Table 1 summarizes the prescribed anthropogenic and biomass burning emissions. Global anthropogenic emissions are taken from the Emission Database for Global Atmospheric Research (EDGAR) v4.2 inventory for CO and NOx. Anthropogenic emissions of NMVOCs use as default the REanalysis of the TROpospheric chemical composition (RETRO) monthly global inventory for 2000, as implemented by Hu et al. (2015). These global inventories are further replaced by regional inventories over Asia, North America and Europe. Emission data include monthly or seasonal variability.

Monthly biomass burning emissions are taken from the Global Fire Emissions Database version 3 (GFED3) (van der Werf et al., 2010). Other natural emissions (lightning NOx, soil NOx, and biogenic NMVOCs) are parameterized and calculated on-the-fly based on model meteorology; these emissions are thus resolution dependent. Soil NOx emissions follow Hudman et al. (2012). Lightning NOx emissions follow the Price and Rind scheme with a further satellite-based adjustment and a backward “C-shape” vertical profile (Price and Rind, 1992; Ott et al., 2010; Murray et al., 2012). Biogenic NMVOC emissions are calculated with the MEGAN v2.1 (PECCA) model (Guenther et al., 2012) driven by monthly mean MODIS leaf area index data.

Table 2 shows slight differences in global total emissions of ozone precursors (CO, NOx, and NMVOCs) between the global model alone and the two-way coupled system. In the coupled system, global emissions from all sources are about 878 Tg yr-1 for CO, 45.5 Tg N yr-1 for NOx and 723 Tg C yr-1 for NMVOCs. These values are larger than those in the global model by about 0.9, 0.7 and 6.5 %, respectively. Greater emission differences are found for biogenic NMVOCs (by 6.9 %) and fertilizer soil NOx (by 25.4 %), reflecting strong resolution dependence.

Figure 2 shows the spatial distributions of annual NMVOCs and NOx emissions in the nested models (first and third columns) and the global model (second and fourth columns). The nested and global models exhibit similar spatial patterns for NMVOCs emissions. Summed over a given nested domain, the nested models have higher emissions of NMVOCs than the global model by 16–48 %, mainly a result of stronger isoprene emissions. The spatial patterns of NOx emissions differ greatly between the nested and global models, with local emission spikes much more obvious in the nested models, although the regional totals are similar (within 5 %).

The differences in model representation of NOx and NMVOCs emissions affect the simulated ozone chemistry. The difference in regional emission magnitude (mainly for biogenic NMVOC in summer) affects the surface ozone simulation within the nested domains (Sect. 5.1), but with a marginal effect on the global tropospheric ozone as a whole (Sect. 4.3). The better resolved emission spatial variability, as well as associated chemical contrast by the nested models, greatly affects both the surface (Sect. 5.1) and the tropospheric ozone (Sect. 4.3, 5.2, and 5.3).

We employ four measurement networks to evaluate the modeled ground-level ozone mixing ratios in 2009. As shown in Fig. 1, these networks contain hourly ozone measurements from a total of 1420 urban, suburban or remote sites from WDCGG (64 sites, http://ds.data.jma.go.jp/gmd/wdcgg/cgi-bin/wdcgg/catalogue.cgi), GMD (12 sites, http://www.esrl.noaa.gov/gmd/), EMEP (130 sites, http://www.nilu.no/projects/ccc/emepdata.html) and AQS (1214 sites, http://aqsdr1.epa.gov/aqsweb/aqstmp/airdata/download_files.html). For model evaluation, we derive the afternoon (12:00–18:00 LT, local time) mean ozone mixing ratios from the hourly data. Modeled afternoon ozone is sampled from the lowest layer (centered at ∼ 0.065 km) in grid cells covering the ground sites, and are sampled from the hourly outputs coincident with available measurements. The afternoon mean ozone is close to the maximum 8 h average ozone in both measurements (36.1 vs. 39.3 ppb averaged over the 1420 sites) and model simulations (46.8 vs. 48.4 ppb for the global model alone; 42.6 vs. 44.5 ppb for the two-way coupled system). Models also capture the diurnal cycle of measured ozone fairly well, although with positive biases in both daytime and nighttime (not shown), consistent with our previous work (Lin and McElroy, 2010).

