Comparison of Chemical Lateral Boundary Conditions for Air Quality

16 The existing National Air Quality Forecast Capability (NAQFC) operated at NOAA provides 17 operational forecast guidance for ozone and particle matter with aerodynamic diameter less than 18 2.5μm (PM2.5) over the contiguous 48 U.S. states (CONUS) using the Community Multi-scale 19 Air Quality (CMAQ) model. Currently NAQFC is using chemical lateral boundary conditions 20 (CLBCs) from a monthly climatology, which cannot capture pollutant intrusion events originated 21 outside of the model domain. In this study, we developed a model framework to introduce the 22 time-varying chemical simulation from the Goddard Earth Observing System Model, version 5 23 (GEOS) as the CLBCs to drive NAQFC. The method of mapping GEOS chemical species to 24 CMAQ CB05-Aero6 species was also developed. We then evaluated NAQFC’s performance 25 using the new CLBCs from GEOS. The utilization of the GEOS dynamic CLBCs showed an 26 overall best score when comparing the NAQFC simulation with the surface observations during 27 the Saharan dust intrusion and Canadian wildfire events in summer 2015: the PM2.5 correlation 28 coefficient R was improved from 0.18 to 0.37 and the mean bias was narrowed from -6.74 μg/m3 29 to -2.96 μg/m3 over CONUS. The CLBCs’ influences depended on not only the distance from 30 the inflow boundary, but also species and their regional characteristics. For the PM2.5 31 prediction, the CLBC’s effect on the correlations was mainly near the inflow boundary, and its 32 impact on the background could reach farther inside the domain. The CLBCs also altered 33 background ozone through the inflows of ozone itself and its precursors. It was further found that 34 aerosol optical thickness (AOT) from VIIRS retrieval correlated well to the column CO and 35 elemental carbon from GEOS, based on which the new CLBCs for wildfire intrusion event was 36 derived. The AOT derived CLBCs successfully captured the wildfire intrusion events in our case 37 study for summer 2018. It can be a useful alternative in case the CLBCs of GEOS are not 38 available. 39 https://doi.org/10.5194/acp-2020-587 Preprint. Discussion started: 5 August 2020 c © Author(s) 2020. CC BY 4.0 License.


16
The existing National Air Quality Forecast Capability (NAQFC) operated at NOAA provides 17 operational forecast guidance for ozone and particle matter with aerodynamic diameter less than 18 2.5µm (PM2.5) over the contiguous 48 U.S. states (CONUS) using the Community Multi-scale 19 Air Quality (CMAQ) model. Currently NAQFC is using chemical lateral boundary conditions 20 (CLBCs) from a monthly climatology, which cannot capture pollutant intrusion events originated 21 outside of the model domain. In this study, we developed a model framework to introduce the 22 time-varying chemical simulation from the Goddard Earth Observing System Model, version 5 23 (GEOS) as the CLBCs to drive NAQFC. The method of mapping GEOS chemical species to 24 CMAQ CB05-Aero6 species was also developed. We then evaluated NAQFC's performance 25 using the new CLBCs from GEOS. The utilization of the GEOS dynamic CLBCs showed an 26 overall best score when comparing the NAQFC simulation with the surface observations during 27 the Saharan dust intrusion and Canadian wildfire events in summer 2015: the PM2.5 correlation 28 coefficient R was improved from 0.18 to 0.37 and the mean bias was narrowed from -6.74 µg/m 3 29 to -2.96 µg/m 3 over CONUS. The CLBCs' influences depended on not only the distance from 30 the inflow boundary, but also species and their regional characteristics. For the PM2.5 31 prediction, the CLBC's effect on the correlations was mainly near the inflow boundary, and its 32 impact on the background could reach farther inside the domain. The CLBCs also altered 33 background ozone through the inflows of ozone itself and its precursors. It was further found that 34 aerosol optical thickness (AOT) from VIIRS retrieval correlated well to the column CO and 35 elemental carbon from GEOS, based on which the new CLBCs for wildfire intrusion event was 36 derived. The AOT derived CLBCs successfully captured the wildfire intrusion events in our case 37 study for summer 2018. It can be a useful alternative in case the CLBCs of GEOS are not 38 available. 39

1
