2.1 CHASER V4.0 model and simulations

CHASER V4.0 (Sudo et al., 2002; Sudo and Akimoto, 2007; Sekiya and Sudo, 2014) is a global chemical transport model developed in the framework of the MIROC-ESM Earth system model (Watanabe et al., 2011), which is coupled online with the MIROC-AGCM atmospheric general circulation model (AGCM) (K-1 model developers, 2004) and the SPRINTARS aerosol transport model (Takemura et al., 2005, 2009). Several updates were made from CHASER V3.0 (Sudo et al., 2002) to CHASER V4.0, which includes the consideration of aerosol species (sulfate, nitrate, ammonium, black and organic carbon, soil dust, and sea salt) and the implementation of related chemistry, radiation, and cloud processes. AGCM was also updated from the NIES/CCSR AGCM 5.7b to the MIROC-AGCM. Detailed information on the AGCM updates are provided by K-1 model developers (2004).

CHASER calculates gaseous, aqueous, and heterogeneous chemical reactions (93 species and 263 reactions), including the O₃–HO_x–NO_x–CH₄–CO system with the oxidation of non-methane volatile organic compounds (NMVOCs). Major chemical reactions related to NO₂ are considered, including (1) the photochemical cycle of NO and NO₂, (2) oxidation of NO₂ with OH, (3) heterogeneous hydrolysis of N₂O₅, (4) the formation, thermal decomposition, and photolysis of peroxyacetyl nitrates (PANs), and (5) the formation of isoprene nitrates. CHASER also calculates stratospheric O₃ chemistry including Chapman mechanisms and catalytic reactions related to HO_x, NO_x, ClO_x, and BrO_x below 50 hPa. Above 50 hPa, prescribed concentrations of O₃, nitrogen, and halogen species are used. Monthly ozone climatology is obtained from UGAMP (Li and Shine, 1995), whereas monthly climatologies of nitrogen and halogen species are taken from the Chemistry–Climate Model Initiative (CCMI) REF-C1SD simulation using NIES CCM (Akiyoshi et al., 2009, 2016; Morgenstern et al., 2016). Dry and wet (rain out and washout) deposition is calculated based on the resistance-based parameterization (Wesely, 1989) and cumulus convection and large-scale condensation parameterizations, respectively. Advective tracer transport is calculated using the piecewise parabolic method (Colella and Woodward, 1984) and the flux-form semi-Lagrangian scheme (Lin and Rood, 1996). The model also incorporates tracer transport on a sub-grid scale in the framework of the prognostic Arakawa–Schubert cumulus convection scheme (Emori et al., 2001) and the vertical diffusion scheme (Mellor and Yamada, 1974).

We evaluated two 1-year global simulations for tropospheric NO₂ in 2008 and 2014 with a 1-year spin-up calculation for each simulation. In each case, three model calculations were conducted at different horizontal resolutions: T42 (i.e., 2.8^∘ × 2.8^∘; hereinafter referred to as the 2.8^∘ simulation), T106 (i.e., 1.1^∘ × 1.1^∘; hereinafter the 1.1^∘ simulation), and T213 (i.e., 0.56^∘ × 0.56^∘; hereinafter the 0.56^∘ simulation); 32 vertical layers from the surface to approximately 40 km altitude were used across the three simulations. To meet the Courant–Friedrich–Levy (CFL) condition, different maximal time steps were used for each resolution: i.e., 20 min at 2.8^∘ resolution, 8 min at 1.1^∘ resolution, and 4 min at 0.56^∘ resolution. Sea-surface temperatures (SSTs) and sea-ice concentrations (SICs) were prescribed by HadISST for the corresponding year (Rayner et al., 2003). Simulated air temperature and horizontal wind were nudged to 12-hourly ERA-Interim reanalysis data (Dee et al., 2011). ERA-Interim reanalysis data at 0.75^∘ × 0.75^∘ horizontal resolution with 37 pressure levels were linearly interpolated to each model grid, possibly degrading simulated meteorological fields at finer resolution (i.e., 0.56^∘). We specified 5 and 0.7 days of nudging time for temperature and horizontal wind, respectively.

NO_x emissions from anthropogenic, biomass burning, lightning, and soil sources were considered. Anthropogenic emissions from the HTAP_v2.2 inventory for the year 2008 (Janssens-Maenhout et al., 2015) were employed for the 2008 simulations, with these originally having 0.1^∘ × 0.1^∘ resolution. For the 2014 simulation, anthropogenic emissions for the latest available year 2010 of the HTAP_v2.2 inventory were used. Biomass burning emissions were taken from the Global Fire Emissions Database (GFED) version 4.1 (0.25^∘ × 0.25^∘ resolution) (Giglio et al., 2013) for the two study years. Soil emissions were obtained from the Global Emission InitiAtive (GEIA) database (1^∘ × 1^∘) (Yienger and Levy, 1995). Model of Emissions of Gases and Aerosols from Nature (MEGAN) version 2 data (0.5^∘ × 0.5^∘) were used for biogenic NMVOCs emissions (Guenther et al., 2006). Annual mean total global NO_x emissions from the surface were 45.3 and 45.9 Tg N yr⁻¹ in 2008 and 2014, respectively. Lightning NO_x sources were calculated as a function of cloud top height in the cumulus convection parameterization (prognostic Arakawa–Schubert scheme) at each time step of CHASER, following Price and Rind (1992).

