COVID-19 perturbation on US air quality and human health impact assessment

Abstract The COVID-19 stay-at-home orders issued in the United States caused significant reductions in traffic and economic activities. To understand the pandemic's perturbations on US emissions and impacts on urban air quality, we developed near-real-time bottom-up emission inventories based on publicly available energy and economic datasets, simulated the emission changes in a chemical transport model, and evaluated air quality impacts against various observations. The COVID-19 pandemic affected US emissions across broad-based energy and economic sectors and the impacts persisted to 2021. Compared with 2019 business-as-usual emission scenario, COVID-19 perturbations resulted in annual decreases of 10–15% in emissions of ozone (O3) and fine particle (PM2.5) gas-phase precursors, which are about two to four times larger than long-term annual trends during 2010–2019. While significant COVID-induced reductions in transportation and industrial activities, particularly in April–June 2020, resulted in overall national decreases in air pollutants, meteorological variability across the nation led to local increases or decreases of air pollutants, and mixed air quality changes across the United States between 2019 and 2020. Over a full year (April 2020 to March 2021), COVID-induced emission reductions led to 3–4% decreases in national population-weighted annual fourth maximum of daily maximum 8-h average O3 and annual PM2.5. Assuming these emission reductions could be maintained in the future, the result would be a 4–5% decrease in premature mortality attributable to ambient air pollution, suggesting that continued efforts to mitigate gaseous pollutants from anthropogenic sources can further protect human health from air pollution in the future.


