City-scale analysis of annual ambient PM2.5 source contributions with the InMAP reduced-complexity air quality model: a case study of Madison, Wisconsin

Air pollution is highly variable, such that source contributions to air pollution can vary even within a single city. However, few tools exist to support city-scale air quality analyses, including impacts of energy system changes. We present a methodology that utilizes regional ground-based monitor measurements to scale speciation data from the Intervention Model for Air Pollution (InMAP), a national-scale reduced-complexity model. InMAP, like all air quality models, has biases in its concentration estimates; these biases may be pronounced when examining a single city. We apply the bias correction methodology to Madison, Wisconsin and estimate the relative contributions of sources to annual-average fine particulate matter (PM2.5), as well as the impacts of coal power plant retirements and electric vehicle (EV) adoption. We find that the largest contributors to ambient PM2.5 concentrations in Madison are on-road transportation, contributing 21% of total PM2.5; non-point sources, 16%; and electricity generating units, 14%. State-wide coal power plant closures from 2014 to 2020 and planned closures through 2025 were modeled to assess air quality benefits. The largest relative reductions are seen in areas north of Milwaukee (up to 7%), though population-weighted PM2.5 was reduced by only 3.8% across the state. EV adoption scenarios lead to a relative reduction in PM2.5 over Madison of 0.5% to 13.7% or a 9.3% reduction in total PM2.5 from a total replacement of light-duty vehicles (LDVs) with EVs. Similar percent reductions are calculated for population-weighted concentrations over Madison. Replacing 100% of LDVs with EVs reduced CO2 emissions by over 50%, highlighting the potential benefits of EVs to both climate and air quality. This work illustrates the potential of combining data from models and monitors to inform city-scale air quality analyses, supporting local decision-makers working to reduce air pollution and improve public health.


Introduction
Mitigating air pollution represents a significant sustainability challenge for many urban centers worldwide. Urban planning documents for large United States (U.S.) cities frequently include air quality initiatives, often targeting industrial sites, power plants, and on-road transportation sources (Hess and McKane 2021). Sometimes goals specifically address disproportionate air pollution exposure for specific population subgroups or demographics (e.g. rerouting trucking and traffic from certain neighborhoods; not siting new industrial facilities in neighborhoods with high preexisting air pollution burdens; investing in trees, greenspace, green buffer zones to mitigate air pollution) (Shandas et al 2016, Hess andMcKane 2021). Cities such as London and Brussels have also implemented low emission zones, or areas where access is restricted However, previous model evaluations of InMAP have shown consistent overestimates of particulate ammonium and underestimates of particulate sulfate (Tessum et al 2017, Paolella et al 2018. These biases may be pronounced when assessing sources in single city rather than aggregating metrics over wider areas (Tessum et al 2021). Here we present a transferable approach to leverage the policy-analysis strengths of InMAP, while reducing biases and increasing relevance to individual cities. We develop and apply these methods to policy-relevant emissions scenarios in Wisconsin, with a focus on the state capitol, Madison. This study was performed in consultation with Madison Gas and Electric, a local utility company that is interested in reducing their environmental footprint. Through this partnership, we aimed to maximize our impact on stakeholder-engaged local decision-making.

Model description
InMAP predicts annual-average changes in PM 2.5 based on annual-average emissions. The model utilizes chemical, physical, and meteorological information derived from a CTM, which is also used to calculate simplifying assumptions regarding secondary pollutant formation (Tessum et al 2017). In our scenarios, we used the preprocessed data from a Weather Research and Forecasting model coupled with Chemistry (WRF-Chem) run that has been previously described (Tessum et al 2015), reflecting 2005 meteorological and chemical data. Output variables from a CTM that are used in InMAP include chemical reaction rates, gas and particle phase partitioning coefficients, wind vectors, mixing coefficients, dry and wet deposition rates, mixing coefficients, and variables related to emissions plume rise. An important difference between WRF-Chem and InMAP is that WRF-Chem requires temporally explicit input data while InMAP requires only annual-average inputs to speed processing time. Although InMAP is an annual-average model, it attempts to account for the impact of time of day and season on air pollution by using temporally explicit data to calculate annual-average inputs when possible (e.g. the average of one year of hourly reaction rates, rather than the explicit annual average reaction rate).

