All models used the same anthropogenic emissions dataset, which is called ECLIPSE (Evaluating the Climate and Air Quality Impacts of Short-Lived Pollutants) v6B. These emissions were created using the IIASA-GAINS (International Institute for Applied Systems Analysis – Greenhouse gas – Air pollution Interactions and Synergies) model , which provides emissions of long-lived greenhouse gases and shorter-lived species in a consistent framework. These historical emissions were provided for the years 1990 to 2015 at 5-year intervals, as well as the years 2008–2009 and 2014. Those models that simulated the 1990–2015 time period linearly interpolated the emissions for the years in between. The ECLIPSEv6b emissions include many pollutants, such as CH4, CO, NOx, BC, and SO2. They include the significant sulfur emission reductions that have taken place since the 1980s . Global anthropogenic BC emissions are estimated to be 6.5 Tg in 2010 and 5.9 Tg in 2020, and global anthropogenic SO2 emissions are estimated to be 90 Tg in 2010 but declined significantly over the subsequent decade to 50 Tg . The reductions are mainly due to stringent emissions standards in the energy and industrial sectors, as well as reduced coal use in the residential sector . Global anthropogenic methane emissions were 340 Tg in 2015 and 350 Tg in 2020, and they are expected to continue to increase, unlike BC and SO2. The largest methane sources in 2015 were agriculture (42 % of total emissions), oil and gas (extraction and distribution) (18 %), waste (18 %), and energy production (including coal mining) (16 %) . CO and NOx emissions have been declining steadily and are expected to continue declining in the future.

In comparison to the CMIP6 emissions , ECLIPSEv6b emissions have additionally taken into account the recent declines in emissions from Asia of SO2, BC, and NOx due to recent control measures, whereas those declines in the CMIP6 emissions were unrealistically small . The inclusion of emissions from the flaring sector in Russia was a significant improvement, which was not present in the previous version of ECLIPSE emissions that was used in the report.

For non-agricultural fire emissions, many models utilized the CMIP6 fire emissions, which are based on monthly GFED (Global Fire Emissions Database) v4.1 . About half of the models included volcanic emissions or stratospheric aerosol concentrations from the CMIP6 dataset or other sources, and the other half did not include volcanic emissions, which mainly impact SO2 and thus modeled SO42-. The emissions from the October to December 2014 Honoluraun volcano eruption were included by six models in a separate set of simulations. Similar differences in biogenic and agricultural waste emissions appear in these model simulations, and all are summarized in Table .

This section contains a summary of models' chemistry schemes, and we refer the reader to Appendix and references therein for more details.

Satellite observations are useful for evaluating models on larger horizontal spatial scales and for evaluating the three-dimensional atmosphere – not the surface concentrations. Observations from three satellite instruments were used to evaluate model trace gas distributions in the free troposphere and, when appropriate, the lower stratosphere. These were the Tropospheric Emission Spectrometer version 7 (TES; ), the Atmospheric Chemistry Experiment–Fourier Transform Spectrometer version 4.1 (ACE-FTS; ), and the Measurements of Pollution in the Troposphere version 8 (MOPITT; ). The vertical profiles of trace gas volume mixing ratios are derived or retrieved from the satellite-measured emission or absorption spectra, with varying degrees of vertical sensitivity. These remote techniques typically have about a 15 % uncertainty in the measurements (e.g., ), though this depends on the specific instrument and the species retrieved (e.g., ).

Note that while TES and MOPITT have global spatial coverage, their coverage does not extend up into the high Arctic. The TES instrument on NASA's Aura satellite measures vertical profiles of trace gases such as O3, CH4, NO2, CO, and HNO3- from 2004–present. After interpolating all models and TES results to a 1∘ × 1∘ horizontal grid, the monthly mean CH4 and O3 from the TES lite products were matched in space and time with models. Models were smoothed with the TES monthly mean averaging kernels prior to comparisons with satellite data. TES measurements started in 2004, stopped in late 2015, and had poorer coverage in the last few years.

A similar comparison method was used for MOPITT data. The MOPITT instrument on NASA's Terra satellite measures CO from 2000 to the present.

The ACE-FTS instrument on CSA's SCISAT satellite has measured the trace gases O3, CO, NO, NO2, and CH4, among over 30 others from 2004–present. SCISAT has a high-inclination orbit, giving its instruments better coverage in the Arctic. ACE-FTS is a limb-sounding instrument measuring the solar absorption spectra of dozens of trace gas concentrations from the upper troposphere to the thermosphere. This gives us the opportunity to evaluate the 3-D model output in a region of the atmosphere where the radiative forcing of ozone is at its highest. Evaluating models with ACE-FTS measurements also provides insight into models' transport and upper-tropospheric chemistry. As was shown in , 3-hourly model output (rather than monthly mean output) is required for accurate comparisons to ACE-FTS data; thus, only models that provided output at this time frequency were compared to ACE-FTS measurements. The model output was sampled to match the times and locations of ACE-FTS measurements. We used an updated version of the advanced method in . Instead of taking the model output at the closest time to the ACE-FTS measurement time, the model output was linearly interpolated onto the ACE-FTS time. This reduces the bias introduced by diurnal cycles, which can cause certain volume mixing ratios (VMRs; e.g., that of NO and NO2) to vary significantly between model output times. As in , the model output is also interpolated vertically in log pressure space and bilinearly in latitude and longitude to account for spatial variation between model grid points.

