Data‐ and Model‐Based Urban O3 Responses to NOx Changes in China and the United States

Urban air pollution continues to pose a significant health threat, despite regulations to control emissions. Here we present a comparative analysis of urban ozone (O3) responses to nitrogen oxide (NOx) changes in China and the United States (US) over 2015–2020 by integrating various data‐ and model‐based methods. The data‐based deep learning (DL) model exhibited good performance in simulating urban air quality: the correlation coefficients (R) of O3 daily variabilities with respect to independent O3 observations are 0.88 and 0.79 over N. China, 0.87 and 0.90 over S. China, and 0.87 and 0.49 over E. United States by the DL and GEOS‐Chem chemical transport models, respectively. Furthermore, the data‐based methods suggest volatile organic compound (VOC)‐limited regimes in urban areas over northern inland China and transitional regimes over eastern US urban areas; in contrast, GEOS‐Chem model suggests strong NOx‐limited regimes. Sensitivity analysis indicates that the inconsistent O3 responses are partially caused by the inaccurate representation of O3 precursor concentrations at the locations of urban air quality stations in the simulations, while the data‐based methods are driven by the variabilities in local O3 precursor concentrations and meteorological conditions. The O3 responses to NOx changes reported here provide a better understanding of urban O3 pollution; for example, reductions in NOx emissions are suggested to have resulted in an increase in surface O3 by approximately 7 ppb in the Sichuan Basin in 2014–2020.


MEE and AQS Surface NO 2 and O 3 Measurements
We use surface in situ hourly NO 2 and O 3 concentration data from the China Ministry of Ecology and Environment (MEE, https://quotsoft.net/air)and US Environmental Protection Agency Air Quality System (AQS, https://aqs.epa.gov/aqsweb/airdata/download_files.html#Raw)monitoring networks for the period 2015-2020.Concentrations were reported by the MEE in units of ug/m 3 under standard temperature (273 K) until 31 August 2018.This reference state was changed on 1 September 2018 to 298 K.We converted the O 3 and NO 2 concentrations to ppb and rescaled the postAugust 2018 concentrations to the standard temperature (273 K) to maintain consistency in the trend analysis.Following K. Li et al. (2020), O 3 and NO 2 observations over all stations, including those with partial records, are used in our analysis because of the limited influence on the derived O 3 trends (K.Li et al., 2020).

The Hybrid DL Model
As shown in Figure 1, the hybrid DL model is a combination of CNN and LSTM.The convolutional layers and max-pooling layers on the left side form the encoder for extracting features from the input data.The encoding process not only reduces the resolution of features but also compresses the high-resolution information into latent vectors that flow through the hidden layers.The decoder on the right side reconstructs the compressed latent features to high resolution through transposing convolution and upsampling layers.We add residual connections to the model, which can forward the high-resolution information to the decoder, contributing to the accuracy of model localization and speeding up the convergence of the training process.We use the squared error loss function.Since the model is a supervised architecture, the summertime maximum daily 8-hr average (MDA8) O 3 from the MEE and AQS monitoring networks are used as true values in the training process.The model obtains the predictions by forward propagation during each iteration and then iteratively updates the weights in the network using the backward propagation algorithm (LeCun et al., 1989;Rumelhart et al., 1986).In addition, we employ the adaptive gradient Adam optimizer, with high computational efficiency and low memory requirements, which can also accelerate the convergence of the cost function.Here are the major hyperparameters of the model: learning rate 0.0001; batch size 50; epochs 400; early stopping patience 20.
The input variables include nine meteorological variables (0.5° × 0.625° horizontal resolution) from Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA-2): sea level pressure (SLP), surface incoming shortwave flux (SWGDN), surface temperature (TS), 2-m air temperature (T2M), 10-m eastward wind (U10M), 10-m northward wind (V10M), 2-m specific humidity (QV2M), total precipitation (PRECTOT) and total cloud area fraction (CLDTOT); and NO 2 concentrations from the MEE and AQS networks.The daily meteorological variables are obtained by averaging hourly data within 12:00-19:00 local time.The observed NO 2 and O 3 concentrations are interpolated into GEOS-Chem grids and processed as daily averages to match the spatial and temporal resolution of the input meteorological variables.The grids without air quality stations are defined as blank data in the model training.To enhance the efficiency and stability of neural network training, we normalize the input meteorological variables so that each variable has a similar range of values.The normalization is performed specifically by subtracting the mean from the original data and then dividing by the standard deviation so that the processed data are approximated to the standard normal distribution.The data from 2015 to 2018 are used as the training set to train the model, and the data from 2019 to 2020 are employed to evaluate the performance of the model as a test set.