We take ozone vertical profiles in 2009 at 11 sites of the MOZAIC program (pink squares in Fig. 1; data available at http://www.iagos.fr/web/) to evaluate the modeled vertical and seasonal distributions of tropospheric ozone. Located in major cities, these sites are representative of the polluted environment. Since 1994, the MOZAIC program has employed five commercial aircrafts to measure ozone concentrations throughout the troposphere and lower stratosphere (Marenco et al., 1998). Ozone is measured with an accuracy estimated at ±(2 ppbv + 2 %) and a 4 s time response (< 50 m vertical resolution) (Thouret et al., 1998). We use measurements taken during both takeoff and landing of the aircrafts to represent the vertical profiles over the associated airports (Zbinden et al., 2013). Each of the 11 sites chosen here has at least 40 profiles in 2009. Measurements are available from the ground level (0.075 km) to the upper troposphere and lower stratosphere (UTLS) at 0.15 km intervals. Model results are sampled at times and locations consistent with the measurements.

For model evaluation in the remote areas, we use 282 ozone vertical profiles over the Pacific Ocean from two HIPPO (HIPPO-1 and HIPPO-2) aircraft campaigns conducted in 2009. The HIPPO campaigns were conducted in the remote troposphere over the Pacific, Arctic and near-Antarctic regions to facilitate atmospheric chemistry analysis (Wofsy, 2011). During HIPPO, ozone was measured by the NOAA O3 photometer using direct absorption at 254 nm (Proffitt and McLaughlin, 1983; Kort et al., 2012). We use the merged data set that has a vertical resolution of 0.1 km (data available at http://hippo.ornl.gov/data_access/). To ensure spatiotemporal consistency with the HIPPO data, model ozone is sampled at the times and locations of the measurements.

We use two monthly OMI tropospheric column ozone (TCO) products that have been used to study the tropospheric ozone variability and sources (Ziemke et al., 2011; Kim et al., 2013). The first product is based on an optimal estimation technique by Liu et al. (2007, 2010) with modifications as described in Kim et al. (2013), and is referred to as OMI/LIU hereafter. For OMI/LIU, errors for individual TCO retrievals are typically 2–5 DU (Liu et al., 2010). Validation against ozonesonde data shows that mean OMI/LIU TCO agrees with ozonesonde data to within 2 DU for both the tropics (30∘ S–30∘ N) and northern mid-latitudes (30–60∘ N), but with season-dependent biases, varying from -0.8 DU in summer (JJA) to 2.1 DU in winter (DJF) for 30∘ S–30∘ N, and varying from -0.1 DU in JJA to 3 DU in DJF for 30–60∘ N (X. Liu, personal communication, 2015). The second product is the OMI/Microwave Limb Sounder (MLS) data set that subtracts the OMI total column ozone by the MLS stratospheric ozone (Ziemke et al., 2011). Ziemke et al. (2011) validated the OMI/MLS data against the Southern Hemisphere Additional OZonesondes (SHADOZ) and the World Ozone and Ultraviolet radiation Data Center (WOUDC) ozonesonde measurements. They found that, on average, the monthly mean OMI/MLS tropospheric ozone mixing ratio is smaller than the ozonesonde data by about 1 ppb (2 %), with large seasonal dependence and a root mean square error at 6–8 ppb. For the present analysis, we average these two independent TCO data sets to reduce data uncertainties; this leads to a third data set referred to as OMI_MEAN.