Finally, there were air- and ship-based measurement campaigns of black carbon that were used for model evaluation. Aircraft campaigns allow vertical profiles of chemical species to be evaluated, and ship campaigns allow for in situ measurements in the remote Arctic seas.

The 2014–2015 average modeled percent biases for surface concentrations of SLCFs are shown in Fig.  for each model and the multi-model mean (mmm). This figure is based on the model comparisons at the surface observation locations that will be shown in subsequent sections (Figs. , , , , , , , , and and additional American observations from the IMPROVE network for BC, SO42-, and OA).

Figure b shows that, for surface Arctic concentrations, no one model performs best for all species but that the mmm performs particularly well. It also shows that the model biases vary quite a bit among SLCF species for both the midlatitudes and the Arctic. It is important to note that there are many more measurement locations at midlatitudes compared to in the Arctic. BC, CH4, O3, and PM2.5 have the smallest model biases out of the SLCFs of this study, whereas OA, CO, and NO2 have larger model biases.

We find for half of the SLCF species that the mmm percent bias decreases with latitude (Fig. ). O3, NO2, BC, and SO4 have a negative slope in the bias vs. latitude figure. So if the mmm bias was high at the midlatitudes, it is close to zero in the Arctic (O3), and if the mmm bias was near zero at midlatitudes, it is negative in the Arctic (NO2, BC, SO42-). This implies that there is not enough long-range transport from the midlatitude source regions to the Arctic. That said, the mmm CH4 bias stays relatively constant with latitude, and we will see in Sect.  that this result is model-dependent. The CO, PM2.5, and OA mmm biases have an increasing trend with latitude. However, both CO and PM2.5 have no observation locations in the high Arctic, so those results cannot represent long-range transport. OA only has one observation location in the high Arctic, and its bias is very large overall, so issues other than long-range transport are at play, as we will see in the following discussion (Sect. ).

Of course, there are spatial, temporal, and model differences in the results, so we will now explore model performance for each SLCF in more detail in the next subsections.

Measured annual mean surface methane is shown in the top left panel of Fig. , along with model biases in the rest of the panels. Recall that unlike the rest of the SLCF species in this study, CH4 concentrations were prescribed in these models from the same CH4 dataset . That said, the different decisions by modelers on how those CH4 global concentrations are distributed make differences in how these models compare to measurements. The mean model biases are small and mainly positive; in the midlatitudes, the multi-model mean bias is +145 ppbv (or +8.5 %), and in the Arctic, the mmm bias is 24 ppbv (or 1.3 %), which means that the models simulate the magnitude of surface CH4 well – though still outside the < 1 % measurement uncertainty range. There is a gradient in CH4 VMRs (higher in the Northern Hemisphere and lower in the Southern Hemisphere) that is seen in the measurements (Fig. , top left) and reported in the literature (e.g., ), though it is not well captured by CMAM, MRI-ESM2, and UKESM1 models, which are all biased low in the Northern Hemisphere and biased higher towards the south. That is because of the simplifications made in these models' distributions of CH4. For example, CMAM used a single global average CH4 concentration that is interpolated linearly in time from once-yearly values.

Figure also shows that observed annual mean surface CH4 ranges geographically from about 1500 to 2100 ppbv depending on location; however, the models have a much smaller range due to their prescribing CH4 concentrations as a lower boundary input. For example CMAM CH4 volume mixing ratios only span about ±3 ppbv around 1836 ppbv. The span of MRI-ESM2 surface CH4 is even smaller. GEOS-Chem, GISS-E2.1, and OsloCTM have a more realistic range of 1700–2000 ppbv, though they still do not get the full variability that is seen in surface CH4 mixing ratios close to CH4 sources. However, in the free troposphere (above the boundary layer), we have TES satellite measurements of CH4 that show that CH4 is much more smoothly distributed aloft. Thus, the simplification of prescribing CH4 concentrations in the models is more realistic there (Fig. , showing the 600 hPa level in the mid-troposphere). Additionally, Fig.  better illustrates the latitudinal gradient in CH4 over the globe and its lack in some models, which have more negative biases in the Northern Hemisphere and more positive biases in the Southern Hemisphere. Other models, such as GISS-E2.1, do a good job of capturing the global distribution of CH4.

In the Arctic, the vertical cross section of CH4 VMRs over time as measured by the ACE-FTS in the middle to upper troposphere and in the stratosphere is shown in Fig. S1. There is a large decrease in CH4 above the tropopause at around 300–100 hPa. The models are all biased low around 300 hPa and high around 100 hPa. This pattern is true for midlatitudes as well as in the Arctic and may imply that the altitude of the modeled tropopause is too low. This same conclusion was also found in via comparisons of these models' simulations to ozonesonde measurements and in our satellite O3 comparison in the next section. The CH4 model–measurement correlation coefficients for ACE-FTS are relatively high (e.g., R=0.48 to 0.86 depending on the model).

Therefore, the general model evaluation for CH4 indicates that because models do not explicitly model CH4 emissions, they do not simulate the surface-level variability of CH4 VMRs. Models differ in their global distribution of CH4; thus, only some contain the north–south CH4 gradient. Those that do not have the largest underestimations of Arctic tropospheric CH4. The CH4 evaluation also implies that the modeled tropopause height may be too low.