We considered diurnal cycles of surface NO_x emissions following Miyazaki et al. (2012). Different diurnal cycles were assumed depending on the dominant source category of each region: anthropogenic-type diurnal cycles (with maxima in the morning and evening, with a factor of about 1.4) over Europe, eastern China, Japan, and North America; biomass-burning-type diurnal cycles (with a rapid increase in the morning and maximum midday, with a factor of about 3) over Central Africa and South America; and soil-type diurnal cycles (with maxima in the afternoon, with a factor of about 1.2) in the grasslands or sparsely vegetated areas of Australia, Sahara, and western China. Miyazaki et al. (2012) confirmed that the application of this scheme leads to improvements in global tropospheric NO₂ simulations at 2.8^∘ resolution. Improvements were commonly found in the 1.1^∘ resolution simulation, whereas we did not evaluate the impact at 0.56^∘ resolution. Over biomass burning regions, the emission diurnal variability applied in this study is generally similar to variability from 3-hourly GFED4.1 data, while distinct differences in relative magnitude around the GOME-2 overpass time suggest that model performance could differ in comparison to the GOME-2 retrievals when using the 3-hourly GFED4.1 data.

The CTM–AGCM online coupling framework used in this study has advantages over the off-line CTM framework driven by meteorological analysis or reanalysis data. First, the online framework is able to simulate short-term nonlinear variations in chemical and transport processes at every time step of the model (1–20 min in this study) in contrast to off-line CTMs driven by meteorological data, typically with 6-hourly intervals. Second, grid-scale and sub-grid-scale transport processes (e.g., convection, turbulent mixing) are represented in a consistent manner based on AGCM physics (e.g., mass balance) at short time intervals. Third, the online framework allows for a flexible choice of CTM resolution, whereas the off-line framework requires matching (or interpolations without physically meaningful variations) between the CTM and meteorological data resolutions.

2.2 Observations

4.1 Global and regional distributions

Figure 2 compares the simulated annual mean tropospheric NO₂ column with satellite retrievals. Both OMI and GOME-2 retrievals showed high tropospheric NO₂ columns over eastern China, the United States, Europe, India, Southeast Asia, Central and South Africa, and South America. For most of these regions, observed concentrations were higher in GOME-2 than OMI, reflecting the difference in overpass time and diurnal variations in tropospheric chemistry (Boersma et al., 2008). All model simulations captured the observed global spatial variation well, with r>0.9 in comparison to both OMI and GOME-2 for annual mean concentration fields. In terms of global averages, the 2.8^∘ simulations were biased on the low side by 40 % compared to OMI and by 47 % compared to GOME-2. This negative global mean bias has commonly been reported using other global CTMs (van Noije et al., 2006; Huijnen et al., 2010b; Miyazaki et al., 2012). As summarized in Table 1, the negative annual global mean bias compared to OMI (GOME-2) was slightly reduced by 5 % (3 %) at 1.1^∘ resolution and by 2 % (1 %) at 0.56^∘ resolution compared to the 2.8^∘ resolution. Global RMSE was reduced by 15 % compared to OMI and GOME-2 by increasing model resolution from 2.8 to 1.1^∘. The improvement when increasing resolution from 1.1 to 0.56^∘ was limited.

Figure 3Monthly time series of tropospheric NO₂ column (× 10¹⁵ molecules cm⁻²) averaged in E-China (110–123^∘ E, 30–40^∘ N), E-USA (95–71^∘ W, 32–43^∘ N), W-USA (125–100^∘ W, 32–43^∘ N), Europe (10^∘ W–30^∘ E, 35–60^∘ N), India (68–88^∘ E, 8–35^∘ N), Mexico (115–90^∘ W, 15–25^∘ N), N-Africa (20^∘ W–40^∘ E, 0–20^∘ N), C-Africa (10–40^∘ E, 20^∘ S–0), S-Africa (26–31^∘ E, 28–23^∘ S), S-America (70–50^∘ W, 20^∘ S–0), and SE-Asia (96–105^∘ E, 10–20^∘ N). The black dots are OMI retrievals, the red dashed line is the model simulation at 2.8^∘ resolution, the yellow dashed-dotted line is the model simulation at 1.1^∘ resolution, and the blue dotted line is the model simulation at 0.56^∘ resolution. The vertical bars indicate mean OMI retrieval errors.

Download

Figure 4Same as Fig. 3, but for GOME-2.

Download

Figure 5Same as Fig. 3, but for root mean square error (RMSE) of tropospheric NO₂ column in comparison to OMI.

Download

Figure 6Same as Fig. 3, but for RMSE of tropospheric NO₂ column in comparison to GOME-2.

Download

Figures 3 and 4 (5 and 6) compare seasonal variations in the regional and monthly mean tropospheric NO₂ column (regional RMSEs) against OMI and GOME-2 in 2008 using data incorporated at a 0.5^∘ bin grid. Because the validation results are similar for OMI and GOME-2 for most cases, the results using OMI are discussed below.

Over eastern China, negative model biases at 2.8^∘ resolution were reduced at 1.1 and 0.56^∘ resolutions from February to July. In December, model bias varied with model resolution: −14 % at 2.8^∘ resolution, +23 % at 1.1^∘ resolution, and −7 % at 0.56^∘ resolution. Negative annual mean bias was reduced by 90 % from 2.8 to 1.1^∘ resolution and by 64 % from 2.8 to 0.56^∘ resolution, with increasing spatial correlations (from r=0.80 at 2.8^∘ resolution to r=0.86 at 1.1^∘ resolution and 0.91 at 0.56^∘ resolution). Annual mean RMSE was also reduced by 32 % from 2.8 to 1.1^∘ resolution and by 9 % from 1.1 to 0.56^∘ resolution.

Over the eastern United States, negative annual mean bias was reduced by 87 % at 1.1^∘ resolution and by 65 % at 0.56^∘ resolution compared to 2.8^∘ resolution. The seasonal bias reduction reached 95 % at 0.56^∘ resolution in summer. Annual RMSE was reduced by 37 % at 1.1^∘ resolution and by 40 % at 0.56^∘ resolution compared to 2.8^∘ resolution. The larger monthly RMSE at 1.1^∘ than at 2.8^∘ resolution during December is attributed to large positive biases over New Jersey and Ohio, although the reason for these is unclear. The spatial correlation for annual mean concentration fields increased from r=0.83 at 2.8^∘ resolution to r=0.93 at 1.1^∘ resolution and 0.96 at 0.56^∘ resolution.

