Limitations in representation of physical processes prevent successful simulation of PM 2 . 5 during KORUS-AQ

High levels of fine particulate matter (PM2.5) pollution in East Asia often exceed local air quality standards. Observations from the Korea–United States Air Quality (KORUS-AQ) field campaign in May and June 2016 showed that development of extreme pollution (haze) occurred through a combination of longrange transport and favorable meteorological conditions that enhanced local production of PM2.5. Atmospheric models often have difficulty simulating PM2.5 chemical composition during haze, which is of concern for the development of successful control measures. We use observations from KORUS-AQ to examine the ability of the GEOS-Chem chemical transport model to simulate PM2.5 composition throughout the campaign and identify the mechanisms driving the pollution event. At the surface, the model underestimates sulfate by −64 % but overestimates nitrate by+36 %. The largest underestimate in sulfate occurs during the pollution event, for which models typically struggle to generate elevated sulfate concentrations due to missing heterogeneous chemistry in aerosol liquid water in the polluted boundary layer. Hourly surface observations show that the model nitrate bias is driven by an overestimation of the nighttime peak. In the model, nitrate formation is limited by the supply of Published by Copernicus Publications on behalf of the European Geosciences Union. 7934 K. R. Travis et al.: Limitations in simulation of PM2.5 during KORUS-AQ nitric acid, which is biased by +100 % against aircraft observations. We hypothesize that this is due to a large missing sink, which we implement here as a factor of 5 increase in dry deposition. We show that the resulting increased deposition velocity is consistent with observations of total nitrate as a function of photochemical age. The model does not account for factors such as the urban heat island effect or the heterogeneity of the built-up urban landscape, resulting in insufficient model turbulence and surface area over the study area that likely results in insufficient dry deposition. Other species such as NH3 could be similarly affected but were not measured during the campaign. Nighttime production of nitrate is driven by NO2 hydrolysis in the model, while observations show that unexpectedly elevated nighttime ozone (not present in the model) should result in N2O5 hydrolysis as the primary pathway. The model is unable to represent nighttime ozone due to an overly rapid collapse of the afternoon mixed layer and excessive titration by NO. We attribute this to missing nighttime heating driving deeper nocturnal mixing that would be expected to occur in a city like Seoul. This urban heating is not considered in air quality models run at large enough scales to treat both local chemistry and long-range transport. Key model failures in simulating nitrate, mainly overestimated daytime nitric acid, incorrect representation of nighttime chemistry, and an overly shallow and insufficiently turbulent nighttime mixed layer, exacerbate the model’s inability to simulate the buildup of PM2.5 during haze pollution. To address the underestimate in sulfate most evident during the haze event, heterogeneous aerosol uptake of SO2 is added to the model, which previously only considered aqueous production of sulfate from SO2 in cloud water. Implementing a simple parameterization of this chemistry improves the model abundance of sulfate but degrades the SO2 simulation, implying that emissions are underestimated. We find that improving model simulations of sulfate has direct relevance to determining local vs. transboundary contributions to PM2.5. During the haze pollution event, the inclusion of heterogeneous aerosol uptake of SO2 decreases the fraction of PM2.5 attributable to long-range transport from 66 % to 54 %. Locally produced sulfate increased from 1 % to 25 % of locally produced PM2.5, implying that local emissions controls could have a larger effect than previously thought. However, this additional uptake of SO2 is coupled to the model nitrate prediction, which affects the aerosol liquid water abundance and chemistry driving sulfate–nitrate–ammonium partitioning. An additional simulation of the haze pollution with heterogeneous uptake of SO2 to aerosol and simple improvements to the model nitrate simulation results in 30 % less sulfate due to 40 % less nitrate and aerosol water, and this results in an underestimate of sulfate during the haze event. Future studies need to better consider the impact of model physical processes such as dry deposition and nighttime boundary layer mixing on the simulation of nitrate and the effect of improved nitrate simulations on the overall simulation of secondary inorganic aerosol (sulfate+ nitrate+ ammonium) in East Asia. Foreign emissions are rapidly changing, increasing the need to understand the impact of local emissions on PM2.5 in South Korea to ensure continued air quality improvements.

heterogeneous chemistry in aerosol liquid water in the polluted boundary layer. Hourly surface observations show that the model nitrate bias is driven by an overestimation of the nighttime peak. In the model, nitrate formation is limited by the supply of nitric acid, which is biased by +100% against aircraft observations. We hypothesize that this is due to a missing sink, which 40 we implement here as a factor of five increase in dry deposition. We show that the resulting increased deposition velocity is consistent with observations of total nitrate as a function of photochemical age. The model does not account for factors such as the urban heat island effect or the heterogeneity of the built-up urban landscape resulting in insufficient model turbulence and surface area over the study area that likely results in insufficient dry deposition. Other species such as NH3 could be similarly affected but were not measured during the campaign. Nighttime production of nitrate is driven by NO2 hydrolysis in 45 the model, while observations show that unexpectedly elevated nighttime ozone (not present in the model) should result in N2O5 hydrolysis as the primary pathway. The model is unable to represent nighttime ozone due to an overly rapid collapse of the afternoon mixed layer and excessive titration by NO. We attribute this to missing nighttime heating driving deeper nocturnal mixing that would be expected to occur in a city like Seoul. This urban heating is not considered in air quality models run at large enough scales to treat both local chemistry and long-range transport. Key model failures in simulating nitrate, 50 mainly overestimated daytime nitric acid, incorrect representation of nighttime chemistry, and an overly shallow and insufficiently turbulent nighttime mixed layer, exacerbate the model's inability to simulate the buildup of PM2.5 during haze pollution. To address the underestimate in sulfate most evident during the haze event, heterogeneous aerosol uptake of SO2 is added to the model which previously only considered aqueous production of sulfate from SO2 in cloud water. Implementing a simple parameterization of this chemistry improves the model abundance of sulfate but degrades the SO2 simulation implying 55 that emissions are underestimated. We find that improving model simulations of sulfate has direct relevance to determining local vs. transboundary contributions to PM2.5. During the haze pollution event, the inclusion of heterogeneous aerosol uptake of SO2 decreases the fraction of PM2.5 attributable to long-range transport from 66% to 54%. Locally-produced sulfate increased from 1% to 46% of locally-produced PM2.5, implying that local emissions controls would have a larger effect than previously thought. However, this additional uptake of SO2 is coupled to the model nitrate prediction which affects the aerosol 60 liquid water abundance and chemistry driving sulfate-nitrate-ammonium partitioning. An additional simulation of the haze pollution with heterogeneous uptake of SO2 to aerosol and simple improvements to the model nitrate simulation results in 30% less sulfate due to 40% less nitrate and aerosol water, and results in an underestimate of sulfate during the haze event. Future studies need to better consider the impact of model physical processes such as dry deposition and boundary layer mixing on the simulation of nitrate and the effect of improved nitrate simulations on the overall simulation of secondary inorganic aerosol 65 (sulfate+nitrate+ammonium) in East Asia. Foreign emissions are rapidly changing, increasing the need to understand the impact of local emissions on PM2.5 in South Korea to ensure continued air quality improvements. emphasizes the urgent need to improve our understanding of the sources and conditions driving haze events.