Figures S1 to S15
Tables S1 to S5 SI References FIVE also includes emissions for non-road engines in a similar manner.Off-road distillate fuel sales are reported by state by the Energy Information Administration (https://www.eia.gov/petroleum/fueloilkerosene/) and allocated to end uses following Kean et al. (13).Non-highway use of gasoline is reported by the Federal Highway Administration (https://www.fhwa.dot.gov/policyinformation/statistics/2020/mf24.cfm).Fuel sales for off-road activity are also able to be adjusted regionally on a near real-time basis (12).Emission factors of co-emitted air pollutants (in g/kg fuel) are taken from the EPA NONROAD model (14).Non-road engine emissions are mapped spatially and temporally using surrogates from the NEI 2017 (1).
The VOC speciation profiles for gasoline and diesel engines are reported in McDonald et al. (9) and based on tunnel and laboratory studies, including profiles for liquid gasoline and headspace vapors distinct from exhaust (15).The FIVE mobile source inventory has been rigorously evaluated in previous modeling studies over Los Angeles (16), US (4), and New York City (17), and with satellite NO2 datasets (18).Updates due to the COVID-19 pandemic are accounted for, including rebounding of traffic after COVID-19 lockdown efforts (12).Adjustments to fuel sales are made using monthly gasoline and diesel sales from the EIA Prime Supplier Sales Volume report (https://www.eia.gov/dnav/pet/pet_cons_prim_dcu_nus_m.htm).Spatial refinement of adjustment factors, to the state-level, for some components of on-road traffic is performed using the US Federal Highway administration Traffic Volume Trends report (https://www.fhwa.dot.gov/policyinformation/travel_monitoring/tvt.cfm).Fig. S1A-B shows the ratio of mobile source gasoline and diesel consumption in 2020 relative to 2019 and 2021 relative to 2020.This shows large decreases in gasoline consumption in 2020 and strong rebounding in 2021.Diesel consumption also shows decreases in 2020 and rebounding in 2021 but the magnitude of each is much smaller than for gasoline, likely due to diesel use in goods transport and machinery.
Volatile Chemical Products.Following McDonald et al. (9), VCP emissions are estimated for coatings, inks, adhesives, personal care products, cleaning agents and pesticides.Briefly, VCP emissions were estimated by first performing a mass balance of chemical feedstocks and their distribution across a variety of products manufactured by the chemical industry.Average daily usage and VOC emission factors are reported in McDonald et al. (9) across the US.Long-term trends are taken into account using the same mass balance approach over time following Kim et al. (19).The VCP inventory reflects continuous efforts to lower the VOC content of chemical products, including architectural coatings and phasing out of solvent to waterborne formulations (20).
Nationally, around ~60% of VCP emissions are for consumer uses and ~40% for agricultural and industrial uses McDonald et al. (9).Agricultural pesticides are spatially and temporally allocated according to agricultural pesticide VOC emissions from the NEI17.Industrial uses are similarly spatially and temporally allocated according to the point source VOC inventory from the NEI17.Consumer product emissions are spatially allocated using population density.Past NOAA CSL measurements in New York City and elsewhere have shown a strong population density dependence of consumer VCP emissions (17,21).Diurnal profiles for personal care product emissions are shown to peak in the morning and exponentially decay across the day (22).Other VCP sectors use diurnal profiles from the NEI17, which exhibit a midday peak.Detailed VOC speciation profiles were compiled in McDonald et al. (9), and updated to the latest California Air Resources Board surveys of consumer products and architectural coatings in Coggon et al. (17).
VCP emissions are adjusted using trade statistics from the US Census Bureau.Industrial VCP emissions are adjusted using monthly wholesale trade value for chemical manufacturing from the US Census Bureau (NAICS #4246, https://www.census.gov/wholesale/index.html).We adjust industrial VCP usage using wholesale chemical manufacturing production as a trends surrogate, equal to monthly Sales + Change in Inventory.Trade statistics are reported in value per month and must be adjusted to account for inflation and variation in price per amount of product.For wholesale production, we adjust the value to constant 2017 dollars using the Producer Price Index for the same North American Industry Classification System (NAICS) code, from the US Bureau of Labor Statistics (https://www.bls.gov/ppi/databases/). Finally, a rolling 3-month average is applied to the adjusted production value to reduce noise due to statistical sampling variability.Industrial VCP emissions and their data sources are further described in Table S1.Fig. S1C-D shows the ratio of wholesale chemical production in 2020 relative to 2019 and 2021 relative to 2020, which shows slight decreases throughout 2020 and a rebounding in 2021.
Consumer product VCP emissions are adjusted using monthly retail sales from the US Census Bureau (https://www.census.gov/retail/index.html) for related product types.Retail sales are used rather than wholesale production, as sales are more closely tied to consumer product VCP usage.Again, trade statistics are reported in value per month and must be adjusted to account for inflation and variation in price per amount of product.Retail sales are adjusted to constant 2017 dollars using the Urban Area Consumer Price Index less Food and Energy (https://www.bls.gov/cpi/data.htm).Finally, a rolling 3-month average (1 month to +1 month) is applied to the adjusted retail sales value to reduce statistical sampling variability.Consumer product VCP emissions and their data sources are further described in Table S1.Fig. S1C-D also shows the two retail sales types used in the consumer VCP emission adjustments, in 2020 relative to 2019 and 2021 relative to 2020.Building materials sales increases throughout the 2020-2021 period while personal care products sales decrease sharply in 2020 and rebound sharply in 2021.