We use the monthly mean OMI products for 2009. The OMI/LIU data set is on a 2.5∘ long. × 2∘ lat. grid. The OMI/MLS product provides data at 1.25∘ long. × 1∘ lat. from 60∘ S to 60∘ N. We calculate the OMI_MEAN TCO after re-gridding the OMI/MLS data to match OMI/LIU. Data polarward of 60∘ are discarded due to higher uncertainty. Modeled monthly mean TCO is calculated from all daily data at the OMI overpass time (13:00–15:00 LT) applied with the monthly mean OMI/LIU averaging kernel; daily averaging kernel data are not available, and the modeled global annual average TCO with and without applying the averaging kernel differ by 0.6 %. These OMI products and model simulations differ between each other in definitions of tropopause height and days of valid data, whose effects are found to be small. To examine the effect of different tropopause heights, we re-calculated in a test analysis the OMI/LIU, OMI_MEAN and model TCO by applying the OMI/MLS tropopause. The resulting bias of the global model relative to OMI_MEAN (2.8 DU, 8.9 %) is similar to the bias without adjusting the tropopause (2.9 DU, 9.2 %). The differences in days of valid data also have a marginal effect, as confirmed by examining the TCO difference between OMI/MLS and global model simulation sampled from days with valid OMI/MLS data (note that the OMI/MLS product also provides daily data for such analysis). The calculated TCO difference (3.9 DU; 12.8 %) is close to the difference (4.0 DU; 13.1 %) without sampling model results.

Table 3 contrasts the global tropospheric O3 budgets in 2009 simulated by the two-way coupled system against those by the global model alone. The chemical production and loss are calculated for the odd oxygen family (Ox= O3+ O + NO2+ 2NO3+ 3N2O5 + PANs + HNO3+ HNO4), following Wu et al. (2007). The chemical production of Ox is mainly driven by reactions of NO with peroxy radicals, and the chemical loss is mostly due to the O(1D) + H2O reaction and reactions of ozone with OH and HO2. The coupled system produces slightly higher (by ∼ 1.0 %) chemical loss and production of Ox than the global model alone. Ozone dry deposition in the coupled system (867 Tg) is smaller by 1.7 % than the global model alone (882 Tg). The STE of ozone in the coupled simulation (478 Tg) is also lower than the global model alone (488 Tg) by 2.0 %, partly compensating for the weaker deposition. This small difference in STE affects the simulated global tropospheric mean ozone by 1.1 % (see Sect. 4.3).

Table 3 shows that the coupled system produces a tropospheric ozone burden at 348 Tg, about 9.5 % lower than the burden simulated by the global model alone (384 Tg). Correspondingly, the lifetime of tropospheric ozone in the coupled system (burden divided by sink = 23.5 days) is shorter than that in the global model (26.1 days) by 9.9 %. The large reduction in ozone burden and lifetime, despite the small change in chemical production and loss of Ox, reflects a faster chemical evolution of ozone on a per molecule basis. Although both lifetimes calculated here are broadly consistent with previous studies (19.9–25.5 days from ACCMIP, Young et al., 2013; and 17.3–25.9 days from ACCENT, Stevenson et al., 2006), the reduction due to our model coupling indicates a significant effect of small-scale processes resolved by the finer resolution, especially the fine-scale spatial variability of emissions and associated chemistry.

Table 4 shows the seasonal dependence of ozone burden and Ox chemical loss and production. The global model alone produces the largest chemical loss in the Northern Hemisphere (NH) summer (1252 Tg) and the smallest loss in winter (1036 Tg). The coupled model reduces the chemical loss by 1.2 % (to 1237 Tg) in NH summer, due to a lower ozone abundance overcompensating for a higher per-molecule loss rate. In winter, the coupled model enhances the loss by 2.3 % (to 1060 Tg), because a higher per-molecule loss rate from reactions with NOx more than offsets a lower ozone abundance. By comparison, the coupled model slightly increases the chemical production by 0.3–1.3 % in all seasons.

