We obtained the in situ measurements of surface SO2, NO2, and O3 from the China National Environmental Monitoring Center for model evaluation. SO2 and NO2 are measured by various commercial instruments using the ultraviolet fluorescence method and the chemiluminescence method, respectively (Zhang and Cao, 2015). In the chemiluminescence method NO2 observations are obtained by measuring NO from decomposed NO2. This can result in a positive bias because NOz (all compounds that are products of the atmospheric oxidation of NOx) will also be reduced to NO. Steinbacher et al. (2007) showed that the ratio of NO2 to NOz (rNO2) depends on the distance that NO2 plumes transport from the source. In other words, the longer the distance, the more the potential for oxidation of NO2 and hence the smaller rNO2; only 43 %–76 % and 70 %–83 % of real NO2 contribute to the measured value (NO2)m for rural and urban sites, respectively (Steinbacher et al., 2007). For this study, as observational sites are in cities, a maximum value of 0.83 is used to convert (NO2)m measurements to the NO2 concentrations, which are subsequently used for evaluating the model results. Additionally, we also test values for rNO2 in the range of 0.7 to 1.0.

The VIIRS on board the Suomi National Polar-orbiting Partnership (Suomi NPP) satellite was launched on 28 October 2011, and its day/night band (DNB) provides observations of nighttime lights with a spatial resolution of 750 m (Miller et al., 2013). Here, we use the VIIRS nighttime lights product that has excluded background noise as well as solar and lunar contamination and has screened out the data degraded by cloud cover and features unrelated to electric lighting (Elvidge et al., 2017a). This product is regridded to 0.05∘×0.05∘ for October 2013 and to 0.05∘×0.05∘ and 0.25∘×0.25∘ for April 2018.

The TROPOspheric Monitoring Instrument (TROPOMI) on board Sentinel-5 Precursor was launched on 13 October 2017, with a nadir footprint of 7km×3.5km, which is finer than that of all its predecessors. The TROPOMI NO2 tropospheric VCDs from the Royal Netherlands Meteorological Institute (KNMI) were retrieved using a differential optical absorption spectroscopy (DOAS) algorithm and validated with Pandora NO2 retrievals (Griffin et al., 2019). We grid the product to the 0.05∘×0.05∘ resolution for April 2018 to investigate the relationship between VIIRS nighttime lights and NO2 tropospheric VCDs.

MIX (Li et al., 2017a) is a mosaic of Asian anthropogenic monthly emissions developed for the years 2008 and 2010 to support the Model Inter-Comparison Study for Asia and the Task Force on Hemispheric Transport of Air Pollution. SO2, NOx, and NH3 emissions in MIX come from the Regional Emission inventory in ASia version 2.1 (REAS2.1) (Kurokawa et al., 2013), replaced by the Multi-resolution Emission Inventory for China (MEIC) SO2 and NOx and the PKU NH3 (Huang et al., 2012) for mainland China, the ANL (Lu et al., 2011; Lu and Streets, 2012) SO2 and NOx of some source sectors for India, and the CAPSS SO2 and NOx for the Republic of Korea (Li et al., 2017a). In spite of variations among spatial resolutions of these emission inventories, they are regridded to 0.25∘×0.25∘ to form the MIX emissions inventory (Li et al., 2017a). In our study, not only is MIX used in the posterior simulations and forecasts, but it also provides information for downscaling the posterior emission inventories from Part 1 (as in Wang et al., 2020).