The chemical lateral boundary condition (CLBC) is one of the most important factors affecting 2 the prediction accuracy of regional chemical transport models (Tang et al., 2009;Tang et al., 3 2007). It mainly plays two roles in the regional modeling system: 1) to impose the constraints 4 with background concentrations and 2) to represent the external influence for intrusion events. 5 The climatological static CLBCs can provide the first role for some long-lived pollutants, such as 6 CO and O3. Models like the Community Air Quality Multi-scale Model (CMAQ) hemispheric 7 version (Mathur et al, 2017) can also get this constraint with its LBC along the equator. The 8 second role of the CLBC, representing the influences of external intrusion events, can only be 9 made with the dynamic (time-varying) CLBCs. Such CLBCs can only come from a global 10 model, a regional model with bigger domain (Tang et al., 2007), or observed profiles (Tang et 11 al., 2009). 12 As a regional chemical forecast system, the existing National Air Quality Forecast Capability 13 (NAQFC) operated at NOAA needs proper CLBCs for its daily prediction. The current NAQFC 14 uses the dust aerosol LBC from the NOAA Environmental Modeling System (NEMS) Global 15 Forecast System (GFS) Aerosol Component (NGAC) (Lu et al, 2016;Wang et al, 2018), which 16 is the GFS model coupled with Goddard Chemistry Aerosol Radiation and Transport 17 (GOCART) aerosol mechanism (Chin et al., 2000(Chin et al., , 2002Colarco et al., 2010). Before the 18 implementation of NGAC LBC, NAQFC used the background static profile LBC for aerosols 19 described in Lee et al. (2017). For gaseous species, NAQFC uses the modified monthly 20 averaged LBCs from the GEOS-Chem (Bey et al., 2001) simulation for year 2006(Pan et al., 21 2014. To alleviate surface ozone over-predictions, the upper tropospheric ozone LBC from 22 GEOS-Chem have been limited ≤ 100 ppbV. 23 The static gaseous LBC cannot capture the signals of some intrusion events, such as the biomass 24 burning plumes from the outside of the domain, which could affect ozone and particle matter 25 with aerodynamic diameter less than 2.5µm (PM2.5). Tang et al. (2007) investigated the 26 sensitivity of the regional chemical transport model (RCTM) to LBCs, and found that the 27 background magnitude of the pollutant concentrations sometimes were more important than the 28 variation of LBCs for the near-surface prediction over polluted areas, or the first role of the 29 CLBC was more critical. Over the Contiguous United States (CONUS) domain, the prevailing 30 inflow lateral boundary includes northern and western USA, where Canadian emission and long-31 rang transported Asian air-masses can affect the CONUS background. Southeastern States could 32 encounter the Saharan dust intrusion during summer time, which usually resulted in a surface 33 PM2.5 increase (Lu et al, 2016). In order to assess their impact, a proper CLBC from a global 34 model that carries those signals is needed. In this study, we extracted the CLBC from the GEOS 35 global chemical circulation model (GCCM) (Strode et al. 2019;Molod et al., 2012) in static 36 (monthly average) and dynamic (3-hour varying) modes. The CMAQ runs with the GEOS 37 CLBCs were then compared to the CMAQ base case and another run with the NGAC aerosol 38 LBC for the summer 2015. During this period, the Canadian wild fire and Sahara dust affected 1 the CONUS domain, which affected the Northern and Southern USA, respectively, and different 2 CLBCs showed their impacts on the CMAQ regional predictions. In addition, we will investigate 3 the method of using historical CLBCs with a certain indicator to derive a new CLBC for the 4 future pollutant intrusion events in case an appropriate global CLBC is not available. 5

Model Configuration and Experiment Design
6 Current NAQFC is using CMAQ version 5.0.2, which includes CB05 gaseous chemical 7 mechanism (Yarwood et al., 2005) with updated toluene (Whitten et al., 2010) and chlorine 8 chemistry (Tanaka et al., 2003;Sarwar et al., 2007), and Aero6 (Sonntag et al., 2014)  LBC) is full chemistry for both gaseous and aerosol species. We also tested its corresponding 23 monthly mean LBC (GLBC-monthly) for the temporal variation. Besides the normal global 24 LBCs, an aerosol optical depth (AOT) derived Northern LBC (AOT-NLBC) is developed, which 25 will be discussed later. These runs used the same settings except the CLBCs. The two CMAQ 26 runs with dynamic CLBCs, the NGAC-LBC and GEOS-LBC, imported the corresponding LBC 27 every 3 hours. The NGAC-LBC only updates the aerosol LBC from the NGAC global model and 28 its gaseous LBC are the same as the CMAQ base case. GEOS-LBC provides both the gaseous 29 and aerosol LBCs. GLBC-monthly is the static CLBC generated from the monthly mean GCCM 30 results. The AOT-NLBC is the same as GLBC-monthly except that its northern LBC is 31 generated from the relationship of VIIRS (Visible Infrared Imaging Radiometer Suite) AOT and 32 GEOS LBC for the wildfire intrusion events, which will be described later. 33 An interface between NAQFC and GEOS has been developed to transfer CLBCs. It is based on 34 the existing Global-to-Regional interfaces developed by Tang et al (2008Tang et al ( , 2007 for MOZART,35 RAQMS, and NGAC global models with the enhancement to support GEOS's NetCDF4 format, 36 vertical layers and chemical species. The interface includes two major functions: spatial mapping 37 and species mapping. Spatially, GEOS's concentrations from its 576×361 grid in the 0.625º×0.5º 38 horizontal resolution with 72 vertical layers are 3-dimensionally interpolated into CMAQ's 1 CONUS lateral boundary periphery in the 12km horizontal resolution. Since the different 2 chemical mechanisms have been employed in global chemical transport models and CMAQ, the 3 species mapping are required to link both models. 4

