To simulate ozone variability over a complex coastal urban environment, a high-resolution air quality forecasting system has been developed for LIS and surrounding areas. The forecasting system is comprised of state-of-the-science weather, emission and chemical transport models. The model domain covers eastern Pennsylvania, New Jersey, southern New York, Connecticut and Rhode Island. While this model domain is large enough to capture key physical and chemical processes within the LIS area, such as sea breeze circulation and photochemistry, the influence of regional transport outside this domain cannot be adequately represented. Therefore, real-time forecasts from the operational NAQFC (Lee et al., 2017), produced by the NOAA National Weather Service, are used to provide dynamic boundary conditions to investigate the effect of this model input on forecasting performance. We also explore the effects of emission adjustment and chemical data assimilation on forecasting performance.

Five groups of simulations are designed to evaluate the performance and effectiveness of different adjustments of the CMAQ model (Table 1). The first group (Control run) applies no adjustment, using the default profile as lateral boundary conditions (LBCs). It serves as the reference case to allow the effectiveness of each adjustment method to be quantified. The second experiment, named BCON, is similar to the Control run, except that dynamic boundary conditions from the NOAA NAQFC with a horizontal resolution of 12 km were applied to replace the default BCs. In the optimal interpolation (OI) run, the initial concentrations in CMAQ are adjusted with three observation interpolation methods, including area-average (OI_avg), inverse distance weighting (idw) and CMAQ concentration gradients (OI_bias) (details of each OI approach provided in Sect. 2.3.2). The best performer of these approaches will be used in the subsequent analyses. Next, a group of emission adjustment experiments are designed to update NOx emissions using observed changes from satellite and ground sensors (Tong et al., 2016). These emission adjustment factors are applied either uniformly across the domain (EmisAdj_whole) or separately for each subdomain (EmisAdj_sub). In the latter case, the domain was divided into five regions based on city areas: New York City (NYC), City of Philadelphia (PH), New Haven–Hartford (NHH) and Providence–Pawtucket (PP) and the areas other than these four regions (OTHR) (Fig. 1). Finally, three simulations with the combination of these three techniques were conducted in search of the best performer. All simulations were conducted for a high-ozone episode, which lasted 168 h from 00:00 UTC on 25 August to 23:00 UTC on 31 August 2018.

The high-resolution air quality forecasting system used here is a new research prediction system deployed during the 2018 LISTOS field campaign period which is comprised of three major components: meteorology, emission and chemical transport models. The Weather Research and Forecasting (WRF) model version 4.0 (Skamarock et al., 2019) is used to generate hourly meteorological fields to drive emission and air quality modeling. The WRF model was configured with the Thompson graupel microphysics scheme, the RRTMG long- and short-wave radiation scheme, the Mellor–Yamada–Janjic PBL (planetary boundary layer) scheme, the unified Noah land-surface model and the Tiedtke cumulus parameterization option. No data assimilation was applied in the WRF simulation. The model is conducted in a single domain with 132 × 122 grid cells, with one grid more on each boundary compared to that of the chemical transport model. There are 41 vertical layers with 20 layers below 1 km and a top layer at 50 hPa. The forecast fields of Global Forecast System (GFS) version 4 products with a horizontal resolution of 0.25∘ × 0.25∘ (available every 6 h) were employed to drive the WRF model.

The emission input was provided using a hybrid emission modeling system that utilized the Sparse Matrix Operator Kernel Emissions (SMOKE) model (Houyoux et al., 2000) version 4.7 to process anthropogenic emissions, and a suite of emission models to estimate emissions from intermittent and/or meteorology-dependent sources. Anthropogenic emissions from area and mobile sources were taken from US EPA 2011 National Emissions Inventory (NEI) version 2 (NEI2011v2). The Motor Vehicle Emissions Simulator (MOVES) was used to generate county-level emission factors for the on-road and off-road sources. SMOKE uses a combination of vehicle activity data, MOVES emission factors, meteorology and other ancillary data (spatial, temporal and speciation information) to generate hourly speciated model-ready emission data. Point sources were processed in two steps. In the first step, emission inventories of point sources were processed with SMOKE to generate intermediate input files. Next, these intermediate files were used to drive inline calculation of plume rise to distribute point source emissions vertically in the CMAQ model domain. Two natural sources are included in this forecasting system: biogenic and sea salt. Biogenic emissions from terrestrial plants were predicted using the inline version of the Biogenic Emission Inventory System (BEIS) (Pierce et al., 1998). The emissions of sea spray aerosols are calculated using an updated version of the Gong (2003) sea spray emission parameterization (Gantt et al., 2015).