Source apportionment
We used the most recent input emissions dataset available for InMAP, derived from the 2014 National Emission Inventory (NEI) and originally developed for a study quantifying 'pollution inequity' in the United States ; a version of the dataset was also used in a paper assessing health implications of renewable energy policies in the United States (Dimanchev et al 2019). Because NEI emissions are aggregated at the county scale, area source emissions were processed using the Air Emissions Processor which allocates them to the InMAP grid based on spatial proxies (Tessum et al 2017). Not included in our emissions dataset are wildfire emissions, biogenic emissions, and international emissions (i.e. Canada and Mexico). The impact of these omitted emissions on exposure and health is assessed by Tessum et al (2019).
For our model simulations, we first ran InMAP with no changes to the emissions input files and a continental United States (CONUS)-wide spatial domain to evaluate the baseline concentrations over Madison and Wisconsin. For all analyses and calculations referring to Madison, we zoomed in on a 24 km × 24 km spatial domain centered on the latitude and longitude coordinates 43.130, −89.426. This encompasses the City of Madison, Middleton, Waunakee, and parts of Fitchburg, McFarland, and Sun Prairie.
To understand the source contribution from each source over this spatial domain, we ran each of our 15 input emissions files separately. The 15 emissions files included on-road transportation, non-road mobile sources, rail, electricity generating units (EGUs), non-EGU point sources, agriculture, area fugitive dust, residential wood combustion, non-point oil and gas, point oil and gas, point agricultural fires, non-point agricultural fires, flaming fire sources, smoldering fires sources, and non-point sources not otherwise classified. In the real atmosphere, pollutants interact and influence the formation and destruction of one another. However, InMAP assumes that the chemical formation of each species of secondary PM 2.5 is both linearly related to precursor emissions and formed independently of all other species of PM 2.5 (Tessum et al 2017). This linear approach to chemical formation impacts sulfate concentration predictions most heavily as sulfate formation is nonlinear (Tessum et al 2017); this leads to the underpredictions of sulfate and total ambient PM 2.5 by the model prior to applying the bias correction discussed below.

Bias correction
We evaluated InMAP PM 2.5 concentration biases by comparing 2014 monitor measurements from the United States Environmental Protection Agency's (EPA) Air Quality System (U.S. EPA, 2022b) to the coincident InMAP-calculated pollutant concentrations for particulate sulfate (pSO 4 ), ammonium (pNH 4 ), and nitrate (pNO 3 ). InMAP also calculates concentrations of primary PM 2.5 (PPM 2.5 ) and secondary organic aerosol (SOA), but these specific species are not explicitly measured by the monitors and are thus omitted from bias correction.
Across the U.S., there are 358 monitors that measure ambient sulfate, 349 monitors that measure ambient nitrate, and 198 monitors that measure ambient ammonium. These monitors provide regionally representative speciation data for large swaths of the country.
Scaling factors are calculated as shown in equation (1), where O i is the monitor measurement, P i is the model-calculated concentration, and n is the number of monitors: We used the average bias from n monitors as our fractional bias corrections directly. The spatial scope of monitor measurements with which to correct model data, and thus the number of monitors, could range from the city to national level, and depends on the interested stakeholders and research questions. These methods may be directly applied to any region with monitors and results in the genericized equation below where a, b, and c are the computed biases described below in equation (2): (2) Here we used five speciation monitors located across Wisconsin to scale InMAP noting that no monitor is located in our primary study site, Madison. The five monitors are located in downtown Milwaukee, Green Bay, Waukesha, Horicon National Wildlife Refuge, and the town of Perkinstown in north-central Wisconsin. Because we are assessing state-scale policies, we chose to incorporate all available data across Wisconsin by calculating the average bias of the five aforementioned monitors. Alternative methodologies to make these corrections include the use of satellite data, especially to improve bias correction in rural areas that may not have a monitor nearby , or the use of the closest monitor or monitors to a given grid cell. We chose not to utilize these methods for several reasons. First, this work is intended to be accessible for city-scale managers and decision-makers. Correcting bias using different scaling factors across a single domain adds additional expertise to this methodology, thus making it less accessible. Second, based on the model evaluations conducted by Tessum et al (2017) and Paollela et al (2018), we concluded that biases for sulfate and ammonium are relatively similar across the entire state of Wisconsin, thus justifying a single correction factor based on the average bias of the five Wisconsin monitors. An increase in the number and spatial coverage of these monitors would likely refine this scaling factor but we are limited by the chemical speciation network already installed state-wide.
We found model-calculated pSO 4 concentrations to be 240% lower than the monitor measurements on average; pNH 4 , 30% too high; and pNO 3 , 30% too low. Without scaling, the underestimate of pSO 4 would lead to significant underestimates in the impacts of coal-fired electricity and sulfur-emitting industrial sources. Similarly, without scaling, the overestimate of pNH 4 would overestimate the PM 2.5 attributable to agriculture. We multiplied each InMAP-calculated pSO 4 concentration by 3.4 and each InMAP-calculated pNH 4 concentration by 0.7 and added to original calculations of pNO 3 , SOA, and primary PM 2.5 to get the new value for total PM 2.5. This is shown by equation (3) below: We did not scale pNO 3 based on the calculated bias. Nitrate is not as highly associated with a single sector. Although the transportation is a major contributor to pNO 3 and its gaseous precursors, nitrogen oxides (NO X ), these pollutants also come from other sources such as industry, residential heating, power plants, as well as wildfires, which are omitted from our input emissions dataset. The omission of the wildfire emissions along with the variety of sources and activities that contribute to ambient concentrations of this pollutant led us to conclude that scaling would not be appropriate and could in fact introduce more uncertainty because of the variety of contributing sectors.
Biases in our input emissions and meteorological data may also have impacted the biases seen in our output concentrations. However . This suggests that these input data do not overwhelm other sources of error in the model or preclude it from use as a screening tool.