Over the western United States, negative annual mean bias was 13 % lower at 1.1^∘ resolution and 14 % lower at 0.56^∘ resolution compared to 2.8^∘ resolution. In summer, the negative seasonal mean bias at 0.56^∘ resolution was slightly larger, reflecting negative biases over rural areas. RMSE for annual mean fields was reduced by 20 % from 2.8 to 1.1^∘ resolution and by 23 % from 1.1 to 0.56^∘ resolution. The spatial correlation for annual mean fields increased from r=0.65 at 2.8^∘ resolution to r=0.82 at 1.1^∘ resolution and 0.91 at 0.56^∘ resolution.

Over Europe, negative model bias for annual mean concentrations was reduced by 23 % from 2.8 to 1.1^∘ resolution, but was 46 % larger at 0.56^∘ resolution than at 1.1^∘ resolution. Large negative bias over the Po Valley at 2.8^∘ resolution was reduced by 13 % at 1.1^∘ resolution and further reduced by 10 % from 1.1 to 0.56^∘ resolution. In contrast, negative bias over London was larger at 0.56^∘ resolution than at 1.1^∘ resolution by a factor of 4, leading to larger negative regional mean bias at 0.56^∘ resolution. Simulated planetary boundary layer (PBL) height in the 0.56^∘ simulation was substantially higher (by 20 %) than ERA-Interim over London, which may partially contribute to the large NO₂ bias. Annual RMSE was also lower by 16 % at 1.1^∘ resolution and by 9 % at 0.56^∘ resolution than at 2.8^∘ resolution. The spatial correlation for annual mean fields increased from 0.87 at 2.8^∘ resolution to 0.91 at 0.56^∘ resolution.

Over India, negative model biases were smaller at 1.1 and 0.56^∘ resolutions than at 2.8^∘ resolution during January–May, but were larger during June–September. RMSE for annual mean fields was reduced at 1.1 and 0.56^∘ resolutions (except during summer), with 16 and 6 % reductions, respectively. When comparing against OLR and precipitation observations, we found an increased error at 0.56^∘ resolution over India during summer. This suggests the need to further optimize model parameters relevant to tropical convection (see Sect. 2) in order to improve high-resolution NO₂ simulations.

Over Mexico, the spatial correlation for annual mean fields increased substantially at 1.1^∘ (r=0.82) and 0.56^∘ resolutions (r=0.93) compared to 2.8^∘ resolution (r=0.61). Increasing model resolution was important to reduce negative biases around Mexico City, reducing annual RMSE by 17 % at 1.1^∘ resolution and by 38 % at 0.56^∘ resolution compared to 2.8^∘ resolution.

Over South Africa, the negative annual mean bias was reduced by 37 % at 1.1^∘ resolution and by 43 % at 0.56^∘ resolution compared to 2.8^∘ resolution, while annual RMSE was reduced by 46 and 56 % at 1.1 and 0.56^∘ resolutions, respectively. The spatial correlation was 0.93 and 0.97 at 0.56^∘ resolution in contrast to 0.61 at 2.8^∘ resolution. Model resolution higher than 1.1^∘ was thus important for reproducing megacity-scale air pollution over the Highveld region of South Africa, which is a complex source area of coal mining, thermal power generation, metal mining, and metallurgical industry as discussed by Duncan et al. (2016).

Over the selected biomass burning regions (South America, North Africa, Central Africa, and Southeast Asia), all model simulations showed negative biases throughout the year. In most cases, bias reduction with increasing model resolution was limited because most forest fires burn over large extents. Over South America, negative bias for the annual mean concentration was 15 % lower at 1.1^∘ resolution and 12 % lower at 0.56^∘ resolution than at 2.8^∘ resolution. Annual RMSE was reduced by 15 % at 1.1^∘ resolution and by 12 % at 0.56^∘ resolution. The smaller spatial correlation at high resolutions reflects an increased positive bias over a major biomass burning hot spot (12^∘ S, 50^∘ W). Over North Africa, annual RMSE was smaller by 9 % at 1.1^∘ resolution and by 3 % at 0.56^∘ resolution (compared to 2.8^∘ resolution), whereas changes in mean bias and spatial correlation were small. Over Central Africa, negative annual mean bias was reduced by 24 % at 1.1^∘ resolution and by 30 % at 0.56^∘ resolution, while RMSE increased during the biomass burning season (by 11 % at 1.1^∘ resolution and by 24 % at 0.56^∘ resolution). The increased RMSE is associated with increased positive biases around 10–20^∘ S. Over Southeast Asia, RMSE for the annual mean fields was reduced by 7 % at 1.1^∘ resolution and by 5 % at 0.56^∘ resolution compared to 2.8^∘ resolution. The increased errors over strong biomass burning hot spots in high-resolution simulations could be a result of more pronounced influences of largely uncertain inventories for individual burning points (e.g., Castellanos et al., 2015).

Negative biases with respect to GOME-2 were larger than to OMI in all simulations over most regions. The differences suggest that all model simulations underestimated high NO₂ concentrations in the morning. The underestimations could be associated with insufficient vertical model resolution for capturing thin nocturnal boundary layers and uncertainties in HO_x–NO_x–CO–VOCs chemistry, NO₂ photolysis rates, and emission diurnal cycles. The different model biases between OMI and GOME-2 could also be attributed to the bias between these retrievals. Irie et al. (2012) concluded that the bias between these retrievals is small and insignificant for East Asia, whereas the bias between these retrievals is unclear for other regions. For the most anthropogenically polluted regions, bias reductions at 0.56^∘ (compared to 2.8^∘ resolution) were similarly found for OMI and GOME-2. For South America and Central Africa, reductions of the negative bias at 0.56^∘ resolution were larger in the comparison against OMI than GOME-2 during the biomass burning season, suggesting that the high-resolution simulation improves the representation of daytime photochemistry in the presence of enhanced biomass burning emissions.