Tropospheric O3 is the third most important greenhouse gas , and it is a regional pollutant that causes damage to human health and ecosystems. O3 is a secondary pollutant formed in the troposphere via photochemical oxidation of volatile organic compounds in the presence of nitrogen oxides (NOx= NO + NO2). As such, models must simultaneously simulate the meteorological conditions, precursor species distributions, and photochemistry correctly in order to accurately simulate O3. That said, since surface O3 is an important contributor to poor air quality, there is significant pressure for models to simulate it accurately, particularly in the heavily populated midlatitudes (e.g., for air quality forecasting). Only models with prognostic O3 are included in this section.

Figure shows the in situ summertime mean O3 measurements (top left panel), and the model biases (remaining panels) and the same for the annual mean is shown in the Supplement (Fig. S2). These include averaging O3 from hourly observations (day and night) and 3-hourly or monthly modeled O3 depending on which were available for each model. Surface O3 is overpredicted by most models, which has been documented previously . It has been shown that models can have problems producing low O3 overnight . In the Arctic, simulated surface O3 has more mixed results. Annual mean concentrations are of the order of 40 ppbv, and individual model biases range from -20 % to +52 % globally on average for 2014–2015. The multi-model mean has a bias of +11 % for the Arctic, but this is not uniformly spatially distributed. All models overestimated surface O3 in Alaska (mainly due to the overestimation of summertime concentrations, discussed below), and most models have too little O3 at the Greenland location and in northern Europe.

The models were all able to represent the summertime peak in the midlatitude seasonal cycle (not shown). In contrast to the more polluted midlatitudes, where surface O3 peaks in the summertime due to photochemical production being at a maximum, Arctic O3 is more influenced by the Brewer–Dobson circulation, bringing a maximum of tropospheric O3 in the springtime due to photochemical production , descent from the stratosphere, and more long-range transport of O3 to the Arctic. Figure , shows this springtime peak in both the western (a, longitude < 0∘) and eastern (b, longitude > 0∘) Arctic in the measurements. However, the models only capture that seasonal cycle in the eastern Arctic (Fig. b), implying that the models represent large-scale circulation and possibly stratosphere to troposphere exchange well. But it is interesting to note that the models that have sophisticated representation of stratosphere–troposphere exchange (such as CMAM, MRI-ESM2, UKESM1) do not particularly stand out as better performers in Fig.  compared to models that do not simulate the stratosphere (such as DEHM, MATCH, MATCH-SALSA). Thus, its impact on surface O3 may be very small.

In the western Arctic (Alaska mainly, Fig. a), models overestimate summertime Arctic O3, likely due to overpredicting the impact of wildfire emissions on tropospheric O3 concentrations, which is a research topic with high uncertainty . Another possibility is that modeled O3 dry deposition over boreal vegetation is underestimated .

Some Arctic locations are more inclined to get springtime surface O3 depletion due to bromine explosions and halogen chemistry . None of the model simulations in this study contain the necessary tropospheric halogen chemistry to simulate those events, which partly explains why some models in Fig.  (bottom) overestimate springtime O3 concentrations. That particular feature is explored further on a site-to-site basis in .

The next subsection shows that both the O3 precursors CO and NO2 are underestimated compared to measurements at all global locations. This has implications for simulated tropospheric O3 chemistry.

Free-tropospheric O3 – satellite comparisons. Aircraft-based measurements and ozonesondes can provide insight into the vertical distribution of O3, and these have been well documented (e.g., ). However, model grid boxes may not be representative of those fine-spatial-scale measurements. In this study, we examine how the model biases change in the vertical when compared to satellite measurements, which have a larger, “smoothed out” spatial sensitivity due to their viewing geometry and retrieval methods. Specifically, we compare modeled O3 to TES and ACE-FTS satellite-based retrievals. These satellite instruments also have better global coverage than aircraft and sonde-based measurements.

The model fractional biases compared to TES measurements from near the surface up to 100 hPa are shown in Fig.  for the Arctic (left) and midlatitudes (right). All models' simulated fractional biases have similar vertical profiles for both the Arctic and midlatitudes, with greater negative values at lower levels and a more positive “bulge” of about 10 % around 300 hPa in the Arctic and about 5 % around 200 hPa at midlatitudes. That bulge in model biases at 300 hPa was also seen to a greater degree (50 %–70 %) when comparing these model simulations to Arctic ozonesonde measurements in . Compared to TES, which has much lower vertical resolution, the results are not as striking but are consistent with ACE-FTS measurements. On average, the models have a negative bias at all vertical levels in the Arctic region and in lower troposphere in the midlatitude region, whereas a positive bias is seen in the upper troposphere below 60∘ N. This is consistent for the two time periods (2008–2009 and 2014–2015).

Given that Fig.  shows positive biases near the midlatitudes, while Fig.  shows lower O3 in the free troposphere, these results imply that there is not enough vertical lifting and/or mixing of tropospheric O3 in most of the models. However, the TES measurements have been shown to be biased high by approximately 13 % throughout the troposphere , which is the same amount that the mmm is low. Similarly, ACE-FTS O3 has an uncertainty range of +5 %–10 % when compared to O3 from other satellite limb-view observations . The ACE-FTS comparison for O3 can be found in the Supplement (Fig. S3), showing higher model biases around 300–100 hPa (except for GEOS-Chem) and good agreement below that. Therefore, overall, participating models simulate the free-tropospheric O3 reasonably well and within the uncertainly limits of the observations.