Case study 1: coal power plant retirements
In our first case study, we assessed the air quality impacts of changes in the electricity sector. Due to recent coal plant closures across the country, emissions inventories published in the last few years are already overestimating pollutant contributions from the electricity sector. The contribution from this sector is overestimated in the most recent emissions dataset available for use with InMAP, as it was developed from the 2014 NEI. In order to update this dataset to more accurately represent the share of pollution coming from the energy sector now and in the future, we considered three scenarios. First, we developed an emissions scenario that takes into account coal power plant retirements that occurred in Wisconsin between 2014 and 2022, the 'present scenario' . Next, we developed a scenario that also incorporates planned coal power plant retirements in the state (before 2030), the 'planned scenario' . Finally, we assessed the air quality impacts of the Columbia Energy Center retirement on Madison, the 'Columbia scenario' . This scenario is hypothetical, as it does not take into account past retirements or other concurrent planned retirements, but it is useful for evaluating the individual impact of a nearby plant. The coal-fired power plants that retired between 2014 and 2022 included Alma, Bay Front, two units of Edgewater, Genoa, Milwaukee County, Nelson Dewey, Pleasant Prairie, Pulliam, Valley, and one unit of Weston. Emissions from these power plants were zeroed out to create the 'present scenario' . The coal power plants in Wisconsin that have plans to retire by 2030 include Columbia Energy Center, South Oak Creek Power Plant, and the final unit of Edgewater. Emissions from these power plants were also zeroed out to create the 'planned scenario' . The final scenario we developed was the hypothetical 'Columbia scenario' which only zeroed out emissions from the Columbia Energy Center.
To apply these scenarios to InMAP, we compared the shapefile points in our EGU emissions input shapefile with the locations and current operating status of coal power plants in the state of Wisconsin in the year 2014. We compared the points and associated SO X emissions from our input file to 2014 power plant data which we acquired from the Clean Air Markets Division (CAMD) (U.S. EPA, 2022c). Two coal-fired power plants, Weston Power Plant and Edgewater Power Plant, retired partially between 2014 and 2022, meaning we could not entirely zero out emissions from these sources in our input file. To address this, we changed the emissions to reflect approximate 2020 CAMD emissions, the most recent emissions inventory available. One of the Weston coal units retired in 2015 and two of the three remaining units are coal-fired with no public retirement plans. At the Edgewater plant, two coal units have retired, one each in 2015 and 2018; the remaining unit will retire by the end of 2022.