For the evaluations, we used simulated and observed concentrations interpolated to a 0.5^∘ bin grid. To identify the main drivers of improvements in the high-resolution simulation, we conducted further comparisons using two concentration fields interpolated to 2.8 and 0.5^∘ bin grids. The drivers consist of (1) closer spatial representativeness between observations and simulations (up to approximately 0.5^∘ resolution) and (2) better representation of large-scale (i.e., at 2.8^∘) concentration fields through the consideration of small-scale processes. Error reductions at the 0.5^∘ bin grid include the effects of both drivers. In contrast, error reductions at the 2.8^∘ bin grid are mainly attributed to the latter effect (2). When error reductions at the 2.8^∘ bin grid are about half those at the 0.5^∘ bin grid, the contributions of the two effects should be identical. For annual RMSE reductions, the contributions of the two effects were almost identical over the eastern United States, the western United States, and South Africa (by up to −1.9 and −0.9 × 10¹⁵ molecules cm⁻² at 0.5 and 2.8^∘ bin grids, respectively). In contrast, over eastern China, improved representations on the large scale (2) contributed up to 90 % (i.e., reductions by 1.1 and 1.0 × 10¹⁵ molecules cm⁻² at 0.5 and 2.8^∘ bin grids, respectively, at 1.1^∘ resolution). In this region, the large contribution of the second effect reflected spatially homogeneous error reductions over Hebei and Henan provinces. Over most biomass burning areas, improved representations on the large scale (2) dominated improvements in high-resolution modeling, with RMSE reductions of up to 0.072 × 10¹⁵ molecules cm⁻² for the 2.8^∘ bin grid and 0.071 × 10¹⁵ molecules cm⁻² for the 0.5^∘ bin grid. These results imply that, even for areas with homogeneous concentration and emission fields, high-resolution modeling can have significant impacts through a better representation of large-scale fields.

4.2 Tropospheric NO₂ over strong local sources

Figure 7 compares the detailed spatial distribution of the tropospheric NO₂ column in summer, as represented by OMI measurements and model simulations over four selected polluted areas: East Asia, South Asia, the western United States, and South Africa. Over East Asia, high concentrations were observed over the North China Plain, the Yangtze River Delta, the Pearl River Delta, Seoul, and Tokyo, which could mainly be attributed to emissions from traffic (Zheng et al., 2014) and large coal-fired power plants in the North China Plain (Liu et al., 2015). The 2.8^∘ simulation underestimated these high concentrations and overestimated low concentrations over surrounding areas, probably associated with artificial mixing at coarse model resolution. The 1.1 and 0.56^∘ simulations reduced negative biases over central eastern China, the Pearl River Delta, Seoul, Tokyo, and the western part of Japan. Over the Yellow Sea, the East China Sea, and off the Pacific coast of Japan, the positive biases at 2.8^∘ resolution were mostly removed at 1.1 and 0.56^∘ resolutions. Consequently, regional RMSE was 32 % lower at 0.56^∘ resolution. In contrast, high-resolution simulations led to overestimation over Beijing and the Yangtze River Delta.

Figure 7Tropospheric NO₂ column (× 10¹⁵ molecules cm⁻²) from OMI retrievals (first column; a, e, i, m) and differences between the model simulation at 2.8^∘ (second column; b, f, j, n), 1.1^∘ (third column; c, g, k, o), and 0.56^∘ (fourth column; d, h, l, p) resolutions and OMI retrievals over East Asia (first row; a–d), South Asia (second row; e–h), and the western United States (third row; i–l) during JJA and over South Africa (forth row; m–p) during DJF 2008. Observed and simulated fields are mapped onto a 0.5^∘ bin grid. Regional mean bias (MB) and RMSE are also shown.

Download

Over South Asia, high concentrations were observed over large cities such as New Delhi, Chennai, Mumbai, and Kolkata in India, over Lahore and Multan in Pakistan, and around reported coal-based thermal power plants at 24^∘ N, 83^∘ E and 22^∘ N, 83^∘ E in India (Lu and Streets, 2012; Prasad et al., 2012). The 2.8^∘ simulation was biased on the low side by up to 50 % over these areas, except westward of New Delhi, as commonly reported using another coarse-resolution model at 2.8^∘ resolution (Sheel et al., 2010). These negative biases were reduced by up to 50 % at 1.1 and 0.56^∘ resolutions, whereas high-resolution simulations reveal excessively high concentrations over New Delhi. Over rural areas, negative biases were larger at 1.1 and 0.56^∘ resolutions, resulting in larger regional RMSE than at 2.8^∘ resolution (see Sect. 4.1).

Over the western United States, high concentrations were observed around Los Angeles, San Francisco, Seattle, Phoenix, Salt Lake City, Denver, and the Four Corners and San Juan power plants. Negative biases were reduced with increasing model resolution over most of these regions. In contrast, negative biases remained at 0.56^∘ resolution around strong local sources. Over rural areas, negative biases increased with model resolution, partly reflecting suppressed artificial dilution from strong local sources. As a result, regional RMSE was reduced by 18 and 27 % at 1.1 and 0.56^∘ resolutions compared to 2.8^∘ resolution. Errors, for instance, in soil NO_x emissions in summer (e.g., Oikawa et al., 2015; Weber et al., 2015) could contribute to underestimations over rural areas.

