Effect of grid resolution and spatial representation of NH3 emissions from fertilizer application on predictions of NH3 and PM2.5 concentrations in the United States Corn Belt

Ammonia (NH3) emissions from fertilizer application is a highly uncertain input to chemical transport models (CTMs). Reducing such uncertainty is important for improving predictions of ambient NH3 and PM2.5 concentrations, for regulatory and policy purposes and for exploring linkages of air pollution to human health and ecosystem services. Here, we implement a spatially and temporally resolved inventory of NH3 emissions from fertilizers, based on high-resolution crop maps, crop nitrogen demand and a process model, as input to the Comprehensive Air Quality Model with Extensions (CAMx). We also examine sensitivity to grid resolution, by developing inputs at 12 km × 12 km and 4 km × 4 km, for the Corn Belt region in the Midwest United States, where NH3 emissions from chemical fertilizer application contributes to approximately 50% of anthropogenic emissions. Resulting predictions of ambient NH3 and PM2.5 concentrations were compared to predictions developed using the baseline 2011 National Emissions Inventory, and evaluated for closure with ground observations for May 2011. While CAMx consistently underpredicted NH3 concentrations for all scenarios, the new emissions inventory reduced bias in ambient NH3 concentration by 33% at 4 km × 4 km, and modestly improved predictions of PM2.5, at 12 km × 12 km (correlation coefficients r = 0.57 for PM2.5, 0.88 for PM-NH4, 0.71 for PM-SO4, 0.52 for PM-NO3). Our findings indicate that in spite of controlling for total magnitude of emissions and for meteorology, representation of NH3 emissions and choice of grid resolution within CAMx impacts the total magnitude and spatial patterns of predicted ambient NH3 and PM2.5 concentrations. This further underlines the need for improvements in NH3 emission inventories. For future research, our results also point to the need for better understanding of the effect of model spatial resolution with regard to both meteorology and chemistry in CTMs, as grid size becomes finer.


Introduction
Ammonia (NH 3 ) is the most abundant alkaline trace gas in the atmosphere, and precursor to PM 10 and PM 2.5 (particulate matter (PM) with diameters10 μm and 2.5 μm, respectively) (Bassett andSeinfeld 1983, Baek et al 2004). PM 10 and PM 2.5 are regulated for impacts on human health and welfare (e.g., visibility) by the current Clean Air Act Amendments in the United States of America (US) (US EPA 2016). Additionally, they influence the Earth's radiative budget by having a net cooling effect (Pinder et al 2012, Smith andBond 2014). Deposition of NH 3 and other reactive nitrogen species exacerbates environmental impacts such as soil acidification, surface water eutrophication and reduced ecosystem productivity (Krupa 2003, Erisman et al 2013. In view of its important role in the atmosphere, the US EPA Science Advisory Board has recommended NH 3 be considered as a harmful PM 2.5 precursor (US EPA 2011).