GEOS-Chem Model Simulations
The GEOS-Chem CTM (http://www.geos-chem.org,version 12-8-1) is driven by assimilated meteorological data of MERRA-2 with nested 0.5° × 0.625° horizontal resolution.The GEOS-Chem model includes fully coupled O 3 -NO x -VOC-halogen-aerosol chemistry.The chemical boundary conditions are updated every 3 hr from a global simulation with 4° × 5° resolution.Emissions in GEOS-Chem are based on the Harvard-NASA Emission Component (HEMCO).Global default anthropogenic emissions are from the Community Emissions Data System (Hoesly et al., 2018).Regional emissions are replaced by Multiresolution Emission Inventory for China (MEIC) in China, MIX in other regions of Asia (M.Li et al., 2017) and NEI2011 in the US.The total anthropogenic NO x and VOC emissions in the MEIC and NEI2011 inventories are further scaled to obtain the annual emissions in 2019 following X.Chen et al. (2021).Open fire emissions are from the Global Fire Emissions Database (van der Werf et al., 2010).Natural emissions of O 3 precursors, including NO x from lightning and soil and VOCs from vegetation, are calculated on the basis of the assimilated MERRA-2 meteorology.The biogenic emissions of VOCs are calculated according to the Model of Emissions of Gases and Aerosols from Nature (MEGAN v2.1) (Guenther et al., 2006).

Photochemical Box Model
The photochemical box model (OBM) is configured with master chemical mechanisms (MCM v3.3.11,http:// mcm.york.ac.uk/home.htt).The MCM-OBM model was designed to investigate the atmospheric oxidation processes of VOC species (X.Liu et al., 2019;Xue et al., 2013Xue et al., , 2016)).The concentrations of sulfur dioxides (SO 2 ), CO, NO x , and VOC, as well as meteorological parameters (atmospheric pressure, temperature, and relative humidity) from two monitoring sites in Chengdu city (in the Sichuan Basin [SCB]), were used as constraining parameters in the model.The MCM-OBM model simulations start at 12:00 local time for 8 hr, by inputting the observed O 3 concentration at the initial time.MCM-OBM simulations have been widely used to calculate the relative incremental reactivity to describe the response of O 3 to individual precursors (Z.He et al., 2019;J. Li et al., 2018;Tan et al., 2018;Wang et al., 2020).

