An Improved Representation of Aerosol Mixing State for Air-Quality-Weather Interactions

. We implement a detailed representation of aerosol mixing-state into the GEM-MACH air quality and weather forecast model. Our mixing-state representation includes three categories: one for more-hygroscopic aerosol, one for less-hygroscopic aerosol with a high black carbon (BC) mass fraction, and one for less-hygroscopic aerosol with a low BC mass fraction. This is the first model with a mixing-state representation of this type simulating a continent-scale domain. The more-detailed representation allows us to better resolve two different aspects of aerosol mixing state: differences in hygroscopicity 5 due to aerosol composition, and the amount of absorption enhancement of BC due to non-absorbing coatings. Notably, this three-category representation allows us to account for BC thickly coated with primary organic matter, which enhances the absorption of the BC but has a low hygroscopicity. We compare the results of the three-category representation (1L2B) with a simulation that uses two categories, split by hygroscopicity (HYGRO), and a simulation using the original size-resolved internally mixed assumption (SRIM). We (cid:58)(cid:58)(cid:58)(cid:58)(cid:58)(cid:58)

Estimates of the factor by which the absorption of BC increases due to coatings (the absorption enhancement) vary from 1 (no 35 enhancement) to 4, with the majority of the studies reporting values between 1 and 2.5 (Adachi et al., 2010;Zhang et al., 2008;Khalizov et al., 2009;Cappa et al., 2012;Lack et al., 2012;Liu et al., 2015;Peng et al., 2016;Schnaiter et al., 2005;Wang et al., 2014a;Xu et al., 2018;Zanatta et al., 2018;Zhang et al., 2018b). Differences in experimental methods and regional and seasonal variations in BC coating thickness both likely contribute to this diversity. This absorption enhancement leads to a local heating of the atmosphere and a cooling of the surface, potentially increasing stability and affecting cloud cover and 40 precipitation (Bond et al., 2013;Boucher et al., 2013). We refer the reader to two recent reviews (Stevens and Dastoor, 2019;Riemer et al., 2019) for more details about aerosol mixing state.
Many previous representations of aerosol mixing-state have been implemented in models to predict CCN concentrations and aerosol optical properties : , ::: and ::: we :::::: include : a :::::: partial ::: list :: of :::: these :: in ::::: Table :: 1. These include representing each particle individually 45 (PartMC-MOSAIC, Riemer et al., 2009;Zaveri et al., 2010); multiple mixing-state categories separated by BC mass fraction, including MADRID-BC (Oshima et al., 2009b, a), ATRAS (Matsui et al., 2014;Matsui, 2017), MADE-soot (Riemer et al., 2003;Vogel et al., 2009), MADE-in (Aquila et al., 2011) and MADE-3 (Kaiser et al., 2019; two categories for at least BC and organic carbon based on hygroscopicity, implemented in GEOS-Chem (Bey et al., 2001;Wang et al., 2018Wang et al., , 2014b, the GLObal Model of Aerosol Processes (GLOMAP) in both its bin Spracklen et al., 2005Spracklen et al., , 2011 and 50 modal Bellouin et al., 2013) configurations, GMXe (Pringle et al., 2010), the M3+ module (Wilson et al., 2001), M7 Stier et al., 2005;Vignati et al., 2004;Zhang et al., 2012), MAM4 (Liu et al., 2016), MAM7 (Liu et al., 2012), the Model for Ozone and Related chemical Tracers (MOZART, Emmons et al., 2010) and the Sectional Aerosol module for Large Scale Applications (SALSA, Bergman et al., 2012;Andersson et al., 2015;Kokkola et al., 2008;Tonttila et al., 2017;Kokkola et al., 2018); and representing all aerosol within the same size bin or mode as internally-mixed, 55 including the Canadian Aerosol Module (CanAM, Gong et al., 2006;Moran et al., 2012;Gong et al., 2003Gong et al., , 2015, CHIMERE (Menut et al., 2013), the Community Multiscale Air Quality (CMAQ, Binkowski and Roselle, 2003;Appel et al., 2013;Elleman and Covert, 2009;USEPA, 2017) model, the Modal Aerosol Dynamics module for Europe (MADE, Lauer et al., 2005) and the Modal Aerosol Module with three lognormal modes (MAM3, Liu et al., 2012). We refer the reader to Stevens and Dastoor (2019) for more detail on previous model representations of aerosol mixing state, including mixing-state representations 60 that did not specifically target resolving CCN concentrations and optical properties, such as detailed categorizations based on chemical composition and source-oriented approaches. Previous studies using the model approaches listed above have found that if all aerosol in the same size bin or mode is assumed to be internally-mixed, CCN concentrations will frequently be overestimated by 10-20 % and absorption coefficients of BC will be overestimated by 20-40 % (Stevens and Dastoor, 2019, and containing references). 65 However, it still remains unclear how best to efficiently represent aerosol mixing state in atmospheric models. In this study, we implement a detailed representation of aerosol mixing-state into the Global Environmental Multiscale -Modelling Air quality and CHemistry (GEM-MACH) (Moran et al., 2010) air quality model with online air-quality-weather interactions. We refer to this new configuration of GEM-MACH as GM-MixingState. Our approach was inspired by the results of Ching et al. 70 (2016): We independently account for both changes in hygroscopicity and BC mass fraction, as aerosol hygroscopic properties and optical properties do not necessarily co-vary. The existing air-quality-weather interactions in GEM-MACH include aerosol-radiation interactions and changes in cloud droplet activation based on CCN concentrations (Gong et al., 2015;Majdzadeh et al., 2022). We perform a case study focused on biomass-burning over North America to evaluate GM-MixingState.
We investigate the interactions between the representation of aerosol mixing state and air-quality-weather interactions. 75 The paper is structured as follows: In Sect. 2, we describe the GEM-MACH model and the GM-MixingState configuration, as well as the experiments performed. In Sect. 3, we present our results and analysis. In Sect. 4, we summarize our study and present our conclusions.