95
Improving model representation of sulfate chemistry cannot be considered entirely separately from model nitrate biases. In the atmosphere, aqueous-phase chemistry is a major source of sulfate, where clouds provide the dominant source of liquid water (Herrmann et al., 2015). Recent studies have hypothesized that the high aerosol liquid water content (ALWC) associated with PM2.5 during extreme pollution events in East Asia allows for significant sulfate production not considered in most models (Wang et al., 2014;Zheng et al., 2015aZheng et al., , 2015bShao et al., 2019). Levels of ALWC are very sensitive to aerosol nitrate (Ge 100 https://doi.org/10.5194/acp-2021-946 Preprint. Discussion started: 7 January 2022 c Author(s) 2022. CC BY 4.0 License.
Heterogeneous aerosol uptake of HO2 produces H2O2 (Mao et al., 2013), with a reactive uptake coefficient ( ) of 0.2 (Jacob, 2000). We implement aromatic chemistry from Yan et al. (2019) for the simulation of KORUS-AQ. We use the model "simple scheme" for organic aerosol (OA) where OA is generated using fixed empirically derived yields 170 from isoprene, monoterpenes, biomass burning, and anthropogenic fuel combustion (Pai et al., 2020). This scheme includes an emitted hydrophobic component (OCPO) with an assumed organic-mass-to-organic carbon (OM:OC) ratio of 1.4 that is aged to a hydrophilic oxygenated component (OCPI) with an OM:OC ratio of 2.1. Secondary organic aerosol (SOA) is a lumped product (SOAS) with a molecular weight of 150 g mol -1 . For comparison to observations, primary organic aerosol (POA) is defined as OCPO and SOA is the sum of OCPI and SOAS. The sulfate-nitrate-ammonium (SNA) aerosol simulation 175 (Park, 2004) includes the addition of metal-catalyzed oxidation of SO2 (Alexander et al., 2009), sulfur oxidation by reactive halogens (Chen et al., 2017), and improved implementation of aerosol cloud-processing and revised uptake coefficients for NO2 (Holmes et al., 2019). Uptake of N2O5 on SNA includes dependence on aerosol water, organic coatings, nitrate aerosol fraction, and particulate chloride (McDuffie et al., 2018). SNA partitioning is calculated with ISORROPIA v2.2 (Pye et al., 2009). The model includes accumulation mode (SALA) and coarse mode (SALC) sea salt aerosol (Alexander et al., 2005;180 Jaeglé et al., 2011) and dust in four size bins (DST1 to 4) (Fairlie et al., 2010), where the first bin and 38% of the second bin are included in PM2.5. The recommended definition of dry PM2.5 is given by Eq 1. 185 The AirKorea PM2.5 observations provided by NIER are obtained using the beta-ray attenuation method (BAM-1020, Table   1). We do not adjust modeled PM2.5 for any measurement relative humidity effects as the BAM-1020 has been shown to perform well against federal reference method monitors (Le et al., 2020).
Specific details of production of model nitric acid (HNO3), the gas-phase precursor to aerosol nitrate ( & % = pNO3), are 190 provided below as KORUS-AQ provides detailed observations of this chemistry. Reactions R1-R6 describe model production of HNO3 from oxidation of NO2 (R1), aqueous uptake and reaction of N2O5, NO2, and NO3 on aerosol (R2, R4, R5), aqueous uptake and reaction of N2O5 and NO3 in cloud water (R3, R5), heterogeneous halogen chemistry (Table S2), and oxidation of VOCs by the nitrate radical (R6). In R2, aqueous uptake and reaction of N2O5 with particle chloride (Cl -) produces nitryl chloride (ClNO2) with a yield (∅) of 1 on sea salt aerosol and zero on all other aerosol types. 195 4 Simulation of PM2.5 during KORUS-AQ Figure 1a shows the model simulation of daily average PM2.5 (Eq. 1) compared to the observed average of the 15 AirKorea sites within the GEOS-Chem grid box containing the major SMA monitoring sites (KIST and Olympic Park). These two sites are in close proximity to the AirKorea monitors ( Fig. 1b). Campaign average PM2.5 is 29 µg m -3 , but this increases to 53 µg 205 m -3 during the Transport/Haze period. The model reproduces the low PM2.5 during the Dynamic period, the increase during the Transport/Haze period, and the variable concentrations during the Blocking period. Across the campaign, the model underestimates PM2.5 (NMB = -15%) due to a low bias during the Stagnant period and the initial build-up during the Transport/Haze period. This model performance is similar to Choi et al. (2019) using a different GEOS-Chem configuration.  (Table 1), representative of PM1 (Guo et al., 2021), is used to speciate daily average PM2.5 from the AirKorea sites (Fig. 1a). Jordan et al. (2020) showed that speciated PM1 was generally representative of PM2.5 mass throughout KORUS-AQ, except for the Transport/Haze period when PM2.5 significantly exceeded PM1. The strong correlation between PM2.5 and PM1 during the campaign implied growth of PM1 to larger sizes. Dust is not a major component of PM2.5 at the surface after May 9 th , as further discussed in Section S1. Therefore PM1 composition likely 215 represents the composition of PM2.5 with the exception of a small contribution from primary aerosol species. Sun et al. (2020) showed that PM2.5 can be up to 50% greater than PM1 in polluted, humid environments and the mass at sizes >PM1 is secondary (not BC or POA). We remove BC and POA from observed PM2.5 and scale the remaining components (SNA, SOA) to the remaining PM2.5. The resulting speciated PM2.5, derived from KIST PM1 composition and AirKorea PM2.5 mass, is provided for each meteorological period in Table 3. Figure 2 and Table 3 include the ALWC associated with PM2.5, calculated for the 220 observations using the E-AIM IV thermodynamic model (Clegg and Brimblecombe, 1990;Clegg et al., 1998;Massucci et al., 1999;Wexler and Clegg, 2002), and ISORROPIAv2.2 (Pye et al., 2009) in GEOS-Chem. During KORUS-AQ, Kim et al. (2022) found that ISORROPIAv2.2 provided similar results as the E-AIM model, reproducing E-AIM pH within ~0.4 units.
The primary campaign average model biases are underestimated sulfate (-64%), overestimated nitrate (+36%), and underestimated SOA (-43%). The excess model nitrate is the primary driver of overestimated ALWC (+82%). During the 225 Stagnant period, the model low bias is due to underestimated SOA (-9 µg m -3 ). This may be due to missing local production from emissions of semi-and intermediate-volatility volatile organic compounds (S/IVOCs, McDonald et al., 2018) and aromatics (Nault et al., 2018), primarily attributable to solvents and vehicle emissions (Shin et al., 2013a(Shin et al., , 2013bSimpson et al., 2020). During the Dynamic and Blocking periods, the model PM2.5 bias is within 20% of the observations but with overestimated nitrate and underestimated sulfate. The model severely underestimates sulfate during the Transport/Haze period 230 (-11 µg m -3 , Table 3) suggesting that the model fails to reproduce the processes driving the pollution episode. As described by Jordan et al. (2020), both ground and aircraft observations during KORUS-AQ showed that cloudy and humid conditions during the Transport/Haze period increased PM2.5 through heterogeneous production of SNA.
The KORUS-AQ aircraft observations included detailed daytime (available from ~8am to 4pm KST) aerosol and gas-phase observations that we use to determine the cause of model sulfate and nitrate biases and their regional extent. Model SOA biases 235 will be the subject of future work as here they do not contribute to PM2.5 exceedances (50 µg m -3 daily average in 2016). The KORUS-AQ campaign included frequent sampling along a repeated flight pattern or "stereoroute" over the SMA up to three times a day, supplemented by less frequent flights to investigate specific source regions or transport events . Figure S2 shows the high data density in the SMA compared to the rest of the study region. We use the 55 descents over Olympic Park from the SMA stereoroute to compare against the daily surface observations shown in Fig. 2. 240 If the model RH simulation was unbiased, we would expect an improved simulation of nitrate as the minimal RH bias during 250 the Dynamic period corresponds to the best nitrate simulation (Fig. 3, Fig. S3). Model aerosol dry deposition may also be too fast but this effect would increase model concentrations by only ~10% (Emerson et al., 2020).
There is no available measurement of PM2.5 from the aircraft to provide a similar scaling from PM1 to PM2.5 as was done in Fig. 2. However, any increase to the observed profiles of PM1 sulfate or nitrate to account for possible growth to larger sizes would exacerbate the model underestimate of these species. The discrepancy between the model low to minimal bias against 255 daytime aircraft nitrate observations (Fig. 3) and the overestimate against daily average nitrate at the KIST ground site (Fig.  2) implies a failure of the model to represent nighttime chemical production. We investigate the possible causes of overestimated daily average model nitrate in Section 5 and underestimated model sulfate in Section 6.