The CMAQ model ingests emissions and meteorology to predict spatial and temporal variations of O3, NO2 and their precursors. In this study, version 5.3.1 of the CMAQ model was configured to include detailed implementation of inline emission processes for biogenic, sea salt and elevated anthropogenic emissions, horizontal and vertical advection, turbulent diffusion, dry/wet deposition and full gas, aqueous and aerosol chemistry using a revised Carbon Bond 6 gas-phase mechanism and the AE6 aerosol mechanism (CB6r3_AE6_AQ) (Byun and Schere, 2006; Luecken et al., 2019). Both the meteorological and air quality models have a 3 km horizontal resolution over the LIS region and its surrounding areas (Fig. 1).

We implement and test three forecasting improvement techniques to assess their effectiveness in enhancing the simulation performance of the CMAQ model. Details of each update are described below.

In this study, a suite of observational datasets was used either as inputs for emissions and chemical data assimilation or to evaluate model performance. These datasets include surface O3 and NO2 measurements from the US EPA Air Quality System (AQS) surface network, the NO2 vertical column density (VCD) from the OMI satellite data, NO2 VCD from the GeoCAPE Airborne Simulator (GCAS) on the NASA Langley Research Center B200 aircraft and the O3 vertical profile from the NASA Langley Mobile Ozone Lidar (LMOL). Detailed information of each dataset is provided below.

Surface concentrations of O3 and NO2 are used for emission adjustment and chemical data assimilation, as well as evaluation of model performance. AQS is a routine monitoring network established to collect ambient air pollution data in urban, suburban and rural areas. AQS monitors determine O3 concentrations according to the Federal Reference Method promulgated in the 2015 revisions to the National Ambient Air Quality Standards (Long et al., 2014) and NOx concentrations using the chemiluminescence instruments described by McClenny et al. (2002). AQS measures both O3 and NO2 at hourly intervals. Note that NO2 measurements are typically biased high due to interference in the chemiluminescence measurement (Dunlea et al., 2007). As the goal of this study is to improve forecasting performance, a near-real-time version of the AQS data was used, called AirNow. This is a preliminary dataset for the purpose of real-time air quality reporting and forecasting; it is not fully verified and provides fewer measured species. The data used in this study are downloaded from the AirNow data portal maintained by the US EPA.

NO2 VCD measurements were provided by the Ozone Monitoring Instrument (OMI) standard product (version 4), available from the NASA Goddard Earth Sciences Data and Information Services Center (GES DISC). OMI is a nadir-viewing hyperspectral imaging spectrometer that measures solar backscattered radiance and solar irradiance in the ultraviolet and visible regions (270–500 nm) (Levelt et al., 2006). The Aura spacecraft has a local equator-crossing time of 13:45 in the ascending node. OMI views the Earth along the satellite track with a swath of 3600 km on the surface in order to provide daily global coverage. In the normal global operational mode, the OMI ground pixel at nadir is approximately 13 km × 24 km, with increasing pixel sizes toward the edges of the orbital swaths. Multi-year OMI NO2 data are further aggregated to calculate state-level emission adjustment factors using a mass conservation approach (Tong et al., 2015).

The high-resolution NO2 observations from the GCAS (Kowalewski and Janz, 2014) are used for a direct comparison against model simulations of the NO2 VCD. GCAS is an ultraviolet–visible spectrometer used in air quality field studies to map the spatiotemporal distribution of NO2 and HCHO VCDs at high spatial resolution (Nowlan et al., 2018; Judd et al., 2020). During LISTOS, this instrument flew on 11 flight days collecting between two and four gapless raster datasets at spatial resolutions for NO2 as fine as 250 × 250 m. More information about the retrieval can be found in Judd et al. (2020). During LISTOS, NO2 from GCAS was validated using coincident Pandora measurements and had a median percent difference of -1.2 %, with 95 % of the most temporally homogeneous points within ± 25 % or 0.1 DU.

Finally, O3 vertical profiles from the NASA LMOL are used to evaluate the CMAQ prediction of O3 profiles during the LISTOS field campaign. LMOL is part of a NASA-sponsored ozone lidar network called the tropospheric ozone lidar network (TOLNet; Sullivan et al., 2017), which is a mobile ground-based ozone lidar platform equipped with a pulsed UV laser and all associated power and lidar control support units (De Young et al., 2017; Gronoff et al., 2019). In this study, we use LMOL lidar observations at the Westport site (41 118∘ N, 73 337∘ W). All available field measurement data during this campaign were obtained from the LISTOS Data Archive (available at https://www-air.larc.nasa.gov/missions/listos/index.html, last access: 1 November 2021).