Gaseous Species Mapping 5
The GCCM outputs 122 gaseous chemical species and 15 aerosol species. For the species such 6 as O3, CO, NO, and NO2, an explicit one-on-one mapping can be achieved. However, some 7 voltaic organic compounds (VOCs) need special treatment during the conversion as GCCM uses 8 different lumping approaches from the CMAQ CB05tucl (carbon bond 5 mechanism with 9 toluene and chloride species). Table 2 lists the VOC species map used to convert GCCM's 10 gaseous species to CMAQ's CB05tucl species. Two methods were employed for VOCs' 11 speciation mapping: one was based on the carbon bond structure, e.g. ALK4  4 PAR ( can also be mapped to the CB05 species based on their carbon bonds, e.g. R4N2 (GEOS's C4-5 23 alkyl nitrates) can be mapped to NTR + 2.0 PAR in the CB05tucl mechanism. 24

Aerosol Species Mapping 25
Both GEOS and NGAC use the GOCART aerosol scheme though in different versions (Bian et 26 al, 2017 andColarco et al 2010, respectively), and GEOS has additional species of ammonium 27 and 3-bin nitrates (NO3an1, NO3an2 and NO3an3). Table 3 lists the aerosol species mapping 28 from GEOS aerosols to CMAQ Aero6 species used in this study. GEOS aerosols have fixed size 29 bins defined by their diameters, while CMAQ aerosols use 3 size modes: Aitken (ATKN), 30 accumulations (ACC) and coarse (COR) modes (i, j, k modes) (Appel et al., 2010) and each size 31 mode has its own lognormal size distribution (Whitby and McMurry, 1997). To convert the 32 aerosol species from GEOS to CMAQ's Aero6, we need to consider not only the aerosol 33 composition and the conversion from GEOS size bins to the CMAQ size modes, but also the size 34 distribution within each CMAQ size mode that is controlled by the CMAQ aerosol number 35 concentrations (the 3 rd column of Table 3). Dust is converted to AOTHRJ (other unreactive 1 aerosol in accumulation mode) and ASOIL (soil particles in coarse mode). They do not 2 participate in any aerosol reaction, but are just counted in total PM2.5 and PM10. Although the 3 CMAQ Aero6 has explicit elemental ions, like Ca and Mg, which are possible dust ingredients, 4 we do not consider the reaction effect due to these ions. Tang et al. (2004) studied the dust 5 outflow during ACE-Asia field experiment and found that only small portion of cations in dust 6 particles are available for aerosol uptake or reactions, which was nearly none for aged dust air 7 masses. 8

9
To evaluate the impact of LBCs on the model simulations, we chose the period that covered the 10 intrusion events. During summer 2015, two intrusion events occurred in the Southeastern USA 11 and Northern USA, respectively. The Southeastern intrusion was brought by the long-range 12 transported dust storm from the Saharan desert. The northern intrusion was caused by the 13 Canadian wildfire and its southward transport into the CONUS. Figure 1 shows the aerosol 14 optical thickness retrieved from Suomi-NPP satellite's VIIRS instrument from later June to early 15 July, 2015, which highlights these two intrusion events. 16

Dust Storm Events in Summer 2015 17
As shown in Figure 1, a dust storm was originated from the Saharan desert, and brought to the 18 Southeastern USA via the trans-Atlantic transport. The two global models, GEOS and NGAC,19 captured this dust intrusion, and provided the signals of aerosol increments via their CLBCs to 20 NAQFC. Figure 2 shows the NAQFC domain and its southeastern corner covered the Bermuda 21 and Bahamas Islands. Figure 3 shows the corresponding three LBCs for ASOIL and AOTHRJ 22 along the model's boundary locations on July 2, 2015 as the GOCART dusts have been mapped 23 into two CMAQ aerosol species (Table 3). The base run (CMAQ_BASE) used the clean 24 background for these two CMAQ aerosols. All three LBCs show enhanced ASOIL and AORTHJ 25 near the domain's southeastern corner and central Southern boundary. The GLBC-Monthly is the 26 monthly average of GEOS-LBC for July 2015, and has the lowest increments for the two types 27 of aerosols. The two dynamic LBCs, the GEOS-LBC and NGAC-LBC, show the similar aerosol 28 increments over similar locations. However, the NGAC aerosols tended to spread broader than 29 those of the GEOS-LBC, especially for ASOIL, which could reach above the altitude of 10km 30 with concentrations > 5 µg/m 3 (Figure 3e). The NGAC-LBC also showed some signals over the 31 western boundary, where the GEOS-LBC did not show any dust-related aerosols. Another 32 difference between these two LBCs is their ratio of AORTHJ versus ASOIL. The dynamic 33 NGAC-LBC had the higher ASOIL, the coarse-mode dust, than that of GEOS-LBC (Figure 3a,  34 3e), but its AOTHRJ (accumulation-mode dust) was lower than the latter (Figure 3b, 3f), 35 especially over the central southern boundary, where the GEOS-LBC had AOTHRJ up to 30 36 µg/m 3 . It implied that these two global models could have some difference on their dust size 37 distributions, besides their difference on transport patterns due to their dynamics or physics. 