Surface O 3 Simulated by the DL Model
The hybrid DL model in this work is an autoencoder with the latent space represented by three LSTM cells (Figure 1).The input variables include nine meteorological variables as well as NO 2 concentrations from MEE and AQS monitoring networks.The optimization of the model is supervised by the "ground truth," which is the summertime MDA8 O 3 concentrations measured by the MEE and AQS networks.Only urban stations are considered in this work, and background stations are excluded.The DL model was trained to be a CTM-independent emulator with the NO 2 -O 3 relationship observed by the MEE and AQS networks.Due to the lack of widely distributed long-term VOC observation data sets, VOC observations were not included in the model training by assuming limited changes in anthropogenic VOC emissions in the 6-year period because of their flatter trends with respect to changes in NO x emissions (H.He et al., 2020;M. Li et al., 2019).We also assume limited influences from land usage changes on surface O 3 in the 6-year period.As the variabilities in biogenic VOC emissions are modulated by meteorological conditions such as temperature and solar radiation (Guenther et al., 2006), it is expected that our DL model, driven by meteorological variables, can predict the impact of biogenic VOC variabilities on surface O 3 .
The observed surface MDA8 O 3 in the training period (2015-2018) exhibits higher O 3 in N.China and lower O 3 in S.China and E. US, which was well captured by our DL model (Figure S1 in Supporting Information S1).The DL model also captures the daily variabilities in summertime MDA8 O 3 concentrations over these three domains during the training period, with a Pearson correlation coefficient (R) larger than 0.95 (Figure S2 in Supporting Information S1).Furthermore, we successfully simulated the summertime MDA8 O 3 in the test period (2019-2020) using the DL model (Figure 2), with correlation coefficients slightly lower than those in the Table S1 in Supporting Information S1 provides more information to evaluate the model performance.The normalized mean biases are 0.2%, 0.2%, and −5.2% in the training period, and −6.6%, −8.9%, and −4.8% in the test period for N. China, S. China, and E. US, respectively.The root mean square error are 2.14, 1.77, and 2.68 ppb in the training period, and 6.85, 5.45, and 3.50 ppb in the test period for N. China, S. China, and E. US, respectively.Furthermore, we find better model performance in China than in the E. US, for example, the slopes are 0.99, 0.73, and 0.58 in 2019-2020 for N. China, S. China, and E. US, respectively, which could be associated with the smaller number of observation stations in the US.It should be noted that our DL model validation replies to 20% (US) and 10% (China

Consistent O 3 Responses in Data-Based Methods
Our DL model is driven by the inputted NO 2 concentrations and meteorological variables.We can thus predict the responses of surface O 3 to changes in NO x emissions by adjusting the inputted NO 2 concentrations, as the responses of NO 2 concentrations to changes in NO x emissions are broadly linear (Jiang et al., 2022).Figure 4a shows the summertime NO 2 concentrations from the MEE stations in 2019, with the same 8-hr periods as MDA8 O 3 .Increasing the 2019 summertime NO 2 concentrations by 20%, the DL model predicts a widespread decrease in surface O 3 over N.China inland provinces (Figure 4b).The decreases in NO 2 concentrations before 2019 have thus led to increases in summertime surface O 3 over N.China inland provinces.The 2018-2020 Chinese Clean Air Action plan called for a 9% decrease in NO x emissions (CSC, 2018), and thus, Figure 4c predicts surface O 3 changes driven by a 10% decrease in NO 2 concentrations.Continuous NO x controls, as shown in Figure 4c, are predicted to result in the exacerbation of O 3 pollution over N.China inland provinces.Similarly, Figure 5a shows the summertime NO 2 concentrations from the AQS stations in 2019, which are much lower than NO 2 concentrations in China.The DL model predicts the widespread distribution of transitional regimes over E. US (i.e., slight red or blue shown in Figures 5b and 5c).
In a recent study, X. Chen et al. (2021) evaluated O 3 nonlinear chemistry regimes via lognormal fits of O 3 and NO 2 observations.Following X. Chen et al. (2021), Figure 6a shows the summertime O 3 -NO 2 relationship over the SCB in 2014-2020.The data (dots) are regional averages of MDA8 O 3 and NO 2 concentrations, binned into 1 ppb NO 2 increments.Figure 6a indicates a VOC-limited regime over the SCB: reductions in NO 2 concentrations (d  6b).Following the approach shown in Figure 6a, lognormal fitting lines are produced for each 0.5° × 0.625° grid, which allows us to analyze O 3 nonlinear chemistry regimes in 2019.We find broadly good consistency between the DL-based prediction and the lognormal fit over China (Figures 4b and 4c vs. Figures 4d and 4e) and the US (Figures 5b and 5c vs. Figures 5d and 5e).