The CTM used for the simulations and forecasts of surface SO2 and NO2 concentrations is GEOS-Chem version 12.0.0 (GCv12.0.0), which is driven by GEOS-FP meteorological fields from the Global Modeling and Assimilation Office (GMAO). Horizontal resolutions are set as 2∘×2.5∘, the same as that of the posterior emissions from Part 1 (as in Wang et al., 2020), and 0.25∘×0.3125∘, which is the finest resolution available for this version of GEOS-Chem, to investigate the impacts of downscaling on simulations and forecasts. There are 47 vertical layers, the lowest (box height is in the range of 115 to 135 m, as shown in Fig. S1) of which represents surface concentrations evaluated with in situ observations. We use the MIX 2010 emissions for October 2013 prior simulations as well as November 2013 prior forecasts. Posterior SO2 and NOx emissions for October 2013 from Part 1, i.e., Wang et al. (2020), are used for October 2013 simulations and November 2013 forecasts at 2∘×2.5∘ resolution but need be downscaled for 0.25∘×0.3125∘ simulations, as described in Sect. 3.3.

It is worth noting that the GEOS-Chem adjoint model (Henze et al., 2007) used in Part 1 of this study by Wang et al. (2020) is v35m, which is developed based on GEOS-Chem version 8.2.1, updated through version 9. Here we use GCv12.0.0 rather than GC adjoint v35m to investigate if the model-dependent posterior emission inventory can be applied to other models to improve simulations and forecasts. With the same MIX emissions used, GCv12.0.0 surface SO2 and NO2 concentrations are in general larger than those from v35m, with differences of up to 15 µg m-3 for SO2 and 10 µg m-3 for NO2 (Fig. 1). The difference is due to differences in chemical mechanism and boundary layer parameterization schemes between the two models. Therefore, by using two different versions of GC, we can study the degree to which the posterior emissions derived from one model (in this case a global model with coarser resolution) can be applied for another (here a regional model with finer resolution).

GEOS-Chem surface SO2 and NO2 concentrations at a resolution of 2∘×2.5∘ are not expected to be able to capture hot spots due to the CGS effect, and thus we aim to downscale them to finer resolutions. The prior emissions are MIX for October and November 2010. The posterior emissions are from separate inverse emission estimates in Part 1 (e.g., E-SO2 and E-NO2 experiments as described in Wang et al., 2020), unless it is specifically stated. The downscaling methods here should be distinguished from interpolation approaches to simply increasing spatial resolutions.

To simulate or forecast surface SO2 and NO2 concentrations at a resolution of 0.25∘×0.3125∘ through the GEOS-Chem model, the posterior emissions at a resolution of 2∘×2.5∘ should be downscaled to fit the model resolution. The prior MIX 2010 emission inventory has a spatial resolution of 0.25∘×0.25∘, which is slightly finer than 0.25∘×0.3125∘, and it can be easily processed to fit 0.25∘×0.3125∘ simulations through HEMCO – the GEOS-Chem emission processing package (Keller et al., 2014). Thus, all the posterior emissions at a resolution of 2∘×2.5∘ are downscaled to 0.25∘×0.25∘, which are further regridded to 0.25∘×0.3125∘ with HEMCO. We introduce two emission downscaling approaches with prior MIX 0.25∘×0.25∘ emissions and 0.05∘×0.05∘ VIIRS nighttime lights used as spatial proxies. The two methods are referred to as downscaling emissions with MIX (MIX-DE) and downscaling emissions with nighttime light (NL-DE).

We use the linear correlation coefficient (R), mean bias (MB), normalized mean bias (NMB), normalized centered root mean square error (NCRMSE) (Wang et al., 2020), and normalized mean squared error (NMSE) as measures to evaluate GEOS-Chem SO2 and NO2 surface concentrations with in situ observations. NCRMSE measures the spatial distribution difference between forecasts, and in situ observations are similar to the root mean squared error, but the impact of bias is removed. NMSE is defined as Eq. (6), where Mi and Oi are the ith GEOS-Chem simulation and in situ observation, respectively, O‾ is mean of the observations, and N is number of the observations. 6NMSE=1N∑i=1N(Mi-Oi)21N∑i=1N(Oi-O‾)2