Oil & Gas.Upstream emissions from the oil and gas sector come from both the NEI17 and the FOG inventory (Fuel-based Oil and Gas) (23,24).The FOG inventory includes oil and gas emissions in production basins, for NOx, CH4, and non-methane VOCs.The NOx emissions resulting from oil and gas engines (e.g., drilling rigs, compressor stations, dehydrators, etc.) are estimated using bottom-up methods and fuel statistics of energy usage by oil and gas companies that are downscaled using Enverus DrillingInfo well-level production data of oil and natural gas.Fugitive leaks of CH4 and non-methane VOCs are estimated by ratio to NOx using aircraft field data over oil and gas fields from the Southeast Nexus (SENEX: https://csl.noaa.gov/projects/senex/)2013 and Shale Oil and Natural Gas Nexus (SONGNEX: https://csl.noaa.gov/projects/songnex/)2015 studies.The emissions have been gridded nationally, and described by Francoeur et al. (24).Other co-emitted species (e.g., PM2.5, CO, etc.) from oil and gas production regions are taken from the NEI17.Note that midstream (e.g., interstate pipelines) and downstream (e.g., refineries, fuel storage/transport facilities) oil and gas emissions are taken from the NEI17.Data sources and adjustment factor sources for oil and gas, refining and storage can be found in Table S1.
Because it is difficult to separate oil and gas production emissions between oil and natural gas components individually, since a well can produce both, we adjust upstream oil and gas emissions from the NEI using an average of monthly trends in natural gas consumption (EIA, https://www.eia.gov/totalenergy/data/monthly/) and wholesale production of petroleum (NAICS #4247, https://www.census.gov/wholesale/index.html).Wholesale petroleum production scaling factors are calculated to take into account inflation, following the same procedure as is described in the previous section.These adjustments and their average are shown in Fig. S1E, where it can be seen that the average is less responsive to fluctuations in either metric.Downstream emissions from the NEI17 are adjusted using wholesale production of petroleum from the US Census Bureau.This is because these sources primarily focus on oil refining rather than natural gas processing.
Other Anthropogenic.Other point and areawide emissions are taken from the NEI17 and adjusted with near real-time scaling factors in a similar fashion.Table S1 lists how source sectors are subset by Source Classification Codes (SCC), and the datasets used to adjust individual sectors in near real-time.For point sources, stack parameters and plume-rise are taken into account in WRF-Chem.
Electricity Generation.Emissions from electricity generation units (EGUs) are updated to include monthly facility-level Continuous Emissions Monitoring Systems (CEMS) data (https://campd.epa.gov/data) for species where emissions are available (NOx, SOx, CO2).For pollutants where CEMS data is not available, monthly EIA energy consumption by fuel type for electricity generation are used to develop national scaling factors, which are taken as a 3-month rolling average.This is shown in Fig. S2A-B, where changes in electricity generation due to COVID lockdowns are not immediately apparent and trends likely reflect year-to-year differences in electricity demand resulting from weather and economic activity.
Fuel Combustion.For fuel combustion emissions in residential, industrial, and commercial settings (e.g., in boilers), national near real-time scaling factors are developed from US monthly fuel consumption data from the EIA (https://www.eia.gov/totalenergy/data/monthly/index.php) by fuel and sector (see Table S1).The breakdown by these sectors and fuel types is similar to an approach employed by Xing et al. (25), which constructed a long-term air quality inventory for model simulations.Boiler demand for commercial and residential buildings can vary greatly with weather.The national scale adjustments used here would miss the variations occurring between different regions due to weather.Currently there is not finer spatial scale monthly data available, which is a limitation of the datasets used for NRT adjustments.These fuel combustion trends are shown in Fig. S2C-D, for 2020 relative to 2019 and 2021 relative to 2020 fuel consumption.Changes are generally small but for industrial combustion, observed reductions in 2020 and rebounding in 2021 may be related to COVID lockdown effects on economic activity.Industrial Processes.For industrial process emissions (e.g., chemical manufacturing, paper production, etc.), near real-time scaling factors are developed from monthly wholesale trade statistics from the US Census Bureau (https://www.census.gov/wholesale/index.html).Industrial processes (differentiated by their SCC codes) are grouped by product type and are adjusted using relevant wholesale production statistics for the group.For example, industrial emissions associated with metals manufacturing and mining are scaled using inflation adjusted wholesale production of Metals & Minerals, except Petroleum (NAICS #4235).Wholesale production scaling factors are adjusted to take into account inflation using the Producer Price Index.Key wholesale production groups used for NRT adjustments to industrial process emissions are shown for 2020 relative to 2019 and 2021 relative to 2020 in Fig. S2E-H.Some sectors show evidence of reductions in 2020 and rebounding in 2021 related to COVID lockdowns impacting on economic activity (petroleum, textiles, automotive), while others do not show obvious evidence of this (food, electronics).
Rail and Shipping.Railroad emissions are adjusted using monthly carload and intermodal unit traffic from the US Bureau of Transportation Statistics (https://data.bts.gov/stories/s/m9eb-yevh). Emissions at airports (in-flight emissions are not included) are adjusted using monthly air carrier revenue miles flown from the US Bureau of Transportation Statistics (https://www.transtats.bts.gov/TRAFFIC/).Shipping emissions are adjusted using monthly cargo weight (imports + exports) of international shipping from the US Census Bureau (https://www.census.gov/data/developers/data-sets/international-trade.html).
These monthly adjustments are applied uniformly at the national-scale.
Miscellaneous.Finally, for some types of emissions we do not have appropriate economic or energy statistics available to make near real-time adjustment factors.For these sectors, we rely on monthly variations from the baseline NEI17 inventory, and do not apply any year-to-year variation.These emission sources (e.g., waste disposal, agriculture, and dust) are generally not expected to vary significantly year-to-year or due to COVID-19 lockdowns.