Table 3 shows that the two-way coupling also significantly affects the tropospheric burdens of ozone-related species. Burdens of NMVOCs (10.2 Tg C; see footnote of Table 3 for species included), NOx (0.176 Tg N) and CO (398 Tg) in 2009 are higher than those simulated by the global model alone by 1.0, 4.1 and 10.8 %, respectively. Table 3 also shows that the global annual mean air-mass-weighted tropospheric OH simulated by the two-way coupled system is lower by 5.0 % than that simulated by the global model alone (1.12 vs. 1.18 × 106 cm-3). The sensitivity of OH to model resolution is broadly consistent with previous studies (Yan et al., 2014; Wild and Prather, 2006; Valin et al., 2011). In particular, Yan et al. (2014) showed a similar OH reduction by 4 % via the two-way coupling based on an earlier version of GEOS-Chem (v8-3-02).

Table 3 further presents methane and MCF lifetimes due to reactions with the tropospheric OH. The lifetime calculation follows the formulae used by Yan et al. (2014); it accounts for the grid-box-specific air mass, temperature-dependent reaction constant, OH content and vertical gradients of methane and MCF with an adjustment coefficient of 0.97 for methane (Predoi-Cross et al., 2006) and 0.92 for MCF (Prather et al., 2012). The coupled system leads to longer lifetimes than the global model alone, by about 5.2 % (from 5.58 to 5.87 yr) for MCF and 5.1 % (from 9.63 to 10.12 yr) for methane. These results are closer to the observation-based estimates of MCF lifetime (6.0 ± 0.4 yr from Prinn et al., 2005; 6.3 ± 0.4 yr from Prather et al., 2012) and methane lifetime (10.2 ± 0.8 yr from Prinn et al., 2005; 11.2 ± 1.3 yr from Prather et al., 2012).

Compared to the global model alone, the two-way coupled system produces lower global tropospheric mean ozone by 9.5 % (Table 3). This difference is driven by four factors including the sub-coarse-grid chemical variability resolved by nested resolution (i.e., emission spatial variability and associated chemical contrast), the sub-coarse-grid variability of non-chemical factors (such as topography), a slight difference in the magnitude of natural emissions (mainly for biogenic NMVOC emissions, Sect. 2) and a slight difference in the magnitude of STE (Sect. 4.1). To delineate the individual effects of these factors, we conducted three additional sensitivity simulations from July 2008 to December 2009 as follows. Results are summarized in Table 5.

The first test simulation was conducted with the global model alone, by adopting at each time step the emissions outputted from the two-way system. Here the global model has the same emission magnitude as the two-way model, which is slightly larger than the original global model simulation (Sect. 2). As a result, the simulated global tropospheric mean ozone was enhanced by 1.1 % relative to the original global model simulation. By linear subtraction, we determine that factors other than emission magnitude lead to an ozone reduction by 10.6 % from the global model alone to the two-way system.

The second test is the counterpart of the first test, by re-running the two-way system and adopting the emissions outputted from the global model simulation. Here the actual resolution of emissions is 2.5∘ long. × 2∘ lat.; thus, the sub-coarse-grid variability of emissions and associated chemical contrast is not resolved. The resulting tropospheric ozone is lower than the original global model by 2.0 %. This difference represents the combined effect of the difference in the magnitude of STE and the sub-coarse-grid variability in non-chemical factors.

The third test addresses the slight difference in STE. The test re-run the global model but with a reduction in the STE by 1.0 %, by scaling down the Linoz stratospheric ozone production rate. As a result, the global tropospheric mean ozone is reduced by 0.55 %. By linear scaling, we determine that a 2.0 % reduction in STE from the global model to the two-way system (Table 3) would lead to a 1.1 % reduction in the global tropospheric mean ozone. Combining the result here and from the second test implies that the sub-coarse-grid non-chemical processes would reduce the global tropospheric mean ozone by 0.9 % from the global model alone to the two-way system.