Case study 2: electric vehicle adoption
In our second case study, we developed emissions scenarios simulating a significant increase in electric vehicle (EV) adoption in the Madison-area. We focused on light-duty vehicles (LDVs) within the transportation sector as well as emissions reductions that are anticipated as a result of the planned retirement of coal-fired power plants over the next several years. The on-road input emissions file we used, however, did not distinguish between emissions from LDVs and heavy-duty vehicles (HDVs). To separate LDVs emissions from total on-road emissions, we used the 2017 National Emissions Inventory. Using Dane County as a proxy for the Madison-area, we estimated the percent contribution of LDVs to individual pollutant emissions from the transportation sector overall (primary PM 2.5 and PM 2.5 precursors). We estimate that LDVs in Dane County contributed to 57% of NO X , 80% of sulfur oxides (SO X ), 91% of ammonia (NH 3 ), 91% of volatile organic compound, and 51% of PPM 2.5 emissions from the transportation sector.
Using these percentages, we then developed three scenarios representing reductions in emissions from LDVs in Dane County: • 25% EV scenario; represents 25% decrease in emissions from LDVs and aligns with 'high-market penetration' scenario for ∼2033 from Electric Power Research Institute (EPRI, 2017). • 50% EV scenario; represents 50% decrease in emissions from LDVs and aligns most closely with President Biden's target to make 50% of all LDV sales electric by 2030 (Executive Order #14037). • 100% EV scenario; represents a 100% decrease in emissions from LDVs.
These scenarios were developed using the following equation, where 'EV Scenario' is calculated for each of the aforementioned species: EV Scenario = on road emissions -(on road emissions × % LDV × % reduction) .
Note that the input emissions file does not distinguish between tailpipe primary PM 2.5 emissions, which would be reduced by EV penetration scenarios, and primary PM 2.5 attributable to brake and tire wear, which would not be reduced by our EV scenarios. As a result, we do not account for this in our EV scenarios. Additionally, each EV-penetration scenario assumes that additional electricity demand associated with increased EV use in Madison was generated with non-emitting sources. Various factors go into determining the electricity sector emissions generated from EV use, including the type of generation (e.g. coal, natural gas, renewable sources), the time and location of vehicle charging, type of vehicle, and the vehicle miles traveled (VMT). These details are beyond the scope of this assessment. However, the potential increase in emissions from the electricity sector that may be associated with increased EV adoption in Wisconsin has been explored by other studies (Meier and Holloway 2022).
To update emissions from the electricity sector to be in line with our coal retirement scenarios, we create a reference case scenario that omits emissions from Wisconsin coal power plants that have retired since 2014 or plan to retire by 2030. Because we omit the emissions from these coal plants and because of the Biden Administration's EV goal, these scenarios theoretically represent the year 2030. President Biden's goal, however, refers to sales of EVs, rather than percent of the on-road fleet. Our scenarios assume an instantaneous replacement of 50% of LDVs with EVs in 2030. In reality, 50% EV sales in 2030 will take time to be reflected in the overall vehicle fleet due to the turnover time of existing internal combustion engines as LDVs are estimated to have a useful lifetime of about 17 years (Keith et al 2019).
In addition to the impact on air quality, the transportation sector is the leading contributor to greenhouse gases in the United States (U.S. EPA, 2021). Though InMAP does not explicitly include CO 2 , we estimated CO 2 emissions from the transportation sector for each scenario using NO X emissions. To do this, we used data from the EPA's Motor Vehicles Emissions Simulator (MOVES) (U.S. EPA, 2022d). MOVES includes estimates of VMT per county as well as emissions factors (grams CO 2 emissions per mile driven) for both LDVs and HDVs. We assigned VMT to the InMAP grid proportionally based on our input NO X emissions and then translated that into CO 2 using the appropriate emissions factor. We then totaled CO 2 over the spatial domain. MOVES estimates VMT for LDVs in Dane County to be 4.24 × 10 9 miles; for HDVs, that value is an order magnitude lower at 8.88 × 10 8 miles. While LDVs have higher VMT, HDVs release far more CO 2 per mile traveled. MOVES estimates an emissions factor of 1411.6 grams CO 2 per mile drive for HDVs and only 409 grams CO 2 per mile drive for LDVs.

Source apportionment: Madison, Wisconsin
We source-apportioned the scaled model results to quantify the magnitude with which each source contributes to total annual ambient PM 2.5 concentrations in Madison, shown in figure 1. We find the largest contributor over Madison is on-road transportation at 21% of total PM 2.5 . This is followed by non-point sources (an amalgamation of all non-point sources not otherwise classified; 16%), EGUs (14%), and agriculture (13%). The transportation sector, including non-road and rail sources, is the largest overall contributing sector, at 30%. Point and non-point sources contribute 23% and 22%, respectively.