In this section, we examine the effects of using the dynamic boundary conditions on O3 and NO2 predictions. As a reference, we also compare these simulations to the NAQFC results, extracted for the same region, during the 29 August high-ozone event. Figure 2 shows the O3 and NO2 24 h average concentrations simulated by Control (static BCs), BCON (dynamic BCs) and the NOAA NAQFC over the LIS region. Compared to the underestimated O3 concentrations simulated by Control run, the concentration level using dynamic boundary conditions increases considerably and is closer to the observations. High O3 concentrations appear over near-coast areas but are lower in the northwest of the domain. This spatial pattern illustrates the ozone river in a northeastward direction along the I-95 corridor, extending from Philadelphia to NYC and then to Connecticut, where the worst air quality is often observed. Although it overestimates surface O3 in Philadelphia and central New Jersey, the BCON simulation can reproduce O3 hourly variations during this episode well in comparison with the observed data (see the time series in Fig. 2d). Note the peak O3 simulated in the Control run is nearly the same on all days during the simulation period. The comparisons between the peak O3 with the default profile and dynamic LBC case indicate relatively large regional contributions on these days. Compared to the Control run, the BCON run performed better not only in bias, but also with higher correlation coefficients between prediction and observations (Table S2), especially during the 26–27 August high-O3 days. As the profile BCs are static and lack spatial–temporal variations, the Control run mainly reflects the local contributions of emissions, transport and chemical processes within the domain (Tang et al., 2007). The underprediction suggests that these processes are insufficient to produce the observed O3 levels and that the transport of air pollutants from upwind is important to predict the high-O3 episodes. It highlights the significant influence of dynamic BCs on the simulations over this region during high-pollution events. In comparison, the influence of BCs is less important during the cold season. There is a smaller difference between upwind concentrations and the background concentrations used in the default BCs, compared to that during a hot season when the upwind photochemical production is more active, resulting in better agreement between the prediction and observations (Fig. S2a, d). Note that other studies have shown the influence of BCs can become more prominent during the cold season when “local” pollution production is slow (e.g., Fiore et al., 2009). The magnitude of the actual influence is determined by several factors, such as emission density, photochemical production and sink, and spatial distribution and gradients of the concerned species.

The performance of the high-resolution simulation is next compared to that by the NAQFC. The NAQFC simulation, which has been used to provide national numerical guidance for O3 and PM2.5 (Lee et al., 2017), is run at a coarser resolution (12 km), using a different CMAQ version (a revised CMAQ5.0), driven by different emission and meteorology datasets. Regardless of these differences, the NAQFC and BCON runs predict similar surface O3 distribution patterns. Compared to that in the NAQFC prediction, the O3 prediction from the 3 km BCON run demonstrates more detailed spatial distribution. For instance, the O3 concentration over the Long Island Sound is lower than its surrounding areas during this episode, which is better resolved by the 3 km simulation than the 12 km NAQFC (Fig. 2a–c). The O3 distribution along the coastal area, such as the coasts of Connecticut and Rhode Island, also agrees better with the observations than the 12 km NAQFC prediction. This suggests that the high-resolution simulation can better reproduce the pollutant variability over this coastal urban area during this episode. In addition, the BCON run performs better over southern New Jersey, and northeast of the LIS domain, with considerably reduced biases in the LIS downwind areas as well. As for the diurnal variations, the BCON run overestimates the peak O3 concentrations on 28 and 29 August, while the NAQFC run performs well and is closer to the measurements (Fig. 2d). The use of coarser resolution NAQFC predictions as BCs substantially improves the capability of the 3 km forecasting system to reproduce the O3 variability. Compared to the Control run, the correlation coefficient between BCON and observed O3 concentrations increases from 0.81 to 0.93, and the root mean square error (RMSE) decreases from 14.97 to 8.22 ppbv with a reduction of 45 %, resulting in a comparable performance with the NOAA NAQFC predictions with correlation of 0.91 (Table S2).

The spatial patterns of predicted NO2 concentrations from the Control, BCON and NAQFC runs are quite similar, with high values over the NYC area (Fig. 3). The simulated NO2 concentrations by the 3 km forecasting system, either with static or with dynamic BCs, agree better with the observations than those from the 12 km NAQFC simulation, highlighting the importance of using high-resolution inputs to better represent the emission sources in the model. The correlation coefficient and RMSEs are 0.69 and 4.12 ppb for the Control run, 0.71 and 3.82 ppb for the BCON run, and 0.67 and 4.98 ppb for the NAQFC run, respectively (Table S2). In addition, the improvement of simulated NO2 using dynamic BCs was much smaller compared to that of O3. This is because the lifetime of NO2 is relatively short (1–7 h in summertime; Lu et al., 2015), and its budget in urban areas is mainly influenced by local emissions and chemistry and less by regional transport, indicating the effectiveness of dynamic BCs depends not only on the downwind/upwind gradients, but also on the lifetime of concerned species.