Model errors representing the nitrate diurnal cycle
The discrepancy between the model daytime and daily performance for nitrate demonstrates the need to compare the model 260 mean nitrate diurnal cycle against the nitrate fraction of PM2.5 derived from KIST observations as described in Section 4. Figure 4a shows that between 6am and 6pm KST (daytime) the model bias minimal (< -1 µg m -3 ) while the bias from 6pm to 6am KST (nighttime) is +3 µg m -3 . As described in Section 3, the model has a newly revised treatment of wet scavenging that significantly reduces the model nitrate and nitric acid biases present in previous model versions (Luo et al., 2019). Without this improvement, the model would have an average nighttime bias of +7 µg m -3 . Figure S4 shows daily precipitation in Seoul 265 from the Korea Meteorological Administration (KMA, 2021) which is infrequent and negligible in the later part of the campaign. The model underestimate in total precipitation across the campaign is minimal (121 vs. 112 mm). Insufficient wet scavenging is unlikely to be the cause of the remaining model nitrate bias.
We perform a sensitivity test to determine the relative impact of daytime (R1) vs. nighttime (R2-R5, Section 3) production of HNO3 on the model bias by shutting off the nighttime reactions. Figure 4c shows that the main model nighttime pathway is 270 aerosol uptake of NO2 (R4) with a small contribution from N2O5 hydrolysis (R2/3) in the early morning hours. Figure 4a shows that removing nighttime chemistry results in improved early morning agreement (1am to 8am KST) but the evening overestimate (8pm to 1am KST) is less affected. Jordan et al. (2020) showed observational evidence for significant nighttime production of nitrate by N2O5 hydrolysis (R2). We use the removal of nighttime chemistry to hypothesize that part of the model nighttime bias is due to excess daytime HNO3 that has not yet been lost to deposition and is converted to nitrate as conditions 275 become thermodynamically favorable for partitioning to the aerosol-phase. The dominance of NO2 uptake over N2O5 hydrolysis in the model suggests that there are additional errors in simulated nighttime chemistry.