Therefore, the general model evaluation for O3 indicates that all models overestimate surface O3 at midlatitudes, and that, combined with a lack of O3 transport to the Arctic, results in modeled Arctic O3 VMRs having relatively little bias (the right answer for the wrong reason). The summertime evaluation implies that models overestimate the O3 produced and transported by wildfires in the western Arctic. The O3 evaluation also implies that the modeled tropopause height may be too low.

Figures  and show the comparisons of the multi-model medians (MMMs) to the surface in situ measurements. The figures for each model appear in the Supplement (Figs. S4 and S5), but only the MMMs are shown here since the spatial patterns were very similar for all models. The multi-model annual mean underpredicts both CO and NO2 by approximately -55 % in the Northern Hemisphere for 2014–2015. The 2015 AMAP report showed similar findings for simulated surface CO, as have other studies , pointing to a likely underestimation of CO emissions and possibly shorter modeled lifetimes of CO due to an overestimation in OH . The annual mean surface CO underestimation is mainly dominated by the wintertime (e.g., the mmm bias in DJF is -92 %), when it has been reported that CO emissions from combustion are too low (e.g., ). All the models exhibit a large negative bias over China, which is consistent with the study by and is attributed to the enhanced destruction of CO by OH radicals, but it was also found in and that bottom-up CO emission inventories in Asia are greatly underestimated.

In the free troposphere, we compare modeled CO to that measured by MOPITT. Figure  shows these comparisons for the summertime (JJA) mean at the 600 hPa level, and Fig. S6 in the Supplement shows the same for the springtime (MAM) mean. Unlike in the winter and spring when all models are biased low at midlatitudes, there is more variability in the summertime biases, with about half the models overestimating free-tropospheric CO and the other half underestimating it. In the Arctic region, all models are biased high in the summer but low in the springtime. In the winter, MOPITT does not have coverage in the Arctic. discussed the fact that models had high biases in the tropospheric column of CO compared to MOPITT measurements in the outflow from Asia and low biases north of there due to lack of transport. The study also suggested that summertime CO transport out of Asia is too zonal when comparing models to IASI CO columns. At 600 hPa, where CO concentrations are lower and the atmosphere is more mixed, these features do not appear.

In the upper troposphere and stratosphere, modeled CO and NOx monthly time series are compared to measurements from the ACE-FTS satellite instrument (where NOx= NO + NO2, which are measured separately), and those results are shown as Taylor diagrams in Fig. S7, along with O3 and CH4 at 150 hPa, which is in the upper troposphere–lower stratosphere (UTLS) region. The contours show the model's overall skill as defined in . Only the models that simulate the stratosphere were included, and the results show that there is a range in model performance by SLCF species, with no one model performing best for all. Comparison statistics for CO were poorer than those for O3, CH4, and NOx.

In this section, we examine the spatial and seasonal distributions of BC using ground-based measurements, which are primarily available in North America, Europe, and several locations in the Arctic, but are also available from two ship-based campaigns and several aircraft campaigns. Given the limited global data available for both BC and SO42- (e.g., we could find none freely available for Asia), we focus the plots on the Arctic region here, and given that the magnitude of BC and SO42- does not span a wide range throughout the Arctic, we show model biases as percent rather than absolute differences as was done in previous sections for trace gas species shown globally. We also analyze the BC model–measurement comparisons, keeping in mind that because there are various definitions and measurement types for BC, we consider agreement within a factor of 2 to be within the uncertainty range (Sect. ).

Figure  (top left panel) shows annual mean surface-level concentrations of black carbon (BC) at nine Arctic observation stations and (remaining panels) the model percent biases there. The annual mean BC concentrations are of the order of less than 1 µg m-3, and most models tend to underestimate BC in the high Arctic while overestimating in Alaska and Scandinavia. This result could be partially explained by the discrepancy caused by the use of EC and eBC data, which are not the same (Sect. ). As aerosols age during transport to high-Arctic locations, their eBC (based on absorption converted to mass) gets more and more of a positive bias compared to EC. As models are more representative of EC, they will not be able to agree with eBC measurements in aged air at high-Arctic remote stations, such as Gruvebadet, Zeppelin, Alert, and Utqiaġvik. This is in contrast to the Alaskan and European stations, which are closer to sources where BC is more fresh; thus, the eBC measurements have lower uncertainty.

That said, a few models (CanAM5-PAM, DEHM, and FLEXPART) overestimate BC concentrations in the high Arctic. Overall individual model biases range ±100 % at individual sites.

The underestimation of high-Arctic atmospheric BC concentrations may be related to excessive BC deposition further south; however, there are very few BC deposition measurements. In the Arctic, we can evaluate total (wet + dry) modeled deposition via derived ice core measurements. There were six ice cores on Greenland and one in the European Arctic in Spitsbergen (Lomosovfonna). Figure  shows that all models overestimate BC deposition fluxes at the ice core locations. While the ice cores contain BC data starting in the year 1750, only data after 1990 have been used to match the modeled time period (1990–2015, 1995–2015, 2008–2009, and 2014–2015, depending on the model).