In summary, of the -9.5 % tropospheric mean ozone change from the global model to the two-way coupled simulation, -0.9 % is related to sub-coarse-grid non-chemical processes, -1.1 % is related to the lowered STE, +1.1 % is associated with the increased natural emission magnitude, and the remaining -8.6 % represents the effect of sub-coarse-grid emission–chemical variability. Thus, the small-scale nonlinear chemical processes (resolved by the nested resolution but not by the coarse resolution) are the dominant driver of the overall ozone difference.

As shown in Fig. 1, most ground measurement sites are located in the US (1214 sites from AQS) and Europe (130 sites from EMEP). Averaged over the US AQS sites, the measured annual mean afternoon (12:00–18:00 LT) mean ozone is 35.8 ppb in 2009. The afternoon ozone is slightly higher over Europe, about 37.7 ppb averaged over the EMEP sites. The ozone level is highest over Asia, with a value of 43.1 ppb averaged over the seven WDCGG sites. The afternoon ozone from the 17 WDCGG sites worldwide is about 33.8 ppb on average.

Figure 3 shows the horizontal distributions of annual mean afternoon ozone biases simulated by the global model alone (Fig. 3a, c and e) and by the two-way coupled system (Fig. 3b, d and f), relative to the four ground networks. All model results are derived from the global model component, i.e., from the 2.5∘ long. × 2∘ lat. grid cell covering a given site. The global model tends to overestimate the ozone concentrations (biases range from -5 to 25 ppb), with a mean bias at +10.8 ppb globally, +10.5 ppb over the US, and +12.1 ppb over Europe. The positive biases exceed 15 ppb at several high-elevation sites of the western US (Fig. 3c) and some coastal sites of Europe (Fig. 3e). These results are broadly consistent with previous multi-model evaluation for the Hemispheric Transport of Air Pollution (HTAP) (Reidmiller et al., 2009) and Atmospheric Composition Change European Network of Excellence (ACCENT) (Dentener et al., 2006) projects that showed an ensemble mean positive bias at 10–20 ppb over the summertime eastern US and 15–20 ppb over South Asia, respectively. Similar model biases are also found from our previous evaluation of MOZART and GEOS-Chem over the US (Lin et al., 2008; Lin and McElroy, 2010). Compared to the global model alone, the two-way coupled system generally reduces the ozone bias worldwide (Fig. 3b, d, and f). The positive bias is reduced to 6.7 ppb globally, 6.6 ppb over the US and 7.5 ppb over Europe. The bias reduction is apparent at several WDCGG sites over the North Pacific and North Atlantic (comparing Fig. 3a and b) and over the eastern US (comparing Fig. 3c and d). The two-way simulation biases against the EMEP measurements are larger than those for the EMEP/MSC-W regional CTM at a horizontal resolution of 50 km × 50 km driven by year-specific emissions (within 10 %) (http://emep.int/publ/reports/2014/sup_Status_Report_1_2014.pdf). Our higher biases are partly because the 2005 EMEP NOx emissions used here are higher than those in 2009 by 25.3 % (http://webdab.umweltbundesamt.at/official_country_trend.html).