The CGS effect of surface SO2 concentrations in the coarse-resolution (2∘×2.5∘) simulations (C-PRI and C-POS experiments) is shown in Fig. 4a–d. The GEOS-Chem 2∘×2.5∘ simulation of every grid cell is the average of surface SO2 at ∼5×104 km2 area, while in situ SO2 observations can only represent average concentrations of a much smaller area. Considering that all sites are in cities, where emission sources are located, GEOS-Chem 2∘×2.5∘ simulations are expected to be smaller than in situ observations. In this study, the NMB is -26.7 % (Fig. 4c) in the C-PRI simulation, while the C-POS simulation shows an even stronger negative NMB of bias of -65.3 % (Fig. 4d), as the posterior SO2 emission is 35.8 % smaller than prior MIX 2010.

To decrease the impact of CGS on surface SO2 simulations, both the prior and posterior GEOS-Chem surface SO2 simulations at 2∘×2.5∘ resolution are downscaled to 0.25∘×0.3125∘ through the MIX-DDC approach (MIX-DDC-PRI and MIX-DDC-POS experiments). Zheng et al. (2017) showed that surface SO2 concentration simulations from WRF-CMAQ, when evaluating with in situ observations, have a NMB of -23 %, 7 %, and 41 % at the resolutions of 36 km (∼ 0.36∘), 12 km (∼ 0.12∘), and 4 km (∼ 0.04∘), respectively, which suggests that (1) the CGS effect and other nonlinear resolution-dependent processes can affect SO2 simulation results and (2) these problems are alleviated at the resolution of 0.25∘×0.3125∘ but are not completely avoided. The downscaled prior (MIX-DDC-PRI) and posterior (MIX-DDC-POS) GEOS-Chem surface SO2 concentrations at 0.25∘×0.3125∘ are shown in Fig. 4e–h. MIX-DDC-PRI and MIX-DDC-POS SO2 simulations show hot spots of up to 270 µg m-3 (Fig. 4e) and 80 µg m-3 (Fig. 4f), respectively, compared with the largest value of less than 60 µg m-3 (Fig. 4a) and 35 µg m-3 (Fig. 4b) in the C-PRI and C-POS simulations, respectively.

MIX-DDC-POS SO2 simulations are in better agreement with in situ observations than MIX-DDC-PRI. The NMSE decreases from 4.63 in MIX-DDC-PRI to 1.50 in MIX-DDC-POS, and the linear correlation coefficient (R) increases from 0.32 to 0.36 (Fig. 4g–h). The NMB changes, however, from 43.4 % to -35.3 % (Fig. 3g–h), which implies the CGS effect may not be completely avoided at a resolution of 0.25∘×0.3125∘, and other factors that may affect the result should be investigated in the future. We also separately compare MIX-DDC-PRI and MIX-DDC-POS simulations with in situ observations over provincial capital cities, as SO2 hot spots in smaller cities may still be difficult to capture using the 0.25∘×0.3125∘ MIX-DDC-PRI and MIX-DDC-POS simulations. The NMB is 115.0 % in the MIX-DDC-PRI simulation and reduces to -5 % in the MIX-DDC-POS simulation. Additionally, the MIX-DDC-POS simulation shows better spatial pattern than the MIX-DDC-PRI simulation in terms of NCRMSE, although linear correlation decreases slightly. In spite of the improvement of capturing hot spots in term of NMB using the MIX-DDC approach, we should also note that the coarse-resolution simulations (Fig. 4c, d) have larger linear correlation coefficients and smaller NCRMSEs than the MIX-DDC simulations. Thus, for SO2 simulations, MIX-DDC helps to capture hot spots but can make the spatial distribution worse than the original coarse-resolution simulations in terms of NCRMSE. The spatial pattern degradation implies that current chemistry transport simulations of surface SO2 concentrations can capture the regional spatial pattern (coarse resolution) well, but it is difficult to simulate the local spatial pattern (fine resolution); the weakness for describing the local spatial pattern simulation suggests the uncertainties of either bottom-up SO2 emission estimates at fine resolution or locally resolved meteorological fields (Ge et al., 2017), or both. This uncertainty in bottom-up emission inventories can further stem from distributing SO2 emissions at the provincial level to a fine-resolution grid. MIX-DDC also performs well when GC adjoint v35m is used for coarse resolution, and results show better spatial pattern (smaller NCRMSE) than using GC v12.0.0 (Fig. S2).

