Global air quality is simulated from 2010 to 2017 using the Community Atmosphere Model with Chemistry (CAM-chem) model (version 4; Lamarque et al., 2012) at a horizontal resolution of 1.9∘ (latitude) × 2.5∘ (longitude), using 56 vertical levels that span the surface to 4 hPa (≈ 40 km), and driven by NASA GEOS5 Global Atmosphere Forcing Data (Tilmes 2016, http://rda.ucar.edu/datasets/ds313.0/, last Accessed 20 April 2020). Additional modeling configurations, including the lower boundary conditions for long-lived species such as CO2 and CH4, the online biogenic emission inventory, and other natural emissions, are described in Zhang et al. (2016) and Tilmes et al. (2016). In this version of CAM-chem, the bulk aerosol model was applied based on the work of Tie et al. (2001, 2005), in which the sulfate aerosol is formed by the oxidation of SO2 in both the gas and aqueous phases (Lamarque et al., 2012; Tilmes et al., 2016). Ammonium nitrate formation is dependent on the amount of sulfate present in the air mass, following the parameterization of gas and aerosol partitioning by Metzger et al. (2002). Secondary organic aerosols (SOAs) are derived using the two-product model approach, with laboratory-derived yields for monoterpenes, isoprene, and aromatic photooxidation (Heald et al., 2008; Times et al., 2016). Recent research has suggested that anthropogenic SOA may be a dominant contributor of health impacts globally (Nault et al., 2021). As our simulations lack representation of important anthropogenic SOA precursors, such as intermediate-volatility organic compounds (IVOCs; Zhao et al., 2014; Q. Lu et al., 2020; Pennington et al., 2021), our simulated PM2.5 concentrations may be low biased. PM2.5 is calculated as the sum of SO4 + NO3 + NH4 + OC + BC + SOA + 0.2 ⋅ Dust + Sea salt (West et al., 2013; Silva et al., 2016). For dust and sea salt, only the size fractions relevant for PM2.5 (size bins 1–3) are used. Dust in desert regions was found to be too high in the model, so global dust concentrations were multiplied by 0.2 to achieve rough consistency with the PM2.5 concentrations estimated with Brauer et al. (2012). A comprehensive evaluation of model performance in simulating temporal and spatial distribution of global ozone and aerosols was carried out in previous studies (Tilmes et al., 2015, 2016; Zhang et al., 2016). Here, we evaluate the model using surface ozone and PM2.5 concentrations in China from 2013 to 2017. The lowest modeled grid cell (∼ 58 m above sea level) is taken to indicate ground-level concentrations.

Our base case simulation (CEDS_MEIC in Table 1) spans 2010 to 2017, with a 1-year spin-up in 2009, and used a time-varying global anthropogenic emission inventory. The base inventory utilized the Community Emissions Data System (CEDS, v2017-05-18; Hoesly et al., 2018), with Chinese emissions supplemented by the Multi-resolution Emission Inventory (MEIC; http://www.meicmodel.org/, last access: 20 April 2020). Since the version of the CEDS inventory used here stops after 2014, global anthropogenic emissions for all other regions outside China repeat 2014 for 2015–2017. McDuffie et al. (2020) updated the global CEDS anthropogenic emissions through 2017 with continued updates into 2019 (https://github.com/JGCRI/CEDS, last access: 6 May 2021). However, this would have a negligible effect on our conclusions since our focus is the emission changes from China, which do feature year-specific emissions for 2015–2017, and on the influence of domestic and international air pollution. We then performed a sensitivity simulation from 2010 to 2017 (CEDS_MEIC_ChinaFix) that kept Chinese emissions constant at 2010 levels. The differences between the base case simulation and this sensitivity simulation represent the influence of anthropogenic emission changes in China on global air quality and human health (Table 1). Finally, we performed a sensitivity simulation with global anthropogenic emissions frozen at 2010 levels (CEDS_MEIC_GlobalFix, Table 1). This simulation isolates the influence of meteorological changes on PM2.5 and O3 changes.

It has been reported that the CEDS emissions inventory used here is high biased for the magnitude of the Chinese emissions and underestimates the decreasing emissions trend in China since 2013 (Zheng et al., 2018b; Paulot et al., 2018). Indeed, we see that in 2014 (Fig. S1 in the Supplement) the emissions in the CEDS inventory are at least 20 % higher than the emissions from the MEIC inventory for most air pollutants. More specifically, the SO2, OC, and BC emissions in CEDS are 84 %, 81 %, and 58 % higher, respectively, in 2014 than those reported in the MEIC inventory. In addition, the CEDS inventory estimated a continued increasing trend for several pollutants, whereas the MEIC inventory often peaked prior to 2012 for pollutants (Liu et al., 2016; Zheng et al., 2018b). We also find emissions in the CEDS inventory to be higher in the western and southern parts of China when compared to the MEIC inventory and lower in eastern China (Fig. S2). To test the influence of these differences in Chinese emissions, we performed another sensitivity simulation from 2010 to 2014 that applied the CEDS inventory globally (CEDS_Global). This enables an evaluation into the model's performance when using a variable inventory and allows for a discussion on the relative air quality changes when applying a different emission inventory in China.