Relative source contribution in Wisconsin
At the state-level, we focus on three sectors that were among the largest relative contributors to ambient PM 2.5 concentrations in Madison, including on-road transportation, EGUs, and agriculture. Figure 2 shows the relative contributions from these sectors in Wisconsin. In general, the relative contribution from on-road transportation is highest-17% or higher-in large cities such as Madison, Milwaukee, Green Bay, and Appleton (figure 2(c)). Across the rest of the state, the relative contribution is over 11% almost everywhere except the northernmost part of the state.
We observe a contrasting pattern when evaluating the relative contributions from agriculture (figure 2(b)) and EGUs (figure 2(a)), as these sources contribute more significantly in rural areas compared to urban areas. Agriculture is a prominent contributor in the southwestern part of the state where it is modeled to contribute over 17% of total ambient PM 2.5 . In the northern part of the state and in large cities, the relative contribution is 14% or lower. It is likely that Madison's location directly adjacent to an agriculture heavy region leads to a slightly higher contribution from this sector than in other large cities in the eastern part of the state (e.g. Green Bay and Milwaukee). Finally, the contribution from EGUs is less in cities but greater elsewhere in the state, where this source is responsible for an estimated 15% to 23% of total PM 2.5 concentrations.

Case study 1: coal power plant retirements
In our first case study, we evaluate the air quality impacts of coal power plant retirements in Wisconsin relative to 2014 baseline emissions.
As shown in figure 3, the retirement of coal-fired power plants in Wisconsin has led to substantially decreased emissions from EGUs, a trend that is expected to continue with additional retirements. Figure 3 shows Wisconsin power plant emissions of NO X (green), SO X (blue), and PPM 2.5 (yellow) in 2014; after  accounting for closures between 2014 and 2022; and after accounting for planned closures. SO X shows the largest declines due to the power plant closures. SO X emissions have decreased by over two-thirds since 2014 and are projected to decrease by nearly 90% when taking into account emissions reductions from planned retirements. Emissions of NO X declined almost 50% between 2014 and 2022, with 70% reductions relative to 2014 expected. Note that figure 3 reflects only emissions reductions due to past and planned coal power plant retirements in Wisconsin, so any other emissions-changing actions associated with other types of EGUs are not reflected here (e.g. conversion to natural gas). Figure 4(a) shows the PM 2.5 concentrations over the state in 2014. Figures 4(b) and (c) indicate the percent reduction in ambient PM 2.5 concentrations attributable to the 'present' and 'planned' scenarios, respectively; the relevant power plants are marked as pink circles in 4(b) and (c). Figure 4(b) suggests that the greatest relative benefits attributable to plant retirements in the last eight years were on the eastern border of Wisconsin as well as the northeastern quadrant. We model a relative reduction in PM 2.5 concentrations of 3%-4.5% in the majority of this region, though a small area in the southeastern corner of the state, home to the 2018-retired Pleasant Prairie Power Plant, boasts a slightly higher percent reduction. In figure 4(c), we incorporate the planned retirements. Once again, the greatest benefits are seen in the eastern part of the state, especially in the area north of Milwaukee where percent reductions of over 7% are possible. We also calculate  population-weighted concentrations as a proxy for human exposure. We calculate a population-weighted concentration of 6.4 µg m −3 taking into account the entire population of Wisconsin. We find that this value decreases 2.2% in response to coal plant closures that occurred between 2014 and the present. An additional 1.7% reduction is estimated to be attributable to power plant closures between 2022 and 2025.
A limitation of these scenarios is that we do not incorporate the emissions that would be generated from the conversion of these plants from coal-fired to natural-gas fired or the reduction in emissions as a result of utility-scale renewable energy projects replacing these plants. Rather, the percent reductions discussed here are strictly those attributable to these coal power plant retirements, rather than the explicit changes that may occur taking into account generation replacements. Thus, we may be overestimating the overall air quality benefits accrued in reality.
Our final scenario assesses the air quality impacts of the Columbia Energy Center retirement on Madison, Wisconsin. In 2021, it was announced that unit one would be retired by the end of 2023 and unit two by the end of 2024. We assess the impact of this single retirement on Madison for two reasons. The first is the plant's proximity to Madison: at 40 miles north of Madison, it is the closest coal power plant to the city; the second is that through our baseline model run we found that the speciation of ambient PM 2.5 attributable to EGUs in Madison resulted in 89% sulfate, 8% nitrate, and 3% primary PM 2.5 . Because sulfate is a fingerprint for coal-fired power plants and Madison does not currently have any coal power plants operating in its boundaries, we could infer that it is not local power plants contributing to the majority of PM 2.5 from that sector in Madison. In figure 5, we show the results of a scenario that omits emissions from only the Columbia Energy Center to assess the plant's impacts on Madison. We find that the city-wide percent reductions in ambient PM 2.5 range from 1% to around 2.3%, where larger relative reductions can be seen in the northern half of our spatial domain both because Columbia is located north of Madison, and because the northern region of our Madison domain has fewer PM 2.5 sources than the downtown area.