Figure 8Scatter plots of observed and simulated tropospheric NO₂ column (× 10¹⁵ molecules cm⁻²) over strong local sources in East Asia (left column; a, c) and the western United States (right column; b, d) during JJA 2008 for the OMI retrievals (upper row; a–b) and GOME-2 retrievals (lower row; c–d). The red marks are the model simulation at 2.8^∘ resolution, the yellow marks are the model simulation at 1.1^∘ resolution, and the blue marks are the model simulation at 0.56^∘ resolution. The horizontal bars indicate mean retrieval errors in OMI and GOME-2. For East Asia, the results are shown for Beijing (116.38^∘ E, 39.92^∘ N), Tianjin (117.18^∘ E, 39.13^∘ N), Shanghai (121.47^∘ E, 31.23^∘ N), Nanjing (118.77^∘ E, 32.05^∘ N), Guangzhou (113.27^∘ E, 23.13^∘ N), Shenzhen (114.10^∘ E, 22.55^∘ N), Seoul (126.96^∘ E, 37.57^∘ N), and Tokyo (139.68^∘ E, 35.68^∘ N). For the western United States, the results are shown for Los Angeles (118.25^∘ W, 34.05^∘ N), San Francisco (122.42^∘ W, 37.78^∘ N), Seattle (122.33^∘ W, 47.61^∘ N), Salt Lake City (111.88^∘ E, 40.75^∘ N), Phoenix (112.07^∘ W, 33.45^∘ N), Denver (104.88^∘ W, 39.76^∘ N), and the Four Corners and San Juan power plants (108.48^∘ W, 36.69^∘ N). The values and mean retrieval errors are averages within a 50 km distance from each strong source, with weighting function application based on the inverse of the distance from each location.

Download

Over South Africa, high concentrations were observed over the Highveld region of South Africa, a complex source area, as noted in Sect. 4.1. The large negative bias (92 % at 2.8^∘ resolution) in peak concentration over a power plant region in Mpumalanga Province (29.5^∘ E, 26.2^∘ S) was reduced to 69 % at 1.1^∘ resolution and 53 % at 0.56^∘ resolution. Negative bias (75 % at 2.8^∘ resolution) in the Johannesburg–Pretoria megacity area (28^∘ E, 25.7–26.2^∘ S) was also reduced to 54 % at 1.1^∘ resolution and 50 % at 0.56^∘ resolution. High-resolution simulations are thus important for regions with complex and strong local sources. At the same time, the remaining negative bias at 0.56^∘ suggests that power plant and industrial emissions are underestimated, as suggested by Miyazaki et al. (2017), or that a model resolution higher than 0.56^∘ is essential.

Figure 9Regional distributions of tropospheric NO₂ column (× 10¹⁵ molecules cm⁻²) from satellite retrievals (first column; a, e) and differences between the model simulation at 2.8^∘ (second column; b, f), 1.1^∘ (third column; c, g), and 0.56^∘ (fourth column; d, h) resolutions and satellite retrievals from OMI (upper row; a–d) and GOME-2 (lower row; e–h) over Colorado state during July–August 2014. The observed and simulated fields are mapped onto a 0.5^∘ bin grid. The DMA area is shown by the blue square in (a).

Download

Figure 8 compares simulated high NO₂ concentrations with satellite retrievals at selected megacities. Eight strong source points were selected from East Asia and seven points from the western United States during June–July–August (JJA). We consider the summertime to be suitable for evaluating local NO₂ pollution because of the short NO₂ lifetime. For the comparisons, retrieved and simulated tropospheric NO₂ columns were averaged within a 50 km distance from the selected points while applying a distance-based weighting function (i.e., the inverse of the distance was applied to each retrieval).

In comparison to OMI retrievals, with increasing model resolution, the slope for East Asia became closer to 1 (0.67 at 0.56^∘ resolution and −0.19 at 2.8^∘ resolution), and the intercept number became smaller (2.8 at 0.56^∘ resolution and 6.4 at 2.8^∘ resolution). The correlation coefficient also increased (r=0.36 at 0.56^∘ resolution in contrast to $r=-\mathrm{0.31}$ at 2.8^∘ resolution). Large negative biases were reduced at 0.56^∘ resolution by 67 % over Beijing, by 73 % over Tianjin, by 18 % Shanghai, by 90 % over Nanjing, by 62 % over Guangzhou, by 48 % over Shenzhen, by 47 % over Seoul, and by 62 % over Tokyo (compared to 2.8^∘ resolution). The estimated biases at 0.56^∘ resolution are within mean OMI retrieval errors. Reductions in negative biases at 0.56^∘ resolution against GOME-2 were also observed: by 91 % over Beijing, by 70 % over Tianjin, by 76 % over Shanghai, by 67 % over Nanjing, by 32 % over Guangzhou, by 50 % over Shenzhen, by 40 % over Seoul, and by 58 % over Tokyo. However, there is more degradation of slope and intercept against GOME-2 than against OMI, reflecting large negative biases over Guangzhou, Shenzhen, Seoul, and Tokyo.

Over the western United States, NO₂ columns in all model simulations were in agreement with OMI retrievals (r>0.9). The 0.56^∘ model reduced negative biases with respect to OMI by 30 % over Los Angeles, by 74 % over San Francisco, by 98 % over Seattle, by 58 % over Salt Lake City, by 83 % over Phoenix, by 44 % over Denver, and by 78 % over the Four Corners and San Juan power plants (compared to the 2.8^∘ model). These bias reductions resulted in an improved slope number at 0.56^∘ resolution (0.31) compared to 2.8^∘ resolution (0.15). In this region, comparison results were generally similar between OMI and GOME-2. These validation results demonstrate the capability of the 0.56^∘ simulation to represent high concentrations over strong local sources.