38 Figure 4 shows the regional PM2.5 comparisons with the observations from the U.S.EPA 1 AIRNow stations. The CMAQ_Base represented the clear background situation, which 2 obviously missed this dust intrusion event, and underestimated the PM2.5 over Southern and 3 Southeastern USA. The two dynamical LBCs, GEOS-LBC and NGAC-LBC, well captured the 4 intrusion signals and yielded the best results. Their performance were similar in Florida, which 5 was much better than the CMAQ_BASE, but still underpredicted the PM2.5 over central Florida. 6 Over Texas, the further downwind region of this dust intrusion, the GEOS-LBC yielded broader 7 and higher PM2.5 increments than that of the NGAC-LBC, and agreed better with observations, 8 though it had some overprediction over Northern Texas. The monthly averaged GLBC-Monthly 9 had moderate PM2.5 enhancement and still underestimated the dust intrusion, ranking between 10 the CMAQ_BASE and two dynamic LBCs. Figure 5 shows a similar story for the scenario of 3 11 days later. The GEOS-LBC yielded the best overall results, though it still underpredicted the 12 PM2.5 over Florida and Northern Texas. Figure 6 illustrates the time-series comparison for this 13 dust intrusion case over Florida and Texas. In general, the performance ranking of these 14 simulations had GEOS-LBC > NGAC-LBC > GLBC-Monthly > CMAQ_Base, except the 15 NGAC-LBC's underprediction over Florida in June. Even though these dynamic LBCs had 16 overall better results than the static LBCs, they still missed some intrusion peaks, such as June 17 30 th over Texas, and had some inconsistent time-variation patterns compared with the 18 observations, e.g. July 1 st over Florida, and July 8 th over Texas. The two dynamic LBCs had 19 similar performance over Florida in July. However, in the further downwind area, such as Texas, 20 the GEOS-LBC showed better result than that of the NGAC-LBC. These model-observation 21 comparisons showed the advantage of the dynamic LBCs for capturing intrusion events. It 22 should be noted that the PM2.5 spike at July 4 th night (July 5 th in UTC time) was not related to 23 the dust intrusion, but caused by firework emissions at night for the Independence Day, and that 24 emission was not included in our anthropogenic emission inventory. 25

The Wildfire Event in Summer 2015 26
During the same period of summer 2015, a wildfire event occurred in Canada and the biomass 27 burning plume was transported to the United States and affected the Northern USA, as shown in 28 Figure 1. Differing from the dust storm intrusion that mainly affected the particle matter (PM) 29 concentrations, the biomass burning plumes also included gaseous pollutants, such as enhanced 30 level of CO, NOx, and volatile organic compounds (VOCs), which could contribute to the 31 photochemical generation of ozone. For aerosol species, the biomass burning airmass were 32 mainly represented with the enhancement of elemental carbon (EC) and primary organic carbon 33 (POC), or AECJ and APOCJ in CMAQ (Table 3). Figure 7 shows a snapshot of the LBCs along 34 the domain boundaries for AECJ+APOCJ and CO. The GEOS-LBC showed the highest aerosol 35 and CO concentrations with AECJ+APOCJ up to 300 µg/m 3 , and CO up to 3000 ppbV along the 36 domain's northern boundary. Another important feature of GEOS-LBC was that its CO 37 enhancement appeared at elevated altitudes up to 12km (Figure 7b). The monthly averaged 38 GLBC-monthly showed the similar features to the GEOS-LBC, but with much lower 39 concentrations (Figure 7c, 7d). The NGAC-LBC had the similar AECJ+APOCJ profiles to 1 GLBC-monthly, and it used the static profile CO boundary condition (same as the CMAQ_base) 2 that did not reflect the wildfire influence (Figure 7e, 7f). 