NO2 has an even shorter lifetime than SO2; thus the CGS effect also exists in the C-PRI and C-POS simulations. Figure 5a–d show that almost all in situ NO2 observations are larger than the GEOS-Chem simulations, regardless of whether the prior MIX 2010 (C-PRI) or the posterior (C-POS) NOx emissions are used. GEOS-Chem surface NO2 averages from the C-PRI and C-POS simulations, sampled according to in situ observational sites, are 49.2 % and 54.5 % smaller than the average of in situ observations, respectively.

The MIX-DDC approach is also applied to downscale NO2 surface simulations (MIX-DDC-PRI and MIX-DDC-POS experiments), and the results are evaluated with in situ observations. As discussed, according to Zheng et al. (2017), the CGS effect and other nonlinear resolution-dependent problems can be alleviated at the resolution of 0.25∘×0.3125∘ (such as in MIX-DDC-PRI and MIX-DDC-POS experiments) but cannot be completely avoided. The MIX-DDC-POS simulation is better than the MIX-DDC-PRI simulation in terms of spatial pattern (NCRMSE), although the CGS effect may still exists and other nonlinear resolution-dependent problem should be investigated in the future. The NMB is -19.3 % and -31.8 % for the MIX-DDC-PRI and MIX-DDC-POS simulations, respectively (Fig. 5e–h), which implies that 0.25∘×0.3125∘ may be still too coarse to capture hot spots due to the short lifetime of NO2. The larger negative bias in the MIX-DDC-POS simulation than in the MIX-DDC-PRI also leads to the former showing large NMSE. Despite the negative bias, R between observations and the MIX-DDC simulations increases from 0.53 in MIX-DDC-PRI to 0.75 in MIX-DDC-POS, and the NCRMSE reduces from 0.96 to 0.67, which is only slightly larger than 0.64 in the C-POS simulation. Thus MIX-DDC-POS can better capture NO2 hot spots and shows spatial pattern as good as C-POS. MIX-DDC also performs well when GC adjoint v35m is used for coarse-resolution simulations, and results show better spatial pattern (smaller NCRMSE) than using GC v12.0.0 (Fig. S3).

To further alleviate the CGS effect, we downscale GEOS-Chem surface NO2 simulations at 2∘×2.5∘ to 0.05∘×0.05∘ according to VIIRS nighttime light distributions through the NL-DC approach (NL-DC-PRI and NL-DC-POS experiments), and the results are evaluated with in situ surface NO2 observations (Fig. 5i–l). The largest GEOS-Chem surface NO2 values are less than 35 µg m-3 in both the coarse C-PRI and C-POS simulations (Fig. 5a, b), while they are larger than 100 µg m-3 at 0.05∘×0.05∘ in the NL-DC-PRI and NL-DC-POS simulations (Fig. 5i, j). The scatter plots of the NL-DC-PRI (Fig. 5k) and NL-DC-POS (Fig. 5l) simulations versus in situ surface NO2 observations show that R increases from 0.61 in the NL-DC-PRI simulation to 0.75 in the NL-DC-POS simulation, and NMSE decreases from 3.69 to 1.80, which is smaller than that in the coarse-resolution simulations and the MIX-DDC downscaled simulation. It suggests that NL-DC has the advantage of downscaling surface concentrations (without evoking any CTM simulation and its associated needs of computational resources). The biases of surface NO2 concentrations are 45.3 % and 25.5 % for NL-DC-PRI and NL-DC-POS, respectively, which could come from total emission bias as well as the downscaling process through the NL-DC approach. The sites used for validation are mainly over urban regions, and we lack sites that are located over rural regions to evaluate if positive or negative bias persists over rural regions. Thus, we are not able to determine how much of the positive bias in NL-DC-PRI and NL-DC-POS is caused by the NL-DC approach. When NL-DC is applied to downscale GC adjoint v35m coarse-resolution simulations, results show better spatial pattern (smaller NCRMSE) than that of GC v12.0.0 (Fig. S3).