Figure 10Vertical profiles of NO (pptv) (a), NO₂ (pptv) (b), NO₂ during mornings (09:00–12:00 LT) (c) and afternoons (13:00–16:00 LT) (d), OH (pptv) (e), HO₂ (pptv) (f), O₃ (ppbv) (g), specific humidity (g kg⁻¹) (h), the photolysis rate of O₃ (s⁻¹) (i), and the OH chemical production rate (molecules cm⁻³ s⁻¹) from O¹D and H₂O (j) over the Denver metropolitan area (39–41^∘ N and 103–105.5^∘ W) during the FRAPPÉ period (from 16 July to 18 August 2014). The black dots represent the measurements, the red dashed line is the model simulation at 2.8^∘ resolution, the yellow dashed-dotted line is the model simulation at 1.1^∘ resolution, and the blue dotted line is the model simulation at 0.56^∘ resolution. The horizontal bars represent the standard deviation of the measurements.

Download

4.3 Validations using FRAPPÉ aircraft measurements

In this section, we evaluated model performance in relation to O₃–HO_x–NO_x chemistry over the Denver metropolitan area (DMA; defined as 39–41^∘ N and 103–105.5^∘ W) in the western United States using the FRAPPÉ campaign observation data and satellite retrievals from July–August 2014. Figure 9 compares the spatial distribution of the tropospheric NO₂ columns between simulations and satellite retrievals around the FRAPPÉ locations. OMI and GOME-2 observed high tropospheric NO₂ columns over the DMA at around 40^∘ N, 105^∘ W. All models underestimated high concentrations by about 50 % at 2.8^∘ resolution compared to OMI, with this declining by 37 % at 1.1^∘ resolution and by 56 % at 0.56^∘ resolution. The negative bias over the DMA was larger for GOME-2 than OMI, suggesting larger underestimations in simulated fields during mornings compared to afternoons. Outside the DMA, negative biases increased by 16 % for OMI and by 11 % for GOME-2 at 0.56^∘ resolution compared to 2.8^∘ resolution. As a result, RMSE against OMI and GOME-2 for the entire domain area was almost constant with varying model resolution.

Figure 11(a) Probability distribution functions of NO, (b) OH, and (c) HO₂ as a function of NO (pptv). The black dots represent measurements, the red dashed line is the model simulation at 2.8^∘ resolution, the yellow dashed-dotted line is the model simulation at 1.1^∘ resolution, and the blue dotted line is the model simulation at 0.56^∘ resolution. The vertical bars represent the standard deviation of the measurements.

Download

Figure 10 compares mean vertical profiles of trace gases and reaction rates over the DMA. Large negative biases of NO and NO₂ at 2.8^∘ resolution were mostly removed at 1.1 and 0.56^∘ resolutions below 650 hPa (by up to 88 %), except at 800 hPa during daytime (09:00–16:00 LT). All simulations revealed large negative biases at 800 hPa during mornings (09:00–12:00 LT), but the bias was greater by 30 % at 2.8^∘ resolution than at 1.1 and 0.56^∘ resolutions. Afternoon lower tropospheric high concentrations (13:00–16:00 LT) were captured well in high-resolution simulations. Strong morning–afternoon variations in the lower troposphere were underestimated by 32 % at 0.56^∘ resolution, by 48 % at 1.1^∘ resolution, and by 62 % at 2.8^∘ resolution. The remaining large bias in the morning at 0.56^∘ resolution could be associated with insufficient vertical model resolution to represent mixing within nocturnal thin boundary layers.

Large negative OH biases at 2.8 and 1.1^∘ resolutions at 850 hPa were reduced by 81 % at 0.56^∘ resolution. From 800 to 750 hPa, the 1.1^∘ simulation showed the closest agreement with observations (0.5–7 %), whereas the 2.8 and 0.56^∘ simulations underestimated OH by 7–21 % and overestimated OH by up to 27 %, respectively. Above 700 hPa, all simulations overestimated OH with a factor of up to 2. All simulations also underestimated HO₂ by 10–32 % below 650 hPa, except at 800 hPa.

OH and HO₂ concentrations depend greatly on NO_x concentrations through O₃–HO_x–NO_x chemistry, as well as HO_x production and OH conversion reactions to peroxy radicals (HO₂ and RO₂) with CO and VOCs. Figure 11a shows the probability distribution function of NO from the FRAPPÉ aircraft observation and the model simulations at 800 hPa over the DMA. The observation revealed a wide range of NO concentrations from 10–10 000 pptv. The 2.8^∘ simulation overestimated the occurrence of concentrations < 100 pptv and underestimated the occurrence of concentrations > 100 pptv. The 1.1 and 0.56^∘ simulations captured the observed probability distribution function, although they slightly overestimated the peak frequency concentration and underestimated the occurrence of low (< 100 pptv) and high (> 1000 pptv) concentrations. Figure 11b shows the OH–NO relationship used to validate O₃–HO_x–NO_x chemistry. The observation showed OH increase with increasing NO to 350 pptv and a decrease with increasing NO from 350 pptv; all simulations captured the lower part (NO < 800 pptv) of the observed NO–OH relationship, suggesting that the model realistically simulates nonlinear O₃–HO_x–NO_x chemistry. The lack of high NO (> 800 pptv) with low OH resulted in an overestimation of mean OH concentrations at 0.56^∘ resolution.