In the US, dominant anthropogenic contributors of NH 3 emissions include livestock operations including manure management and land application (55%) and chemical fertilizers (26%) (US EPA 2015). However, relative source contributions vary based on regional agricultural practices. In the Corn Belt region in the Midwest US (henceforth Midwest), chemical nitrogen fertilizers are the dominant source of NH 3 emissions (55%, year 2011) (US EPA 2015). This region spanning the States of Illinois, Indiana, Iowa, Missouri, Nebraska and Kansas is dominated by corn and soybean cultivation (Green et al 2018) that accounts for nearly 60% of the US fertilizer sales (AAPFCO and The Fertilizer Institute 2014). In view of decreasing emissions of major PM 2.5 precursors of sulfur and nitrogen oxides, emission controls for NH 3 from agriculture have been proposed as potentially cost-effective for further reduction of PM 2.5 concentrations (Pinder et al 2007. However, sensitivity to NH 3 emission reductions vary seasonally and by region (Holt et al 2015, Bauer et al 2016. State-of-the-art chemical transport models (CTMs) can improve our current understanding of linkages between NH 3 emissions from agriculture and air quality . Predictive capabilities of CTMs to model PM 2.5 formation depend on accurate representations of spatial and temporal distributions of precursor emissions. As compared to other PM 2.5 precursors, there are large uncertainties in NH 3 emissions (>50%) at the global scale and even higher uncertainties at regional and local scales , Sutton et al 2013. While field measurements of NH 3 emissions provide insight into inherent spatial and temporal trends, such studies in the US have largely focused on livestock waste . Measurements over fertilized cropland are limited to a few locations (e.g., ) due to the resource-intense and technical challenges of NH 3 flux measurements (Norman et al 2009). Hence modeling approaches have been adopted to characterize NH 3 emissions at the regional scale as inputs to CTMs (Goebes et al 2003). Regarding NH 3 emission inputs, current CTM predictions are limited by errors in grid-scale representation (Bray et al 2017) and timing of agricultural NH 3 emissions (Appel et al 2011), and lack or limited representations of bidirectional NH 3 exchange within CTMs (Cooter et al 2012, Battye et al 2016, Schiferl et al 2016. In particular, studies have recognized the need to better characterize NH 3 emissions from chemical fertilizers at high spatial (<32 km×32 km) and temporal (hourly) resolutions (Appel et al 2011, Paulot et al 2014 to capture the episodic nature of this emission source and influence of local climatic and soil parameters. Yet, there is no conclusive evidence if increase in spatial and temporal resolution of NH 3 emissions improves CTM predictions, given the multitude of inputs that need to be developed at similar resolutions Zhang 2008, Schaap et al 2015).
In the National Emissions Inventory (NEI) up to year 2011, NH 3 emissions from chemical fertilizers in the US are estimated using the emission-factor approach (Goebes et al 2003, US EPA 2014a. These county-scale, annual emissions are allocated to spatial and temporal resolutions required by CTMs using emission processing models (Pouliot et al 2015) such as the Sparse Matrix Operator Kernel Emissions (SMOKE) (CMAS 2016). In the case of NH 3 emissions from chemical fertilizers, spatial allocation from coarser to finer CTM grid resolution is achieved using land cover maps as the spatial surrogate (Goebes et al 2003). Recent work has additionally accounted for crop nitrogen application rates, as demonstrated by the Improved Spatial Surrogate (ISS) approach (Balasubramanian et al 2015a) and the Magnitude and Seasonality of Agricultural Emissions model (Paulot et al 2014). Temporal allocation of annual emissions is based on seasonal nitrogen management data and crop planting and harvesting schedules that are averaged to hourly scale (Goebes et al 2003) or using semiempirical models that estimate emissions as functions of hourly temperature and wind speed (Gyldenkaerne et al 2005, Paulot et al 2014. Such approaches introduce errors in NH 3 emissions due to approximations in fertilizer usage at the county scale and use of European emissions factors that require evaluation for climatic and field management conditions in the US (Balasubramanian et al 2015a), homogenization across CTM grid cells (Walker et al 2012) and lack of accounting of local meteorological influences (Hendriks et al 2016). Subsequently, implementations of the NEI emissions within CTMs have been linked to underpredictions in modeled NH 3 and PM 2.5 concentrations in comparison with observations (Rodriguez et al 2011, Schiferl et al 2016. Inverse modeling approaches to optimize NH 3 emissions have identified large underestimates in total NH 3 emissions for California and the Midwest, and errors in seasonality for emissions from chemical fertilizers in the Midwest (Paulot et al 2014), especially during spring (Heald et al 2012, Banzhaf et al 2013. Process-based models present an alternative approach to developing emission inventories for agriculture. Such models advantageously predict spatial and temporal variations in NH 3 emissions, based on local variability in climate and soil conditions, and regional crop management practices (Brilli et al 2017).