3 As enhanced gaseous pollutants brought by the full-chemistry LBCs would increase the 4 photochemical generation of ozone, the higher ozone also appeared along the northcentral 5 boundary ( Figure S1a, S1b), where the GEOS-LBC showed 10 ppbv or higher O3 concentration 6 below 4km more than those in the static NGAC-LBC or CMAQ_Base ( Figure S1c). The wildfire 7 induced ozone enhancement appeared not only in the lower troposphere, but also at higher 8 altitudes, e.g. 11km, where the high ozone did not solely come from the stratosphere ( Figure  9 S1a). Figure S2 showed the other species from GEOS-LBC, in which the short-lived NOx had < 10 1 ppbv increment ( Figure S2a) due to the wildfire intrusion. However, its NOz (sum total of all 11 NOx oxidation products, NOz=NOy-NOx) enhancement could be up to 30 ppbv ( Figure S2b) 12 along the northern boundary around 10-12km altitude, where the CO increment also co-existed 13 ( Figure 7b). NOz is a good indicator for NOx's photochemical formation of ozone (Sillman et 14 al., 1997) and the O3/NOz ratio is used as the ozone photochemical efficiency per NOx. The CO 15 and NOz appearance in the high altitudes reflected that the GEOS injected the wildfire emissions 16 to upper troposphere from the strong fire case. Besides these species, the VOCs also showed 17 increment due to the wildfire plume, such as ethane ( Figure S2c) and HCHO (S2d). HCHO is 18 short-lived species, and an indicator of VOC oxidation (Arlander et al., 1995). Considering the 19 magnitudes of CO, VOC and NOx increments in this LBC, the GEOS-LBC mainly provided the 20 VOC and CO rich airmass with limited NOx to the regional CMAQ model. When this kind of 21 airmass arrived at NOx-rich region, such as the urban areas, it would contribute to the 22 photochemical generation of ozone. 23 than that of GEOS-LBC. Figure 9 shows the similar predictions but for ozone. Again, the 31 GEOS-LBC yielded the highest ozone increment due to its relatively high ozone concentration 32 from the wildfire plume, which, however, still underestimated the ozone over North Dakota 33 (Figure 9b). The monthly mean LBC, GLBC-Monthly, systemically underestimated the ozone 34 over these regions. The CMAQ_Base and NGAC-LBC used the same static gaseous LBC, 35 including that for ozone, and they underestimated more. Since the NGAC-LBC had more 36 wildfire-induced aerosol loading than that of CMAQ_Base, the former's photolysis rate was 37 lower than the latter. As both of NGAC-LBC and CMAQ_Base carried the "clean" airmass with 38 low-concentration ozone precursors over the Northern USA, the photolysis reduction due to 39 aerosols mainly led to the reduced ozone's photolytic destruction, such as O3 →O 1 D + O2 or O3 1 → O 3 P + O2, instead of its photochemical generation. For the same reason, ozone's lifetime in 2 winter is longer that in summer (Janach, 1989). On the contrary, over polluted regions, the 3 photolysis reduction would cause lower ozone concentration by limiting its photochemical 4 production. Overall, this effect of photolysis rates on ozone was relatively small. Figure 10  5 shows the time-series comparison over the Northcentral and Northeastern USA for PM2.5 and 6 ozone. Except the systematic PM2.5 underestimation on the night of July 4 th due to the missed 7 firework emissions, the GEOS-LBC showed better PM2.5 prediction than the others, especially 8 from June 29 to July 2 over Northern USA. It should be noted that this run was still not perfect, 9 as it underestimated PM2.5 in the further downwind, the Northwestern USA. The GEOS-LBC 10 also better captured the peak ozone concentrations, e.g. July 1 st and July 2 nd , though it sometimes 11 overpredicted ozone especially during nighttime. The small ozone difference between the 12 CMAQ_Base and NGAC-LBC reflected the impact of wildfire aerosols on photolysis rates, 13 which was very small with regional averages < 1 ppbv throughout this period (Figure 9c, 9d). 14 showed significant improvements for almost all scores over these regions as compared to the 18

Statistics and Discussion 15
CMAQ_Base. The GLBC-Monthly was also better than the base case, though its improvement 19 on correlation coefficient R and index of agreement (IOA) was relatively moderate compared to 20 the dynamic LBCs, as its time-averaging removed the temporal variations. For the further 21 downwind regions of the intrusion events, the LBCs' improvement depended on the regional 22 characteristics of pollutant concentrations. For instance, since the Rocky Mountain region was 23 relatively clean due to its low local PM sources, the external influence weighed more, and the 24 LBCs also showed more significant impact there. Over more polluted regions where relatively 25 strong local PM sources existed, such as Pacific Coast and Northeastern USA, the LBCs mainly 26 changed the background concentration for PM2.5, and their impact on R or IOA were very 27 limited. Overall, the GEOS-LBC yielded the best prediction by reducing the mean bias (MB), 28 root mean square error (RMSE) and increasing the R and IOA. The other dynamic LBC, NGAC-29 LBC, ranked second. All these LBCs showed better performance than the base case for PM2.5 30 prediction. 