95
By default, GEM-MACH uses a size-resolved internally mixed representation of the aerosol population: the aerosol population within each size bin is internally mixed, but the population of aerosol in each size bin is externally mixed with respect to each other size bin. The operational version of GEM-MACH uses two size bins ( :::::: aerosol ::: dry :::::::: diameters : 0-2.5 µm and 2.5-10 µm, Moran et al. (2010)), but for this study we use twelve size bins spanning 10 nm to 10 µm. The 12-bin configuration has 100 been shown to yield results that more closely resemble observations (Akingunola et al., 2018).
For this study, we implemented a more detailed representation of the aerosol mixing state into GEM-MACH. Within each size bin, we separate the aerosol into up to three mixing-state categories based on hygroscopicity and BC mass fraction: 1.

130
To calculate the hygroscopicity of aerosol in the model, we assume that sulphate, nitrate, ammonium, sea-salt, dust, SOA, POA, and BC have κ values of 0.65, 0.65, 0.65, 1.1, 0.03, 0.1, 0.001, and 0, respectively (Ching et al., 2016;Zieger et al., 2017;Koehler et al., 2009). Following the volume Zdanovskii-Stokes-Robinson (ZSR) mixing rule (Petters and Kreidenweis, 2007), we assume that the hygroscopicity of a particle is the volume-weighted average of the component species. We therefore do not account explicitly for coating of insoluble components by soluble components, nor do we consider how particle size or 135 shape may affect the mass fraction of coating material necessary for a particle to be rendered "hydrophilic". However, other studies have shown that neither CCN concentrations (Liu et al., 2016;Lee et al., 2013) nor aerosol effective radiative forcing, either through aerosol-cloud interactions or through aerosol-radiation interactions (Regayre et al., 2018), are sensitive to the threshold amount of soluble material needed to render a particle hydrophilic. However, global burdens of BC and POA, especially in remote regions, have been shown to be sensitive to this parameter (Liu et al., 2012(Liu et al., , 2016. This volume-weighted 140 hygroscopicity is only used to determine the proper mixing-state category for aerosol. It is not used to determine cloud droplet activation. Instead, cloud droplet activation is calculated using the parameterization for sectional models described by . Particle hygroscopicity is calculated separately for each mixing-state category based on molecular weights 145 and ion dissociation, as per eq. 7 from Abdul-Razzak and Ghan (2002). Properties of SOA are assumed to be those of adipic acid; BC, POA and dust are assumed to be insoluble. We assume that aerosol in the lo-κ mixing-state categories does not participate in aqueous chemistry, and is not removed by cloud-to-rain conversion and subsequent wet deposition. It :: We ::::::: discuss ::: this :: in ::::: more ::::: detail :: in :::: the :::::::::: supplement. ::::::: Aerosol :: in ::: the :::: lo-κ ::::::::::: mixing-state ::::::::: categories : is still removed from the atmosphere by below-cloud impaction by rain, as this process is not expected to depend strongly on aerosol composition.