Emissions estimated using process models have been incorporated into CTMs through different approaches. As an example, the Volt'Air model that simulates influences of meteorology and slurry application on transfer of nitrogen in the soil and exchange between soil and atmosphere to estimate NH 3 emissions (Génermont and Cellier 1997), has been implemented to improve sub-grid variability of emissions within the CHIMERE CTM (Hamaoui-Laguel et al 2012). The Environmental Policy Integrated Climate (EPIC) model has been adopted to simulate ammonium content in applied fertilizer, and coupled with the Community Multiscale Air Quality (CMAQ) CTM to subsequently capture bidirectional NH 3 exchange between soil and atmosphere inline at each time step (Cooter et al 2012. More recently, the DeNitrification DeComposition (DNDC) model that simulates daily NH 3 emissions after fertilizer usage (Li et al 1992, Li 2000 was coupled with NEI to characterize daily NH 3 emissions for the Midwest (Balasubramanian et al 2015a). While it is highly desirable to integrate representations of field-scale processes within CTMs, extensive effort is required to develop parametrizations that adequately capture emission potentials of different trace gas fluxes from soils (Flechard et al 2013).
In this study, we undertook an evaluation of the Comprehensive Air Quality Model with Extensions (CAMx) (Ramboll ENVIRON 2016), for two different approaches for spatially allocating NH 3 emissions from chemical fertilizer application and at two model spatial resolutions. CAMx is a state of the art CTM that has been widely implemented for air quality modeling studies focussing on multiple regions worldwide, as a part of academic and regulatory assessments (Ramboll ENVIRON 2016, Emery et al 2017. For spatial and temporal allocation of NH 3 emissions, we used the ISS spatial allocation method, to develop an NH 3 inventory at two spatial resolutions (4 km×4 km and 12 km×12 km), combined with the process DNDC model for temporal resolution at the daily scale (Balasubramanian et al 2015a). In an earlier study, we showed that DNDC successfully captured the temporal occurrence of NH 3 flux peaks following chemical fertilizer application, by comparison to NH 3 flux measurements in a typically fertilized corn field in Central Illinois (Balasubramanian et al 2017). An US EPA regulatory modeling guidance document has recommended implementation of CTMs at spatial resolutions finer than 4 km×4 km for regions with strong emission gradients (US EPA 2014b). This is of importance to the Midwest that has dominant contributions of NH 3 emissions from chemical fertilizers. Region-specific CTM studies linking agricultural NH 3 emissions to air quality have been conducted over the Eastern US using CMAQ (Appel et al 2011 and CAMx (Zhang et al 2013), California using CMAQ (Bray et al 2017, Lonsdale et al 2017, and Colorado using CMAQ (Battye et al 2016) and CAMx (Rodriguez et al 2011, Thompson et al 2015. However, fewer studies have been conducted for the Midwest, with focus on modeling wintertime PM formation using CMAQ (Spak et al 2012, Kim et al 2014 and CAMx (Spak et al 2012). While CAMx has been identified as an appropriate CTM for PM 2.5 prediction (Baker and Scheff 2007), it has not been as widely implemented as CMAQ for linking agricultural NH 3 emissions to air quality (Rodriguez et al 2011, Thompson et al 2015, thereby providing aditional motivation to study CAMx performance for this application.
Based on these research motivations, we present a first investigation of sensitivity of CAMx predictions to the method used for representation of NH 3 emissions and to the spatial resolution of gridded emission inputs. This study aims to bridge a gap in our understanding of CTM performance in a region of strong NH 3 emissions gradient that are predominantly a result of chemical fertilizer application in the US Corn Belt, by examining the sensitivity of CAMx to NH 3 emission representations. Such understanding is of scientific as well as of policy interest, as model predictions are often used to guide policy decisions.