Figure 4 compares the modeled and measured day-to-day time series of regional mean afternoon ozone in 2009 for six regions in the US (from AQS) and two regions in Europe (from EMEP). The regions are defined in Fig. 3c–f, as separated by blue lines. In general, the measured ozone levels are the highest in spring and summer (Fig. 4, black lines), due to stronger STE and/or higher chemical production. Both the global model and the two-way coupled system capture the seasonal variation of measured ozone (Fig. 4, blue and red lines). The global model alone tends to overestimate the observations; the annual mean bias is 9–15 ppb for any given region. Seasonally, the overestimate is the largest in winter over the western US (Fig. 4a and b), in summer over the eastern US and northern Europe (Fig. 4c–g) and in fall over southern Europe (Fig. 4h). The two-way coupled simulation reduces the ozone biases in most days and regions (Fig. 4, red lines). On a seasonal mean basis, the largest reductions occur in winter (2–8 ppb for individual regions), due mainly to much enhanced titration by NOx (not shown). The bias reductions are smallest in summer (< 3 ppb). It is partly because the enhanced ozone production from the increased natural precursor emissions (Table 2) compensate to some extent for a stronger chemical ozone loss; a sensitivity global model simulation adopting emissions in the two-way system produces more summertime ozone than the original global model by 1.7 ppb over the eastern US (100–70∘ W, 30–50∘ N) and by 2.1 ppb over Europe (10∘ W–30∘ E, 35–70∘ N). Furthermore, although the nested models reduce the net chemical production of ground-level ozone (Sect. 4.3), the effect is partly offset by stronger vertical transport that brings more high-ozone air aloft down to the ground (Roelofs et al., 2003; M. Lin et al., 2012b). The persistent large summertime bias may also be due to some non-resolution-dependent factors such as isoprene nitrate chemistry and dry deposition (Lin et al., 2008; Fiore et al., 2014; Monks et al., 2015). Although the two-way coupling leads to a relatively small improvement in summertime ground-level ozone simulations over the US and Europe (Fig. 4), the coupling results in large error reductions for tropospheric ozone (see Sect. 5.2 and 5.3 below).

Figure 5 compares the day-to-day time series of modeled afternoon ozone against the observations at 12 background sites from WDCGG (panels a, b for Europe, c, d for US, e–g for Asia, h for North Pole, i for Mauna Loa in the North Pacific and j for the Southern Hemisphere). Each observation site provides a nearly complete hourly data set for model evaluation. Although the global model alone and the two-way coupled system generally overestimate the observed ozone, both simulations reproduce the observed temporal patterns fairly well. At 11 sites, the correlation between modeled and observed ozone time series exceeds 0.61 and 0.55 for the coupled system and the global model, respectively. At six sites, the correlation exceeds 0.75 and 0.71, respectively. Compared to the global model alone, the coupled simulation is closer to the observations with a lower bias and higher correlation. At MLO (outside the nested domains, Fig. 5i), the coupled system produces a positive bias of 3.0 ppb with a correlation coefficient at 0.61, compared to the values at 8.2 ppb and 0.59 for the global model alone. Over the three Asian sites (Figs. 5e–g), the coupled system reduces the biases by 7.1, 5.1 and 6.1 ppb. Although both simulations capture the observed temporal variability at the two tropical Asian sites (Fig. 5e and f) with correlation coefficients exceeding 0.66, their performances are poorer at the mid-latitude mountain site (Fig. 5g) due to large overestimates in the cold seasons and much smaller biases in the warm seasons. Nevertheless, the spring–summer high values at these Asian sites are captured fairly well by both simulations. The spring–summer peaks are also found for other Asian regions (Lin et al., 2009; Wang et al., 2011).

Figure 6 further presents for individual sites the day-to-day correlation and mean bias of simulated afternoon ozone relative to the observations. Figure 6a presents the results for all 1420 sites. It shows that compared to the global model alone, the two-way coupled simulation increases the correlation for 1179 sites and decreases the bias for 1221 sites. Averaged over all sites, the correlation is increased from 0.53 to 0.68, and the bias is reduced from 10.8 to 6.7 ppb. Figure 6b further shows the evaluation results at the 25 sites outside the nested domains from WDCGG and GMD. The two-way coupled simulation results are within 5 ppb of the observations at 21 sites, compared to 17 sites for the global model alone. Averaged across all the 25 sites, the coupled simulation has a mean bias at 2.2 ppb and correlation at 0.74, compared to the global model bias at 4.6 ppb and correlation at 0.61. These results again indicate the improvement by the two-way coupling for ozone simulations both within and outside the nested domains.

