The differences between inventories are substantial enough to influence simulated aerosol burdens (i.e., column-integrated aerosol mass, in mg m-2) at the global mean level. For 2014, i.e., the most recent common year for all three emission inventories, we estimate 4 % and 8 % lower global mean burdens of total BC when using CEDS21 and ECLv6 (6 % and 11 % if considering aerosols only from fossil fuel and biofuel combustion), respectively, compared to CEDS (Table S1). For primary organic aerosol (POA), the corresponding numbers are 11 % and 13 % (30 % and 40 %), while the global mean total sulfate burden is 8 % and 15 % lower with CEDS21 and ECLv6. Smaller reductions of the order of 3 %–4 % are also seen in the global mean SOA burden. Biogenic VOC emissions, the main source of SOA, are the same in all simulations. However, the SOA abundance is affected by the lower emissions of anthropogenic VOCs in both CEDS21 and ECLv6 than in CEDS (Fig. 1), as well as by the lower amount of POAs, which serve as substrates for SOA formation.

For all these aerosol species, the burden differences are consistently the largest over East Asia, followed by South Asia, and larger for ECLv6 than for CEDS21. Figure 2 shows the absolute regional mean burden (with corresponding relative changes given in Fig. S3). Regions considered are East Asia (EAS), South Asia (SAS), sub-Saharan Africa (SAF), North America (NAM), South America (SAM), North Africa and the Middle East (NAF), Europe (EUR), Southeast Asia (SEA), and Russia (RBU) (see also Fig. S2). For EAS, the new simulated burden of BC and POA is 30 %–40 % lower, depending on inventory, compared to simulations using CEDS, following 50 %–60 % lower BC and OC emissions. The 40 %–50 % lower SO2 emissions translate to a 20 %–30 % lower regional sulfate burden in our simulations. A similar relationship between emission and burden differences is simulated for SAS, where the burdens of BC, POA, and sulfate are 6 %, 27 %, and 30 % lower, respectively, in experiments with ECLv6 than with CEDS. Lower burdens of sulfate and POA are simulated for all other regions as well and in particular over NAF with ECLv6. In some regions, like SAM, NAF, and SAF, the new inventories estimate 20 %–30 % lower BC emissions than CEDS; however, due to the lower absolute magnitudes, the simulated burden differences are small compared to other aerosols. We note that regional burdens can be influenced by long-range transport and thus affected by inventory differences outside the main source region. We also note that we find differences in surface concentrations between simulations that are broadly similar to the burden changes. While beyond the scope of the present study, this may have implications for assessments of air-pollution-related health impacts.

The only species that is globally more abundant in simulations with the two new inventories is nitrate. There is considerable regional heterogeneity, where the burden is lower compared to the CEDS experiments in South Asia and on the US east coast but higher in the US Midwest; parts of Africa and South America; and, especially, over East Asia (Figs. 2 and S3). While absolute differences are small in many regions compared to other species, the net effect is nevertheless a 15 % and 24 % higher global mean nitrate burden with CEDS21 and ECLv6, respectively, compared to using CEDS emissions. Changes in the atmospheric nitrate distribution result from a complex interplay between differences in emissions of NOx, NH3, and SO2. Studies have also shown that nitrate formation can be influenced by background concentrations of VOCs (e.g., Womack et al., 2019) We find the largest absolute difference in nitrate in EAS and SAS, however, of opposite sign. In EAS, emissions of SO2 and NOx are both lower in ECLv6 and CEDS21 than in CEDS, whereas NH3 emissions are higher (Figs. 1 and S1). This results in lower chemical competition for available sulfate and, in turn, enhanced formation of nitrate aerosol. In SAS, SO2, NOx, and NH3 are all lower in the two new inventories than in CEDS, as is the nitrate burden. Differences in concentrations of VOCs in the simulations with different inventories is a further complicating factor. Studies have suggested that nitrate formation can be more sensitive to changes in VOCs than NOx; however, this is highly site specific (Yang et al., 2022). Further delineating the role of individual factors in nitrate differences would require simulations beyond what is available for the current study. The potential for an increasing relative role of nitrate in air pollution and climate in a world with concurrent declines in SO2 and NOx emissions but little in NH3 has also been discussed in previous studies (e.g., Bauer et al., 2007; Bellouin et al., 2011; Zhai et al., 2021). However, while more studies have focused on local air pollution impacts of nitrate and associated mitigation strategies, nitrate is still missing from many global climate models. Moreover, when included, the model diversity in simulated distributions is large (Bian et al., 2017). Our results suggest that uncertainties in emissions and choice of inventory can contribute to spread in simulated nitrate aerosols and confound the comparison of conclusions across modeling studies. Moreover, the complexity of the nitrate response demonstrates that the impact of inventory differences on simulated aerosols cannot be understood from scaling with the changes in individual emissions but requires explicit modeling.