The surface NO2 concentrations used for evaluation are derived from measurements of (NO2)m assuming rNO2 of 0.83 as stated in Sect. 2.1. Due to the lack of information on rNO2, we also test the values in the range of 0.7 to 1.0, and the derived NO2 concentrations are used to validate the NL-DC-PRI and NL-DC-POS simulations at the 0.05∘×0.05∘ resolution. Figure 6 shows that the NL-DC-POS simulation has NMSE in the range of 1 to 4, which is always better than the NL-DC-PRI simulation with NMSE in the range of 2 to 8.

Instead of downscaling simulation results as shown in Sect. 4.1, we directly simulate surface SO2 concentrations at 0.25∘×0.3125∘ resolution through GEOS-Chem over China in October 2013 using the prior MIX 2010 emissions and the posterior emissions. The posterior SO2 emissions at 2∘×2.5∘ resolution are downscaled to 0.25∘×0.3125∘ through the MIX-DE approach. The posterior MIX-DE SO2 emissions are smaller than the prior MIX 2010 SO2 emissions over northern China and southwestern China (Fig. 7).

The 0.25∘×0.3125∘ GEOS-Chem simulations of surface SO2 for October 2013 when using the prior MIX (F-PRI experiment) and the posterior MIX-DE (MIX-DE-POS experiment) emission inventories are shown in Fig. 8. When evaluating with all in situ SO2 observations, NMSE decreases from 3.73 in F-PRI to 1.55 in MIX-DE-POS, but bias changes from 15.76 to -14.98 µg m-3. For the same reason, in the MIX-DDC-PRI and MIX-DDC-POS assessment in Sect. 4.1, we also focus on provincial capital cities. In this scene, NMSE of the MIX-DE-POS simulation is 1.85 (Fig. 8d), which is much smaller than 15.07 in the F-PRI simulation (Fig. 8c), but it is slightly larger than 1.76 in the MIX-DDC-POS simulation (Fig. 4h). Moreover, NMB decreases from 101.2 % in the F-PRI simulation to -8.4 % in the MIX-DE-POS simulation (Fig. 8). R is 0.23 and 0.14 in F-PRI and MIX-DE-POS, respectively, neither of which is significant at the 95 % confidence level. In Sect. 4.2, we have shown that good spatial distribution is captured in coarse-resolution rather than fine-resolution simulations, which implies that large uncertainty of bottom-up SO2 emission estimates at fine resolution may be introduced when distributing SO2 emissions at the provincial level to a fine-resolution grid. It is not surprising that the correlation coefficients are small for F-PRI and MIX-DE-POS, as both are in fine resolution.

Posterior NOx emissions at 2∘×2.5∘ resolution are downscaled through MIX-DE and NL-DE approaches. Figure 9 shows the prior MIX, posterior MIX-DE, and posterior NL-DE NOx emissions at 0.25∘×0.3125∘ resolutions. All three emission inventories show NOx emission hot spots over metropolises (Fig. 9a–c). Compared with prior MIX, posterior MIX-DE is larger over Hebei Province but smaller over most other areas of the North China Plain and eastern China (Fig. 9d). As the posterior NL-DE emission inventory is downscaled according to the VIIRS nighttime light distribution, the difference (Fig. 9e) between posterior NL-DE and prior MIX and the difference (Fig. 9f) between posterior NL-DE and posterior MIX-DE show scattered positive and negative values.