Sensitivity of model nitrate bias to gas-phase precursors
Inorganic aerosol ammonium nitrate (NH4NO3) is formed by dissolution of HNO3, which reacts in the aqueous phase with ammonia (NH3) to establish a thermal equilibrium with NH4NO3. The conditions that favor NH4NO3 are generally cool and 280 humid (i.e., nighttime) and characterized by high NH3 and HNO3 concentrations relative to sulfate (Guo et al., 2016). We calculate that average nighttime RH (temperature) in the SMA is 74% (290K) compared to the model value of 71% (288K), indicating that significant errors in RH or temperature are not the cause of nighttime biases. Overproduction of model nighttime nitrate could be due to overestimated NH3 if this species limits NH4NO3 production. In South Korea, and generally East Asia, NH4NO3 is limited by availability of HNO3. This due to high levels of NH3 (~10 ppb) observed in East Asia, attributable to 285 non-agricultural sources such as transportation (Song et al., 2009;Phan et al., 2013;Link et al., 2017;   . 290 Few datasets exist to further test the performance of HNO3-pNO3 partitioning in the model but KORUS-AQ observations provide this opportunity. This partitioning is described by Eq. 2, where the ratio of pNO3 to total nitrate (TNO3 = HNO3 and pNO3), known as εNO3, is impacted by temperature, relative humidity, and aerosol composition (Guo et al., 2016(Guo et al., , 2017. Accurate simulation of εNO3 is critical to regulating the deposition of TNO3 as HNO3 deposits more rapidly than pNO3 (Nenes et al., 2021). Figure 5 shows ɛNO3 as a function of RH for the observations and the model for the same domain as Fig. 3 below 1.5 km. While the model represents the increase of ɛNO3 with RH, model ɛNO3 is generally underestimated, particularly at lower RH (<50%). This low bias in ɛNO3 could be due to overestimated HNO3, as the lower RH and associated higher temperatures generally prevent excess HNO3 (denominator of Eq. 2) from partitioning to the aerosol-phase. We discuss the 300 possibility of overestimated model HNO3 below. As ɛNO3 is underestimated in the model, excess partitioning to the aerosolphase is not a cause of the model nitrate overestimate shown in Fig. 2. The successful performance of ISORROPIAv2.2 during KORUS-AQ is also evident from the comparison against the E-AIM model in Kim et al. (2022). Figure 6a shows vertical profiles of observed and modeled HNO3 for the Olympic Park descents. The model overestimates 305 HNO3 in the lowest bin (0.5 km) by +1600 ppt or +100%. This high bias persists across most of the study domain except over the ocean south of 34 o N ( Fig. 7) where local emissions have a small impact and loss to deposition is slow. During average daytime conditions (~50% RH, 295K), model ɛNO3 is ~0.3, indicating that while the aerosol is HNO3-limited, higher temperatures and low RH also prevent the excess model HNO3 from partitioning to aerosol. A simulation turning off South Korean emissions shows that local sources contribute ~50% to model HNO3 concentrations below 0.5 km (Fig. 6a). Thus while 310 model errors in emissions or chemistry could be a cause of the bias, an overestimated lifetime of HNO3 against dry or wet deposition could also play a role. We evaluate these possibilities further in Section 5.2.

Causes of overestimated daytime HNO3
KORUS-AQ provides aircraft and surface observations that provide additional constraints on the model HNO3 bias of +100% described in Section 5.1. We use observations of NO2 and OH from aircraft to evaluate whether NOx emissions or production 315 from R1 (NO2 + OH) are overestimated. Figure 6b shows that model NO2 is underestimated by -40% below 0.5 km. This is partially due to the expected model inability to resolve the highest observed levels of NO2 in an urban region, illustrated by the larger standard deviation in the observations compared to the model. However, given the same emissions inventory used here (KORUSv5), a set of eight models varied in their biases for NOx against KORUS-AQ aircraft observations from a minimal underestimate (-7%) to a large overestimate (+56%) depending on model configuration . Thus, model biases 320 could be due to a range of factors including underestimated emissions, inaccuracies in the emission diurnal cycle, or overestimated mixed layer heights. Errors in any of these factors that could increase model NO2, such as decreased mixed layer heights or increased emissions, would be expected to increase the model overestimate of HNO3. Fig 6c shows that the model bias in OH is small (+20%) and well within measurement uncertainty (+32%) and therefore it is unlikely that model errors in R1 could cause the model HNO3 bias +100%. 325 The fastest removal pathways for HNO3 are wet and dry deposition. The model implementation of these processes is described in Section 2. The revised wet scavenging scheme has improved annual average model simulations of HNO3, but the effect on HNO3 during KORUS-AQ is limited as precipitation was infrequent after the beginning of the campaign as discussed above.
Section S2 further discusses the impact of this scheme on KORUS-AQ nitrate and HNO3 but errors in wet deposition are unlikely to be the cause of overestimated model HNO3. Section S2 also describes other possible loss pathways to dust, seasalt, 330 or production of ClNO2 from N2O5 hydrolysis that have negligible effects on the model HNO3 and nitrate.
Previous attempts to improve model nitrate invoked an unknown sink of HNO3 in the model (Heald et al., 2012;Weagle et al., 2018), as uncertainties in precursor emissions, the rate of N2O5 hydrolysis (R2/R3) or gas-phase production (R1), OH concentrations, and HNO3 dry deposition velocity (VdHNO3) could not explain model nitrate biases. We similarly conclude that 335 an unknown loss process must be a main cause of the daytime model overestimate in HNO3 and associated evening nitrate bias during KORUS-AQ that occurs as conditions become more favorable for partitioning HNO3 to pNO3. This unknown loss process could be a larger underestimate in dry deposition than has been previously considered, as constraints from KORUS-AQ show that uncertainties in emissions, nighttime production (R1-R5), and wet deposition are not the cause. Heald et al.
(2012) ruled out dry deposition after assuming an uncertainty of a factor of two. Here, the increase in VdHNO3 required to 340 reproduce observed HNO3 (Fig. 6a) is a factor of five. A similar increase in VdHNO3 was invoked by Itahashi et al. (2017) in their model study of wintertime nitrate in East Asia based on the finding from Shimadera et al. (2014) that VdHNO3 (as well as NH3 emissions and dry deposition) were the main factors driving model nitrate performance.
The increase in VdHNO3 suggested above would result in an average value of 7.5 cm s -1 compared to the standard model value 345 of 1.5 cm s -1 . This corresponds to a maximum midday rate of 15.4 cm s -1 cm s -1 compared to the original value of 3.1 cm s -1 ( Fig S8). Deposition of HNO3 is limited only by aerodynamic resistance (and available surface area), as it readily adheres to surfaces. While the increase to VdHNO3 we suggest here is large, this could arise from factors such as increased surface area in urban or heavily forested regions and increased vertical mixing over cities due to turbulence induced by the urban heat island effect. These factors are not accounted for in the limited existing deposition velocity measurements that have been compared 350 against models (Nguyen et al., 2015). Increased turbulence over forested regions results in higher deposition velocities (Sievering et al., 2001;Yazbeck et al., 2021), which would also be expected in an urban environment (i.e. Keuken et al., 1990).
The model does not account for increased available surface area for deposition contributed by urban buildings, or the elevated vertical mixing over cities due to the urban heat island effect (Hong and Hong, 2016;Halios and Barlow, 2018). Dry deposition rates thus may be much higher than in model parameterizations that do not include a specific treatment of the urban canopy 355 (Cherin et al., 2015) and this is the case in GEOS-Chem.
NOx (0) is the initial NOx mixing ratio normalized to CO (Fig. 8, 0.24 ppbv / ppbv CO), ꞵ is the first order loss rate for TNO3, c is the first order production rate for TNO3 (pTNO3 = pHNO3 = kR1[OH]), and TNO3(t) is observed TNO3 as a function of photochemical age (t). As the production of TNO3 was constrained by observed OH, and assuming the main loss of TNO3 (ꞵ) 365 is from deposition of HNO3, the unknown for TNO3 evolution is the deposition rate. The full details of this calculation are provided in Section S3. Figure 8 shows NOx, TNO3, and the other NOx oxidation products of total peroxy nitrates (ΣPNs) and the sum of alkyl-and multi-functional nitrates (ΣANs) as a function of photochemical age. All species are normalized by background subtracted 370 CO. NOx is continuously depleted at a rate of 0.31 hr -1 , implying continued production of TNO3, ΣPNs, and ΣANs. This loss rate corresponds to a lifetime of 3.2 hrs that is similar to the lifetime of 4.8 hrs for NO2 against conversion to HNO3 (R1) using the SMA average OH of 5.2×10 -6 molec cm -3 . From Eq. 3, we derive a loss rate (ꞵ) of 13.9 cm s -1 that best fits the observed change in TNO3 with aging. As deposition of pNO3 is slow, we assume that VdHNO3=VdTNO3. All three NOx oxidation products (TNO3, ΣPNs, ΣANs) exhibit similar behavior with production outpacing loss until approximately three hours of aging, where 375 loss appears to balance production and concentrations remain relatively constant. There is likely large uncertainty in the derived photochemical ages shown in Fig 8, as the aircraft did not follow plumes as in Neuman et al. (2004). However, our derived NOx lifetime is consistent with average SMA conditions and is not affected by our choice of observed altitude range, suggesting that the aging represents true chemical processing.
380 Figure 8 shows that the slower value for midday VdHNO3 in the original model (3.1 cm s -1 ) poorly represents observations compared to the faster value obtained in Fig. 6 (15.4 cm s -1 ). We calculate that the original deposition rate would correspond to a first order loss rate for TNO3 of only 0.07 hr -1 (assuming a 1.5 km boundary layer height) and thus observed TNO3 should increase with photochemical age, which is not supported by the observed relationship in Fig. 8. The factor of five increase in VdHNO3, constrained only using observed HNO3, implies a similar loss rate of TNO3 as derived in Fig. 8 and leads to the 385 observed behavior where after initial production, the normalized mixing ratio remains constant. This analysis supports the hypothesis given above, that existing observations supporting lower values for VdHNO3 (Nguyen et al., 2015) may underrepresent deposition in regions with greater turbulence and available surface area such as in cities like Seoul. Deposition of atmospheric pollutants such as nitric acid on buildings generates 'urban grime' that may photolyze and produce NOx and HONO Donaldson, 2013, 2016). This urban grime could be a source of HONO (Zhang et al., 2016) and may be 390 larger than previously thought if models underestimate nitric acid deposition. Figure 4a shows the impact to the diurnal cycle of model nitrate from increasing model VdHNO3 by a factor of five. The rapid late afternoon /early evening increase in model nitrate (Fig. 4a) is largely resolved and the model simulation of HNO3 is now in good agreement with aircraft observations (Fig. 6) due to a significant dampening of the HNO3 diurnal cycle (Fig. 4b). This 395 reduction in the HNO3 diurnal cycle is better supported by observations of TNO3 as discussed above. We conclude that a key reason for the overestimated daily average model nitrate shown in Fig. 2 is overestimated daytime HNO3 that results in excess nighttime nitrate when conditions become favorable (cool and humid) for gas to aerosol partitioning. The model overestimate is due to insufficient loss, likely underestimated dry deposition. This finding does not address possible errors in model nighttime production pathways (NO2 vs. N2O5), and KORUS-AQ provides detailed ground observations that can be used to 400 constrain the model representation of nighttime chemistry.