Case study 2: electric vehicle adoption
In our second case study, we assess a 25%, 50% and 100% decrease in emissions from LDVs in Madison, Wisconsin. These broadly represent a 25%, 50%, and 100% EV adoption rate. The results of our scenarios are shown in figure 6, along with an overlay displaying the major highways and interstates in the area. Figures 6(b)-(d) show the percent reduction from the reference case (figure 6(a)) for each of our three EV-penetration scenarios.
Our results show PM 2.5 % reductions that range from 0.5% to 13.7%. Overall, a 100% decrease in emissions from LDVs (figure 6(d)) resulted in a 9.3% decrease in total ambient PM 2.5 concentrations over the Madison-area. The largest percent reductions are seen downtown and across the east side of Madison, followed by reductions in south Madison, as shown by all three scenarios.
Our 50% EV scenario resulted in approximately a 7% decrease in PM 2.5 concentrations over the population-dense downtown, specifically the portion of highway that runs through downtown Madison. Percent reductions of 4%-5% are seen in grid cells bifurcated by the interstate south of Madison, while reductions of just over 5% are seen over the grid cells in the east that contain multiple interstates. We also calculate population-weighted metrics for each EV scenario. In the reference scenario (figure 6(a)), the population-weighted PM 2.5 concentration was 7.1 ug m −3 over the Madison-area. Next, we calculated the percent reduction in population-weighted PM 2.5 resulting from each EV-penetration scenario. Because of the linear construction of InMAP, uniform percent reductions in LDV emissions result in linear percent reductions in PM 2.5 and thus PM 2.5 exposure. Subsequently, the 25% EV scenario resulted in percent reductions of approximately 2.2% across the total population; the 50% EV scenario resulted in reductions of 4.4%; and the 100% EV scenario resulted in reductions of 8.8%.
We observe larger relative reductions for NO X concentrations than for PM 2.5 concentrations as a result of these emissions scenarios: the 50% EV scenario resulted in NO X concentration percent reductions of up to 14.2%. This suggests that LDVs contribute more to ambient NO X than ambient PM 2.5 .
EV adoption would also impact carbon dioxide emissions from the transportation sector. We estimate that LDVs account for about 58% of on-road CO 2 emissions in the reference case. In the 25% scenario, LDVs emit 51% of CO 2 and total CO 2 decreases 14.5% from the reference scenario. In the 50% scenario, LDVs emit 41% of CO 2, and total CO 2 decreases by 29% compared to the reference scenario. Finally, in the 100% EV adoption scenario total CO 2 emissions from the on-road sector decrease by 58% as only CO 2 emissions from HDVs remain.