To place the range in estimates between simulations with different inventories into more context, we compare the differences in simulated aerosol burdens in 2014 to the difference in burdens over the 5-year period from 2014 to 2019 using CEDS21. Both globally and regionally, the spread in burdens between simulations with different inventories and the 2014–2019 burden changes are of the same order of magnitude. In other words, at least in this case, the changes resulting from inventory differences are as large as those due to the recent overall change in anthropogenic emissions.

Combined, these burden differences translate to a 3 % and 6 % lower global annual mean AOD with CEDS21 and ECLv6, respectively, compared to CEDS in 2014 in our simulations. As expected, the differences are most pronounced over China and India (Fig. S4), where we estimate a 20 % and 30 % lower regional mean AOD in 2014 using the two new emission inventories, respectively, compared to using CEDS. For context, Fig. S4 also shows the interannual variability in AOD from MODIS Aqua (see Sect. 2): in these regions the differences between inventories are markedly larger than what can be expected from natural year-to-year variations.

Next, we take a closer look at differences in the simulated temporal trend, focusing on total AOD. Figure 3 shows the global and regional mean AOD from 1990 to 2019. Also shown is the linear trend from 2005 to 2017 for each of the time series. This period overlaps with the availability of remotely sensed AOD measurements, discussed in Sect. 3.1.3, as well as the most pronounced differences in the inventories. However, as there is a certain extent of inventory differences prior to 2005, we also provide corresponding linear trends over the full 1990–2017 period in Table S2.

The simulated AOD is consistently lower when using CEDS21 and ECLv6 emissions compared to CEDS over the full period studied, with increasing divergence over time, especially after 2005. We estimate a significant (at the 0.05 level – see Sect. 2) negative linear trend in the global mean AOD of -0.005 and -0.006 per decade in simulations with CEDS21 and ECLv6, respectively. This trend strengthens when extended to 2019 based on simulations with CEDS21. A negative global trend is also found when using the first CEDS release; however, it is smaller and not significant over the period 2005–2014. Extending the time series to 2017 by assuming that emissions follow SSP2-4.5 after 2014 (see Sect. 2), as in Fig. 3, the negative trend strengthens and switches to significant as per our definition, but it remains weaker than for the other two inventories. Considering the full period, we estimate a significant negative trend in simulations with CEDS21 and ECLv6 but no trend when using CEDS (Table S2). This long-term decline in the total AOD is primarily driven by the decline in sulfate AOD, following the emission decline after the introduction of air quality policies, first in the US and Europe, then in China, and the collapse of the Soviet Union (e.g., Aas et al., 2019). Over the full period, we simulate increasing trends in BC and nitrate AOD, significant at the 0.05 level, with all three inventories (not shown); however, their contributions to the total AOD are much smaller than that of sulfate. Robust evidence of a declining influence of aerosols on climate since 1990 was recently found from observables (Quaas et al., 2022). Our model simulations capture this overall trend, and the findings reinforce the role of changes in anthropogenic emission, particularly since 2005. Furthermore, we suggest that if using the original CEDS emissions, models may have failed to capture this trend. We note that biomass burning emissions also change over time in our simulations, but we do not find any significant trend in the AOD of biomass aerosols (BC and POA) on the global mean scale over this period. We do note that years of high biomass burning activity, such as in 2019 where GFED4 emissions are 25 % higher than in 2018, can lead to marked jumps in simulated AOD. We have limited possible influence of such years on the linear trend calculated (see Sect. 2).