Initial concentrations are an important input to air quality forecasting. Adjusting initial concentrations through chemical data assimilation has been shown to significantly improve air quality forecasting (Tang et al., 2015; Chai et al., 2017), although the impacts wane with increasing forecast length. Here we compare the results using various OI methods with the simulations without any BC adjustment (same as the Control run) and study the effects of adjusting initial conditions on O3 and NO2 prediction. Figure 4 illustrates the initial concentrations of surface O3 adjusted by OI_avg, OI_idw and OI_bias, respectively. In the initial concentrations, the areas influenced by OI_avg are primarily limited to the ground-based sites and the regions within five model grid cells in each direction of the observations compared to the Control run (Fig. 4a, b). The rest of the domain is not affected by the adjustment, resulting in significant differences between adjusted and unadjusted areas. The O3 fields adjusted by OI_idw (Fig. 4c) and OI_bias (Fig. 4d) show similar horizontal distributions, but the concentration level of OI_bias is relatively higher over NYC and northern New Jersey. Furthermore, in contrast to the localized changes by OI_avg, those of OI_idw and OI_bias show smoother changes over larger parts of the domain.

Next, the initial concentration files after adjustment are used to feed CMAQ simulations. The O3 prediction by the Control run and three OI runs at 00:00 UTC on 26 August 2018 (the first hour after OI adjusting) is depicted in Fig. 5. The adjusted O3 fields show different patterns compared to that in the Control run with no IC adjustment. The predicted O3 field with the OI_avg method shows a distribution with localized high-value areas near the observational sites. As for the other two OI methods, the distribution using OI_bias has similar patterns with that of OI_idw, while the concentrations over the high-O3 area are further elevated. Biases between observed and predicted concentrations are reduced in most areas. The statistical metrics calculated from hourly simulated and observed data from 26 to 31 August 2018 are reported in Table 2. The RMSEs for O3 are reduced from 14.97 ppbv in the Control run to 13.72 ppbv in the OI_bias run, to 13.79 ppbv in the OI_idw run and to 14.30 ppbv in the OI_avg run. The correlation for O3 also slightly increases from the Control run to the OI runs (Table 2). In comparison, NO2 prediction is less influenced by this adjustment, with insignificant changes in the model performance (Table 3). In addition, the effects of this adjustment on the modeling results decrease with the simulation time and display no discernible difference from the Control run after 12 h (Fig. S1 in the Supplement). Generally, the improvement of the simulated results due to OI data assimilation over the study domain is smaller than that from the dynamic BCs. Among the three OI methods, the simulation with OI_bias shows the best performance, so this method is chosen for subsequent analyses, in which multiple techniques are combined to improve forecasting skills.

The ICs for each day were adjusted by OI using real-time observations; it is interesting to note that the duration of OI influence on O3 simulation varies from place to place. Figure 6 shows the time series of the averaged differences in predicted hourly O3 concentrations between the Control run and each of the three OI runs from 26 to 31 August 2018 in three urban areas (NYC, Philadelphia, New Haven–Hartford) and other (OTHR) areas. The differences illustrate the effect of adjusting initial concentrations on O3 prediction. In large metropolitan areas, OI adjustments result in spikes that indicate larger model errors at the time of OI adjustment, with the mean errors up to 14 ppbv in surface hourly O3 concentrations over NYC and 16 ppbv over Philadelphia, respectively. In comparison, the spikes in non-urban areas are much smaller, reflecting the fact that there are smaller biases between observations and predictions (Fig. 6). The New Haven–Hartford region sees a smaller change of O3 concentration compared to between that in large cities. The OI effects in large cities remain for a shorter time than in non-urban area or smaller cities. For example, the differences between OI runs and the Control run decrease to ∼0 ppb in 4 to 8 h in two metropolitan areas, NYC and Philadelphia (Fig. 6a, b). Meanwhile, in the New Haven–Hartford region (Fig. 6c), Providence–Pawtucket region (not shown) and the non-urban areas (Fig. 6d), the differences could last 12 to 16 h. The different durations indicate the influence time of OI-adjusted ICs, not necessarily the improvement in model skill, which is determined by both initial concentrations and other processes (chemical production and transport, etc.). The improvement using OI adjustment is comparable over different subdomains (Table S3). This difference reflects the dependence of O3 level on the initial concentrations in the air quality model. In general, the influence of OI adjustment lingers for a longer period in an area with low emission density, where emissions and chemical reactions make a smaller contribution to the O3 budget than that in the area with high emission density.