Discussion
In this study, we present an approach for correcting ambient PM 2.5 concentration data from the InMAP reduced-complexity air quality model, using regional ground-based measurement data to maximize relevance of model results to an individual city or state.
We find that scaling sulfate and ammonium affects the source apportionment in rural areas more than urban areas due to high transportation NO X and pNO 3 concentrations in urban areas (21% in Madison). When assessing how scaling ammonium and sulfate impacts source apportionment in Madison, we find that the percent contribution from agriculture decreases from 20% to 13% when ammonium is scaled down, where EGU contributions, which account for almost two-thirds of the sulfate concentrations in the city, increases from 6% to 14%. The other major sources of ambient PM 2.5 in Madison stay close to constant after scaling: on-road transportation remains the largest contributor decreasing from 22% to 21% and 'other non-point sources' remain at a 16% relative contribution.
Missing from our estimates is the relative contribution from wildfires. One study found that during the fire season of 2014 (May 1 through September 30), the average contribution from wildfires to total ambient PM 2.5 was between 0.2 and 1 µg m −3 across Wisconsin (Munoz-Alpizar et al 2017). This would constitute a range of relative contribution from wildfires of 3% to 12% in Madison, suggesting that it may have been a major source of ambient PM 2.5 . Including these input emissions in future work would more accurately reflect ambient concentrations, especially as wildfires are expected to become a larger relative source of total PM 2.5 across much of the country (Ford et al 2018).
Understanding how ambient PM 2.5 is apportioned to various sources can inform sustainability initiatives in cities, especially those designed to protect the health of those most vulnerable to air pollution. Focusing on a single city also allows for further analyses to be conducted based on location-specific factors by simulating scenarios that reflect city-or state-wide policies and planning decisions that impact air quality.
Our analysis supports previous findings that the replacement of internal combustion engine vehicles with EVs has the potential to have a localized effect on PM 2.5 and NO X due to the prevalence of cars in urban areas (Ferrero et al 2016, Rizza et al 2021. In Dane County, LDVs, specifically, contribute around two-thirds of all on-road emissions (both PM 2.5 precursor emissions and primary PM 2.5 ) emphasizing the importance of electrifying passenger vehicles. Because LDVs are prevalent in urban areas, they also directly impact human exposure to PM 2.5 . When calculating total percent reductions in ambient PM 2.5 versus population-weighted PM 2.5 we find similar percent reductions. In reality, emissions reductions would depend on which internal combustion engine vehicles were being replaced; EVs replacing old and inefficient internal combustion engine vehicles would have the largest impact, a replacement which can be accelerated by incentivization programs such as the 'cash-4-clunkers' program (Naumov et al 2022).
The retirement of coal power plants has varied impacts across Wisconsin. In general, it appears that coal power plant retirements have a less of an impact on human exposure in Wisconsin compared to EVs because they contribute less to ambient PM 2.5 in large cities (illustrated in figure 2). While ambient concentrations are reduced up to 7% in specific locations (north of Milwaukee), population-weighted concentrations for the total population are only reduced 3.8% across the state. This study supports previous findings that states and cities are impacted by coal power plant retirements differently based on their proximities to these plants, with communities proximate to coal plants experiencing the largest health benefits from their closures (Gallagher and Holloway 2020).
This scaling methodology helps address biases that make InMAP results less reliable for a single city and more relevant for patterns among many cities. However, sources of uncertainty remain in both our input data and scaling methodology. First, although we attempt to manually update the electricity sector, other outdated sectors within the emissions inventory may have impacted our results. Second, air quality models, in general, are limited by the resolution and accuracy of input emissions. Disaggregating county level emissions traditionally consists of using spatial proxies like highways and major roads to allocate the emissions to grid cells, but this can introduce positive urban biases and negative urban biases at finer resolutions (Zheng et al 2017). These issues are being ameliorated by research on developing more accurate emissions inventories (Wu et al 2021. Lastly, additional uncertainty is introduced based on the number and distribution of monitors used to calculate the scaling factor for each pollutant because this varies by spatial domain. In other words, a similar analysis for Minnesota would utilize a different number of monitors that are distributed differently across the state. While this inconsistency may impact our scaling factors, it is necessary for analyses of cities like Madison, Wisconsin that do not have an speciated PM 2.5 air quality monitor located in or around city boundaries, and must rely on other sources of data. Recognizing these uncertainties, InMAP should be considered a screening tool, with the conclusions drawn considered guidelines rather than prescriptive policy recommendations.

Conclusion
This approach offers a transferable model for other U.S. cities that are near chemical speciation monitors. This methodology could be applicable to sustainability planning in cities of various sizes, both spatially and in terms of population. Although we worked alongside a local utility, other community-scale stakeholders including the Mayor's office, community groups, or private companies could also play a role in making this methodology relevant to decision-makers. We employ a RCM here, a choice that represents a trade-off between the complexity of a CTM and the ability of an RCM to quickly address the needs of decision-makers. Despite the simplicity of InMAP, it has been shown to be an appropriate tool to support policy analyses with only a modest loss of accuracy compared to full-form models (Gilmore et al 2019).
Local officials need information specific to their city in order to know which energy policies to prioritize, especially because cities have limited resources to implement these policies. Furthermore, understanding to what extent sources outside city boundaries are contributing is an important consideration (e.g. we simulated the contribution of the Columbia Power Plant to PM 2.5 in Madison). This study demonstrates how information from air quality models can be made policy-relevant at a scale as granular as a city. In employing this information, improvements in local public health can be pursued most effectively.

Data availability statement
The data that support the findings of this study are available upon reasonable request from the authors.