Hourly surface observations for PM2.5 and O3 were retrieved from the China National Environmental Monitoring Center (CNEMC) network (http://106.37.208.233:20035/) from 2013 to 2017. Data prior to 2013 are not available. We then evaluated the model's performance in simulating annual PM2.5 and the annual average maximum daily 8 h average (MDA8) O3 as these two metrics have been reported to be associated with the health impacts in epidemiological analyses.

To calculate health impacts, we applied the relative risk associated with long-term air pollution exposure from various epidemiological studies, baseline mortality rates, population, and modeled exposure concentrations as follows: 1ΔMort=Y0⋅AF⋅Pop, where ΔMort is mortality burden attributed to long-term PM2.5 and O3 exposure, Y0 is the baseline mortality rates for cause of specific disease, AF is the attribution fraction of mortality associated with air pollution exposure, which is calculated as 1-1RR (RR signifies relative risk), and Pop is the exposed population with ages greater than 25 years old. The RR for long-term PM2.5 exposure is calculated using the integrated exposure response model (IER; Burnett et al., 2014) from the Global Burden of Disease 2017 (GBD2017) study (Stanaway et al., 2018). The RR for long-term O3 exposure is retrieved from Turner et al., (2016), which reports a RR of 1.12 (95 % confidence interval (CI): 1.08, 1.16) for respiratory disease. Country age-specific baseline mortality rates (Y0) in 2010 were retrieved from the GBD2017 project (Stanaway et al., 2018) and remapped to match the 10th International Statistical Classification of Diseases and Related Health Problems codes used in the cohort study (Turner et al., 2016; Seltzer et al., 2020). The theoretical minimum risk exposure level for PM2.5 exposure assessment is drawn from a uniform distribution with a lower bound of 5.8 µgm-3 and an upper bound of 8.8 µgm-3, and for O3 exposure it is 26.7 ppbv. Previous studies have shown that coarse-resolution global CTMs (chemical transport models), e.g., 1.9∘ × 2.5∘ , likely generate low biases in estimating health effects, especially in urban areas (Li et al., 2016; Punger and West, 2013; Silva et al., 2013, 2016). However, less is known of how these underestimates would affect the relative contributions of downwind transportation (Liang et al., 2018). Jin et al. (2019) concluded that the uncertainties in estimating the ambient PM2.5-related mortality burden is dominated by the uncertainties in the underlying exposure response function and less influenced by the uncertainties associated with the PM2.5 concentration estimates.

In the base CAM-chem simulation (CEDS_MEIC scenario), predicted annual PM2.5 concentrations in China are overestimated, with a mean bias (MB) of 19.3 µgm-3 and normalized mean bias (NMB) of 37.2 %. The MB fluctuates around 20 µgm-3 from 2014 to 2017 and is lowest in 2013 (MB of 7.6 µgm-3). The lower MB and NMB in 2013 could be due to the fact that there are much less data available in 2013. The consistent positive NMBs indicate that the overestimates are systematic. This should not affect our main conclusions since we focus on the changes among years. The high modeling bias for CAM-chem-predicted surface PM2.5 has also been reported in other studies (e.g., He and Zhang, 2014). These studies attribute CAM-chem modeling bias in surface PM2.5 to predictions of SO42-, NH4+, and organic aerosols, as well as missing major inorganic aerosol species such as nitrate and chloride (He and Zhang, 2014; Tilmes et al., 2016). By including advanced inorganic aerosol treatments, such as condensation of volatile species, explicit inorganic aerosol thermodynamics for sulfate, ammonium, nitrate, sodium, and chloride (He and Zhang, 2014) and more comprehensively representing secondary organic aerosol formation (e.g., using the volatility basis set scheme; Times et al., 2019; Liu et al., 2020), modeling performance could significantly improve. We also find that the bias metrics exhibit small interannual variability, with NMB generally around 40 % and NME around 50 %. Exceptions include the years 2013 and 2014, which have smaller NMB values due to the limited number of the observations (Table 2). CAM-chem can generally reproduce the spatial patterns of the annual PM2.5 distributions, with a correlation coefficient (R) greater than 0.7. Notably, simulations using only CEDS emissions (i.e., the CEDS_Global scenario) generate better performance in both 2013 and 2014. Part of the reason is that we have less available data in these 2 years. However, we suspect the main reason for this variable bias is due to spatial changes in emissions between the inventories. While total emissions in China are higher in the CEDS inventory than in the MEIC inventory, CEDS tends to allocate more emissions in western and central China (Fig. S2). As such, predictions of annual PM2.5 concentrations using the CEDS inventory are lower in eastern China and higher in western and northwestern China (Fig. S3). CAM-chem NMB in simulating the annual MDA8 O3 is lower than 20 % for all evaluated years (Table 3). From Table 3, we can also see that CAM-chem overestimates the annual MDA8 O3 in China, which means our estimates for the O3-related mortality burden will likely be biased high, too. A high bias of about 10 ppb can be attributed to the coarse model resolution, which leads to an overestimate of ozone production due to diluted emissions of ozone precursors (Tilmes et al, 2015). In contrast to the PM2.5 performance, the CEDS simulation (CEDS_Global scenario) generates poorer model performance for MDA8 O3. For both PM2.5 and ozone, we also find that the NMBs are lower in eastern China compared with other inland regions (Figs. S5 and S6).