31 Table 5 shows the similar statistics for ozone. It should be noted that the CMAQ_Base had a 32 systemic O3 overprediction, especially over the Southcentral region, which affected the 33 improvement of LBCs. Differing from PM2.5, ozone had strong diurnal variation during the 34 summertime, which made the LBCs' impact on R and IOA less significant. It should also be 35 noted that the NGAC-LBC did not change any precursor concentrations related to ozone 36 production, and just affect the ozone formation by reducing photolysis rates. Therefore, as 37 compared to CMAQ_Base, the NGAC-LBC had very weak influence on O3 by generally 38 reducing the regional O3 by around 0.2 ppbV, and had almost no impact on R or IOA. The 39 GEOS-LBC tended to increase ozone concentrations in most regions, except the Southcentral 1 USA, where the GEOS-LBC showed general improvement for all scores. It had the weakest 2 impact on ozone over Pacific Coast and Rocky Mountain regions, or the farther downstream 3 areas. The GLBC-monthly had the highest ozone increment over most region except the 4 Southcentral, and its RMSE was also slightly higher. This result showed that removing temporal 5 variation of a LBC might not affect ozone prediction linearly. Except the mean bias, the GEOS-6 LBC got better scores over most regions, though the improvement on O3 was not as significant 7 as that on PM2.5. As discussed above, the LBC impact on ozone inside the domain was realized 8 through changing inflow concentration of O3 itself and/or O3 precursors, such as NOx, VOC or 9 CO. The distance or depth of LBC's effective impact from the inflow boundary depended on the 10 lifetime of these species. All these species have longer lifetime in winter than those in summer. 11 Our other study showed that the LBC's impact on ozone in winter was stronger than that in 12 summer. 13 Figure 11 further illustrated the impact of LBCs (using GEOS-LBC as an example) on prediction 14 statistics and their relations to the distance from the domain boundary during the intrusion 15 events: Southern USA for the Saharan dust intrusion (Figure 11 a,b), and Northern USA for the 16 wildfire intrusion case (Figure 11 c,d). As discussed before, the CLBC has two roles in the 17 regional predictions: provide a constraint for background concentrations, represented by the 18 mean biases, and introduce the variational external influence, represented by the correlation 19 coefficients. Both the background and the variation of CLBC affected the RMSE of predictions. 20 Over the Southern USA, the Saharan dust storm intruded through States of Texas and Louisiana, 21 or -100°W to -86°W, and moved northwardly ( Figure 4). Figure 11a showed that the GEOS-22 LBC's improvement on the correlation coefficient R for the PM2.5 prediction reached the 23 highest near the southernmost near-boundary region, and gradually reduced along the latitude for 24 the inland region. On the other hand, the corresponding MB improvement for PM2.5 did not 25 show significant reduction along the distance from the influenced boundary. The second role of 26 the CLBC, constraining background concentrations for PM2.5, can affect farther inside of the 27 domain. The PM2.5 RMSE change reflected the combined changes of MB and R, and its 28 improvement brought by the GEOS-LBC also reduced along the distance from the influenced 29 boundary since the MB improvement did not vary much and the change trend of RMSE mainly 30 followed the change of R along the latitude. The spatial variations of O3 statistics differed 31 obviously from those of PM2.5 statistic (Figure 11b), and the most significant R's improvement 32 for O3 was not near boundary, but in the some middle latitudes (29°N to 32°N) before reducing 33 in the farther inland. Its MB and RMSE improvements had the similar spatial variations, and 34 they were the highest near the boundary and reduced along the latitude increment. One reason for 35 this difference between PM2.5 and O3 statistics is that the O3 usually has stronger local diurnal 36 variation in summer driven by the photochemical activities, and that influence on R could be 37 stronger than the external influence over polluted areas. So, for this event in which O3 was not 38 the key species, the GEOS-LBC's influence on O3 prediction was mainly about changing its 39 background concentration. Figure 11b also showed that the O3 MB of the GEOS-LBC run could 40 change from lower to higher than of that the reference run (CMAQ_base) along with the 1 latitudinal increment. Although the ozone concentration of the GEOS-LBC over the south 2 boundary was lower than that of the CMAQ_base in low altitudes, it had higher ozone values in 3 the altitudes higher than 14000 m ( Figure S1). That high ozone concentration could reach surface 4 after a certain distance of downward transport in the model system with strong vertical mixing 5 (Tang et al., 2009), which resulted in the higher ozone MB of the GEOS-LBC over the deeper 6 inland region. 