolution on the interpretation of satellite NO 2 observations

Inference of NOx emissions (NO+NO2) from satellite observations of tropospheric NO 2 column requires knowledge of NOx lifetime, usually provided by chemical transport models (CTMs). However, it is known that species subject to non-linear sources or sinks, such as ozone, are susceptible to biases in coarse-resolution CTMs. Here we compute the resolution-dependent bias in predicted NO 2 column, a quantity relevant to the interpretation of space-based observations. We use 1-D and 2-D models to illustrate the mechanisms responsible for these biases over a range of NO 2 concentrations and model resolutions. We find that predicted biases are largest at coarsest model resolutions with negative biases predicted over large sources and positive biases predicted over small sources. As an example, we use WRFCHEM to illustrate the resolution necessary to predict 10 AM and 1 PM NO2 column to 10 and 25 % accuracy over three large sources, the Four Corners power plants in NW New Mexico, Los Angeles, and the San Joaquin Valley in California for a week-long simulation in July 2006. We find that resolution in the range of 4–12 km is sufficient to accurately model nonlinear effects in the NO 2 loss rate.


Introduction
NO x (NO + NO 2 ) is emitted to the troposphere by fossil-fuel combustion, biomass burning, soil microbial processes, and lightning.In the troposphere, NO x affects ozone production, aerosol formation and atmospheric composition (e.g.CH 4 ) through feedback on HO x (OH + HO 2 + RO 2 ).The concen-Correspondence to: R. C. Cohen (rccohen@berkeley.edu)tration of OH radical, the main daytime sink of NO x , depends strongly on NO x concentration.As a result, the removal rate of NO x (i.e.k NO 2 +OH [OH]) depends strongly on its own concentration.Thus, to accurately quantify this removal rate, a model must accurately resolve NO x from its source (10-100 ppb) to background (10-100 ppt).Column NO 2 (e.g., Heue et al., 2008;Valin et al., 2011) and in situ observations (e.g., Ryerson et al., 2001;Russell et al., 2011) show that this transition can occur at scales as small as 10-20 km, a length scale similar to those reported from calculations (e.g., Cohan et al., 2006;Loughner et al., 2007).
Satellite-based observations of tropospheric NO 2 have provided unique insights into spatial and temporal patterns on regional scales of soil (e.g., Bertram et al., 2005;van der A et al., 2007;Hudman et al., 2010), biomass burning (e.g., Jaegle et al., 2005;Mebust et al., 2011), lightning (e.g., Beirle et al., 2010), and urban NO x emissions (e.g., Kim et al., 2009;Russell et al., 2010;Beirle et al., 2011).These satellite observations have also provided constraints on inverse models that are used to validate emission inventories (e.g., Martin et al., 2003;Konovalov et al., 2006;Napelenok et al., 2008;Kim et al., 2009).Most of the inverse modeling studies used to validate these emission inventories adjust NO x emissions with the assumption that model chemistry is accurate.However, errors in our understanding of atmospheric reactions (e.g., Thornton et al., 2002;Mollner et al., 2010) or their representation in models may compromise the accuracy of these inferred emissions.For example, extensive research has demonstrated that modeled ozone production depends strongly on model resolution due to its nonlinear dependence on NO x concentration (e.g., Sillman et al., 1990;Kumar et al., 1994;Gillani et al., 1996;Cohan et al., 2006;Wild and Prather, 2006).OH, which has the same NO x -dependence as ozone production, will also vary a Alkyl-nitrate formation rate has been set to 0 in these simulations.Increases in this rate will slightly affect NO x lifetime (Farmer et. al., 2011).
with model resolution affecting NO x lifetime and concentration.As a result, NO x emissions derived from an inversion of satellite NO 2 observations will have resolution-dependent biases, the magnitude of which is not well known.
We investigate the biases in predicted NO 2 column due to model horizontal resolution.To illustrate the mechanisms at work, we first show how model resolution, OH, and NO 2 interact in 1-D and 2-D plume models.We then use WRF-CHEM, a fully-coupled regional 3-D chemical transport model (CTM), to evaluate the effect in some realistic model situations.We use the simpler models to understand the source of resolution-dependent biases and use WRF-CHEM to determine the model resolution necessary to predict 10 AM and 1 PM column NO 2 to 10 % and 25 % accuracy over the Four Corners and San Juan power plants, the city of Los Angeles, and the San Joaquin Valley in California for a week-long simulation in July 2006.