Errors in model nighttime production of HNO3
Figure 4c shows that model nighttime production of HNO3 by aerosol uptake of NO2 (R4) is approximately twice as large as R2 (N2O5 hydrolysis). This contradicts the calculation from Jordan et al. (2020) that R2 is the driver of nitrate production during KORUS-AQ, particularly during the Transport/Haze period due to sufficient nighttime ozone concentrations that allow 405 for production of the nitrate radical and N2O5 through R8 and R9.
Production of nitrate by N2O5 hydrolysis is supported by observations of ClNO2, thought to be produced primarily by this 410 reaction (Thornton et al., 2010). As discussed above in Section 5.2, observations of ClNO2 at Olympic Park are elevated at night (Fig. S7). Despite recent large reductions of the uptake coefficient ( ) for NO2 in the model (Holmes et al., 2019), NO2 uptake still is the dominant nighttime pathway for HNO3 production in the model. We use observations of ozone, NO, and NO2 at Olympic Park to determine whether errors in R7-R9 are impacting model ability to produce N2O5. Figure 9 shows the mean modeled and observed diurnal cycles of ozone and NO2 for the AirKorea sites in the model grid box (Fig. 1b) and for ozone, NO, NO2, and NOx at Olympic Park. Ozone might be expected to be titrated in an urban area by R7 as the mixed layer collapses in the evening, resulting in elevated NO and shutting down production of the nitrate radical (R8). This is the case in the model where nighttime ozone is <2 ppb approximately 20% of the time but this never occurs in the observations (Fig. S9). As a result, average observed nighttime ozone is 24 ppb but only 13 ppb in the model (Fig. 9). The 420 time series of observed and modeled ozone in Fig. S9 shows while the model does succeed in simulating high nighttime ozone concentrations during the Dynamic Period, characterized by higher windspeeds, ozone is incorrectly titrated at other times particularly during the buildup of the haze pollution following a frontal passage on May 24 th . The implications of this excess ozone titration for the simulation of PM2.5 specifically during haze conditions will be further discussed in Section 6.