Sensitivity studies
An important question remains regarding the minimum level of complexity required to well represent aerosol-weather feed-205 backs in air quality models. We therefore perform several simulations with diverse representations of the aerosol mixing state.
We consider a configuration with two mixing-state categories, split based on particle hygroscopicity (denoted as representation HYGRO; categories hi-κ and lo-κ). We also consider a mixing-state representation with three mixing-state categories: we use one mixing-state category for all high-hygroscopicity particles and two mixing-state categories for low-hygroscopicity particles, split based on BC mass fraction (high-κ, low-κ_hi-BC, and low-κ_lo-BC). We refer to this representation as 1L2B

210
(one hydrophilic, two hydrophobic). We would not expect any improvement over HYGRO in the representation of the radiative properties of hydrophilic particles, but we would expect that 1L2B would better represent the radiative properties of low-hygroscopicity BC-containing particles. In particular, this representation should better distinguish BC thickly coated with POA from BC that is bare or only thinly coated with POA.

215
In addition to performing simulations with different representations of the aerosol mixing state, we also perform simulations with aerosol effects on meteorology (feedbacks) either permitted or disabled. When feedbacks are disabled, cloud droplet nucleation is independent of aerosol concentrations and aerosol interactions with radiation have no effect on atmospheric temperatures or any other meteorological variables. Cloud droplet nucleation is instead determined following Cohard and Pinty (2010) as a function of updraft velocity, temperature, and pressure assuming a pre-specified CCN concentration that does not 220 vary with space or time. The meteorology in these simulations is independent of the aerosol and gas-phase concentrations.
This allows us to directly attribute any differences in results solely to differences in aerosol processes caused by the differences in the representation of mixing state. We designate the simulations where aerosol effects on meteorology are permitted with the suffix "_feedbacks", as these simulations include feedbacks of changes in aerosol concentrations and properties on the meteorology. 3 Results and analysis

Non-Feedbacks Simulations
We present a summary of the domain-averaged, temporally averaged results from all simulations in Table 2. We will start by discussing differences between simulations with aerosol effects on weather disabled, in order to simplify the analysis. We 230 remind the reader that because the meteorology is identical in these simulations, any differences in results can be attributed solely to differences in aerosol processes caused by the differences in the representation of mixing state.