Annual, population-weighted (pop-weighted) PM2.5 concentrations decreased by 17.6 % between 2010 and 2017 (84.7 µgm-3 in 2010 to 69.8 µgm-3 in 2017; Fig. 1). The pop-weighted PM2.5 concentration was highest in 2011 (annual average of 94.1 µgm-3) and subsequently decreased through 2017 (69.8 µgm-3). Between 2013 and 2017, the national annual pop-weighted PM2.5 concentration decreased by 15.9 µgm-3, which is comparable to the reductions (19.8 µgm-3) reported by Q. Zhang et al. (2019), who used a high-resolution regional air quality model. At the national scale, the pop-weighted PM2.5 concentration features a similar trend as the area-weighted average trend but is notably higher (Fig. 1a), indicating that higher PM2.5 concentrations happen in regions with higher population density. The annual average of area-weighted PM2.5 concentration decreased by 7.6 µgm-3 between 2010 and 2017, which is consistent with the estimates reported by Ding et al. (2019a) at 9.0 µgm-3. The surface PM2.5 changes in China due to emission changes usually peak in the fall and winter (Fig. S7a). Spatially, we see that significant PM2.5 changes (increases before 2013 and decreases thereafter) occur in eastern China (Fig. 2), which was the focus region for China in the APPCAP (Ding et al., 2019a, b). We also find that the annual PM2.5 decreases in China are mainly dominated by the changes in emissions and not due to meteorology, which is consistent with prior studies (Dang and Liao, 2019; Ding et al., 2019a; Zhai et al., 2019; Q. Zhang et al., 2019). Relative to 2010, interannual meteorology led to annual PM2.5 decreases as high as 4.4 µgm-3 in 2011 and increases as high as 3.1 µgm-3 in 2015. Meanwhile, the emission changes led to annual PM2.5 decreases as high as 16.7 µgm-3 in 2017 and increases as high as 4.9 µgm-3 in 2015 (Fig. 4a). There were also isolated increases in PM2.5 in northwest China from 2010 to 2013, which were mainly caused by the dust storms (Meng et al., 2019; Luo et al., 2020; Zhao et al., 2020).

In contrast to the PM2.5 trends, the annual average pop-weighted MDA8 O3 has increased relative to 2010 (Fig. 1b), with a peak in 2014 (59.5 ppbv) and subsequent decreases to 2017 (57.1 ppbv). The area-weighted MDA8 O3 was comparable to or larger than the pop-weighted concentration due to the more uniform O3 distribution in China or even higher ozone events in western China from stratosphere–troposphere exchange (Wang et al., 2011; J. Li et al., 2019). For ozone, the emission changes from 2010 to 2013 exacerbate summer ozone pollution in China but alleviate ozone pollution in the other three regions (Fig. S8a). After 2013, the emission decreases in China exacerbate the ozone pollution for all the seasons, especially in winter. The spatial pattern of ozone trends mainly featured increases in the Beijing–Tianjin–Hebei and Yangtze River Delta regions and slightly decreases in the south (Figs. 3 and S5). The anthropogenic emission reductions in China that led to ozone increases (Fig. 4b) could partially be explained by the aerosol sink of hydroperoxyl radicals, which slowed in recent years due to PM2.5 decreases (K. Li et al., 2018). In addition, the effect of interannual meteorological conditions had a profound influence on the annual average MDA8 O3 concentrations. Ozone increased as much as 8.1 ppbv due to meteorology. The meteorology-induced ozone increases can be attributed to increasing temperature, which enhances the ozone production and biogenic NMVOCs (non-methane volatile organic compounds) emissions (Ding et al., 2019b; Liu and Wang, 2020), and increasing solar radiation (Wang et al., 2020; Ma et al., 2021).