Figure 11c compares the HO₂–NO relationship. All simulations underestimated the occurrence of high HO₂ (> 25 pptv) at low NO (< 100 pptv). This implies an underestimation of HO_x chemical production in the simulations. We evaluated HO_x production from the chemical reaction of O(¹D) with H₂O using temperature, specific humidity, O₃ photolysis rate to O(¹D) (J${}_{{\mathrm{O}}_{\mathrm{3}}\to \mathrm{O}{(}^{\mathrm{1}}\mathrm{D})}$), and O₃ concentration with the assumption of O(¹D) equilibrium (Fig. 10g–j). All simulations underestimated HO_x production, with the underestimation being smaller by 13 % at 0.56^∘ resolution at 800 hPa. The underestimation of HO_x production was primarily attributable to a negative bias in O₃ by 11 % and J${}_{{\mathrm{O}}_{\mathrm{3}}\to \mathrm{O}{(}^{\mathrm{1}}\mathrm{D})}$ by 2.5 % at 0.56^∘ resolution at 800 hPa. The negative biases of O₃ and J${}_{{\mathrm{O}}_{\mathrm{3}}\to \mathrm{O}{(}^{\mathrm{1}}\mathrm{D})}$ were reduced by 39 and 58 %, respectively, at 0.56^∘ (compared to 2.8^∘) resolution. Biases in specific humidity also had small impacts on calculated HO_x production. Positive biases of specific humidity at 2.8^∘ resolution above 750 hPa were reduced by up to 83 % at 0.56^∘ resolution. The lack of nitrous acid (HONO) in the model could explain a component of the HO_x production underestimation, especially during mornings (e.g., Kanaya et al., 2001). The underestimation of OH conversion to peroxy radicals could also explain simulated errors in OH and HO₂. Griffith et al. (2016) attributed OH overprediction and HO₂ underprediction in a box model simulation to underestimation of total OH reactivity (i.e., missing OH sink) over the United States.

Figure 12Latitude–pressure distribution of zonal mean (a–c) O₃ (ppbv) and (d–f) OH (× 10⁵ molecules cm⁻³) in the model simulation at 2.8^∘ (left column) during JJA in 2008 and differences between the model simulation at 1.1^∘ (middle column) and 0.56^∘ (right column) resolutions and the model at 2.8^∘ resolution.

Download

The 2014 simulations used the anthropogenic emission inventory for the year 2010 (see Sect. 2.1). The optimized NO_x emissions from an assimilation of multiple species satellite measurements (Miyazaki et al., 2017) suggest that surface NO_x emissions over the DMA in July–August increased by 7 % from 2010 to 2014. The temporal variation, together with large uncertainties in the emission inventories, could explain part of the negative biases of NO and NO₂ at 800 hPa, which also affects OH, HO₂, and O₃ through nonlinear chemistry processes.

6.1 Other model error sources

Various factors other than horizontal model resolution can lead to errors in tropospheric NO₂ simulation. Insufficient vertical model resolution could introduce additional errors in vertical mixing, atmospheric transport, and subsequent chemistry processes, for instance, under stable boundary layer conditions during nighttime (Menut et al., 2013). Such errors could also cause large negative NO₂ biases during mornings in the lower troposphere (see Sect. 4.3). More detailed validation of diurnal variations is required using ground-based observations such as MAX-DOAS and lidar in future work.

Chemical kinetics information could also have large uncertainties. Lin et al. (2012) and Stavrakou et al. (2013) suggested that the uptake of HO₂ on aerosols is the most important factor but remains largely uncertain. CHASER includes simplified NO_x–VOC chemistry related to PANs and isoprene nitrates (Sudo et al., 2002). The incorporation of more detailed NO_x–VOC chemistry would also be needed to improve simulated peroxy nitrates and organic nitrates, as per Ito et al. (2007, 2009) and Fischer et al. (2014).

Surface emissions are another important error source. The total amounts of anthropogenic NO_x emissions in China in 2008 differ by 27 % between two (highest and lowest) bottom-up inventories: EDGAR4.2 and MEIC (Saikawa et al., 2017). Ding et al. (2017a) also discussed large diversity in emission inventories over East Asia. Biomass burning NO_x emissions also differ significantly between inventories: for example, the annual mean emission is 2.293 Tg yr⁻¹ in GFASv1.0 in contrast to 2.700 Tg yr⁻¹ in GFEDv3.1 over the SH Africa, as reported by Kaiser et al. (2012).

Based on data assimilation of multiple species satellite measurements, Miyazaki et al. (2017) investigated large uncertainty in anthropogenic and fire-related emission factors and a significant underestimation of soil NO_x sources in bottom-up emission inventories. Using a similar approach, Miyazaki et al. (2014) optimized lightning NO_x sources and indicated that the widely used lightning parameterization based on the C-shape assumption (Price and Rind, 1992; Pickering et al., 1998) has large uncertainty. Implementing these optimized emissions could improve model performance, although optimal emissions could also be dependent on model resolution.

Representations of meteorological parameters, such as cloud optical depth, temperature, water vapor, PBL height, and relevant transport and chemical processes, are also important in tropospheric NO₂ simulations (Lin et al., 2012). Because we employed an AGCM–CTM online coupling system, meteorological fields are simulated explicitly at each model resolution. This could help to improve the tropospheric chemistry simulation. For instance, we found that simulated regional mean PBL height is sensitive to the choice of model resolution, with the 0.56^∘ simulation showing closer agreements with ERA-Interim reanalysis, as discussed in Sect. 4.3. Resolving small-scale cloud distributions may lead to improved photolysis and convective transport calculations in high-resolution simulations.

Nevertheless, the AGCM meteorological fields still need to be carefully validated and improved. For instance, cumulus convection and cloud parameterization calculations were sensitive to model resolution. Although the relevant model parameters have been optimized separately for each model resolution, there are still some discrepancies against observed OLR and precipitation distributions (see Sect. 2).