The measured BC deposition flux values on Greenland vary with elevation (lower fluxes at higher elevation). Summit (3177 m a.s.l.) has an average of 285 µg m-2 yr-1 in contrast to ACT2 (2461 m a.s.l.) with 676 µg m-2 yr-1. BC deposition is highest in the European Arctic at Spitsbergen with 856 µg m-2 yr-1. For all seven ice cores used in this comparison the averaged model mean is 3 times as high as the observations. At D4 (2728 m a.s.l.) the modeled mean corresponds best to the observation, with a mean bias of +83 %. At ACT11 (2296 m a.s.l.) the models have 4 times the deposition flux compared to the measurements. Generally though, the model mean is skewed higher by FLEXPART and DEHM (Fig. ), which also had higher atmospheric BC concentrations. A few models simulated less BC deposition than observed at these sites, and these models also underestimated BC atmospheric concentrations. Thus, it is difficult to conclude that deposition biases are a cause for atmospheric biases when the two are inter-related parameters.

The seasonal cycles of surface-level atmospheric BC concentrations at several Arctic locations are shown in Fig. . As seen in , some models still underestimate wintertime BC, but many models now show similar seasonality as the observations. The multi-model mean also captures the monthly variations well including the summertime peak at some Alaskan sites caused by fire emissions. The multi-model mean Arctic BC is underestimated in winter (-24 %) and overestimated in the summer (+32 %), though overall, this is an improvement in model performance in simulating Arctic BC since the 2015 AMAP assessment of black carbon and ozone as climate forcers – the latter of which had -59 % winter bias and +88 % summertime bias . However, it is difficult to make direct comparisons to that report as those values were for a smaller set of Arctic locations, different observation periods, and with a different set of models (though many overlapping). The model improvement may be due to the improved anthropogenic emissions of BC, particularly from northern Russia, where flaring emission factors were increased in ECLIPSEv6B compared to those used for the 2015 AMAP report.

Most models have reasonable spatial correlation with the measurements across the Arctic in that they correctly simulate the range of BC concentrations that appear across the Arctic (e.g., higher concentrations at Hurdal, lower concentrations at Zeppelin), as shown in Fig. S8. However, there are still large differences and low R values in the statistics shown in Fig.  S8.

There are positive model biases at midlatitudes (in North America and Europe; not shown) for surface-level BC. The vertical analysis of BC from the aircraft campaigns (below) provides further insight and support for the suggestion that models do not have adequate long-range transport of the pollutants from their sources in the midlatitude; thus, they do not simulate enough pollution in the Arctic.

Vertical profiles of BC – aircraft campaigns. Gridded BC output at 3-hourly intervals was provided by 11 of the participating models and was compared to aircraft campaign measurements of BC. The interpolation of model output to flight track coordinates was carried out by tools from the Community Intercomparison Suite (CIS; ), which co-located the extracted model tracks with their corresponding observational values.

Figure shows the median vertical profiles of BC concentrations from the aircraft measurements and from the models. At midlatitudes, from 0–2 km, all of the models agree well with the measurements. However, BC concentrations decline steeply in a few models (e.g., MATCH-SALSA, EMEP-MSC-W, and GEOS-Chem) above 2 km. It would appear that they do not have enough vertical lifting of BC and/or perhaps too short of a BC lifetime. Indeed, one of these is EMEP MSC-W, in which the short BC lifetime was previously reported in . That said, in , the Oslo-CTM and ECHAM models were shown to overestimate the BC lifetime. In our case, OsloCTM is not shown in Fig.  because it did no provide BC at 3-hourly timescales. But ECHAM-SALSA results are consistent with the study in that they particularly overestimate BC in the upper altitudes in both the midlatitudes and the Arctic, implying too long a lifetime and too much long-range transport into the upper Arctic atmosphere. The measured BC profile at midlatitudes drops off more quickly around the tropopause at 11–12 km, and, except for CanAM5-PAM, the models do not reproduce this drop.

In the Arctic profiles (Fig. ), the modeled and observed profiles do not decline with altitude throughout the troposphere, but the observed median BC concentration does drop sharply around 9 km – again near the Arctic tropopause, and again, the only model to capture that change is CanAM5-PAM. In the Arctic comparisons, the models that did not simulate enough BC aloft at the midlatitudes stand out as having larger underestimates of BC in the Arctic. For example, MATCH-SALSA and EMEP MSC-W have very low BC throughout the Arctic vertical profile. These results are consistent with the surface BC underestimation in Fig. . Therefore, the underestimation seen in the Arctic for those two models is likely due to a lack of long-range transport from the midlatitudes, as well as errors in BC deposition mentioned above. In addition, the study showed that different parts of the Arctic BC vertical profile are sensitive to BC transported from different areas of the world. For example, the lower-tropospheric BC is influenced by emissions transported from North America, Russia–Belarus–Ukraine, Europe, and East Asia, whereas upper-tropospheric Arctic BC is mainly influenced by transport from South Asia. Thus, the differences in the model results could be related to differences in how they simulate these transport pathways.

found that, overall, considerable differences in wet deposition efficiencies in the models exist and are a leading cause of differences in simulated BC burdens. Results from their model sensitivity experiments indicated that convective scavenging outside the Arctic reduces the mean altitude of BC residing in the Arctic, making it more susceptible to scavenging by stratiform (layer) clouds in the Arctic. Consequently, scavenging of BC in convective clouds outside the Arctic acts to substantially increase the overall efficiency of BC wet deposition in the Arctic, which leads to low BC burdens compared to simulations without convective BC scavenging. also found that convective scavenging at middle and subtropical latitudes removes a significant fraction of BC. In contrast, BC concentrations in the upper troposphere are only weakly influenced by wet deposition in stratiform clouds, whereas lower-tropospheric concentrations are highly sensitive – these are consistent with the results we find in this study, wherein the multi-model median is too high above about 9 km and too low from 0–9 km. Indeed, the MATCH and MATCH-SALSA models, for example, assume reduced scavenging of aerosol in mixed-phase clouds following , which increases long-range transport to the Arctic. It is odd then that MATCH is one of the better=performing models in Fig.  and MATCH-SALSA is not. Despite the large range in modeled vertical BC concentrations, the multi-model median is close to the observed throughout the troposphere at both midlatitudes and in the Arctic.