Aerosol Concentrations
We show the mean concentrations of particulate matter with a diameter smaller than 2.5 microns (PM 2.5 ) in Fig. 1, along with 235 the absolute and relative differences in PM 2.5 concentrations between the HYGRO and SRIM simulations, and we show a similar figure for PM 10 concentrations as Fig. S1. We note that PM 2.5 and PM 10 concentrations are nearly identical :::: very :::::: similar in the HYGRO and 1L2B simulations . :::: (Fig. ::: S3). : We find that spatially and temporally averaged surface PM 2.5 concentrations and PM 10 concentrations increase by 23 : % and 41 : %, respectively, from the SRIM simulation to either the HYGRO or 1L2B simulations. These differences are due mostly to increases in less-hygroscopic species, with concentrations of BC, POA, SOA, 240 and dust being increased in the HYGRO and 1L2B simulations by 16 %, 16 %, 21 : %, and 93 %. The concentrations of morehygroscopic species (NH 4 , NO 3 , SO 4 , and sea-spray aerosol) were increased by 3 : % or less.
These changes in aerosol concentrations are due primarily to changes in aerosol wet deposition. In the HYGRO and 1L2B simulations, all aerosol in the low-κ categories are excluded from wet deposition ::::::: in-cloud ::::::::: scavenging : processes. However, di-245 rect comparison of wet deposition fluxes between simulations is complicated because of the greater aerosol mass concentrations in the HYGRO and 1L2B simulations than in the SRIM simulation. Even though the wet deposition process is less efficient for the same air parcel under the same conditions, local wet deposition fluxes can be greater in the HYGRO and 1L2B simulations due to the greater mass concentrations of aerosol in these simulations. For example, a reduced wet deposition flux close to an emissions source can yield an increased wet deposition flux further downwind, as more aerosol mass will be transported 250 further downwind. We attempt to isolate for these effects by dividing the daily wet deposition flux by the daily mean surface aerosol concentrations, to approximate the wet deposition efficiency. This approach is limited in that cloud uptake of gases also contributes to the wet deposition fluxes, and cloud uptake of aerosol and subsequent wet deposition are not necessarily co-located in space and time with surface aerosol concentrations. However, we expect that the relationships between surface concentrations and wet deposition fluxes are similar enough across simulations for the comparison between simulations to be 255 informative.
We can further control for differences in location and timing between wet deposition and surface concentrations by temporally and spatially averaging both the wet deposition fluxes and the surface concentrations before we divide the former by the 275 latter. We therefore show the spatially and temporally averaged wet deposition fluxes normalized by the spatially and temporally averaged surface concentrations of each species in Table 3. After normalizing by the surface concentrations of aerosol, wet deposition rates of BC, POA, SOA, and dust were reduced in the HYGRO and 1L2B simulations by 27 : %, 40 %, 12 : %, and 10 : %. The normalized wet deposition rates of more-hygroscopic species were reduced by less than 5 %.
280 Table 3. Temporally and spatially averaged wet deposition fluxes normalized by temporally and spatially averaged surface concentrations for each simulation. All units are (mol cm -2 day -1 ) / (µg kg -1 ). In the HYGRO and 1L2B simulations, all aerosol in the low-κ categories are excluded from wet deposition ::::::: in-cloud ::::::::: scavenging : processes. Since the low-κ category is defined as having a κ less than 0.1, this excludes large aerosol with low hygroscopicities from participating in wet deposition ::::::: in-cloud ::::::::: scavenging, even if their large size would allow them activate as droplets despite their low hygroscopicity. In particular, this may cause the wet deposition of dust particles to be underestimated in the HYGRO and 1L2B simulations, while it is likely overestimated in the SRIM simulation. A more detailed treatment of 285 cloud uptake of aerosol is beyond the scope of this study, but will be revisited in a future version of GEM-MACH.
The SRIM and 1L2B simulations therefore compare similarly well to observations for these species. This is expected, as these 305 species are more weakly affected by the difference in mixing-state representation. There is an existing high bias in the SRIMpredicted concentrations of BC, organic aerosol, and dust. This high bias is worsened in the 1L2B simulation, due to the slower removal of these species by wet deposition in the 1L2B simulation, and this affects the calculated NMB, RMSE, and Fac2 values for these species. The correlation coefficients for EC and organic aerosol are not strongly affected. This suggests that the variability in BC and organic aerosol concentrations is not primarily controlled by wet deposition at these sites during the 310 case study time period. As discussed, the wet deposition of dust is likely reduced too much in the 1L2B simulation, which may be responsible for the lower correlation between the observed and simulated dust concentrations in the 1L2B simulation. There is a slight shift of the sea salt size distribution to larger sizes in the 1L2B simulation, perhaps due to more coagulation with the larger concentrations EC, organic aerosol, and dust. This reduces the fine sea salt aerosol mass, even while total sea salt aerosol concentrations slightly increase. The NMB and RMSE for sea salt is therefore reduced in the 1L2B simulation compared to 315 the SRIM simulation. The increased concentrations of PM 2.5 and PM 10 in the 1L2B simulation increase the already high bias in PM 2.5 and reduces the underprediction of PM 10 , as compared to the SRIM simulation. However, in both cases the RMSE is reduced, and the R and Fac2 values are either unchanged or slightly improved.