6.2 Nonlinearity in model error reductions

Model performance was clearly better at 0.56 and 1.1^∘ resolutions than at 2.8^∘ resolution in most cases. The 0.56^∘ simulation largely improved spatial variations over eastern China, the eastern and western United States, Mexico, and South Africa, as confirmed by large RMSE reductions, especially for megacities and regions with power plants (see Sect. 4). In most cases, the improvement was smaller from 1.1 to 0.56^∘ resolution than from 2.8 to 1.1^∘ resolution. Meanwhile, regional RMSEs increased at 0.56^∘ from 1.1^∘ resolution for some cases over Europe, India, and the selected biomass burning regions, possibly related to more pronounced errors in meteorological fields for Europe and India and in biomass burning hot spot emissions.

Comparisons to aircraft measurements showed better performance of NO_x simulation at high resolutions. However, the representation of NO variability (i.e., the probability distribution function) was insufficient even at 0.56^∘ resolution. Further improvements could be obtained using a model with resolution finer than 0.56^∘. For instance, Valin et al. (2011) noted that 4 and 12 km resolutions are required for Four Corners and Los Angeles, respectively, to accurately simulate the nonlinear chemical feedback of the O₃–HO_x–NO_x system. Yamaji et al. (2014) reported that errors in the simulated tropospheric NO₂ column at 20 km resolution did not yet approach convergence over Tokyo.

Most previous high-resolution modeling studies have used regional models to simulate NO₂ concentration fields at high spatial resolution, primarily focusing on urban regions, with reduced or equivalent computational costs compared to global models. Several studies have demonstrated that a better representation of the long-range transport of NO_x reservoir species such as peroxyacetyl nitrate (PAN) are important on simulated NO₂ in the free troposphere in remote areas (e.g., Hudman et al., 2004; Fischer et al., 2010, 2014; Jiang et al., 2016). A two-way nesting between regional and coarse-resolution global models (e.g., Yan et al., 2016) is able to consider both small-scale processes inside focusing regions and long-range transport over the globe, which has an advantage over regional models. An important advantage of global high-resolution models over regional models and two-way nesting systems is the ability to simulate NO₂ concentration fields at high resolutions over the entire globe across urban, biomass burning, and remote regions in a consistent framework. Even over remote regions, a high-resolution simulation has the potential to improve model performance through considering the effects of nonlinear chemistry in highly concentrated NO_x plumes emitted from ships and lightning (Charlton-Perez et al., 2009; Vinken et al., 2011; Gressent et al., 2016). These NO_x emission sources in remote regions have significant impacts on climate and air quality (Eyring et al., 2010; Holmes et al., 2014; Banerjee et al., 2014; Finney et al., 2016). It is thus important to clarify the importance of resolving small-scale sources and plumes within a global modeling framework for a better understanding of the global atmospheric environment and chemistry–climate system.

Williams et al. (2017) also showed that the differences in NO₂ profiles between the TM5-MP model at 3^∘ × 2^∘ and 1^∘ × 1^∘ horizontal resolutions are within a few percentage points below 850 hPa over the Pacific in boreal spring. They also showed much larger differences with changing model resolution over Texas in autumn. Model resolution impact thus varies significantly with location and season. Further investigations using other aircraft measurements would be helpful to evaluate model performance in different cases.

High-resolution chemical transport modeling requires huge computational resources: e.g., compared to the simulation at 2.8^∘ resolution (approximately 480 s computer time for a 1-day simulation), the computational cost increased by a factor of 67 at 0.56^∘ resolution (approximately 32 000 s computer time) and by a factor of 14 at 1.1^∘ resolution (approximately 6700 s computer time). High-performance computing (HPC) systems are thus essential for performing high-resolution simulations. At the same time, because the size of a 3-D array is large in the high-resolution model, computational efficiency is important: e.g., efficient data throughputs in memory transfer, network communication between multiple nodes, and file input–output. In the future, further improvements in computational efficiency will be required, together with the development of HPC systems.

6.3 Application for satellite retrieval and data assimilation

An important application of high-resolution tropospheric NO₂ simulations is to provide a priori profile information on satellite retrieval and chemical data assimilation (Liu et al., 2017). Here, we would like to discuss the potentials of the obtained results for these applications.

Current satellite retrievals of the tropospheric NO₂ column use a priori NO₂ profiles obtained from global model simulations at relatively coarse resolutions: from TM5 at 3^∘ × 2^∘ in DOMINO-2 (Boersma et al., 2011) and GEOS-Chem at 2.5^∘ × 2^∘ in OMNO2 (Bucsela et al., 2006; Celarier et al., 2008), whereas the TROPOMI retrieval product will employ 1^∘ × 1^∘ resolution simulation fields from TM5 (Williams et al., 2017). To provide high-resolution (ranging from 4 to 50 km) a priori information, several regional retrievals have employed regional models (Heckel et al., 2011; Russell et al., 2011; Lin et al., 2014), showing improvements in the retrieved fields in comparison to independent observations. High-resolution a priori fields from global CTMs are important in providing consistent global datasets.

To avoid spatial representation gaps between satellite measurements and coarse-resolution global models, super-observation techniques have been employed to produce representative data before assimilation (e.g., Miyazaki et al., 2012). The average of averaging kernels over a number of retrievals within a super-observation grid does not hold any physical meaning. This may inhibit effective improvement by assimilating over regions with varying conditions. High-resolution CTMs allow for the assimilation of satellite measurements, with reduced representation gaps without any averages.

Because of the distinct nonlinearity in chemical reactions, the high-resolution assimilation of satellite measurements considering small-scale variations in background error covariance would be essential in making the best use of observational information. High-resolution chemical data assimilation could also benefit air pollutant emission estimates (e.g., Miyazaki et al., 2014, 2017; Liu et al., 2017), especially using high-resolution measurements from future satellite missions such as TOROPOMI and geostationary satellites (e.g., Sentinel-4, GEMS, TEMPO), even when model resolution is still coarser than the measurement resolution through improved model processes and spatial representativeness for mega-cities as demonstrated by this study.