Surface NO 2 and O 3 Simulated by the GEOS-Chem Model
Figure 7 (blue lines) shows the daily variabilities in the modeled (GEOS-Chem) and observed NO 2 (based on the 8-hr periods as MDA8 O 3 ) in 2019 over China and the US.Here the modeled NO 2 are sampled at the locations and times of NO 2 observations and are adjusted using the ratios of NO y /NO 2 to consider the influences from reactive oxidized nitrogen compounds in the chemiluminescence analyzers on the observed NO 2 concentrations (Lamsal et al., 2008;F. Liu et al., 2018), where NO y = NO 2 + ΣAN + 0.95 × PAN + 0.35 × HNO 3 and ΣAN is the sum of all alkyl nitrate concentrations.We find dramatic underestimations in the modeled NO 2 concentrations: 5.2 and 7.8 ppb (N.China), 2.7 and 6.3 ppb (S.China), and 2.0 and 3.2 ppb (E.US) from GEOS-Chem simulations and observations from MEE and AQS stations, respectively.Figures 4f and 5f further show the distributions of summertime surface NO 2 concentrations from GEOS-Chem in 2019 over China and the US, respectively.The sampled NO 2 concentrations (Figure 4f) are comparable to the surface NO 2 observations (Figure 4a) over the NCP and Yangtze River Delta but lower in the rest of China.It seems that our model simulations (with 0.5° × 0.625° resolution) tend to underestimate urban NO 2 concentrations over areas with low regional population density.
Here we further investigate the consistency between modeled and observed daily variabilities of MDA8 O 3 .As reported in recent studies (Guo et al., 2018;Kerr et al., 2019;McDonald et al., 2018), we find a significant overestimation of surface MDA8 O 3 (blue line in Figure 3c) with respect to the AQS O 3 observations over the US in 2019.The correlation coefficient between the modeled and observed MDA8 O 3 is as low as 0.49 over US in the CTM, which is much lower than the correlation coefficient in the DL model (0.87).However, we find comparable simulation capability of O 3 daily variabilities between GEOS-Chem and DL models in China: the correlation coefficients between the modeled and observed MDA8 O 3 are 0.79 and 0.88 (N.China) and 0.90 and 0.87 (S.China) by GEOS-Chem and DL, respectively.As shown in Figure S3a in Supporting Information S1, the modeled regional background O 3 are comparable with urban O 3 observations in China; the modeled regional background O 3 is higher than the urban O 3 observations in the US (Figure S3b in Supporting Information S1).It should be noted that GEOS-Chem simulations are only performed in 2019 in this work to avoid the complicated perturbations of the 2019 novel coronavirus (COVID-19) on NO x and VOC emissions.