The three emission inventories are used to simulate surface NO2 concentrations at the 0.25∘×0.3125∘ resolution over China in October 2013, that is F-PRI, MIX-DE-POS, and NL-DE-POS experiments in Table 1. All these simulations are evaluated with in situ NO2 observations (Fig. 10). R increases from 0.46 in F-PRI to 0.61 in MIX-DE-POS and 0.58 in NL-DE-POS, and NCRMSE decreases from 0.95 in F-PRI to 0.81 in MIX-DE-POS and 0.85 in NL-DE-POS (Fig. 10d–f). Both MIX-DE-POS and NL-DE-POS show stronger negative NMB and larger NMSE than F-PRI, which should be partly caused by the CGS effect, although emission bias and other nonlinear resolution-dependent processes could play a role.

MIX-DE-POS has improved values of R and NCRMSE compared to NL-DE-POS; here we discuss the possible reasons and propose future works to improve NL-DE. MIX is a mosaic bottom-up emission inventory, and it is actually the MEIC emission inventory for NOx emissions over China (Li et al., 2017a). The MIX (or MEIC) NOx emission inventory over China consists of emissions from four sectors including coal-fired power plant, industrial, transport, and residential sectors. Coal-fired power plant emissions in MEIC are derived through extensively using detailed information (including locations of individual units) of 7657 generation units in China (Liu et al., 2015); coal-fired power plant emissions can be accurately placed in grids according to source location information (Li et al., 2017a). Thus, if we can allocate posterior total anthropogenic NOx emissions into the four sectors, we expect that it is better to use the MIX coal-fired power plant NOx emission inventory rather than nighttime lights to downscale the posterior coal-fired power plant NOx emissions. For the other sectors in MIX (or MEIC) over China, population density is used to allocate industrial and residential emissions to grids (Li et al., 2017a), and transport emissions are distributed according to road networks (Li et al., 2017a). Using population density to downscale industrial and residential NOx could underestimate emissions over urban regions, compared with the approach using nighttime light which could better represent economic development levels (Geng et al., 2017). Whether it is better to use road networks or nighttime lights to downscale NOx emissions from the transport sector requires future investigations. In this study, the posterior NOx emission inventory to be downscaled is total anthropogenic NOx emissions, which is not allocated into different source sectors. Thus, if we assume that the ratios of every sectoral emissions to total anthropogenic emissions do not change between prior and posterior emission inventories, MIX-DE has an advantage for the coal-fired power sector, while NL-DE could benefit the downscaling for the industrial and residential sectors. In future work, we could optimize sectoral emissions rather than total anthropogenic emissions, subsequently downscale posterior coal-fired power emissions through prior MIX coal-fired power emissions, and ultimately use VIIRS nighttime light data to downscale posterior industrial and residential emissions.

To evaluate the posterior SO2 and NOx emissions of joint assimilations with various observation balance parameter (γ) values from Part 1, i.e., Wang et al. (2020), we focus on the sum of NMSE of surface SO2 and NO2 as shown in Fig. 11. The experiment using the prior MIX SO2 and NOx emissions has the largest sum of NMSE, which is followed by the simulation using the prior MIX SO2 emissions and the posterior NOx emissions from separate assimilation. The sum of NMSE using the posterior SO2 and NOx emissions of joint assimilations (JOINT-F-POS) with various observation balance parameter (γ) values (as γ increases, the NO2 species is more emphasized in the cost function) is always smaller than that of the experiment using the prior MIX SO2 emissions and the posterior NOx emission from separate assimilation and decreases as γ increases. When γ is 1500 or 2000, the sum of NMSE using the posterior SO2 and NOx emissions of joint assimilations is smaller than that of the experiment using the prior MIX NOx emission and the posterior SO2 emission from separate assimilation, but it equals that of the experiment using the posterior SO2 and NOx emissions from separate assimilations. The value of γ mainly affects SO2 NMSE (Fig. S4a) rather than NO2 NMSE (Fig. S4b).