7 For the wildfire intrusion event over Northern USA, the PM2.5 statistical difference between 8 GEOS-LBC and CMAQ_Base showed the similar spatial distribution to the dust intrusion event: 9 R and RMSE improvements of the GEOS-LBC appeared the most significant near the boundary, 10 and reduced along the distance from the boundary, and its MB difference could be maintained 11 deeper inland (Figure 11c). For the O3 statistic, the difference between GEOS-LBC and 12 CMAQ_Base became more complex as the wildfire plume also contained the intrusion influence 13 for O3 and its precursors. The GEOS-LBC run generally yielded higher O3, which exaggerated 14 the existing overprediction bias near the boundary, but helped correct the underprediction bias 15 when moving farther inland (Figure 11d). The biggest difference of O3 MB also appeared in the 16 middle latitude as the O3 precursors brought by the full-chemistry LBC took time to contribute to 17 O3 photochemical formation. The spatial variation of O3 RMSE difference was similar to that of 18 O3 MB except for the farther inland region with latitude < 43°N where the GEOS-LBC did not 19 improve the RMSE. The similar issue also appeared for the R difference for the region south of 20 46°N, implying that the wildfire plume represented by the GEOS-LBC could introduce some 21 spatial or temporal biases for O3 precursors. 22

23
The dynamic LBC, such as GEOS-LBC, showed overall better prediction for the intrusion events 24 by capturing the external influence at right time over right locations. However, this full-25 chemistry LBC sometimes is not easy to obtain, especially for the near-real-time forecast. Its 26 event-depended emissions, such as the wildfire emission, also need some time to get relatively 27 accurate estimation, and their impacts on regional domain could lag behind the scene for the 28 forecast. In order to get the intrusion influence when the real-time LBC was not available, we 29 tested the method of developing an alternative LBC based on the historical data with certain 30 indicators. Here we focused on the wildfire intrusion, since it was more difficult to capture the 31 sudden outbreak of wildfire signal than the long-range transport dust intrusion. In addition, the 32 operational NGAC dust forecast has been available to NAQFC (Wang et al, 2018).  Figure S3  37 showed the comparison of extracted VIIRS AOT versus GEOS CO and EC column loading 1 along the northern boundary for June-July, 2015, with their correlation coefficients R > 0.5. The 2 regression relationship derived out of Figure S3 can then be used to resample the historical 3 GEOS LBC data to derive a new LBC for wildfire intrusion events when the corresponding AOT 4 is available. The domain's northern boundary was relatively clean in most time of the summer, 5 unless the wildfire events occurred. During the June and July 2015, the VIIRS AOT data was 6 available once or twice per day around local noontime under cloud-free condition. To get more 7 VIIRS AOT data along the northern boundary, we relaxed the influencing distance up to 300 km 8 when pairing the VIIRS AOT geolocation and the northern boundary location with the nearest 9 neighbor method. In this study, we paired the GEOS's northern LBC (NLBC) for 18UTC with 10 the daily VIIRS AOT along the same location, and made an average of the whole column with 11 AOT interval of 0.2 to build a LBC database sorted in AOT. We only chose to resample the LBC 12 for primarily emitted species from the wildfire sources, including POC, EC, CO, NOx, and two 13 NOz species: PAN and HNO3, but did not include the ozone LBC. When the VIIRS AOT for the 14 new events are available for NLBC, the whole-column species concentration data from that 15 database are chosen to form the new LBC based on the VIIRS AOT value in the nearest 16 neighbor. 17

A Case Study with VIIRS AOT Derived LBC in August, 2018 18
In the middle-later August 2018, a wildfire occurred in western Canada. Figure S4 showed that a 19 high-pressure system controlled the western Canada, and the dry weather made the wild fire 20 easily to spread. There were prevailing northern or northeastern wind, which brought the fire 21 pollutants southward to affect the northwestern and northern U.S. states. Figure 12a shows the 22 VIIRS AOT for this event with the high AOT appearing in the western Canada, the main source 23 region, and the Northern and Northwestern USA. We used this AOT data to derive the new LBC 24 along the northern boundary (Figure 12b, c) for CO and wildfire emitted aerosols 25 (AECJ+APOCJ) by resampling the historical GEOS-LBC database from the Jun-Jul, 2015 26 period. This AOT derived northern LBC (AOT-NLBC) was updated once per day due to the 27 VIIRS data availability, while its western, southern, and eastern boundaries came from the 28 climatologic monthly-mean GEOS-LBC (averaged from 2011 to 2015). The AECJ+APOCJ 29 increment of the AOT-NLBC mainly existed below 3km, but its CO enhancement could reach up 30 to the altitude of 10km, due to the elevated CO plume in the original GEOS-LBC, e.g. Figure 7b. 31 Figure 13 shows the surface ozone and PM2.5 over this region one day later (08/17/2018). The 32 CMAQ_Base underpredicted both species over this region, while the AOT-NLBC greatly 33 reduced the underprediction with increased background concentration from the northern 34 boundary. Since the AOT-NLBC did not include dynamic ozone LBC, its enhanced ozone 35 concentration was mainly brought by the CO and NOx increments from the northern boundary, 36 which sometimes caused the overprediction over further downwind areas, such as Colorado. 