Arctic seas analysis – ship campaigns. From the ship-based measurements, we see that there is a consistent model overestimate of BC in the Pacific region (Japan cruise) where measured concentrations are very low (Fig. ). Indeed, report that BC concentrations were in the range 0–66 ng m-3, with an overall mean value of just 1.0 ± 1.2 ng m-3. The models, possibly due to their coarse resolutions, were not able to simulate such low background BC concentrations. However, even the model with the highest resolution (GEM-MACH at 15 km resolution) overestimated BC in the Pacific – though that limited-area model in that region near the boundary would have been heavily influenced by the upwind, coarser-resolution boundary conditions that were assumed (1∘ × 1∘ MOZART4 chemical boundary conditions). The high bias in the Pacific may be due to all models overestimating the amount of BC that gets transported off the Asian continent. That model bias is consistent with our low-altitude comparisons of the models to the HIPPO aircraft campaign measurements, which were taken over the northwestern Pacific (Fig. ). The BC overestimate over the Pacific was also found in the study looking at simulated BC from the AeroCom global model intercomparison initiative compared to HIPPO measurements.

Conversely, the modeled BC concentrations generally agree with measurements in the Russian Arctic Ocean, though they are biased slightly low for the most part. attribute the higher BC concentrations measured near the Kara Straight (north of 70∘ N) to gas-flaring emissions, and near Arkhangelsk (White Sea), important sources were midlatitude biomass burning, transportation, and combustion (residential and commercial). Since models were able to simulate this well, their improvement is likely due to improved Russian anthropogenic emissions in ECLIPSE v6B (Sect. , ) compared to previous emissions datasets, which did not include enough Russian flaring emissions. The best model results were from ECHAM-SALSA and MATCH when compared to all of the ship campaign data.

Therefore, the general model evaluation for BC indicates that while there is a large variability in models results, they tend to overestimate surface BC at midlatitudes (including over the Pacific Ocean) and underestimate surface BC in the Arctic. Again, these results point to a lack of transport to the Arctic and, in this case, too much BC deposition along the way. While we were only able to evaluate BC deposition in the Arctic in this study, those results support the hypothesis of some models having too much BC deposition. The BC vertical profile evaluation also implies that the modeled tropopause height may be too low.

We used monthly mean surface level observations of SO42- from 18 Arctic sites to evaluate the models. Figure  shows that, similar to BC, the SO42- concentrations in the high Arctic are underestimated by most of the models. A few models overestimate SO42- in Scandinavia and Alaska.

The model underestimations of SO42- could be mainly due to higher efficiencies of models in removing aerosol during the long-range transport to the high Arctic. This is consistent with a previous study based on AMAP 2015 model simulations that found that the convective wet deposition outside the Arctic region may have led to different seasonal cycles of aerosol concentrations in the Arctic . Dimethylsulfide (DMS), a naturally occurring source of sulfur from marine algae emissions, could also be misrepresented in models. However, this source would be more pronounced in the summer when there is less sea ice in the high Arctic, and it does not appear as though models are underestimating only in the summertime (Fig. ). Rather, some models appear too low in the winter and spring – which points towards underestimating local Arctic sources and to a lack of transport from midlatitudes as being the key issues. Despite the individual model differences in representing the seasonal cycle, the multi-model mean compares well with observations at most locations. However, the multi-model mean SO42- is significantly underestimated at Alert and Irafoss sites. Mean model biases for all Arctic sites range from -65 % to +80 % among different models, and correlation coefficients are typically around 0.5 (Fig. S9).

The high GEOS-Chem bias in the summertime seen in Fig.  was first reported in and found to be due to problems with cloud pH and cloud liquid water in the summer over the Arctic. In , the summertime Arctic surface SO42- concentrations are reduced by a factor of 2 by reducing the cloud liquid water content to a uniform value of 1 × 10-7 g m-3 north of 65∘ N in the model. The version of GEOS-Chem in our study is more recent and uses the offline cloud liquid water content from both GEOS-FP and MERRA2. We did not scale this variable down, which may be a reason for the high GEOS-Chem sulfate bias in Fig. .

The seasonal cycle for observations grouped together is shown in Fig. S10, showing a consistent seasonal cycle for 2008–2009 as seen in the observations. Most models (e.g., CanAM5-PAM, DEHM, MATCH, OsloCTM) are able to capture the seasonal cycle well. However, several models (e.g., CESM, CIESM-MAM7, ECHAM-SALSA, and EMEP-MSCW) strongly underestimate observed springtime peak values. Conversely, the models and the measurements showed a weaker seasonal cycle during the 2014–2015 time period (Fig. S11). It may be partly due to the local pollution sources in the Arctic during wintertime (e.g., Fairbanks). Those highly localized pollution events, caused by local emissions getting trapped in a stable boundary layer, occur on scales that are smaller than the model resolutions employed here can represent. Many models are also missing chemical formation processes for SO42- in the absence of sunlight, which may explain underestimations seen in winter (e.g., ). An evaluation of SO2 could help with our understanding but was beyond the scope of this study. A lack of dark chemistry may be true for organic aerosol, as discussed in the next section as well.