320
The Air Quality Health Index (AQHI; Stieb et al., 2008) is used by Environment and Climate Change Canada to communicate adverse health risks due to poor air quality to Canadians. It is formulated as a scale that ranges from 0 (excellent air quality) to 10 (very poor air quality), and is calculated based on the concentrations of PM 2.5 or PM 10 , ozone (O 3 ), and nitrogen dioxide (NO 2 ). While the equations for calculating the AQHI permit values greater than 10 under exceptionally high concentrations of PM 2.5 , PM 10 , O 3 , or NO 2 , we restrict the values of AQHI to a maximum of 10, both because this is the intended range of the 325 AQHI, and to reduce the influence of exceptional, highly concentrated plumes in uninhabited areas (such as those from forest fires) on our results.
We note that the 1L2B simulation is better able to capture regions with large BC mass fractions than the HYGRO simulation because the HYGRO configuration assumes that all low-hygroscopicity species within the same size bin are internally mixed, including BC, dust, and POA. Most dust mass exists in larger size bins than BC. Therefore, even the SRIM simulation does not 370 assume much internal mixing of BC and dust. However, BC and POA are emitted into the same size bins and from the same source regions. When the BC is assumed to be internally mixed with other low-hygroscopicity species, the resulting particles frequently consist of BC thickly coated with POA. The 1L2B simulation is able to distinguish BC thinly coated with POA from BC thickly coated with POA, and it predicts that a large proportion of BC near source regions has only a thin coating of 1L2B.
non-BC species.

Aerosol-radiation interactions
We show the monthly mean AOD from the SRIM simulation and the difference between the HYGRO and SRIM simulations in Fig. 6. We remind the reader that the calculations of aerosol optical properties are restricted to daylight hours in GEM-MACH.
As such, we include only data from between 1300 and 2100 UTC in Fig. 6, in order to exclude times of day when the AOD was 380 not calculated for some part of the domain shown. We also note that the mean AOD in the 1L2B and HYGRO simulations differs by no more than 0.0011 :: 0.9 :: % : for any grid cell in the domain, :: as :::::: shown :: in :::: Fig. :: S8. When using the HYGRO configuration, the AOD is 34 % larger than in the SRIM case. A comparison of the AOD with the absorption aerosol optical depth (AAOD) (see Table 2 and Fig. 7) reveals that the AOD is dominated by aerosol scattering, rather than aerosol absorption. Previous studies have found that the optical properties of non-absorbing aerosol is :: are : not strongly sensitive to the mixing-state of the 385 aerosol (e.g. Zaveri et al., 2010;Klingmüller et al., 2014), and that because AOD is dominated by the scattering component, ambient AOD is not strongly sensitive to mixing-state (e.g. Matsui et al., 2013Matsui et al., , 2014Klingmüller et al., 2014;Han et al., 2013), although a recent study has shown that aerosol scattering can be very sensitive to aerosol mixing-state under certain conditions (Yao et al., 2022). We therefore do not expect our more-detailed representation of the BC mass to yield strong changes in aerosol scattering, but we do expect a decrease in aerosol absorption. We therefore conclude that the differences 390 are due predominantly to the increases in aerosol mass, in turn due to the decrease in aerosol wet deposition. This is supported by the fact that the aerosol AOD and differences in AOD are visibly well-correlated with PM 2.5 and the differences in PM 2.5 Figure 6. left: mean AOD from the SRIM simulation; top right :::: centre: mean difference in AOD between the HYGRO and SRIM simulations; bottom right: relative difference in mean AOD between the HYGRO and SRIM simulations. Only results from the hours of 1300-2100 UTC are included as the AOD is only calculated during local daylight hours.
We show the monthly mean AAOD from the SRIM simulation and the differences between the 1L2B, HYGRO and SRIM 395 simulations in Fig. 7. The AAOD is 39 : % higher in the HYGRO case than in the SRIM case. As shown in Fig. 5, the BC mass fraction in BC-containing particles is only slightly larger in the HYGRO case than the SRIM cases. If the mass concentrations of all aerosol species were equal in both cases, higher BC mass fractions would imply thinner coatings and smaller absorption enhancements for the BC-containing particles. This effect would be expected to reduce the AAOD in the HYGRO case as compared to the SRIM case. The simulated increase in AAOD is due primarily to the increased concentrations of BC in the 400 HYGRO case compared to the SRIM case.