Study region and time period
The Midwest represents a region of dominant chemical fertilizer sales in the US (>60% (AAPFCO and The Fertilizer Institute 2014)), high observed concentrations of ambient NH 3 (>1.8 μg m −3 ) (Puchalski et al 2015); and high wet deposition fluxes of NH 4 + (>3 kg-N ha −1 yr −1 ) (Lehmann et al 2007, Li et al 2016. A nested 4 km×4 km domain was defined over the Midwest surrounding the State of Illinois, within a parent (surrounding) domain over the contiguous US with 12 km×12 km grid resolution (figure 1). May 2011 was chosen for evaluation because it represents the period of active planting and fertilizer usage dates for corn (USDA 2010). As a result, this period includes peak NH 3 fluxes from chemical fertilizers, as verified by field measurements at a site located in central Illinois .
Year 2011 was chosen because it was the year with the most recent and updated NEI data (US EPA 2015), when this study was initiated.

Meteorology
The mesoscale, non-hydrostatic Weather Research and Forecasting-Advanced Research model, thereafter WRF, version 3.5.1 was used to develop 3D gridded meteorological fields (UCAR 2016). WRF was run for the same geographical domains shown in figure 1, at the same 4 km×4 km and 12 km×12 km spatial resolutions. 30 vertical Eta Levels were used (from surface to 50 hPa), interpolated by WRF's default algorithm for the entire model. Two-way nesting was employed to allow feedback of model parameteres from the parent domain to the nested domain, and vice versa. Initial and boundary conditions for WRF simulations were driven by the 12 km North American Model (NAM) data, from the National Operational Model Archive and Distribution System, with 6-hour temporal resolution (NOAA 2016). Terrain input data were obtained from the WRF official website (UCAR 2016). The objective analysis program, OBSGRID was used for observational nudging, using National Centers for Environmental Prediction (NCEP) observational data (NCEP 2008(NCEP , 2015. Following sensitivity analysis including 17 runs with different physical parameterizations (Fu 2016), the parameterization shown in

Emissions
For NH3 emissions from chemical fertilizer application, we used the typical NEI-SMOKE approach (US EPA 2014a) as baseline to compare with the ISS-DNDC approach described in ( A comprehensive accounting of all PM 2.5 precursor emissions is required for accurately modeling ambient NH 3 and PM 2.5 concentrations. Thus, we obtained emissions for sectors apart from the chemical fertilizer application from the NEI2011v6.2 modeling platform (US EPA 2017a). The NEI2011v6.2 platform reports criteria air pollutants and precursor pollutants for 60 emission sectors, that are grouped as point sources, nonpoint sources including agriculture, nonroad and onroad mobile sources, biogenic and fugitive dust sources, US map obtained from https://www.census.gov/geo/maps-data/data/cbf/cbf_state.html. and fires. The agricultural sector is further divided into NH 3 emissions from livestock manure management (classified by animal type and manure management system), and chemical fertilizer application (classified by fertilizer type). In this study, we use the ISS-DNDC approach to spatially and temporally allocate NH 3 emissions from chemical fertilizer application as mentioned previously, and use the spatial surrogates and temporal factors from the NEI2011v6.2 modeling platform for the other emission sectors (US EPA 2017a).
Four scenarios of emission inputs to CAMx were developed (table 1) to investigate how NH 3 spatial allocation influences CAMx predictions of NH 3 and PM 2.5 concentrations. Baseline emissions were first established using NEI 2011 for all emission sectors and allocated using spatial surrogates and temporal factors available within the NEIv6.2 modeling platform. Then NH 3 emissions from fertilizer application were modified based on the ISS-DNDC approach, by modifying spatial surrogate ratios and temporal factors provided in the NEIv6.2 modeling platform, using the ISS method, and daily fraction of annual NH 3 emissions estimated using DNDC, respectively (supplementary material, text S1). These modifications were applied to all States within the Midwest that were considered within the 4 km×4 km domain shown in figure 1. In order to study sensitivity to spatial resolution, both baseline and hybrid ISS-DNDC emissions were processed for the Midwest at two spatial resolutions: (1) as a nested 4 km×4 km grid within a parent 12 km×12 km grid, and (2) on a uniform 12 km×12 km grid.