NO x -OH steady-state chemistry in 1-D and 2-D plume models
Using the analytical solution to the NO x -HO x steady-state relationship to determine OH concentration (Table 1) (Murphy et al., 2006 and references therein), we simulate the removal of NO 2 by OH in the outflow of sources using both 1-D and 2-D plume models.The selected chemical parameters are representative of noontime, mid-latitude, NO x -CH 4 -CO chemistry.
Figure 1 shows the dependence of OH on NO 2 derived from this steady-state relationship with the corresponding NO 2 lifetime indicated (k NO 2 +OH × [OH]) −1 .As is wellknown, the response of OH to changes in NO 2 depends on NO 2 concentration.For example, decreasing NO 2 at high NO 2 (red) results in an increase of OH and a shorter NO 2 The key feature of this relationship is that the maximum in OH feedback (green) corresponds to a NO 2 lifetime of two hours, an e-folding decay length of 54 km in 5 m s −1 winds and of 5-15 km at the slower moving diffusive edges.Poorly resolving these gradients in NO 2 will result in inaccurate OH feedback, biases in NO 2 lifetime, and as a result, modest biases in domain-total NO 2 .For resolutions that grossly misrepresent the distribution of NO 2 , biases in OH feedback, NO 2 lifetime, and domain-total NO 2 will be massive.

Biases in a 1-D plume model
In a 1-D model, NO 2 is emitted at the western end of the domain (x = 0-1024 km) and transported to the east at a constant rate (5 m/s).We run the model with emission rates of 30, 6.0, and 0.30 kmol h −1 .These emission rates are selected to correspond with high, intermediate and low NO 2 -OH feedback regimes depicted in Fig. 1.The model is run at east-west resolutions of 0.5 to 512 km with the dimensions perpendicular to the flow fixed at 1 km (north-south and vertical).
When 1-D simulations are run to steady-state, NO 2 is removed by OH such that the spatial gradient reflects the applied NO 2 -OH feedback (Fig. 2).Failure to accurately resolve these gradients results in inaccurate OH and biases in both the NO 2 lifetime and concentration.For example, when a large source of NO 2 is computed at 2 km resolution, OH is suppressed so strongly that NO 2 decays by only one  e-fold in 400 km, corresponding to a chemical lifetime of 22 h (Fig. 2a, d -solid line).Computed at coarser resolution (128 km), NO 2 is averaged over the entire grid cell such that OH is enhanced, NO 2 is shorter-lived, and NO 2 concentrations are biased low (Fig. 2a, d -dashed line).At this high emission rate, biases in the domain-total NO 2 exceed 50 % at the coarsest resolutions (Fig. 3, red line).
In contrast to a large source, NO 2 emitted from a small source decays rapidly when computed at 2 km resolution, a reflection of high OH concentration (Fig. 2c, f -solid line), but at coarser model resolution (128 km), NO 2 is instantaneously mixed over an entire grid cell resulting in lower OH, a longer NO 2 lifetime, and a corresponding positive bias in domain-total NO 2 concentration (Fig. 2c, f -dashed line, Fig. 3, blue line).
For an intermediate source of NO 2 , the plume decays by an e-fold within 60 km at 2 km model resolution, a gradient corresponding to a chemical lifetime of about three hours and near-maximum OH (Fig. 2b, e -solid line).Model calculations at coarser resolutions (128 km) are not capable of resolving this sharp gradient.Because NO 2 concentrations predicted for an intermediate source are near the maximum OH (Fig. 1, green), biases behave like those of a large source at finer model resolutions and like those of a small source at coarser resolutions (Fig. 3, green line).
In a 1-D model, domain-averaged NO 2 predicted at coarse resolutions is biased (Fig. 3).At intermediate resolutions the biases are moderate (10-30 %) and result from numerical resolution that dilute NO 2 at the leading edge of the plume and shift OH to the left in Fig. 1.When the model resolution becomes much larger than the plume itself, OH feedbacks are shifted by a factor 2-5 times that of the plume simulated in a resolved calculation (e.g., Fig. 2a, d) and gross biases in domain-total NO 2 are predicted (>50 %, Fig. 3).(d-f).The color-scheme corresponds to NO 2 -OH feedback regimes depicted in Fig. 1.The vertical layer is fixed at 1 km thickness for all simulations shown.