VOC measurements at Boulder, Colorado
Volatile organic compounds were monitored in Boulder, CO during the COVID Air Quality Study (COVID-AQS).A full description of the campaign is provided by Rickly et al. (26).Briefly, gas-phase organic and inorganic compounds were monitored from the NOAA David Skaggs Research Center from March 30 -August 31, 2020.Additional measurements were performed in March 2018 at the same location.In both deployments, a proton-transfer-reaction time-of-flight mass spectrometer (PTR-ToF-MS) was deployed to measure mixing ratios of a wide range of VOCs.Here, we use PTR-ToF-MS measurements of select compounds indicative of individual emissions sectors, including D5-siloxane for personal care products (21,22), parachlorobenzotrifluoride for architectural coatings (20,21), and benzene for motor vehicle emissions (22).The PTR-ToF-MS was calibrated for each species using gravimetrically-prepared gas standards, or by liquid calibration as described by Coggon et al. (22).

WRF-Chem model configurations and simulations
The Weather Research and Forecasting (WRF) model coupled with Chemistry (WRF-Chem) (27) version 4.2.2 is applied to simulate emission changes and air quality impacts over the contiguous United States (CONUS).To address the research objectives outlined in the main text, the model is configured at 12 km x 12 km spatial resolution, with total 50 vertical layers that extend up to 50 hPa into the Upper Troposphere-Lower Stratosphere (UTLS).The meteorological initial and boundary conditions for the CONUS domain are from the North American Mesoscale Model (NAM, https://www.ncei.noaa.gov/products/weather-climate-models/north-american-mesoscale).
Chemical boundary conditions are provided from a global model developed by the University of Wisconsin called the Realtime Air Quality Modeling System (RAQMS, http://raqms-ops.ssec.wisc.edu/),which includes data assimilation of satellite ozone and aerosol optical depth (AOD) products.Major physics and chemistry options utilized in the WRF-Chem setup are listed in Table S2.These settings have been well tested and evaluated previously in modeling over the Southeastern US (4), Eastern US (17), and CONUS (18).The oxygenated VOC species and their chemical reactions newly added by Coggon et al. (17) to better represent VCP emissions were added into the racm_soa_vbs scheme (i.e., chem_opt = 108) to create the RACM_ESRL_VCP mechanism.Both anthropogenic emissions and the BEIS biogenic emissions were respeciated to include the newly added oxygenated VOCs.A couple updates to isoprene chemistry were also included.The low-NO OH recycling updates included in Li et al. (18) were added, which included the isoprene hydroxy peroxy radical isomerization reaction to account for OH recycling described in McDonald et al.( 4) and an update to the isoprene hydroxy peroxy radical + HO2 reaction to the following: ISOP+HO2=0.88 ISHP+0.12HO+0.12 MACR+0.12HO2+0.12HCHO with a reaction rate of 7.4E-13 x exp(700.0/T)where ISOP = Isoprene hydroxy peroxy radical, ISHP = isoprene hydroxy hydroperoxide, HO2 = hydroperoxyl radical, OH = hydroxyl radical, MACR = Methacrolein, HCHO = formaldehyde, and T = Temperature.For this study, the products for the isoprene hydroxy peroxy radical + NO reaction were also updated to reflect the latest recommendations for NOx recycling (28) to the following: ISOP+NO=0.87MACR+0.87NO2+0.87HCHO+0.87HO2+0.13ISON with a reaction rate of 2.43E-12 x exp (360.0/T)where NO = nitrogen oxide, NO2 = nitrogen dioxide, ISON = isoprene hydroxy nitrate.
To evaluate emission changes and understand air quality and associated health impacts, we have conducted several sets of model simulations listed in Table S3.Specifically, we simulate emission changes during April to June for 2019, 2020, and 2021.Each year represents different meteorological conditions.We consider 2019 emissions as business-as-usual emission scenario (BAU), 2020 emissions as COVID-induced emission reduction scenario (COV), and 2021 emissions as rebounded emission scenario (REB).Paired simulation with the same anthropogenic emissions but different meteorological inputs (e.g., 2019BAU vs 2020BAU, 2020COV vs 2021COV) are conducted to estimate meteorological impacts.Paired simulation with the same meteorological inputs but different anthropogenic emissions (e.g., 2020BAU vs 2020COV, 2021COV vs 2021REB) are conducted to estimate anthropogenic emission impacts.In addition, paired simulation for April 2020 to March 2021 with the same meteorological inputs but different anthropogenic emissions (BAU vs COV) are conducted to assess health impacts under emission reducing scenario, and paired simulation for April 2019 to March 2020 with the same meteorological inputs but different anthropogenic emissions (BAU vs COV) are conducted to assess meteorological variability impacts on mortality estimates.