One of the major challenges in air quality forecasting is the time lag in updating the emission inputs generated for a specified base year, which is typically different than the year for which the simulation is desired (Tong et al., 2012). Here we test the effects of implementing a new emission update technique, the rapid emission refresh, on forecasting performance. In this study, the NEI2011v2 data are used to represent anthropogenic emissions, while the target forecasting year is 2018. Both the AQS ground monitors and the OMI sensor observed considerable decreases in NOx during summertime (May–September) from 2011 to 2018 (Fig. 7). The largest reduction in ground concentrations appears in the west of NYC. The OMI NO2 observations show an increase primarily over Connecticut and Rhode Island, the region downwind of the Long Island Sound (Fig. 7b). The average AF for the whole domain is -18.6 %. The AFs for each subdomain are -31.9 % for NYC, -12.7 % for Philadelphia, -9.4 % for the New Haven–Hartford region, -28.2 % for the Providence–Pawtucket region and -16.5 % for other regions, respectively. In general, the NOx variations in this study are similar to those between 2005 and 2012 (Tong et al., 2015), indicating that the NOx emissions continued decreasing during the past 14 years. This trend highlights the importance of updating the emissions to the model year, in order to reduce the bias in the emission inputs for model simulations, especially for time-sensitive applications such as air quality forecasting.

The results in Tables 4 and 5 show that the performance for O3 and NO2 prediction is very similar between two simulations using the two emission adjustment methods (a uniform average adjustment factor over the entire domain and spatially varied factors for each subdomain defined in Fig. 1). The correlations in each sub-domains are the same, and the average for both simulations is 0.81 for O3 and 0.69 for NO2, respectively. The biases and errors are also at the same level from the two simulations. Compared to the O3 in the Control run, RMSE changes slightly from 14.97 to 14.71 ppbv (EmisAdj_avg) and 14.55 ppbv (EmisAdj_sub), while the correlation remains the same. The largest differences appeared in NYC with RMSE of 15.54 (EmisAdj_avg) and 14.93 ppb (EmisAdj_sub). This demonstrates that emission adjustment alone results in limited improvement of O3 prediction, due in part to the fact that the O3 production in this region is NOx-saturated (VOC-limited) in urban areas where most AQS monitors are deployed, so the O3 level is less sensitive to the change in NOx emissions. Similarly, the retrievals of satellite observations are also more sensitive to urban plumes. In addition, regional transport of air pollution results in dispersion of emitted NOx and its byproducts/reservoirs. The observations from satellite or ground monitors, based on which the emissions were adjusted, may not accurately capture the temporal evolution of the emission sources. A large geographical range may better reflect the overall changes of NOx emissions in the LIS region. Previous studies either use a coarse model resolution (e.g., 1∘ in Lamsal et al., 2011, or state-level adjustment in Tong et al., 2016). As a result, the simulated concentrations using different methods were very close, and the limited difference can also get averaged out when calculating the averaged statistical metrics. The effect of the emission adjustment method in this study is not as large as BCON or OI adjustments, which directly influence O3 concentrations. A recent study by Jin et al. (2020) showed that the decrease in NOx emissions has shifted the NOx-saturated to NOx-sensitive regime transition zone closer to urban centers, approximately 40 to 60 km from the center (the highest emission point) of New York City. Therefore, it is expected that the effectiveness of emission adjustment will increase over time in this region. For surface NO2, the emission adjustment showed a more significant impact on simulated concentrations. Note that the emission adjustment was only implemented in the LIS system, not in NAQFC, which still uses the 2014 NEI for anthropogenic emissions. Without the emission adjustment, the changes in NOx emissions between the inventory and forecast years are not accounted for. On the high-O3 days, NAQFC overpredicted surface O3 during the study period (Fig. 2c). The NAQFC LBCs are likely associated with a possible overprediction of the regional transport, which can be partially responsible for the BCON LIS simulation overpredicting O3 during high-O3 days (Fig 2d). Considering the similarities of these two emission adjustment methods, they will be both tested in the subsequent multi-adjustment simulations.

After assessing the effects of individual updates, we test how these updates can be combined to optimize forecasting performance. In the preceding sections, three groups of adjustment approaches have been included and evaluated. For each group, the best performing method has been identified, including the dynamic BCs, ICs with OI bias and rapid emission refresh (including EmisAdj_avg and EmisAdj_sub). With these selected updates, we design and conduct two multi-adjustment simulations; the first one used both the dynamic BCs and the OI-bias-adjusted initial concentration files (BO for short), and the other one employed the NOx emission adjustment together with the combination of BCON and OI bias (BOE hereafter). Results of these combined adjustments are compared against the Control, BCON run and the NAQFC prediction.