Biases in a 2-D plume model
While 1-D models are illustrative, 2-D models are a better approximation of NO 2 column and provide some additional insights.For instance, we can consider the effects of horizontal diffusion well as different source distributions.In this 2-D model, we define a point source (2×2 km 2 ) with emission rates of 200, 40, and 2 kmol h −1 and an area source (96 × 96 km 2 ) with emission rates of 1000, 200, and 10 kmol h −1 both located in the far southwest corner (x = y = 0 km) of the domain (x = y = 0-384 km).These emission rates are selected so that the simulated plume concentration corresponds with the high, intermediate and low NO 2 -OH feedback regimes depicted in Fig. 1.In this model, NO 2 is transported with x and y wind speeds of 3 m s −1 , a mean flow of 3 √ 2 m s −1 to the northeast.Diffusion rates are set to 10 m 2 s −1 .Initial and horizontal boundary concentrations are set to 0.5 ppt.The model is computed at six grid resolutions (2, 4, 12, 24, 48, and 96 km), with the vertical layer fixed at 1 km.We run simulations of NO 2 with OH determined by the steady state equation (Table 1, Fig. 1).
When the 2-D model is run to steady-state, a large point source of NO 2 is OH suppressing and long-lived when simulated at 2 km model resolution, but experiences high OH and is short-lived when simulated at coarse resolution (96 km) (Fig. 4a-c).Because the resolved plume is so narrow (∼24 √ 2 km, Fig. 4a), the distribution of NO 2 and the corresponding OH feedbacks are grossly mis-predicted at 96 km resolution with biases in domain-averaged NO 2 as high as 75 % (Fig. 5b).Over a small source (not shown), the oppo-site effect occurs in a model; a small source is OH-enhancing and short-lived at 2 km resolution, but experiences low OH and is long-lived at coarser resolutions.As a result, domainaveraged NO 2 predicted over a small source is biased 100 %.
When the same 2-D model is run to steady-state using emissions from an area source (Fig. 4d-f), domain-averaged biases in NO 2 are relatively modest (∼25 % at 96 km model resolution) but follow the same general pattern as those simulated over a point source (Fig. 5).The biases predicted over an area source are smaller because coarse model resolutions are better able to characterize the distribution of NO 2 and corresponding OH feedbacks over the simulated plume, which is much wider than that predicted over a point source.From previous discussion, it would be expected that simulation of NO 2 at 96 km model resolution would be massively biased since the resolved plume is approximately 96 km wide.However, the alignment of this area source (96×96 km 2 ) on this 96-km grid results in an NO 2 distribution that roughly approximates that simulated in a resolved model, resulting in biases that are relatively small (∼25 %).When the 96×96 km 2 area source is shifted by 48 km and evenly distributed over four 96 km grid cells (not shown), biases predicted at 96 km resolution are as large as those predicted for a point source (>75 %).
Figure 5 summarizes biases predicted in a 2-D model as a function of model resolution, source strength, and proximity to the source.Biases predicted for NO 2 in a constant OH field (OH = 5.5×10 6 molecules cm 3 ), that is without any NO 2 -OH feedback and only subject to transport effects (Fig. 5b-e, black line), are negligible over the entire domain (∼0 %, 5b), significant over the 192 km nearest the source (−10 %, 5c), and large in both the 96 km nearest the source (−30 %, 5d) and in a 12 × 12 km 2 pixel 96 km downwind from the source (−75 %, 5e).Without OH feedback, these biases are due to numerical artifacts in computing a short-lived species on a coarse grid, but are are negligible over regions that are large (e.g., 192 km) compared to the spatial scale of chemical decay (e-fold decay over 84 km) (e.g., Fig. 5c, black).When subject to NO 2 -OH chemical feedbacks (Fig. 5b-e, red, green, and blue lines), biases diverge from that of transport alone (black) with behavior depending, as expected, on the rate of NO x emissions.Over each spatial domain considered, simulation of a large source (red) is biased low versus transport alone (black) whereas the opposite is true for a small source (blue).For both large and small sources, the magnitude of the bias increases as the model resolution coarsens.In both the point and area source examples, the behavior of biases depends on model resolution.
VOC reactivity affects predicted biases by altering the NO 2 concentration at which maximum OH occurs (Fig. 1, Table 1).For simulations with VOC reactivity increased by a factor of 10 and the same NO 2 emissions, we find that the pattern of predicted biases changes in the direction one expects based on the shifts in the NO 2 -OH relationship for an increase in VOC.However, if the NO x emission rates are increased along with VOC reactivity, the patterns of biases returns to those predicted at the original VOC conditions.
Biases in a 2-D model are predicted to be unacceptably large (up to ∼75-100 %) at relevant horizontal resolutions (96 km) with behavior that depends dramatically on source strength (+100 % for small source to −75 % for large source) and distribution (∼ 75 % for point source to ∼ 25-75 % for area source).Biases predicted in a 2-D model are much larger than those predicted at the same model resolution in a 1-D model because the plume can diffuse horizontally.However, the basic resolution dependent effects are the samenumerical dilution shifts the NO x -OH feedback to the left on Figure 1, altering the spatial pattern of NO 2 from a resolved calculation.These effects are modest (0-30 %) if the model resolution is finer than the width of a fully-resolved plume, but are massive (>50 %) if the resolution is comparable to or larger than the plume (Figs. 4, 5).