Regionally, we simulate significant declining trends in AOD over 2005–2017 for EUR and NAM, with this trend extending back to 1990 (Table S2), as expected. This is also consistent with surface observations for both AOD and atmospheric sulfur and is in agreement with other models (Mortier et al., 2020; Aas et al., 2019); we capture the decline regardless of which emission inventory is used. In both regions, and across simulations with all three scenarios, we find a decline in the AOD of BC, OA, and sulfate but an increasing trend in nitrate AOD. Over RBU, we also simulate a significant decline in the area average AOD over the full 1990–2017 period but a flatter evolution when considering only 2005–2017. However, the results are also similar with all three scenarios here. In parts of the RBU region, GFED4 shows an increase in emissions over the latter period, resulting in a positive trend in the AOD of biomass aerosols from 2005. On the African continent, we simulate a negative, albeit weak, trend in AOD over the 2005–2017 period for SAF. In contrast, the trend over the full period is positive. Anthropogenic emissions in SAF have increased (Fig. S1), although less steeply than in Asia, and we find significant increases in the AOD of all the anthropogenic species with all inventories from 1990 to 2017. However, from 2005 onwards, there has been a decreasing trend in GFED4 emissions, following a reduction in the burned area of savannas (Wu et al., 2021). Biomass burning aerosols contribute relatively more to the total AOD here than in the Northern Hemisphere regions and hence impose a stronger effect on the area average trend. A similar pattern is seen for SAM, while for SEA, another biomass-burning-influenced region, we find less clear trends. While diagnosed trends in the total AOD in these regions are mostly of a similar sign across simulations with the three inventories, we find that the trend in sulfate AOD diverges between model runs using CEDS or CEDS21 (positive trend) and ECLv6 (negative trend) in SAF and NAF, pointing to a need to better understand the drivers of emission changes in these regions and homogenize between inventories. As expected, the key differences between simulations with different inventories arise over Asia. Simulations with both CEDS21 and ECLv6 show a significant decreasing trend in the total AOD over EAS between 2005 and 2017. While a decline is found using CEDS, it is much weaker and not significant. Moreover, differences between inventories affect the sign of the simulated trend when considering the full period, owing primarily to the spread in estimated sulfate AOD. For SAS, we simulate a consistent positive trend, but ranging from 0.01 per decade with ECLv6 to 0.03 per decade with CEDS, with increasing divergence in AOD over time. Similar magnitude differences between the sets of experiments exist for the AOD of all anthropogenic aerosols in this region.

To explore whether the model captures observed global and regional trends better with the CEDS21 emissions than with CEDS, we compare simulated AOD to MODIS Aqua retrievals and ground-based AERONET measurement. For this evaluation, we also use simulations where the model is driven by meteorology for the respective years, referred to as CEDSmet and CEDS21met (see Sect. 2), for more realistic comparison with the observations. Using both these, we also estimate negative linear trends in the simulated global mean AOD from 2005 to 2017, strengthening from -0.001 per decade in CEDSmet to -0.003 per decade in CEDS21met. These are, however, weaker than the trends derived from the fixed meteorology simulations in Sect. 3.1.2 (Fig. 3) and not significant at the 0.05 level, demonstrating the notable influence of variability in meteorology and natural aerosols, masking trends due to changes in anthropogenic emissions. This influence is particularly visible for the area average AOD for SAF and NAF, where the diagnosed trend is positive but non-significant in these simulations, in contrast to the negative trend found in simulations with fixed meteorology above. The negative trend over SAM is also not significant at the at the 0.05 level in these runs. For other focus regions, results are similar between fixed and actual meteorology runs, and significant trends arise over the natural variability.

Figure 4a shows the annual global mean simulated AOD from 1990 to 2017 and the MODIS Aqua AOD from 2003 to 2019. Dashed lines show the linear 2005–2017 trends. Figure 4b–d show the spatially explicit trends. We first note that the magnitude of the simulated global mean AOD is lower than that derived from MODIS Aqua by around 20 %. However, the overall geographical pattern of observed AOD is captured by the model (Fig. S5). Furthermore, the AOD simulated by the OsloCTM3 is within, although in the lower range, the spread in AOD between the CMIP and AeroCom models (Vogel et al., 2022). As also shown by Vogel et al. (2022), there can be a notable spread in AOD derived from different satellite products. They found a 13 % standard deviation range in the global mean AOD between eight satellite products, with MODIS retrievals in the upper end. Although again in the lower range, the OsloCTM3 AOD falls within the full range of the satellite-derived annual mean AOD. Overall, this suggests a reasonable OsloCTM3 performance in terms of magnitude and distribution.