Aerosol-Meteorology Feedbacks
In order to examine the interactions between aerosol mixing-state representation and meteorology, we will now describe the results of the aerosol-meteorology feedbacks simulations. In these simulations, the cloud droplet number concentration is parameterized based on the aerosol size distribution using Abdul-Razzak and Ghan (2002), as described in Sect. 2 ::: and ::: the 420 ::::::::: supplement. In the case of multiple mixing-state categories, the distinct composition of aerosol in each mixing-state category is considered, so aerosol in different mixing-state categories will have different critical radii for activation under the same atmospheric conditions. Additionally, aerosol and trace gas concentrations are permitted to reduce incoming radiation, which would subsequently alter atmospheric and surface energy balances.

425
As our focus is on the effects of differences in the aerosol mixing-state representation, we only compare cases with aerosolmeteorology feedbacks to other cases with aerosol-meteorology feedbacks. For comparisons of GEM-MACH results with and without aerosol-meteorology feedbacks, we refer the reader to Gong et al. (2015) and Makar et al. (2015a, b).

Aerosol-cloud interactions
In order to target low clouds most likely to be affected by aerosol emitted from the surface, we restrict our analysis to the 430 clouds with model hybrid levels between 0.807 and 0.962, approximately 35-185 hPa below surface pressure. As all cloud variables were saved as 3-hourly means, which will include transitions between cloudy and cloud-free periods, our reported cloud properties will have smaller values than if we had analyzed instantaneous model output. This includes, most notably, the cloud droplet and raindrop number mixing ratios. However, as our interest is in the comparison between simulations, which are all treated identically, this would not alter our conclusions. Additionally, in order to provide more physically meaningful values, 435 when calculating temporally and horizontally averaged cloud properties we define "cloudy" grid cells as those with 3-hourly cloud water mixing ratios (Q C ) >0.005g kg -1 , and we filter out grid cells with lower 3-hourly Q C values. The mean number of cloudy grid cells differs by less than 0.7 : % between simulations (not shown), with the HYGRO_feedbacks and 1L2B_feedbacks simulations having slightly more cloudy grid cells than the SRIM_feedbacks simulation. Therefore differences between simulations are better explained as changes in in-cloud properties, rather than as changes in the spatial extent of clouds. We note 440 that the cloud fraction over the western United States was low during July of 2016, as evidenced in the MODIS satellite retrievals (NASA Earth Observations, https://neo.gsfc.nasa.gov/view.php?datasetId=MODAL2_M_CLD_FR&date=2016-07-28, last access November 19th, 2021).
We show in Fig. 8 the vertical distributions of the temporally and horizontally meaned in-cloud cloud droplet number mix-445 ing ratios (N C ), cloud water mixing ratios (Q C ), rain drop number mixing ratios (N R ), and rain water mixing ratios (Q R ) in the SRIM_feedbacks, HYGRO_feedbacks, and 1L2B_feedbacks simulations. The HYGRO_feedbacks and 1L2B_feedbacks model simulations predict N C values that are approximately 15 % larger than in the SRIM_feedbacks simulation. The difference in N C is approximately constant with altitude. This difference is due to increased aerosol number concentrations, in turn due to both greater aerosol mass concentrations and smaller aerosol diameter, as shown in Fig. 9. These increased N C values 450 lead to mean Q C values that are about 7 : % greater than in the SRIM_feedbacks simulation. As N C increases more than Q C , the mean cloud droplet size will be decreased in the HYGRO_feedbacks and 1L2B_feedbacks simulations. These reduced cloud droplet sizes would be expected to result in reduced autoconversion and slower drizzle formation. Indeed, both N R and Q R are reduced in the HYGRO_feedbacks and 1L2B_feedbacks simulations relative to the SRIM_feedbacks simulation, by about 20 : % for N R and 9 : % for Q R . The difference in N C is approximately constant with altitude, while the differences in Q C , N R 455 and Q R increase with altitude. For all cloud variables, the differences are slightly larger in the 1L2B_feedbacks simulation compared to the HYGRO_feedbacks simulation.