Illustration of effects in WRF-CHEM: Four Corners, Los Angeles, and San Joaquin Valley
Regional CTMs, like WRF-CHEM (Grell et al., 2005), provide a realistic and fully-coupled description of atmospheric mixing, chemistry, and emissions.We use WRF-CHEM to test for resolution-dependent biases over three distinct source regions: the Four Corners Power Plants, Los Angeles, and the San Joaquin Valley.Air quality control strategies often use 10 or 25 % NO x reductions as a realistic regulatory benchmark.Using WRF-CHEM, we simulate the 10 AM and 1 PM NO 2 column, as would be observed by space-based Four Corners, Los Angeles and the San Joaquin Valley are large enough sources of NO x such that OH is suppressed in a resolved simulation.For example, effective OH, or the NO 2weighted OH concenntration that is simulated at 1 km resolution over Four Corners is low (<5 × 10 6 molecules cm −3 ) where NO 2 concentration is high (Fig. 6a-b, red), enhanced (>1.5 × 10 7 molecules cm −3 ) where NO 2 is intermediate (Fig. 6a-b, green) and low (∼ 5 × 10 6 molecules cm −3 ) where NO 2 is low (Fig. 6a-b, blue).At coarser resolutions, NO 2 is numerically diluted, and prediction of NO 2 column is biased low just as was predicted over large sources in the 1-D and 2-D plume models.Over Four Corners, the 1 PM NO 2 column simulated at 24 km is biased −50 % relative to simulation at 1 km (Fig. 7).Over Los Angeles, the 1 PM NO 2 column is biased −13 % at 24 km resolution relative to simulation at 4 km, but is biased −37 % at 96 km resolution (Fig. 8).Over the San Joaquin Valley, the biases relative to 4 km simulation are −16 % at 12 km resolution, −24 % at 24 km resolution and −36 % at 48 km resolution (Fig. 9).We find that the exact numbers depend on the choice of the boundaries that surround each plume, but that the conclusions are independent of that choice.As a result of these biases, prediction of the domain-averaged NO 2 column to 25 % accuracy at 1 PM requires model resolution of 4 km over Four Corners, 12 km over San Joaquin Valley, and 48 km over Los Angeles under the conditions tested here.For 10 % accuracy, 4 km model resolution is required over Four Corners and the San Joaquin Valley while 12 km resolution is sufficient over Los Angeles.
Biases in NO 2 column at 10 AM (not shown) are consistently smaller than those predicted at 1 PM because the NO 2 column predicted at 10 AM is exposed to much less OH through the nighttime and early morning hours.As a result, prediction of 10 AM NO 2 column to 25 % accuracy as tested here requires slightly coarser model resolution than was necessary at 1 PM: 12 km over Four Corners, 24 km over San Joaquin Valley, and 48 km over Los Angeles.For 10 % accuracy, model resolution of 4 km is necessary over Four Corners, 24 km over San Joaquin Valley and 48 km over Los Angeles.
All three of these regions are biased low at coarse model resolutions (Figs.7-9), which is expected for large, OHsuppressing sources of NO x .For Los Angeles and Four Corners, sources that are directly comparable to the large 2-D area and point sources, biases behave similarly to those predicted in the 2-D plume model.And while the San Joaquin Valley appears to be an intermediate source of NO x (Fig. 9), which according to the 2-D plume model would indicate that coarse resolution prediction of NO 2 should be biased high (Fig. 5h), it is important to consider the differences between the 2-D plume model, which simulates midday summertime chemistry at steady-state, with WRF-CHEM, which integrates the full diurnal cycle.In WRF-CHEM, all of these sources (Figs.7-9), including the San Joaquin Valley, suppress OH through the morning hours when NO x concentrations are higher and HO x production rate is lower, leading to  biases that start small in the morning hours and grow with time of day.