425
As shown in Fig. 9b+c, model ozone titration corresponds to excess model NO and NO2 at night and explains the dominance of NO2 uptake in the model over N2O5 hydrolysis for nighttime HNO3 production. The model bias for NOx is minimal during the day, providing additional support for the level of emissions in the model, but is overestimated by a factor of two at night.
The excess model ozone titration and overestimated nighttime NOx implies an error in nocturnal mixing. Figure 10a shows the mixed layer height (MLH) diurnal cycle measured by ceilometers at Olympic Park and Seoul National University. The aerosol 430 gradients detected by the ceilometer to estimate MLH are less reliable at night due to the possible presence of aerosols in the residual layer (Jordan et al., 2020). We support these measurements with additional calculations of nighttime MLH from radiosonde observations of temperature and RH four times a day (Section S4, Fig. S10), showing that the average MLH at 3 KST could be ~300m compared to 220m in the model. As previously discussed in Section 5.2, in urban regions such as Seoul, the anthropogenic heat island effect and the heterogeneity of the urban land cover increase sensible heat fluxes and turbulence 435 over non-urban areas (Halios and Barlow, 2018) and create an unstable mixed layer even at night. Min et al. (2020) showed that the nighttime mixed layer in Seoul is elevated in all seasons, and that nighttime conditions are generally unstable due to urban heat storage and anthropogenic heat release and this could explain the observed elevated nighttime MLH (Fig. 10a, Fig.   S11). This effect is not captured in many meteorological models including the one used here (GEOS-CF, Section 3). Nighttime sensible heat flux in the model is always negative (stable conditions) (Fig. 10b). 440 Starting at 17 KST, the model mixed layer collapses early, causing a more rapid decline in ozone than in the observations (Fig.   9a, Fig. 10a). The transition from convective daytime mixed-layer to stable nocturnal boundary layer is poorly understood While addressing the shortcomings of the model mixing scheme is beyond the scope of this study, we test the sensitivity of model nitrate production to the main two problems identified above, 1) the overly rapid collapse of the afternoon mixed layer, and 2) insufficient nocturnal mixing. While model meteorology is calculated offline, mixing in the boundary layer is calculated online (Section 2), allowing us to perturb mixing parameters. We increase the nighttime MLH to 500m to examine the impact on model ozone, NO, and NO2. The effect of this change on these species is minimal (Fig. 9), similar to the findings of other 455 model sensitivity studies that performed this same test (Oak et al., 2019;Miao et al., 2020). While the strength of model vertical mixing is sensitive to MLH, the model sensible heat flux and friction velocity have a larger impact (Holtslag and Boville, 1993), and the nighttime mixed layer will remain stable while the sensible heat flux is negative regardless of MLH.
The increase in nighttime mixing in urban vs. rural regions has been addressed in the CMAQ model (Li and Rappenglueck,460 2018) by using a higher value for the minimum mixing strength (eddy diffusivity) over urban areas. However, we find that this approach is insufficient to address model ozone titration without increasing the sensible heat flux to a positive value to produce an unstable mixed layer. This is illustrated in Fig. S12, where we scale the model MLH to match the profile at Olympic Park (Fig. 10a) and raise the model minimum eddy diffusivity from 0.01 m 2 s -1 to 1 m 2 s -1 over the SMA. Reducing the collapse of the evening MLH without a change to the drivers of mixing (i.e., heat fluxes, friction velocity) has negligible impact on 465 decreasing model ozone titration (Fig. S12). In addition, the MLH at Olympic Park in the early morning hours appears inconsistent with observed ozone, likely due to the uncertainties in the measurement technique discussed above and supported by the lower values obtained from radiosonde profiles (Fig. S11). Errors in model nighttime mixing are difficult to remedy without significant revisions to the model mixing parameterizations, including implementing continued mixing from daytime eddies into the evening hours (Blay-Carreras et al., 2014) and parameterizing the excess sensible heat flux in urban areas 470 (Halios and Barlow, 2018). We address the implications of these errors in the simulation of haze pollution events in Section 6.