Impacts of fire emissions on air quality
Warneke et al. (29) summarized the cumulative area burned in the western US for 2017-2021 (Figure 14 in Warneke's paper), which clearly shows the typical peak months for wildfire between July and October.There were no significant increases in the cumulative burned area over the western US during April-June in 2019-2021 (<3% of annual totals) but noticeable increases in the cumulative burned area started in July especially in 2021.As a matter of fact, fire impacts were well captured by TROPOMI NO2 observations in July 2021, which were not much shown in April-June from 2019 to 2021 (Fig. S4).We therefore conducted sensitivity simulations to include fire emissions based on an earlier version of Regional ABI and VIIRS fire Emissions (RAVE) provided by NOAA's Satellite and Information Service for July 2019, 2020, and 2021 (Table S3).We see significant fire impacts over western US in July 2021 (Fig. S5).However, the simulations with inclusion of fire emissions show much larger biases of O3 and PM2.5 compared to the simulations without fire emissions (Table S5).We acknowledge the uncertainties associated with fire emissions and model representations of plume rise, which could partly contribute to the model biases in simulating O3 and PM2.5.To have fair comparisons from 2019 to 2021, we focus on the period of April-June to disentangle air quality changes due to changes in anthropogenic emissions and meteorology.As the focus of this work is to understand the impacts due to COVID-induced anthropogenic emission changes, we therefore do not include fire emissions in this work to avoid the uncertainties associated with fire emissions and fire model representations that could complicate the interpretation of modeling results.Additional work to include fire emissions and to improve fire representations for air quality modeling could be conducted in the future to better understand fire impacts on regional air quality.

Fig. S1 .
Fig. S1.Monthly emission adjustments for mobile sources, VCPs, and oil & gas.A-B: 2020 and 2021 fuel consumption used in mobile source emissions adjustments; C-D: 2020 and 2021 economic activity used in VCP emissions adjustments; E: activity in petroleum production and natural gas consumption from 2019-2021.

Fig. S2 .
Fig. S2.Monthly emission adjustments for other anthropogenic sectors.A-B: 2020 and 2021 fuel consumption by electricity generation units; C-D: 2020 and 2021 fuel consumption by the industrial, residential and commercial Sectors; E-H: wholesale production activity in 2020 and 2021.

Fig. S3 .
Fig. S3.Fractional contributions of source sectors for 2019 (left panel) and 2020 (right panel).Annual total emissions for each species are shown above each pie chart.

Fig. S5 .
Fig. S5.Comparisons of simulated surface ozone (A&B) and fine particles (C&D) against AQS observations for July 2021.A&C: simulations without fire emissions; B&D: simulations with fire emissions.Circles overlaid on model simulated air pollutants represent observed concentrations from AQS surface monitoring sites.

Fig. S6 .
Fig. S6.Comparison of simulated NO2 columns for April-June 2019, 2020, and 2021 over the US.A&B: difference in NO2 columns between 2019 and 2020 (A) and between 2020 and 2021 (B); C&D: difference in NO2 columns due to emission changes from business-as-usual condition (BAU) to COVID condition (COV, C), and from COV to rebounded emission condition (REB, D); E&F: difference in NO2 columns due to meteorological changes from 2019 condition to 2020 condition (E) and from 2020 condition to 2021 condition (F).Black boxes indicate top 10 populated cities.

Fig. S7 .
Fig. S7.Evaluation of simulated tropospheric NO2 column concentrations with multiple satellite observations during April-June between 2020 and 2021.A: Observed NO2 changes based on the average of four satellite data (S5P TROPOMI, Aura OMI, S-NPP OMPS, and NOAA-20 OMPS) with air mass factors or shape factors in the satellite products replaced by the model profiles; B: Model simulated NO2 changes based on the average of resampled model data along each satellite track; C-H: NO2 columns over urban (C&D), industrial/power plant (E&F), and oil & gas (G&H) source regions from satellite data (C, E, and G) and model estimates (D, F, and H) for 2020 COVID scenario (2020COV, x axis) and 2021 rebounded emission scenario (2021REB, y axis).Slope is calculated based on the orthogonal distance regression with 95% confidence interval.