O 3 Responses Simulated by the GEOS-Chem Model
Can the underestimated urban NO 2 concentrations pose a potential barrier to simulating urban O 3 responses?Figures 4g and 4h and Figures 5g and 5h show the model-based responses of surface O 3 to perturbations in anthropogenic NO x emissions.We find different responses of O 3 to NO 2 changes between data-and model-based methods over both continents: VOC-limited regimes (blue in Figures 4b and 4d) over N.China and weak NO x -limited regimes (slight red in Figures 4b and 4d) over S.China in the data-based methods, in contrast to widespread distributions of strong NO x -limited regimes in GEOS-Chem (red in Figure 4g); transitional regimes over the eastern US (slight red or blue shown in Figures 5b and 5d) in the data-based methods, in contrast to widespread distributions of strong NO x -limited regimes in GEOS-Chem (red in Figure 5g).Furthermore, as shown in Figure S4 in Supporting Information S1, the correlations in the grid-based O 3 responses between the DL and lognormal fit methods over E.China are 0.47 (Figure S4a in Supporting Information S1, NO 2 concentration decrease by 10%) and 0.42 (Figure S4b in Supporting Information S1, NO 2 concentration increase by 20%), in contrast to −0.43 and −0.45 between DL and GEOS-Chem simulations.
GEOS-Chem sensitivity simulations (Run 2, Table 1) adjust anthropogenic and soil NO x emissions over urban grids (i.e., grids have MEE or AQS urban stations) using the ratios of observed/modeled NO 2 .The adjustment of NO x emissions over urban grids led to enhancements of sampled NO 2 concentrations in Figures 4i and 5i; however, they are still noticeably lower than observations (Figures 4a and 5a, also shown in Figure 7).Consequently, we further adjust regional background NO x emissions based on the ratios of averaged observed/modeled NO 2 within neighboring grids (Run 3).As shown in Figures 4l and 5l (also shown in Figure 7), the sampled NO 2 concentrations in Run 3 match better with the observed urban NO 2 concentrations.It should be noted that the adjustments of NO x emissions are designed to cover the underestimated urban NO 2 concentrations, which cannot be explained as an underestimation of NO x emissions: it is unlikely that NO x emissions in the emission inventory should be enhanced by 400% to match observations as shown in Figure S5 in Supporting Information S1.
The consistent NO 2 concentrations between the observations and sensitivity simulations lead to improvements in the modeled urban O 3 responses in China.For example, (Figure S6 in Supporting Information S1), the DL-and lognormal fit-based analyses predict −0.8 ± 0.8 ppb (Figure 4b) and −1.5 ± 1.5 ppb (Figure 4d) decreases in surface O 3 in the SCB in 2019, respectively, due to a 20% increase in anthropogenic NO x emissions; in contrast, the modeled responses are increases in surface O 3 by 2.8 ± 0.2 ppb (Figure 4g), 1.2 ± 1.1 ppb (Figure 4j) and a decrease of −0.2 ± 1.2 ppb (Figure 4m).However, the enhanced NO 2 concentrations have a weaker influence on the modeled urban O 3 evolution in the US (Figure 5), and the model-based O 3 chemical regimes still show a large discrepancy with the reported transitional or VOC-limited regimes in recent studies (H.He et al., 2020;Jin et al., 2020).In a recent study, Zhu et al. (2023) suggests larger relative contributions of regional background O 3 to urban surface O 3 observations in the US than in China, which could be associated with the weaker influence of scaling NO x emissions.anthropogenic VOC emissions in 2015-2020.The lognormal fit approach is a simplified version of the DL model that further assumes the mean states of meteorological variables in 2015-2020.Consequently, the three data-based methods have consistent mechanisms: they are all driven by atmospheric composition concentrations and meteorological variables at the local scale, despite more assumptions of mean states being considered in the DL and lognormal fit approaches.However, we may not expect the same accuracy between the DL and photochemical box models because of the assumption of mean states of anthropogenic VOC emissions.
We find noticeable differences between data-based methods and GEOS-Chem simulation.First, the horizontal resolution (0.5° × 0.625°) in our GEOS-Chem simulation could be too coarse to exactly simulate urban air quality at the local scale.For example, Benavides et al. ( 2021) compared the impacts of diesel NO x emissions on urban NO 2 concentrations with both mesoscale (4 and 1 km resolutions) and street-scale models, and found a noticeable underestimation in the mesoscale model.Furthermore, regional transport plays a key role in GEOS-Chem simulations but does not affect data-based methods.For example, Figure S7a in Supporting Information S1 shows the impact of NO x emission adjustment over urban grids (Run 2, Table 1) on regional background NO 2 concentrations, which leads to an increase in background NO 2 concentrations in the SCB by 2.05 ppb; Figure S7b in Supporting Information S1 further shows the impact of regional background NO x emission adjustment (Run 3, Table 1) on urban NO 2 concentrations, which leads to an increase in urban NO 2 concentrations in the SCB by 2.56 ppb.In addition, uncertainties in physical, chemical processes and emission inventories (Jiang et al., 2022;K. Li et al., 2019b) could further affect the accuracy of model simulations.Consequently, the inconsistency in the derived O 3 responses may reflect the discrepancy in O 3 regimes at different spatial scales, that is, O 3 chemical regimes are VOC-limited at the local scale in urban centers; however, are NO x -limited at larger regional scale when urban and rural O 3 regimes are mixed thorough grid average and regional transport.The lack of VOC observations in the DL model and uncertainties in physical, chemical processes and emission inventories in the simulations could also contribute to this inconsistency.