Model simulation of haze buildup
The failure of models to simulate sulfate production in haze in East Asia is a current topic of intensive research and is attributable to missing sulfate production in aerosol water (Wang et al., 2014;Zheng et al., 2015a;Chen et al., 2016;Shao et al., 2019;Miao et al., 2020). There has been less attention paid to the ability of models to simulate nitrate in haze as nitrate-475 dominated haze is a more recent phenomenon due to the reductions in SO2 in East Asia . Figure 2 and Table   3 show that the model can reproduce the increase in the nitrate component of PM2.5 during the Transport/Haze period but overestimates absolute concentrations by ~20%. This contributes to an 80% overestimate in ALWC. Efforts to explicitly simulate SO2 oxidation in ALWC may be hindered by this model bias, which also impacts the rates of all other heterogeneous reactions through the increase in aerosol surface area. 480 Figure 11a shows the hourly time series of observed and modeled nitrate at Olympic Park during the Transport/Haze period.
During the haze buildup, the model initially overestimates nitrate during the day (5/24) followed by large nighttime underestimates (5/24-5/25). This is opposite to the nighttime overestimate but daytime agreement shown in the campaign average (Fig. 4a). During the haze buildup, daytime RH remained elevated (>50%, Fig. S13) and the daytime mixed layer was 485 suppressed (Fig. S14 and Jordan et al. 2020). The model reproduces both conditions, which are favorable for SNA production.
Model nitrate biases here are likely due to the errors identified in Section 5.2 (overestimated daytime HNO3) and Section 5.3 (incorrect representation of nighttime conditions), but here the excess daytime HNO3 in the model results in higher daytime nitrate than in the campaign average. Insufficient model sulfate during the haze event results in overestimated model pH and excess partitioning of HNO3 to the particle phase (Guo et al., 2016). Fig. S15 shows that ɛNO3 (the calculated fraction of TNO3 490 in the aerosol phase) decreases as sulfate increases and the model sulfate bias corresponds to a difference in ɛNO3 of ~0.3.
The model underestimate of nighttime nitrate concentrations during the haze buildup must be because the rate of observed N2O5 hydrolysis (R2) exceeds even the erroneously high model rate of NO2 aerosol uptake (R4). The haze buildup was characterized by a lower daytime MLH and a deeper nocturnal MLH (inferred from the lack of ozone titration) that resulted 495 in higher nitrate production from N2O5 hydrolysis (Jordan et al., 2020). The model overly titrates ozone (Fig. 11c) due to insufficient nighttime mixing. We drive additional nocturnal mixing by increasing the sensible heat flux at night from slightly negative (-4 W m -2 ) to weakly positive (+10 W m -2 ), representative of anthropogenic heat fluxes in this region (Hong and Hong, 2016;Varquez et al., 2021). To reduce the rate of R4 from overestimated NO2 and allow for a high rate of R2, we increase the nighttime MLH over land to 300 m as suggested by the observations. This largely resolves the incorrect model 500 ozone titration and the severe model overestimate of nighttime NO2 on 5/23-5/24 and on 5/24-5/25 but does not remedy the early model collapse of the evening mixed layer (Fig. 11). Extending this sensitivity test past the haze buildup results in excess nighttime ozone. This may be due to the increased cloud cover during the haze buildup (Fig. S16), that could cause additional nighttime mixing over average conditions through enhancement of the urban heat island effect (Theeuwes et al., 2019). 505 Figure 11b shows that increased nighttime mixing allows for N2O5 hydrolysis (R2) to become the main nighttime pathway for HNO3, with a rate three times greater than NO2 uptake (R4) in the base model. The raised mixed layer height of 300 m prevents this high rate from resulting in overestimated model nitrate. Increased model nighttime nitrate corresponds to an increase in ALWC of 40%. We use the simulations shown in in Fig. 11 to illustrate that model errors in simulating mixed layer dynamics (overly rapid collapse of the evening mixed layer and insufficient nighttime mixing) result in errors in model chemistry. 510 Nighttime measurements of the vertical structure of key species such as ozone, NO2, N2O5, and HNO3, complemented by sensible heat flux observations, are needed to further constrain model simulations of nighttime nitrate production.
As discussed in Section 4, in addition to the above difficulties in simulating nitrate, the model fails to reproduce observed sulfate during the Transport/Haze period and this corresponds to a 15 µg m -3 underestimate in PM2.5 (Table 3). Studies have 515 shown a strong relationship between increasing RH and conversion of gas-phase precursors to SNA in haze, indicating the occurrence of heterogeneous chemistry in ALWC (Sun et al. Liu et al., 2015;Quan et al., 2015;2015a;Chen et al., 2016;Wu et al., 2018a). Figure 12 shows the sulfate oxidation ratio, SOR ≡ M N as a function of RH at Olympic Park and from aircraft observations. In the observations, SOR increases with RH, but this is missing from the model. We take the approach of Wang et al. (2014) and implement heterogeneous uptake of SO2 on aerosol (not present in the standard model) as a function 520 of RH according to Eq. 4, where the mass transfer rate (kT) at which a species is lost from the gas-phase is a function of the particle radius (a), the molecular diffusion coefficient (Dg), the mean molecular speed (v), and the reactive uptake coefficient (γ), or the probability of irreversible reaction. The value for γ depends on RH (Wang et al., 2014) according to Eq. 5. 525 The values B6 ,**% = 3 × 10 %$ and B6 )*% = 3 × 10 %# best fit the observations using the model without the aforementioned adjustments for nitrate simulation. These values are two orders of magnitude slower than in the original formulation of Wang et al. (2014) but similar to more recent studies (Zheng et al., 2015a;Chen et al., 2016). During the Transport/Haze period, this improves model agreement with average surface (Table 3,  pollution, as the model attributes ~60% of SO2 to foreign sources and ~40% to local emissions (Fig. S17). We simulate PM2.5 with heterogeneous conversion of SO2 as described above, and then remove South Korean emissions in order to investigate changes to the fraction of transported pollution. Figure 13 shows the model PM2.5 composition for each case during the Transport/Haze period, with an additional 15 µg m -3 of PM2.5 in the model with heterogeneous uptake of SO2. In the original model, foreign transport accounts for 66% of PM2.5 (25 µg m -3 ), but this fraction is reduced to 54% (29 µg m -3 ) in the revised 540 model as the local contribution (13 vs. 24 µg m -3 ) makes up a greater fraction of the increase. Locally produced sulfate increases from only 1% (<1 µg m -3 ) to 25% (6 µg m -3 ) of local PM2.5, implying that local SO2 controls could have an effect on PM2.5 levels. Locally produced nitrate increases from 6 µg m -3 to 8 µg m -3 . The total amount of model nitrate (local + foreign) decreases slightly at the surface and aloft (Fig. S17) which we attribute to the impact of sulfate on reducing ɛNO3 described above and shown in Fig. S15 but this does not resolve the model nitrate biases described in Section 5. 545 The previous calculations only account for the missing model sulfate during the Transport/Haze period, and do not account for the incorrect model representation of nighttime nitrate production or overestimated model HNO3. This accounts for the dramatic increase in ALWC in Fig. 13, which is already overestimated in the original model formulation as shown in Fig 2. Given the uncertainties in revising the model nitrate simulation, we did not assess the policy implications for improving model 550 nitrate on local vs. transported pollution. A simple test however of the haze buildup with the inclusion of a factor of five increase to VdHNO3, increased nighttime mixing, and the addition of heterogeneous SO2 uptake described above, results in 40% less nitrate and ALWC. As a result, sulfate concentrations are 30% less than in the simulation with heterogeneous SO2 uptake alone. Therefore, studies attempting to determine γ to improve sulfate simulations of haze must also consider the impact of model nitrate biases on their parameterization. 555