First, we compare two BOE simulations, one with the EmisAdj_avg emission adjustment and the other with EmisAdj_sub. The statistical metrics of BOE with EmisAdj_avg and BOE with EmisAdj_sub (Tables S4, S5) are quite similar in each sub region and also have the same correlations. On average, the RMSEs of the combined BOE setup using the EmisAdj_avg method are slightly smaller than that using the EmisAdj_sub method during the study period (Tables S4, S5), which is different from that when a single adjustment method was applied (see Tables 4 and 5). Therefore, in the subsequent evaluation we take BOE (EmisAdj_avg) to compare against surface and other observations. Figure 8 compares the predicted hourly O3 and NO2 concentrations against in situ observations from 26 to 31 August 2018 in five subdomains and the overall domain with Taylor diagrams (Taylor, 2001). In the Taylor diagram, the relative skill of each forecasting system to reproduce the O3 and NO2 variability is represented using three statistical metrics: correlation (R) with values on arc of the right angled sector, normalized standard deviation (SD) with values on the y axis and centered root-mean-square difference (RMSD) with values on the x axis. The normalized SD is shown as the dashed-line concentric circles, while RMSD is shown as non-dashed-line concentric circles, with the observation point acting as the center (OBS on the x axis). Their values higher (lower) than 1 indicate biased high (low) of the simulations. In general, the forecasting skill is measured by the distance to the OBS point on these diagrams: the shorter the better. The default (Control) run yielded a correlation coefficient of approximately 0.8 (0.77–0.84) in each subdomain, while those with adjustments show stronger correlations with R all above 0.9. Furthermore, the performance in the OTHER areas is better than that in the five subdomains, with the R value up to 0.97 and SD close to 1 (Fig. 8e). Taylor diagrams also reveal that these adjustments are even more effective over the low emission areas. The three adjusted runs, namely BCON (no. 2), BO (no. 3) and BOE (no. 4) in the diagrams, have reproduced surface O3 concentrations over the NYC region well. The simulations with BOE usually demonstrate a relatively lower O3 concentration level than that with the BCON run or the combined BCON and OI run. This means in the overestimated areas (such as NYC, Fig. 8a), the simulations with emission adjustment show better performance than that without emission adjustment. In addition, these three simulations have similar biases and errors, with NMB ranging from 4 % to 22 % and NME from 15 % to 22 % (Fig. 9a, c). These results illustrate the importance of combining complementary modeling system updates to reduce model uncertainties in a comprehensive way. A single update, such as emission adjustment, may result in a better emission input closer to the “true” level, but its effect can be offset by systematic biases caused by other inputs. Concurrent improvements of boundary conditions and initial concentrations allow for a more realistic initial state and boundary conditions to demonstrate the effectiveness of the emission adjustment in improving O3 forecasting (Fig. 9).

The Taylor diagrams show that the performance of variability of NO2 predictions is generally worse than that of variability of O3 predictions. Overlaid on the same diagrams, the points that represent NO2 performance are all further away from the OBS point compared to that representing O3 from the same simulations (Fig. 8). This is not surprising as O3 has been one of the focal points in air quality modeling in the past decades, while NO2 has not been scrutinized with the same intensity. All of the high-resolution simulations, including the Control run with unrealistic boundary conditions, perform better for NO2 prediction than the NAQFC run (Fig. 9), highlighting the benefit of using a high-resolution modeling system for predicting short-lived chemical species such as NO2. The NAQFC generally underestimates NO2 concentrations in all subdomains. Its bias is the smallest in the NYC subdomain and the largest in the downwind New Haven–Hartford region. The correlation coefficient is between 0.8 and 0.9 in NYC but lower than 0.6 in the New Haven–Hartford region (Fig. 8). Similarly, the NMB is within 10 % in NYC but can be as large as -65 % in the New Haven–Hartford region. Such a contrast suggests either an underestimate of emission sources in Connecticut or an unrealistically short lifetime of NOx due to flawed model chemistry, or a combination of both.