Implications for interpretation of satellite observations
The calculations above show that predicted NO 2 columns will depend on the resolution of the model.As a consequence, any inference of NO x emissions that relies on a model to interpret satellite observations will have biases if the model resolution is too coarse.The calculations above show that the biases are both positive and negative as a result of the interplay of NO 2 and OH.We find that predicted biases are especially large where steep gradients of NO 2 dominate the total NO 2 mass (e.g.2-D point source) and that they are are smaller where shallow gradients dominate the total NO 2 mass (e.g.2-D area source).In general, gradients are shallow where NO x lifetime is long compared to transport timescales (e.g.wintertime, morning hours) and are steep where NO x lifetime is short compared to the timescale of transport (e.g.summertime, afternoon hours).
For situations presented here, model resolution in the range of 4-12 km is sufficient to predict the effects of NO 2 -OH feedbacks on the NO 2 lifetime and column to 10 % accuracy.This value will change depending on the location, season, and time of day according to general guidelines provided above.Nonetheless our analysis suggests that numerical resolution needs to be small compared to the NO x efolding distance.A growing set of observations over both power plants and urban sources at different times of year indicate that boundary layer NO x often exhibit e-folding distances of order 10-30 km (e.g, Ryerson et al., 2001;Loughner et al., 2007;Heue et al., 2008;Russell et al., 2011;Valin et al., 2011) supporting the idea that model resolution in the range of 4-12 km is necessary in a broad range of contexts.For any calculation where accurate computation of boundary layer NO x is important we recommend model calculations to test whether the spatial resolution is sufficient to calculate NO 2 to the desired accuracy.

Conclusions
We investigate the effects of NO 2 -OH chemical feedbacks on predicted NO 2 in a 1-D plume model, a 2-D plume model, and WRF-CHEM, a fully-coupled 3-D CTM.We use 1-D and 2-D plume models to demonstrate that nonlinear NO 2 -OH chemical feedback leads to biases in column NO 2 .As a result, inference of NO x emission inventories from chemical transport models will suffer biases that depend on the horizontal resolution of the model.Using WRF-CHEM, we determine the model resolution necessary to predict NO 2 column to 10 % and 25 % accuracy over Los Angeles, the San Joaquin Valley, and Four Corners for a week-long simulation in July 2006.In this example, we find that prediction of NO 2 column to 10 % accuracy at 1 PM requires model resolution of 4 km over both Four Corners and the San Joaquin Valley while 12 km is sufficient over Los Angeles.Prediction to 10 % accuracy at 10 AM requires model resolution of 4 km over Four Corners, 24 km over San Joaquin Valley, and 48 km over Los Angeles.In these examples, we find that model resolution must be comparable to, or smaller than, the spatial variability of NO 2 to accurately model NO 2 -OH feedbacks on NO 2 column.Thus, simulations aimed at matching satellite observations must be run at sufficient spatial resolution to avoid contamination by numerical artifacts.