From October to December 2014, the Honoluraun volcanic eruption may have elevated SO42- concentrations at some locations in the Arctic. However, in our model–measurement comparisons, there does not appear to be a large underestimate during those months, which implies that model performance was not impeded by not including those volcanic emissions.

As mentioned above, the uncertainty in wet deposition could be a significant factor in atmospheric SO42- model biases. Previous studies have shown that models have too much washout in winter and not enough wet deposition in summer, leading to a “flatter” seasonal cycle than observed (e.g., Fig. ; ). As with BC in the previous section, the SO42- deposition was evaluated here in the same manner. The average measured SO42- deposition fluxes from ice cores for all locations (only Greenland was available here) are 18 mg m-2 yr-1. The lowest observed fluxes are found at D4 (12 mg m-2 yr-1) and the highest at ACT11D (30 mg m-2 yr-1). The model average for all locations is overestimated by around 20 % compared to measured fluxes. This is similar to the model biases in atmospheric SO42- concentrations.

Therefore, the general model evaluation for SO42- indicates that while there is a large variability in models results, as with BC, models underestimate surface SO42- in the Arctic. The evaluation of SO42- deposition in the Arctic is similar to BC, with both overestimating deposition.

Unfortunately, there is only one high-Arctic station with OA measurements (Alert, NV, Canada); however, there are a few additional stations measuring OA in the sub-Arctic (still > 60∘ N). These are all shown in Fig. . The seasonal cycles are shown in Fig. , and the model vs. measurement scatter plot along with some comparison statistics are presented in Fig. S12.

Model biases have a large range of ±200 % at the different locations, but the multi-model mean for the region is +65 %. At midlatitudes (30–60∘ N), measurements are conducted mainly in the US, where the multi-model mean bias is +83 % for the 2014–2015 average (not shown).

Several models (CanAM5-PAM, DEHM, CIESM-MAM7, ECHAM-SALSA, GEOS-Chem, MRI-ESM2, NorESM, OsloCTM, and UKESM1) are able to simulate the summertime peak in Arctic OA concentrations; however, the other seven models in Fig.  simulate a seasonal cycle that is too flat or peaks at the wrong time (e.g., the CESM seasonal cycle peaks too early in the year).

Figures and S11 both show that most models consistently overestimate Alaskan OA and underestimate European OA, consistent with our assessment of other species showing that models are likely overestimating wildfire influence in summertime Alaska. All models, except EMEP-MSC-W, used the GFED fire emissions inventory. EMEP-MSC-W used the FINN fire emissions inventory, which for BC+OC has been shown to be significantly lower than the GFED emissions . As a result, EMEP-MSC-W model biases in Alaska are lower. However, they are not the lowest. MATCH-SALSA has the lowest OA model biases in Alaska, despite that model using GFED emissions. The comparison statistics in Fig. S12 show highly varying results.

PM2.5 is partly connected to direct and indirect climate effects via its interactions with clouds. It is mainly composed of BC, SO42-, OA, NO3-, and NH4+, as well as crustal material (dust), sea salt, and water, though the water component is often dried off during measurements. Model biases of those species will contribute to the total PM2.5 biases.

While BC, SO42-, and OA were discussed above, it is beyond the scope of this project to evaluate the other major PM species, which, aside from water, have a smaller radiative impact. Note that the analysis in this section is focused on sub-Arctic and midlatitude sites closer to human populations rather than remote high-Arctic sites due to a lack of data. PM2.5 is not a typical parameter included in the longer-term remote Arctic observations. Since PM2.5 has important health impacts, it is well measured at air quality monitoring networks.

The model PM2.5 biases at several locations in the United States, Europe, and China are within the 60 %–80 % range. However, some models (CanAM5-PAM, CIESM-MAM7, GEOS-CHEM, GEM-MACH, and Oslo-CTM) show biases larger than 200 %, especially in the western US and Alaska. The large inter-model differences in Fig.  are likely due to uncertainty in mineral dust. CanAM5-PAM has a particularly large contribution of dust to PM2.5. The study showed that dust regions were globally one of the largest areas of diversity in PM2.5 concentrations between different CMIP6 models. The EMEP MSC-W results are consistent with EMEP annual evaluations for Europe, where the model underestimates PM2.5 by 10 %–25 %, including a few Arctic sites in Norway and Finland (, and annual model evaluation reports at https://www.emep.int/publ/common_publications.html, last access: 14 April 2022).

The simulated surface-level SLCFs were quite sensitive to the different meteorological conditions such as boundary layer stability and levels of photochemistry, which differed between the two time periods chosen in this study. 2014–2015 was also a time period with more wildfires compared to 2008–2009. For example, according to the CMIP6 emission data that were used in most of the models, emissions of BC from Canadian wildfires in 2014–2015 were 340 % higher than in 2008–2009, whereas the emissions from the USA and Russia were similar for these years. Given the very intense wildfire emissions and low anthropogenic emissions in northern Canada in 2014–2015, differences in simulated PM2.5 concentrations over Canada and Alaska can be partly attributed to differences in simulations of wildfire aerosols in the models.