CMAQ predictions of NO2 surface concentrations and vertical column density are compared against ground and aircraft observations. NO2 is not only a key precursor to tropospheric ozone, but also a proxy for traffic-related air pollution in many epidemiological studies (e.g., Jerrett et al., 2007). Within the LISTOS CMAQ domain, there are four active ground monitors with valid NO2 readings during the study period. Hourly variations from AQS monitors, the BOE 3 km prediction and the operational NAQFC prediction are illustrated in Fig. 10. Among these sites, the lowest NO2 concentrations were observed at the Flax Pond site in the middle of Long Island, away from the major emission sources. Both BOE and NAQFC are able to reproduce the magnitude and diurnal variations of surface NO2 concentrations at this site. The NO2 concentration at the Queens College site, also located in the Long Island Sound and downtown NYC, is significantly higher than at the Flax Pond site, due to its close proximity to major sources such as the tunnels, harbors and highways. At this site, the BOE 3 km prediction is considerably better than that from the NAQFC prediction. Similarly, the BOE prediction outperforms the NAQFC at the New Haven site in Connecticut, where the surface NO2 concentration reaches 40 ppbv on 28 August and 55 ppbv on 29 August 2018. The NAQFC-predicted concentration is constantly below 10 ppbv, severely underestimating the observations. In comparison, the BOE-predicted concentrations are much closer to the observations, although still underpredicting the latter. Finally, both models missed the first, primary peak on both days at the Westport, CT site, which is strongly influenced by the NYC plume and sea breeze circulation.

Next, the two model simulations are compared against the NO2 VCD measured by NASA GCAS during the LISTOS field campaign. In order to allow for a comparison between simulations and measurements from GCAS, the CMAQ prediction of NO2 mixing ratio is vertically integrated from the surface to the layer, which is the closest to the plane altitude to generate vertical column density (unit: molecules cm-2), with GCAS data averaged over the 3 km grid to provide a spatially representative observation dataset. We also sample the model data to match the actual measurement time. The GCAS observations show higher NO2 VCD in the morning and lower values in the afternoon. This temporal pattern is well captured by both simulations. The GCAS observations depict an NO2 hotspot over lower Manhattan and Brooklyn, which is reproduced by both BOE and NAQFC (Fig. 11). The observed and simulated VCDs are generally at the same magnitude (4–40 × 1015 molecules cm-2), with BOE better capturing the peak values. Moreover, the VCD prediction from the BOE run presents a northeastward pattern, and it was lower over the water area of LIS than that over surrounding lands. In comparison, the VCD from NAQFC shows a high NO2 plume over the land and the water around LIS. Compared to that from NAQFC, the spatial distribution of NO2 VCD from BOE is more consistent with that of GCAS. This is also the case for the prediction of surface NO2 distributions (Fig. 3), indicating the high-resolution system can outperform NAQFC through resolving the fine-scale processes. It should be noted that the VCD levels from both simulations are biased high outside the high-emission-density areas, especially in the morning. The BOE prediction shows a larger area of high-NO2 VCD than that from GCAS, suggesting either a positive bias in NOx emissions or inefficient transformation and removal of emitted NOx in the CMAQ model. The high-NO2 VCD from the NAQFC simulation is lower than the measurements over lower Manhattan and Brooklyn, and the high-NO2 VCD extends to an area larger than that from both GCAS and BOE. The performance is relatively unsatisfactory during the high-pollution period in the morning of 28 August (Fig. 11e, i), with a correlation of only 0.56 for BOE and 0.44 for NAQFC. These low correlations could be in part caused by the high spatial variability of fine resolution measured VCD, so that the averaged VCD is still more variable than either model. In contrast, the spatial patterns of NO2 VCD in the afternoon are better reproduced than in the morning (Table S6). In addition, the NO2 VCD from the simulation with combined adjustments using the EmisAdj_sub method for emission refresh shows a similar spatial pattern to that using BOE (Fig. S3). The NO2 VCD level, however, is lower over the NYC area, suggesting an underestimate over the hotspot but much better prediction over the rest of the area. Besides the uncertainties in the model, an evaluation conducted by Judd et al. (2020) showed that the absolute difference in GCAS from Pandora measurements has an average and standard deviation of -0.2×1015±2×1015 molecules cm-2 and a percent difference on average of -1.5 % ± 20 %. Overall, the BOE simulation at the 3 km resolution is able to reproduce the observed NO2 VCD, and unlike the results of surface NO2, the NO2 VCD using EmisAdj_sub has lower NMB (33 %) and NME (57 %) compared to that using EmisAdj_avg (40 % and 61 %), while their correlation is still the same (0.74). It indicates the advantage of adjusting emission with a finer spatial resolution in simulating NO2 vertical column in this study. Table S6 shows that both 3 km simulations perform better than the 12 km NAQFC (R=0.57, NMB = 45 % and NME = 76 %, respectively).