Conclusion
A comparative analysis by integrating various data-and model-based methods was presented in this work to investigate urban O 3 responses to NO x changes in China and the US.Data-based methods, driven by atmospheric composition concentrations and meteorological variables at the local scale, suggest VOC-limited regimes in urban areas over northern inland China and transitional regimes over eastern US urban areas.In contrast, GEOS-Chem simulations suggest strong NO x -limited regimes.This discrepancy could be partially caused by the inaccurate representation of O 3 precursor concentrations at the locations of urban air quality stations in the simulations Run 3 #1(Base) Urban grids + regional backgrounds 1.0 #2 1.2 #3 0.9 Note.The scaling factors (A1) to enhance anthropogenic and soil NO x emissions over urban grids (i.e., grids have urban stations) and urban grids + regional background, based on the ratios between observed and modeled NO 2 concentrations, are shown in Figure S5 in Supporting Information S1.NO x emissions over the highly industrialized NCP, YRD, and PRD are not adjusted.associated with the coarse model resolution (0.5° × 0.625°) which cannot match the exact locations of urban air quality stations and regional transport of atmospheric compositions which leads to the mixing of urban and rural O 3 regimes.The lack of VOC observations in the DL model and uncertainties in physical, chemical processes and emission inventories in the simulations could also contribute to this inconsistency.
The consistent NO 2 concentrations between the observations and sensitivity simulations lead to better agreement between data-and model-based methods, for example, the DL-based analysis predicts a −0.8 ± 0.8 ppb decrease in surface O 3 due to a 20% increase in anthropogenic NO x emissions in the SCB in 2019 (Figure 4b); in contrast, the modeled responses are increases in surface O 3 by 2.8 ± 0.2 ppb (Figure 4g), 1.2 ± 1.1 ppb (Figure 4j), and a decrease of −0.2 ± 1.2 ppb in surface O 3 (Figure 4m) by gradually enhancing the modeled NO 2 concentrations.However, it should be noted that the adjustments of NO x emissions in the sensitivity simulations are designed to cover the underestimated urban NO 2 concentrations, which cannot be explained as an underestimation of NO x emissions.
The results illustrated in this work indicate the potential of data-based methods as a supplement to CTMs in providing new insights for air quality regulations.The derived O 3 responses to NO x changes are helpful for a better understanding of urban O 3 pollution.It should be noted that the derived O 3 responses in this work are limited by the assumption of mean states of VOC in the DL-and lognormal fit-based analyses, although VOC observations were considered in our OBM model.More studies are suggested in the future to investigate the performance of data-based methods by including a few primary VOC species.We also advise more comparative analyses, particularly using CTMs with higher spatial resolutions, to investigate the advantages and disadvantages of data-based methods, which is critical for exploring better strategies for designing and using atmospheric observations.

Figure 1 .
Figure 1.A hybrid deep learning model used in this paper.Nine meteorological variables and NO 2 observations from Ministry of Ecology and Environment and Air Quality System are used as input variables and the output is the MDA8 ozone.Both have 144 pixels in the latitudinal dimension and 288 pixels in the longitudinal dimension.The orange boxes represent 3 × 3 convolutional layers with ReLU as the activation function.The red boxes represent 2 × 2 max-pooling layers that can extract the most critical features.The blue circle represents 3 stacked long short-term memory cells, which take flattened latent vectors as input.The gray box is a reshaping layer.The light blue box is a transposed convolutional layer for upsampling the latent vectors.Light orange boxes after each transpose convolutional layer represent the features transferred from the encoder by the skip connections.The arrows on the top represent the direction of the skip connections.