37 Overall, the AOT-NLBC showed better PM2.5 prediction over Southwestern Canada and 38 Northwestern USA with its higher background concentrations. 39 Figure 14 shows the corresponding time-series comparison over EPA region 8 (states of 1 Montana, North and South Dokotas, Wyoming, Colorado, and Utah) and region 10 (states of 2 Washington, Idaho, and Oregon). Both observed and predicted ozone showed strong diurnal 3 variation. The AOT-NLBC showed better skill on capturing daytime ozone maximum, and was 4 about 5-10 ppbv higher than the CMAQ_base prediction, though it tended to overpredict ozone 5 at night, especially over the region 8. For PM2.5, the observation clearly showed two peaks 6 related to wildfire plumes over two regions: 08/19-08/21 and 08/24-08/25 for EPA region 8; 7 08/14-08/17 and 08/19-08/22 for EPA region 10. Without the boundary influence, the 8 CMAQ_Base missed all these PM2.5 peaks even though it had the same inside-domain wildfire 9 emissions. AOT-NLBC successfully captured these intrusion signals, though overpredicted 10 PM2.5 before 08/18 over EPA region 8. This result showed that the alterative LBC could be 11 useful for capturing the key intrusion signals in case the global LBC was not available. This 12 alternative approach was especially important for the forecast as the satellite AOT can be 13 obtained in near-real-time. In this case study of summer 2018, the wildfire events were similar to 14 the wildfire case occurred in summer 2015, which made the quantitative derivation of LBC 15 possible. 16

17
In this study, we examined the influence of the CLBC on our regional air quality prediction, 18 verified with surface ozone and PM2.5 monitoring observations. We developed the full-19 chemistry mapping table from the global model GEOS to CMAQ's CB05-Aero6 species. The 20 GEOS dynamic LBC showed the overall best score when comparing with the surface 21 observations during the June-July 2015 while Saharan dust intrusion and Canadian wildfire 22 events occurred. The base simulation (CMAQ_Base) ranked last as it missed all these external 23 influences. The NGAC-LBC only considered the GOCART aerosols, and had the good 24 performance for capturing the dust storm intrusion but missed the ozone enhancement due to the 25 Canadian fire events. The LBC influences on the model performance depended on not only the 26 distance from the inflow boundary but also species and their regional characteristics, as the 27 LBCs' influence on ozone and PM2.5 differed significantly. During the studied events of 28 summer 2015, The CLBCs affected both PM2.5 mean background concentration and its 29 temporal/spatial variation. Their influences on PM2.5's correlation coefficient R mainly 30 appeared near inflow boundary, and reduced along with the distance from the boundary. 31 However, their influence on PM2.5 background concentration could be kept in the further inside 32 domain. The CLBCs' influence on ozone could be more complex, and affected by the boundary 33 inflow of ozone and/or its precursors, and downward transport from the upper troposphere. In 34 this study, the influences with temporal/spatial variation were mainly shown in the aerosol 35 dynamic LBC, e.g. the GEOS-LBC or NGAC-LBC. All other LBCs mainly changed the 36 background concentrations and shifted the mean bias of the corresponding predictions. 37 The AOT-derived LBC can be used as an alternative method to capture the intrusion when a 1 reliable dynamic LBC is not available. Although the VIIRS AOT was updated only once per day 2 and the derived LBC had noisy spatial distribution, this method still showed its value to replace 3 the static LBC in the air quality forecast. It should be noted that other indicators, such as surface 4 monitoring data, can be also used to derive the similar LBC if the historical LBC has good 5 correlation with these data and there are relatively dense station available near the inflow 6 boundary. Geostationary satellites can achieve a near-real-time AOT retrieval in time interval of 7 several minutes, which will provide a better solution for fast capturing the intrusion signals. 8 Currently the main issue for using geostationary AOT is their relatively poor retrieval quality 9 over high latitude or under high zenith angles. Once that issue gets resolved, its AOT can be used 10 as an indicator to derive the LBC or even replace the LBC provided by the global models. 11