The decreases in in-cloud Q R discussed above would be expected to result in decreases in precipitation at the surface. We show the mean precipitation from the SRIM simulation and the effects on precipitation of the HYGRO and 1L2B mixing-460 state representations in Fig. 10. Many of the differences shown in Fig. 10 include large decreases near large increases. These are due in part to small changes in advection patterns, which subsequently alter the locations of precipitation. We can determine the net effect of the difference in mixing-state representation on surface precipitation by averaging across the domain.
When spatially and temporally averaged, the effects ::::: effect :: of : mixing-state representation on precipitation is modest: In the HYGRO_feedbacks and 1L2B_feedbacks simulations, the precipitation is reduced by 0.6 : % relative to the SRIM_feedbacks 465 simulation, much smaller than the differences in in-cloud Q R discussed above. As the decreases in N R are greater than those in Q R , the HYGRO_feedbacks and 1L2B_feedbacks simulations would have larger rain drops than the SRIM_feedbacks simulation, and these larger rain drops would settle to the surface more efficiently, thereby partially offsetting the reduction in Q R .
The increases in N C and Q C shown above, along with the small increases in AOD shown in Sect. 3.1.5, would be expected 470 to reduce the shortwave radiation reaching the surface and to potentially reduce surface temperatures. We show in Fig. 11 the differences in mean surface temperatures between the HYGRO_feedbacks and SRIM_feedbacks simulations and between the 1L2B_feedbacks and HYGRO_feedbacks simulations. Between HYGRO_feedbacks and SRIM_feedbacks, eastern and southern North America shows either small differences or noisy differences that would be consistent with slight changes in the locations of clouds. However, there is a clear increase of about 0.01 K over large areas of the oceans and a clear decrease of 475 about 0.06 K over northern Quebec and eastern Nunavut. We note that this region encompasses the outflow of forest fires that occurred in north-eastern Canada during the simulation, as is visible in the differences in surface BC concentrations :::::: mixing :::: ratios : (Fig. 4). In the HYGRO_feedbacks simulation, the emissions from these forest fires are removed more slowly by wet deposition. Therefore, more aerosol particles remain to act as CCN further downwind from the source. The greater CCN concentration increases both N C and Q C within the cloud, reducing the solar radiation reaching the surface, and reducing surface   the domain, consistent with slight changes in the locations of clouds. We therefore cannot determine any clear effect on surface temperatures due to differences in mixing-state representation between these two simulations.
We note that for all cloud properties, there are only small differences between 1L2B_feedbacks and HYGRO_feedbacks. 485 We remind the reader that the differences in mixing-state representation between 1L2B and HYGRO were designed to capture the effects of correctly resolving the thickness of non-absorbing shells on BC and the subsequent enhancement in aerosol absorption. The effects of these differences in absorption enhancement would be permitted to affect atmospheric temperatures in our simulations, with potential subsequent effects on atmospheric stability. However, we do not find a strong effect on cloud properties, surface precipitation or surface temperatures in this study. This may be, in part, due to our choice of case study.