Fig. S8 .
Fig. S8.Evaluation of VCP emissions with ground-based measurement at Boulder.A: ambient derived emission changes for D5-Siloxane from personal care product (left Y axis, blue bars) and Benzene from mobile sources (right Y axis, solid black lines).B: ambient derived emission changes for PCBTF from solvent-based coatings (left Y axis, tan bars) with emission changes of Benzene from mobile sources (solid black lines) shown on the right Y axis; C: Change in retail sales of Building Material stores (proxy for coatings, brown line) and Health and Personal Care stores (blue line) (https://www.census.gov/retail/index.html).

Fig. S9 .
Fig. S9.Evaluation of simulated tropospheric HCHO column concentrations with multiple satellite observations during April-June between 2019 and 2020.A: Observed HCHO changes based on the average of four satellite data (S5P TROPOMI, Aura OMI, S-NPP OMPS, and NOAA-20 OMPS) with air mass factors or shape factors in the satellite products replaced by the model profiles; B: Model simulated HCHO changes based on the average of resampled model data along each satellite track; C-H: HCHO columns over urban (C&D), industrial/power plant (E&F), and oil & gas (G&H) source regions from satellite data (C, E, and G) and model estimates (D, F, and H) for 2019 business-as-usual scenario (2019BAU, x axis) and 2020 COVID scenario (2020COV, y axis).Slope is calculated based on the orthogonal distance regression with 95% confidence interval.

Fig. S10 .
Fig. S10.April-June changes in MDA8 O3 (upper panel, A, C, and E) and 24-hour averaged PM2.5 (lower panel, B, D, and F) from 2020 to 2021.A&B: circles overlaid on model simulated air quality changes represent observed changes from AQS surface monitoring sites with size in proportion to the absolute changes; site-averaged changes ± standard deviation from AQS observations and model estimates are shown above each figure.C-F: air quality impacts due to emission changes only (C&D) and due to meteorological variability only (E&F); groups of metropolitan areas are shown in black polylines; population-weighted averaged changes ± standard deviation from model grids are shown above on each figure.

Fig. S11 .
Fig. S11.Evaluation of simulated tropospheric column HCHO/NO2 ratios (FNR) with multiple satellite observations during April-June under 2019 business-as-usual scenario (2019BAU) and 2020 COVID scenario (2020COV).A&C: Observed FNR based on the average of four satellite data (S5P TROPOMI, Aura OMI, S-NPP OMPS, and NOAA-20 OMPS); B&D: Model simulated FNR based on the average of resampled model data along each satellite track; A&B: FNR over urban source regions; C&D: FNR over Los Angeles; Urban transitional regime with FNR in the range of 3.0 to 4.5 and in the range of 4.1 to 5.0 over Los Angeles based on Jin et al. (30).The count in the histogram figures (Y axis) represents the number of urban grids that fall into each FNR bin and the density (X axis) represents the probability density.

Fig. S14 .
Fig. S14.Comparison of annual mean MDA8 O3 (upper panel) and 24-hour averaged PM2.5 (lower panel) for the period of April 2020 to March 2021 between business-as-usual emission condition (BAU) and COVID emission condition (COV).Population weighted averages with standard deviation based on all grid cells are shown on each figure.

Fig. 15 .
Fig.15.O3 and PM2.5 attributable deaths (per year per 10 5 people) based on business-as-usual scenario (BAU, A&D) and COVID scenario (COV, B&E) for the period of April 2019 to March 2020.The total attributable deaths due to each air pollutant and scenario are shown above each figure.C&F: difference in attributable deaths between BAU and COV scenarios.The difference in the total attributable deaths are shown above each figure.

Table S1 .
Summary of Anthropogenic Emission Sectors and COVID-19 Pandemic Scaling Factor Data.

Table S2 .
NOAA CSL WRF-Chem Model Configuration a

Table S3 .
Summary of WRF-Chem simulations conducted in this work