Figure 2 .
Figure 2. (a) Summertime MDA8 O 3 in 2019-2020 from the Ministry of Ecology and Environment urban stations.The station-based measurements are averaged and re-grided to 0.5° × 0.625° resolution.(b) Simulated O 3 concentrations by the deep learning (DL) model.(c, d) Differences and Pearson correlation coefficients between DL and observed O 3 .(e-h) Same as panels (a-d), but for (e) United States by using Air Quality System O 3 observations.The black boxes in panel (a) define the domains for N. China, S. China, and the Sichuan Basin (SCB).The star in panel (a) represents the location of Chengdu City in the SCB.
) O 3 observations in 2015-2018.Surface O 3 observations in 2019-2020 are independent observations, which are untouched in the model training, and thus, we can expect consistent model performance outside the period of 2015-2020.

Figure 3 .
Figure 3. Daily variabilities of MDA8 O 3 from observations and simulations.(a, b) N.China and S.China (Ministry of Ecology and Environment, deep learning [DL], and GEOS-Chem) (c) E. United States (Air Quality System, DL, and GEOS-Chem).The domain definitions are shown in Figure 2. The numbers (and numbers in parentheses) of R and root mean square error (RMSE) represent correlation coefficient and RMSE for 2019, and 2019-2020, respectively.

Figure 4 .
Figure 4. (a) Observed summertime surface NO 2 concentrations from the Ministry of Ecology and Environment (MEE) stations in 2019.The station-based measurements are averaged and re-grided to 0.5° × 0.625° resolution.The 8-hr range of surface NO 2 is selected according to the time range of MDA8 O 3 .(b, c) Predicted responses of MDA8 O 3 in 2019 to NO 2 changes by the deep learning model.(d, e) Same as panels (b, c), but from analysis based on the lognormal fit.(f) Surface NO 2 concentrations from GEOS-Chem model standard simulation (Run 1), adjusted using the ratios of NO y /NO 2 to consider the influences from reactive oxidized nitrogen compounds in the chemiluminescence analyzers.(g, h) Responses of MDA8 O 3 in 2019 to NO x emission changes from GEOS-Chem model.(i-n)Similar to panels (f-h), but for sensitivity simulations by enhancing NO x emissions over urban grids (Run 2) and urban + regional background (Run 3, see details in Supporting Information S1).The modeled NO 2 and O 3 are sampled at the locations and times of MEE surface measurements, and then averaged and re-grided to 0.5° × 625° resolution.

Figure 5 .
Figure 5. Same as Figure 4, but for E. United States by using Air Quality System observations.

Figure 6 .
Figure 6.(a) Observed summertime O 3 -NO 2 relationships from Ministry of Ecology and Environment stations with both O 3 and NO 2 measurements.The dots represent regional averages of MDA8 O 3 and NO 2 concentrations, binned into 1 ppb NO 2 increments.The 8-hr range of surface NO 2 measurements is selected according to the time range of MDA8 O 3 .The blue line is lognormal fitting line.The error bars represent standard error.The numbers (0-9) represent the summertime mean O 3 and NO 2 abundances, and a number itself corresponds a year with the year's last digit during 2014-2020.(b) 8-hr averaged responses of O 3 to NO x changes at the SL (103.93°N,30.58°E, 20170801-20170815) and JPJ (104.05°N,30.66°E, 20180601-20180610) sites in Chengdu, Sichuan Province.The simulations are performed using an OBM model with inputs of SO 2 , CO, NO x , and VOC concentrations and meteorological parameters from observations.Positive responses represent increases in O 3 within 8 hr (12:00-19:00 local time) due to a decrease in NO x , indicating a VOC-limited regime.

Figure 7 .
Figure 7. Daily variabilities of NO 2 (based on the 8-hr periods as MDA8 O 3 ) from observations and simulations.(a, b) N.China and S.China (Ministry of Ecology and Environment and GEOS-Chem) (C) E. United States (Air Quality System and GEOS-Chem).