Figure 12 shows evaluations of surface SO2 and NO2 forecasts with in situ observations. AQF-PRI SO2 concentrations are generally larger than in situ observations with MB of 45.07 µg m-3 and NMSE of 7.97 (Fig. 12a). The MB and NMSE reduces to -7.12 µg m-3 and 1.38 (Fig. 12b), respectively, in AQF-MIX-DE-POS. For surface NO2, NCRMSE and R are 0.76 and 0.65 (Fig. 12c), respectively, in AQF-PRI and change to 0.75 and 0.66 (Fig. 12d), respectively, in AQF-MIX-DE-POS. The stronger negative NMB and larger NMSE for NO2 in AQF-MIX-DE-POS than that in AQF-PRI are likely attributable to the CGS effect. The CGS effect is eliminated in both the AQF-NL-DC-PRI and AQF-NL-DC-POS, which show positive bias (Fig. 12e, f). In the 0.05∘×0.05∘ forecasts, NMSE decreases from 4.61 in AQF-NL-DC-PRI to 3.43 in AQF-NL-DC-POS, and R increases from 0.38 to 0.42.

In this study, we show the improvements of GEOS-Chem simulations or forecasts of surface SO2 and NO2 concentrations through posterior emissions constrained by integration of GEOS-Chem adjoint and OMPS observations. All the improvements of SO2 and NO2 forecasts are summarized in the Taylor diagrams (Fig. 13), which includes R, normalized standard deviation (the ratio of forecast standard deviation to in situ observations), NMB, and normalized centered root mean square error (NCRMSE). NCRMSE is shown as the distance between the forecast point and the expected (in situ observation) point. The improvements of forecasting surface SO2 concentrations through MIX-DE and MIX-DDC are comparable (Fig. 13a). For NO2 forecast, MIX-DE shows larger improvement than NL-DE and MIX-DDC (Fig. 13b). NL-DC is able to better decrease the coarse-grid smearing effect than MIX-DE but shows a larger normalized centered root mean square error. In the future, we plan to investigate if the posterior emissions can be applied to other models such as WRF-Chem and WRF-GC at a spatial resolution finer than 0.25∘×0.3125∘. In the case of the global model of chemistry, it is promising to use nighttime light to downscale NO2 simulations so as to obtain a quick look of NO2 air quality at very fine resolution.

In addition to the improvement of SO2 and NO2, AQF-MIX-DE-POS enhances AQF-PRI in the forecast of surface O3 concentrations (Fig. 14). If all O3 in situ observations in the research domain are used for evaluation, a spatial distribution improvement is shown with NCRMSE decreasing from 1.08 for AQF-PRI to 1.05 for AQF-MIX-DE-POS, but NMB changes from -3.1 % to 5.0 % (Fig. 14c). Indeed, whether bias becomes smaller or larger depends on the region. In the North China Plain and eastern China where NOx emissions (or NO2 surface concentrations) are large (the black box in Fig. 14a), forecasts of surface O3 concentration are much lower than in other regions; and the NMB is -16.7 % for AQF-PRI and -6.3 % for AQF-MIX-DE-POS, with NCRMSE decreasing from 1.20 to 1.16 (Fig. 14c). In this relatively NOx-rich region, the increase in O3 concentration in AQF-MIX-DE-POS is caused by the decrease in NO2 concentrations; the change in SO2 concentrations has negligible impacts on O3 concentrations (Fig. S5). This implies that if volatile organic compound (VOC) concentrations remain constant, emission control of NOx emissions will exacerbate O3 pollution. For the region that is outside of the black box, although NCRMSE decreases from 0.82 for AQF-PRI to 0.80 for AQF-MIX-DE-POS, NMB increases from 19.0 % to 23.3 % (Fig. 14c).