Appendix A
We simulate column NO 2 from 1-7 July 2006, over California, Nevada, Northern Mexico, and the Eastern Pacific centered over Southern California (2304 × 2304 km 2 ) at 4, 12, 24, 48, and 96 km resolution.The simulated domain is much larger than the region of interest to ensure that there are no effects of boundary conditions in the coarser resolution model simulations.The first two days of simulation are used as spin-up, and the last five days (3-7 July) are averaged to 1 p.m. LST for all analyses.Emissions are the National Emission Inventory (NEI) 2005 on-road and offroad transportation emissions for a typical July weekday and Continuous Emissions Monitoring (CEMS) averaged point source emissions for a typical August, 2006, weekday.For more information, see ftp://aftp.fsl.noaa.gov/divisions/taq/emissions data 2005/Weekday emissions/readme.txt.Biogenic emissions for all model resolutions were generated by an online module as in (Grell et al., 2005) at 4 km horizontal resolution for a single July day and kept constant throughout the 7-day simulation.We use the Regional Acid Deposition Model, version 2 chemical mechanism (Stockwell et al., 1990).The initial and boundary chemical conditions are derived from idealized profiles that are standard in WRF-CHEM.Radiative feedback from clouds on photolysis rates was disabled in order to simulate column NO 2 under clearsky conditions that are typical of satellite observations.Meteorological initial and boundary conditions for the simulation are derived from the North American Regional Reanalysis for July 2005 (NARR -http://nomads.ncdc.noaa.gov/dods/NCEP NARR DAILY).
We simulate column NO 2 over the Four Corners region in the Western US (384 × 384 km 2 ) at 1, 4, and 12 km resolution and extend the boundaries (1536 × 1536 km 2 ) to simulate the same domain at 24 km resolution.This domain is centered on the Four Corners and San Juan Power Plants, which are approximately 20 km apart.This simulation is run in the same manner as that run over California except that only emissions from point sources are included.For simulation at 24 km resolution, as mentioned, the domain is extended to avoid boundary relaxation effects that occur over the 5 boundary grid cells in WRF-CHEM.

Fig. 1 .
Fig. 1.Steady state OH concentration (left axis) and corresponding NO 2 lifetime (k NO 2 +OH ×OH) −1 on the right axis versus boundary layer NO 2 column (molecules cm −2 ) assuming a 1-km well-mixed boundary layer.The color scheme corresponds to regions of high NO 2 (red), where OH is low, intermediate NO 2 (green), where OH is high, and low NO 2 (blue), where OH is low, and is used throughout the article.

Fig. 2 .
Fig. 2. NO 2 column (molecules cm −2 ) predicted in a 1-D plume model at 2 km (solid), 128 km (dashed), and 512 km model resolutions (dash-dot) for (a) a large, (b) intermediate, and (c) small source of NO 2 and (d-f) the corresponding OH feedback.The color-scheme corresponds to NO 2 -OH feedback regimes depicted in Fig. 1.Horizontal (N-S) and vertical layers are fixed at 1 km thickenss for all resolutions.

Fig. 4 .
Fig. 4. NO 2 column (molecules cm −2 ) predicted in a 2-D plume model for a large point source simulated at (a) 2 km, (b) 24 km, and (c) 96 km resolution, and the same for simulation of a large area source(d-f).The color-scheme corresponds to NO 2 -OH feedback regimes depicted in Fig.1.The vertical layer is fixed at 1 km thickness for all simulations shown.

Fig. 5 .
Fig. 5. (a) NO 2 column (molecules cm −2 ) predicted in a 2-D plume model for a large point source simulated at 2 km resolution with vertical layer fixed at 1 km thickness.Resolution-dependent bias in domain-averaged NO 2 column over the (b) entire domain (x = y = 0-384 km), (c) the 192 km near-field (x = y = 0-192 km), (d) the 96 km near-field (x = y = 0-96 km), and (e) at a pixel downwind (x = y = 96-108) for NO 2 emitted from a large (red), an intermediate (green), and a small (blue) point source of NO 2 with OH = f NO 2 as in Fig. 1 and from a large point source with OH set to 5 × 10 6 molecules cm −3 (black).(f-j) The same for simulation of a 96 × 96 km 2 area source.

Fig. 6 .
Fig. 6.(a) NO 2 column (molecules cm −2 ) and (b) effective OH, or the NO 2 -weighted average of OH (molecule cm −3 ), taken over the entire vertical column as simulated with WRF-CHEM at 1 km resolution over the Four Corners Region of US at 10 a.m. on 5 July 2006.

Fig. 7 .
Fig. 7. WRF-CHEM 3-7 July 2006, 1 p.m. LST average NO 2 column (molecules cm −2 ) simulated over Four Corners Region of US at (a) 1 km, (b) 4 km, (c) 12 km, and (d) 24 km model resolution.NO 2 column is averaged over a sub-domain (white box) and reported in the bottom left corner of each panel.The sub-domain in the 24 km simulation was rotated to include the plume, which was predicted further to the SE than those predicted at 1, 4, or 12 km resolution.

Table 1 .
Parameters used in HO x -NO x steady-state model(Murphy et.al., 2006 and references therein).