Some models (CanAM5-PAM, CESM, CIESM-MAM7, GEOS-Chem, and WRF-Chem) simulate higher PM2.5 and more variable PM2.5 in the summertime (e.g., Fig. S13 in the Supplement). While this is seen to some extent in the observations, this may be due to the way fire emissions and sea salt emissions are treated in these models. Fire emissions, fire plume injection height, plume rise, and wet deposition of fire pollutants are all highly uncertain model processes and a subject of ongoing research (e.g., ). Indeed, the individual model PM2.5 Arctic biases are more tightly clustered for 2008–2009 when there were fewer fires. showed that Arctic PM2.5 seasonal pollution is mainly due to local air pollution in the winter and due to fires in the summer.

Figure  shows that the annual mean simulated PM2.5 concentrations compare well with observations and the correlation coefficients are relatively high (R=0.8 or higher for all models). The high concentrations in China and low concentrations in the US and Europe are captured by the models, providing confidence for health impact assessments that utilize these model results.

Recall that the in situ SLCF measurements had the following reported uncertainties: CH4 1 %, O3 3 %, CO 5 %, NO2 5 %–100 %, BC 200 %, SO42- 20 %, OA 20 %, and PM2.5 1 %–6 %. However, since the variability in measurements from different techniques was only really taken into account for the BC uncertainty, and since we are comparing annual mean results to each other, it is not a fair comparison to say that models and measurements agree with each other if model biases are within the reported measurement uncertainty range. However, we do use those numbers as a rough guideline for “good” model performance in the absence of other quantitative criteria.

Some model annual mean biases were within those uncertainty ranges. For example, CMAM, MRI-ESM2, and UKESM1 simulate Arctic CH4 to within 1 %, thus agreeing with the CH4 measurements. However, at midlatitudes, they are all out of range at around +6 %–10 %. MATCH and WRF-Chem simulated midlatitude O3 to within 2 %, but only MATCH-SALSA was within 3 % in the Arctic region. The Arctic NO2 measurements are highly uncertain at around 100 %, so all of the models agreed with Arctic NO2 measurements. However, in the higher-NO2 midlatitude environment, NO2 measurement uncertainty is at the smaller end of the range. OsloCTM and WRF-Chem midlatitude NO2 biases were within 10 %. All models agree with BC measurements in both midlatitudes and the Arctic, as all biases are less than 80 %. CESM, CIESM-MAM7, DEHM, MATCH, UKESM1, and WRF-Chem all simulate midlatitude SO42- to within 20 %. But only CanAM5-PAM and MRI-ESM2 do the same for the Arctic region. OA had some of the largest model biases (Figs.  and ), though ECHAM-SALSA, EMEP-MSC-W, GISS-E2.1, MATCH, and OsloCTM are all within 20 % in the Arctic but none at midlatitudes. Finally, with the small uncertainty in PM2.5, only CIESM-MAM7 in the midlatitudes and GISS-E2.1 in the Arctic agree within 2 %.

To summarize the mmm annual mean performance, it “matches” surface observations in the Arctic for CH4, NO2, BC, SO42-, OA, and PM2.5 – as such, the mmm has the best overall performance for the Arctic. In the midlatitudes the mmm matches surface observations for BC and SO42- only.

Regarding the comparisons of trace gases to the TES, MOPITT, and ACE-FTS satellite measurements (which have roughly 5 %–20 % uncertainty), models agree well. Free-tropospheric distributions of trace gases are somewhat easier to simulate, as common problems like a boundary layer that is too stable or too much deposition do not negatively impact the free-tropospheric SLCF distributions. The variability in the free troposphere is smaller compared to at the surface as well. It is also because of the previously noted difference between the spatial range that remote measurements cover being more akin to the spatial scale of model grid boxes compared to the point measurements from in situ observations.

In the report, these models go on to be used to simulate future emission scenarios, and from those results, the future temperature changes due to these SLCFs are predicted. They are further used to determine future changes to human health due to the changes in SLCFs. Given the model evaluation of this study, we have determined that using the mmm to predict SLCF climate impacts is generally robust. Considering the SLCFs with the greatest radiative impact (CH4, O3, BC, and SO4), the mmm was within ±25 % of the measurements across the Northern Hemisphere. The mmm also performed very well for PM2.5, which is a main component considered for human health impacts. Thus, for climate studies, wherein relatively large regions and time periods are considered, the model performance is sufficient.

The model evaluation in this study brought about results that have been reported in previous publications, and several notable issues remain. Here we recommend some future work that may help improve model performance.

Models simulate too much surface O3 at midlatitudes, and this may be due inadequate treatment of dry deposition and/or not including parameterizations for the shade provided by vegetation that reduces photochemistry, as reported in . That said, MATCH had the smallest midlatitude surface O3 bias without accounting for canopy shading; hence, precursor emissions, vertical mixing, deposition, and O3 chemistry all have a role in model O3 results, and errors in these may sometimes cancel out.

Therefore, we conclude that sensitivity tests for the abovementioned model processes will be important for further understanding and improving model performance for SLCFs. But just as important is having additional Arctic measurements and the continuation of existing Arctic measurements in order to assess the model improvements.