Conclusions
In this study, we have implemented a detailed representation of aerosol mixing-state into the GEM-MACH air quality and weather forecast model. Our mixing-state representation includes three categories: one for more-hygroscopic aerosol, one for less-hygroscopic aerosol with a high BC mass fraction, and one for less-hygroscopic aerosol with a low BC mass fraction. This 495 is the first model with a mixing-state representation of this type simulating a continent-scale domain. Currently, the HYGRO and 1L2B configurations require approximately 70 : % and 150 : % more running-time, respectively, than the SRIM configuration. We expect to reduce this additional cost through improvements to the efficiency of the model tracer transport scheme in the near future. The more-detailed representation allowed us to better resolve two different aspects of aerosol mixing state: First, differences in hygroscopicity due to differences in aerosol composition, including the change in hygroscopicity with time 500 as less-hygroscopic aerosol becomes coated with hydrophilic material. Second, the thickness of non-absorbing coatings on BC aerosol which enhance the absorption of the BC aerosol.
We compared the results of the three-category representation (1L2B) with a simulation that uses two categories, split by hygroscopicity (HYGRO), and a simulation using the original size-resolved internally-mixed assumption (SRIM). We showed 505 that when we included one or two categories of less-hygroscopic aerosol, wet deposition of BC, POA, SOA and dust was reduced, yielding increases in the mean concentrations of these species of 16-93 : %, and an increase in the mean PM 2.5 concentration by 23 : %. The effect on dust concentrations is likely overestimated, as the current implementation prevents wet deposition ::::::: in-cloud ::::::::: scavenging : of aerosol in the hydrophobic category, even if the aerosol is large. We intend to improve on this in a future version of GEM-MACH. As BC, POA, and SOA mass is more concentrated in smaller aerosol particles, we 510 believe that the reductions of wet deposition in these species is realistic. The increased PM 2.5 concentrations led to an increase in the AQHI 2.5 by 0.05 units on average. The increases in aerosol concentrations also led to increases in both AOD and AAOD.
We briefly compared the results of the SRIM and 1L2B simulations and observations from the IMPROVE, CSN and AQS networks. However, we did not find significant improvement in model-observation agreement with the more-detailed mixing-515 state representation. The reduced wet deposition worsened an existing high bias in BC, organic matter, and dust concentrations, and we saw only small changes in correlation with the observations. It is likely that a more thorough assessment will require observations from sites that are strongly affected by long-range transport of BC and organic aerosol. The CSN network sites in particular are located in urban centres, and would therefore be expected to be weakly affected by changes in wet deposition.
We will investigate this further in future work.

520
However, using two categories to resolve more-hygroscopic and less-hygroscopic aerosol only yielded modest improvements in resolving the amount of coating material on BC particles, which alters their absorption of solar radiation. We found that using three mixing-state categories (more-hygroscopic, less-hygroscopic high BC mass fraction, less-hygroscopic low BC mass fraction) allowed us to distinguish thinly coated BC from BC that was thickly coated with POA. This yielded a mean 525 AAOD that was 3 : % less than when separating the aerosol by hygroscopicity alone. Many sources of BC are also sources of POA, and observations indicate that the BC-containing particles frequently also contain POA, even close to emission sources (Perring et al., 2017;Kondo et al., 2011). We note that we assumed that particles from area sources were externally mixed at emission. This assumption will yield a maximum difference between our sensitivity simulations. Nonetheless, as thinly coated BC particles have been observed in the ambient atmosphere, even far from emission sources (Zanatta et al., 2018;Sharma 530 et al., 2017), it is clear that POA and BC are not evenly distributed across particles in the same size range. The proportion of POA that is emitted as BC-containing particles vs. BC-free particles is currently poorly constrained. We therefore suggest that future observation campaigns record not only the coating thickness on BC-containing particles, but also, when possible, the proportion of organic matter that exists as BC-free particles vs. BC-containing particles. 535 We then performed simulations that included aerosol feedbacks on meteorology in order to determine the effects of mixingstate representation on the forecast meteorology. We found a clear effect due to including two categories of aerosol hygroscopicity: the increased aerosol concentrations due to the decreases in wet deposition increased cloud droplet mixing ratios by approximately 15 %. This led to a reduction in the mean precipitation by 0.6 : %. The increased cloud reflectivity resulted in a decrease in surface temperatures by about 0.06 K over northeastern Canada, in the outflow of large forest fires. When 540 we compared the results of the HYGRO simulation with those of the IL2B ::::: 1L2B simulation, which better resolves BC mass fraction and aerosol absorption, we did not find a strong effect on forecast meteorology. Competing interests. We declare that no competing interests are present.