The black lines in Fig. 7a–k show the measured vertical profiles of tropospheric ozone averaged over 2009 at individual MOZAIC sites. In general, the measured ozone increases with height, from 20–40 ppb in the lower troposphere to 40–70 ppb at 5 km, and to larger values in the upper troposphere. For the HIPPO campaigns (black line in Fig. 7l), the average ozone mixing ratio is between 20 and 50 ppb below 9 km.

The red and blue lines in Fig. 7 show the ozone profiles simulated by the two-way coupled system and the global model alone, respectively. Here the model evaluation is focused on ozone biases below 9 km, the mean tropopause height. Both simulations capture the general vertical structures of MOZAIC and HIPPO ozone. Below 9 km, the global model generally overestimates the measured ozone, with a positive bias by 10.4 ppb averaged vertically and across all profiles. This overestimate is consistent with the positive bias, especially north of 30∘ N, reported from the ACCENT and ACCMIP model ensemble evaluation against ozonesonde data (Stevenson at al., 2006; Young et al., 2013). The coupled system produces lower ozone concentrations in the troposphere (0–9 km) than the global model alone. This translates to ozone bias reductions by 3–11 ppb at most MOZAIC sites (in the polluted areas) and by 5.3 ppb for HIPPO profiles (in the remote areas), averaged over 0–9 km. These improvements are a result of interactions between improved ozone simulations over pollution source regions and improved simulations of background ozone, as initially driven by a higher resolution over the source regions.

Figure 7 shows that for the MOZAIC sites, the observed ozone variability at a particular height of the profile is much larger than the modeled variability. This is because the observation is sampled at every 0.15 km vertically, at a much finer resolution than the vertical resolution of the model. When the observations are mapped to the vertical resolution of the model, the observed variability is greatly reduced to a level comparable to the modeled variability (not shown).

Figure 8 further shows the ozone profiles in individual seasons of 2009 at Frankfurt. With several hundred profiles in each season, this site allows for a detailed seasonal analysis. Again, although both the two-way coupled system and the global model alone capture the general vertical distribution of ozone in any given season, the coupled system leads to much lower biases below 9 km.

Figure 9 presents the horizontal distributions of TCO in individual seasons from OMI/MLS, OMI/LIU, their average OMI_MEAN, the two-way coupled system, and the global model alone. OMI/MLS and OMI/LIU produce similar seasonal and spatial distributions of TCO, with lower values in the tropics but higher values in the northern mid-latitudes (especially in the Northern Hemisphere (NH) summer and fall) and near 30∘ S. In general, OMI/LIU produces higher TCO values than OMI/MLS by 0.8 DU (2.8 %), 1.6 DU (5.3 %), 3.8 DU (11.9 %) and 2.8 DU (9.0 %) in NH spring, summer, fall and winter, respectively. These differences are broadly consistent with the uncertainties in OMI/MLS and OMI/LIU discussed in Sect. 3.3. We thus use their average, OMI_MEAN, for model evaluation.

Figure 9 shows that both the global model alone and the coupled system reproduce the general seasonal and spatial structures of OMI_MEAN TCO. The global model tends to overestimate the seasonal TCO in OMI_MEAN, with a global mean bias of 4.4 DU (15.2 %), 3.4 DU (10.9 %), 2.2 DU (6.5 %), and 1.6 DU (4.9 %) in NH spring, summer, fall, and winter, respectively. The positive bias is more significant in the NH (annual mean bias = 3.6 DU) than in the Southern Hemisphere (SH, bias = 2.2 DU). The large NH overestimate was found also for the ACCMIP model ensemble (Young et al., 2013). Compared to the global model alone, the coupled system reduces the annual average TCO by 3.0 DU (9.5 %) globally, by 3.8 DU in the NH, and by 2.1 DU in the SH. The coupled system also leads to TCO values closer to OMI_MEAN, with a global mean bias of 1.2 DU (4.1 %) in NH spring, 0.1 DU (0.3 %) in summer, -0.7 DU (-2.1 %) in fall and -0.7 DU (-2.2 %) in winter. The model improvements are more significant in the NH.