One key result expected from the improved prediction system is better prediction of high-O3 episodes, especially those events that cause the exceedance of NAAQS. Here we compare the model performance between BOE and NAQFC at the seven sites where the O3 concentrations exceeded the NAAQS. Compared to NAQFC, BOE demonstrates enhanced prediction skills at all sites (Fig. 12). Note the comparisons may be attributed to the differences in meteorology, emission and other factors. Although it is difficult to attribute the improvement quantitatively to each factor, the magnitude of O3 improvement from the base run to the BOE run is comparable to that of the overall reduced O3 bias, suggesting a significant contribution from these improvement techniques. The results show that BOE can better capture peak O3 values than NAQFC in the afternoon, a highly desired feature in predicting O3 exceedances. Hourly surface O3 concentrations reached more than 100 ppbv at four Connecticut sites, including Greenwich, Westport, Middletown-CVH-Shed and Stratford. While neither BOE nor NAQFC is able to predict such high values, BOE reduces the bias by 10–20 ppbv during peak hours at these sites. The improvement of peak O3 prediction is less significant on the other sites with lower observed O3 concentration, but BOE still displays better performance than NAQFC. There are only three sites at which one or both simulations overpredict peak O3 on the 29 August 2018. Compared to NAQFC, BOE shows larger overprediction of the peak O3 at the Greenwich site but smaller overprediction at two other sites (Middletown and Westport).

Besides better peak prediction, BOE has also improved the prediction of the timing of peak O3. The peaks predicted by BOE are 2 to 3 h earlier than that by NAQFC, which agrees better with the timing of the observed peaks (Fig. 12). The BOE peaks are narrower than the NAQFC ones, so that the former follows the observed O3 downslope and avoids the positive biases during late afternoon and early evening. Finally, BOE has improved the prediction of low O3 concentrations and nighttime O3 valleys that are lower than those from NAQFC. Both simulations, however, are unable to reproduce the extreme low nighttime values at several sites. Overall, the BOE simulation performs better in capturing daytime O3 peaks and nighttime valleys, as well as the timing of both, with a mean correlation coefficient of 0.93 compared to 0.88 for the NAQFC simulation. This can be in part attributed to the high resolution of the LIS 3 km system, which can better resolve meteorology and emission variations. As the emissions and meteorological inputs play an important role in determining the magnitude and timing of high peaks (Pan et al., 2017), high-resolution data of both emission and meteorology contributed to the improved the simulation of peak O3 value and its timing, especially over urban areas (Fig. 12).

Vertical profiles of O3 are compared between the Langley Mobile O3 Lidar (LMOL) observations and the CMAQ simulations at the Westport site. As shown in Fig. 13, LMOL observations reveal that the O3 concentration in the planetary boundary layer starts to build up around 16:00–17:00 UTC, and high concentrations (>∼70 ppbv), which extend to a height of about 1.5 km, last until 23:00 UTC on 28 and 29 August. This pattern is reproduced by both the BOE and NAQFC simulations. Above the PBL, the variations of O3 concentrations are also captured by both simulations. O3 concentrations in the free troposphere are more controlled by regional O3 production and transport than in the PBL. Consequently, the structure and magnitude of the O3 profiles are very similar between the BOE and NAQFC simulations, since the BOE simulation is driven by the dynamic boundary conditions derived from the same NAQFC simulation. Compared to that from the LMOL observations, the predicted O3 concentrations from both runs are biased low above 800 hPa but biased high below it. Between the two model simulations, the BOE run not only produces more O3 in the PBL, but also shows a better temporal evolution of the PBL structure, with a short-lived high O3 peak and a PBL height peak between 20:00–22:00 UTC on 28 August and persistent O3 and PBL height plateaus between 16:00–23:00 UTC on 29 August (Fig. 13). The PBL in the BOE simulation extends well above 850 mbar, while the observed high O3 from LMOL generally stays beneath this height, suggesting possible overprediction of the PBL height.

In general, the 3 km BOE simulation performs better to capture the temporal variability of the PBL and O3 production but tends to overestimate both during this episode. In contrast, the NAQFC simulation has produced less pronounced temporal variations in both O3 concentrations and PBL height in the lower troposphere, in particular on 28 August when this region experienced the worst air quality in several states. The NAQFC simulation, however, performed better during the time with lower O3 concentrations, which resulted in an overall lower NMB (9 %) and NME (21 %) comparing to that in BOE (22 % and 26 % respectively). The BOE simulation, however, presented a much better reproduction of the O3 variability in terms of correlation (0.71) than the NAQFC run (0.54). This suggests that the new 3 km BOE system is more responsive to the variations of the controlling factors that shape O3 pollution, although the system needs to be further refined to reduce bias. The model performance for O3 surface concentration and vertical distribution using the AFs from EmisAdj_sub is very close to that using the AFs from EmisAdj_avg in the BOE case (Fig. S4, Table S7).