Conclusions
We used aircraft and surface observations from the NIER-NASA KORUS-AQ field campaign in May and June 2016 to evaluate GEOS-Chem simulations of PM2.5 composition in the Seoul Metropolitan Area, including during a haze pollution event characterized by high levels of secondary inorganic aerosol. Models generally overestimate nitric acid and the gasparticle partitioning of nitric acid to aerosol and underestimate sulfate during haze events across East Asia . 560 This is of concern for using models to determine the fraction of PM2.5 pollution that can be controlled using local policy measures in South Korea, and the level to which exceedances of PM2.5 standards are caused by long-range transport.
The model underestimated PM2.5 in Seoul during the campaign (NMB = -15%) with larger errors in composition. On average, the model underestimated sulfate (-64%) and SOA (-43%) but overestimated nitrate (+36%). Models typically underestimate 565 secondary organic aerosol (SOA, Zhao et al., 2016), and this could be due to missing sources from anthropogenic precursors . This SOA bias will be investigated in future studies. Aircraft observations, only available during daytime hours, showed model underestimates in sulfate comparable to the bias at the surface. However, modeled nitrate was underestimated aloft, contradicting the model overestimate in the campaign average (which includes nighttime observations).
Hourly surface observations showed that this was due to a model overestimate at night. During the campaign, nitrate formation 570 was limited by the supply of nitric acid, which was overestimated against daytime aircraft observations by +100% and contributed to the model nighttime bias. Recent developments to the model wet deposition scheme have significantly improved the simulation of nitrate and nitric acid, but further improvements are unlikely to resolve the model bias.
The model overestimate in nitric acid was not due to overestimated production, insufficient loss to wet deposition, or uptake 575 to dust or seasalt. Increasing the nitric acid dry deposition velocity by a factor of five was required to reconcile the model with observations. Aircraft observations of total nitrate (TNO3 = HNO3 and pNO3) as a function of photochemical age support this increase. The model underestimate in deposition could be explained by missing treatment of turbulence driven by the urban heat island effect and the heterogeneity of the urban landscape, which would also increase the surface area available for deposition. Here, we only consider the effect on HNO3, but these factors would also impact other species that readily deposit 580 to surfaces such as NH3, which was not measured during the campaign.
Observations of ozone, NO2, and ClNO2 showed that N2O5 hydrolysis should be the main driver of nighttime nitrate production while the model primarily produced nitrate through aerosol uptake of NO2. The model overly titrated ozone, with an average nighttime concentration of 13 ppb compared to 24 ppb in the observations. This resulted in excess model NO2 and prevented 585 the production of N2O5. Observations of ozone and of the nighttime mixed layer height implied insufficient nighttime mixing and an overly rapid collapse of the afternoon mixed layer in the model. We attributed these errors to the premature shutdown of afternoon eddies and missing treatment of the urban heat island effect that typically generates a positive nighttime heat flux that is not present in the model. Nighttime measurements of the vertical structure of key species such as ozone, NO2, N2O5, and HNO3, ideally complemented by surface heat flux observations, are needed to further constrain model nighttime nitrate 590 production, and determine the extent to which the model underestimates nighttime heating and mixing depth.
The model errors in simulating nitrate and nitric acid, mainly arising from overestimated daytime nitric acid and excess nighttime ozone titration, are exacerbated in the simulation of haze pollution. Overestimated nitric acid results in larger values of daytime nitrate during the haze buildup. This could be due to the model underestimate in sulfate as overestimated model pH 595 would allow for increased partitioning of nitric acid to the particle phase. Nighttime nitrate in the model is underestimated during the haze buildup likely due to missing rapid N2O5 hydrolysis. Sensitivity simulations showed that raising the nighttime mixed layer and providing a positive nighttime sensible heat flux of +10 W m -2 improved the model simulation of nitrate, ozone, and the N2O5 pathway for nitrate production during haze. Previous studies have simply raised the nighttime mixed layer and found little effect on simulated pollution (Oak et al., 2019;Miao et al., 2020) but this may be due to missing nocturnal 600 heating from anthropogenic heat release. The underestimate in model sulfate during the KORUS-AQ haze event is typical of models that often do not include heterogeneous aerosol uptake of SO2 (Wang et al., 2014;Zheng et al., 2015aZheng et al., , 2015bShao et al., 2019). Observations of the sulfate oxidation ratio (SOR) as a function of RH supported the need for this pathway as the strong increase in SOR with RH 605 was not present in the model. A simple parameterization of this process increased model sulfate levels from 4 to 15 µg m -3 during the haze, in better agreement with observations. However, the success of this parameterization was complicated by model nitrate biases. A simulation of the haze with both improved model nitrate and heterogeneous uptake of SO2 resulted in a 30% reduction in model sulfate over the simulation with heterogeneous uptake of SO2, illustrating the need to consider model biases in sulfate and nitrate simultaneously. GEOS-Chem parameterizations of the urban environment are lacking and cannot 610 be currently adjusted to robustly simulate nitrate during the campaign. However, future studies attempting to simulate sulfate in haze should consider the impact of model nitrate biases on their parameterizations. These studies require models that are able to simulate a large domain to calculate long-range transport but include the detailed parameterizations of the urban environment (urban heat island effect etc.) required to successfully simulate nitrate.

615
Determining the contribution of local vs. transported PM2.5 is essential to the development of successful policy measures to reduce unhealthy pollution levels. Significant effort has gone into this evaluation in South Korea, but with models that have errors in PM2.5 composition (Choi et al., 2019;Kumar et al., 2021). The local PM2.5 contribution may be underestimated without including heterogeneous uptake of SO2 on aerosol to produce sulfate during haze. Locally-produced PM2.5 increased from 13 to 24 µg m -3 , decreasing the fraction of foreign pollution from 66% to 54%. Locally-produced sulfate increased from 620 <1 µg m -3 to 6 µg m -3 , implying that controls on SO2 could have a larger impact than in model formulations without this chemistry. As a consequence of the 2013 Clean Air Action plan implemented in China, emissions of inorganic aerosol precursors have been decreasing (Zheng et al., 2018) and concentrations of PM2.5 in China have declined by approximately -5 µg m -3 per year from 2013-2018 (Zhai et al., 2019). Emission reductions in South Korea may be less rapid , and thus the impact of long-range transport on future PM2.5 pollution events could decline in the future. It is critical for models 625 to improve representations of the interactions between physical processes and chemical production of PM2.5 production to support continued local air quality improvements.

Code Availability
The model code used in this work is available at 10.5281/zenodo.5620667. 630

Data Availability
The KORUS-AQ data archive (KORUS-AQ Science Team, 2019) includes both the aircraft and ground-based measurements from AirKorea, Olympic Park, and KIST. The precipitation data is available at: https://www.ncdc.noaa.gov/cdo-web/datasets.

Competing Interests
The authors have the following competing interests: Some authors are members of the editorial board of Atmospheric Chemistry and Physics. The peer-review process was guided by an independent editor, and the authors have also no other 645 competing interests to declare. the use of his ATHOS OH data. We acknowledge Paul Wennberg and John Crounse for the use of their CIT-CIMS HNO3 data. We acknowledge L. Greg Huey for the use of his SO2 data. We acknowledge James J. Szykman for the use of ceilometer data at Olympic Park. We acknowledge Seogjo Cho for the MARGA data at Olympic Park. We acknowledge Ke Li and Yingying Yan for their help implementing aromatic chemistry in GEOS-Chem. We thank Jerome Fast, Rahul Zaveri, and 655 David Peterson for their helpful discussions. PCJ and JLJ were supported by NASA Grants 80NSSC18K0630 and 80NSSC19K0124. The GEOS-FP data used in this study/project have been provided by the Global Modeling and Assimilation Office (GMAO) at NASA Goddard Space Flight Center.    the AirKorea sites in Figure 1b, and the speciated PM2.5 components from Figure 2. Production of nitric acid c)

Observations
Model +No nighttime production +Old wet scavenging scheme +5x dry deposition All Halogens (Table S3) NO + OH ( Figure 9. Mean diurnal cycle from May 1 to June 7, 2016 for a) ozone and b) NO2 for the AirKorea sites within the GEOS-Chem gridbox (Fig. 1b) and for c) NO and d) NOx at Olympic Park. The gray shading represents the standard deviation across the AirKorea sites. The solid gray line is the AirKorea site closest to Olympic Park, and the dashed line is the measurement from the EPA (Table 2) Figure 11. a) Transport/Haze period timeseries of modeled and observed hourly nitrate fraction of PM2.5, b) modeled 1190 production of HNO3 from N2O5 (R2) and NO2 (R4), c) ozone, d) and NO2. The sensitivity studies are described in Section 6. The gray shaded regions represent 8pm to 8am.