In terms of temporal evolution, MODIS Aqua data indicate a very weak positive linear trend of 0.001 per decade in the global mean AOD over the 2005–2017 period (0.004 per decade when extending the data to 2019). We do not, however, find this trend to be significant. MODIS data are influenced by substantial year-to-year variability, in particular after 2010, which was also pointed out by Vogel et al. (2022). Regions with significant positive observed AOD trends include parts of the ocean in the Southern Hemisphere (Fig. 4b). Here, sea salt aerosols could be causing the increase. However, Quaas et al. (2022) recently showed that this positive trend is not clear in Multi-angle Imaging SpectroRadiometer (MISR) data. While we are focused on the anthropogenically influenced regions in the present analysis, we briefly note that the magnitude of the trends over the Southern Hemisphere oceanic regions is also not captured by the model (Fig. 4c–d). We also simulate weaker trends in the boreal regions of North America and Russia contributing to the model–observation difference.

Over the main anthropogenic emission source regions, there are significant observed declines in AOD over East Asia, the US, and Europe (Fig. 4b). These trends have been confirmed by both ground-based and remote sensing observations of AOD and other variables (Gui et al., 2021; Moseid et al., 2020; Paulot et al., 2018; Quaas et al., 2022). For NAM and EUR, we calculate an area average negative observed trend of -0.006 and -0.009, respectively, from MODIS Aqua. This is of the same sign but weaker than the trend simulated with both emission inventories. For the latter, this contrasts with findings by Mortier et al. (2020), where models in general were found to underestimate the observed decrease in AOD seen in surface observations. Over EAS, where the influence of inventory differences is most pronounced, a significant negative observed trend of -0.044 per decade is calculated. This is in very close agreement with the -0.40 per decade AOD trend simulated with CEDS21, while simulations with CEDS do not show a significant trend. Hence, the model is clearly able to better represent observed trends with the updated inventory. This is further confirmed in Fig. 5, where we show 5-year average deviations from the period 2003–2017 in both MODIS Aqua and simulated AOD. Using CEDS21 results in marked improvements compared to observed AOD trends over China, both for the first and most recent full 5-year periods. However, the opposite tendency is found for AOD over SAS. Here observations suggest a significant positive trend of 0.04 per decade. The diagnosed trends are also positive in simulations using both inventories but somewhat weaker, especially when switching from CEDS to CEDS21 (and even more so when using EClv6 emissions – Fig. 3). Figure 5 suggests that this discrepancy arises in the more recent decade. Furthermore, simulated AOD and underlying emissions suggest a leveling off in recent years, which is not seen from MODIS Aqua. Whether this is due to inaccurate representation of the evolution of anthropogenic emissions in the inventories or poor model representation of other aerosols such as dust from agricultural soils and urban areas (e.g., construction, non-exhaust transport emissions) is however not clear from this analysis. We note that the model underestimates the magnitude of AOD observed by MODIS Aqua in both EAS and SAS. To the extent that the MODIS is accurate, this could support the latter. This type of dust is suggested to give an important contribution to the particulate matter load (e.g., Chen et al., 2019; Xia et al., 2022) but is still missing from many global models. Other contributing factors include the representation of processes related to aerosol transport and scavenging. Finally, we also note that the 5-year deviations in Fig. 5 show quite some variability over the Middle East, with both positive and negative deviations from the baseline period. While anthropogenic emissions in this region increase steadily over the period (by 13 %–40 % depending on species) in the inventories used in the present study, the strong influence from dust emissions in this region likely dominates the temporal variability.

A previous OsloCTM3 study by Lund et al. (2018) found an improved agreement between the year 2010 ground-based AERONET observations and model output, including over Asia, when switching from CMIP5 and ECLIPSEv5 emissions to CEDS, the latter having higher emissions. This seemingly contradicts expectations following the now-known biases in this first release of CEDS. Here we repeat the comparison with AERONET but for the year 2014. Resulting scatter-density plots are given in Fig. S6. The normalized mean bias (NMB) compared to AERONET ranges from -21 to -29 % in the simulations with fixed and actual meteorology. We find higher bias and lower correlation when switching from the original CEDS release to CEDS21 and ECLv6. Hence, while the model is better able to represent observed recent aerosol trends over East Asia with newer emission inventories, these results point to other issues that may have been concealed by too high anthropogenic emissions. Dust and atmospheric processing, as discussed above, are again possible contributing factors.