CAMx configuration and inputs
CAMx v6.4 was implemented with parameterizations recommended by Zhang et al (2013) (table S3). Horizontal and vertical grid structures matched those of the meteorological inputs. Vertical diffusion is modelled with the default K-theory diffusion option. Horizontal advection is modeled with the Piecewise Parabolic Method (PPM) approach (Ramboll ENVIRON 2016). Gas phase chemistry mechanism CB6r4 is used. CB6r4 developed with a focus on ozone photochemistry, it also results in production of PM precursors such as sulfuric and nitric acids and semi-volatile organic compounds (Yarwood et al 2010). For aerosols, the 2-mode (coarse-fine) CAMx PM size distribution is used, in which all secondary particles are modelled as fine only. Aqueous sulfate and nitrate formation in cloud water is modeled with the RADM algorithm (Chang et al 1987). The thermodynamic partitioning of inorganic aerosol species (sulfate, nitrate, ammonium, and natural minerals) is based on the routine ISORROPIA v.1.6. (Nenes et al 1998). In CAMx, ISORROPIA is also called in every time step for cloudy grid cells to ensure that rapidly evolving (through aqueous chemistry) sulfate, nitrate and neutralizing cations, are in balance with the local environment. Secondary organic aerosol (SOA) is modeled with the routine SOAP (Secondary Organic Aerosol Partitioning) (Strader et al 1999). Aerosol and gas dry deposition is modeled according to resistance model Zhang et al (2001) for aerosols and Zhang et al (2003) for gases. Wet deposition is estimated by the scavenging model for gases and aerosols (Seinfeld and Pandis 2012). Wet deposition is invoked only if precipitation is reaching the surface, assuming cloud and precipitation exist in steady state during the hourly model time step. The model assumes that all gases are dissolved in cloud water in equilibrium with ambient air concentrations according to Henry's Law solubility, aqueous dissociation, cloud water temperature and acidity; and all particles are taken up into cloud water. It is also assumed that all particle species within a particular size range (fine or coarse) are internally mixed and hydrophilic. Cloud water gas scavenging may be reversible but particle scavenging is irreversible (Ramboll ENVIRON 2016). CAMx was initialized during May 2011 with a three-day spin up starting May 1. While a longer spin up of about 2 weeks is ideal, a 3-day spin up period was reported to increase PM 2.5 concentrations by ∼0.1 μg m −3 in Samaali et al (2009). Initial and boundary conditions for CAMx were generated using outputs from the Model for Ozone and Related Chemical Tracers (MOZART) model (NCAR 2017) and photolysis rates were generated using ozone column data and PM-nitrate (PM-NO 3 )) concentrations were extracted for further evaluation. The CAMx modeling framework adopted in this study, with linkages between various models, pre-processor programs and model inputs are shown in figure 2.

Evaluation of CAMx predictions
Operational evaluation of CTM predictions is recommended using graphical techniques and statistical evaluation in comparison with ground-based observation data (US EPA 2014c). In this study, predicted hourly concentrations of ambient NH 3 , total PM 2.5 , PM-NH 4 , PM-SO 4 and PM-NO 3 were evaluated graphically for all four emissions scenarios, using the Visualization Environment for Rich Data Interpolation (VERDIv1.5) tool (CMAS 2014). Observations were mapped using ArcGISv10.4.1 (ESRI 2016). A custom program was developed using MATLAB version R2017a (The MathWorks Inc. 2017) to pair CAMx predictions with observations in space and time, for statistical evaluation. NH 3 is the most challenging pollutant to evaluate due to scarcity of measured data, at time resolutions matching the hourly resolution of CAMx. Observations of ambient NH 3 concentrations were obtained from the Ammonia Monitoring Network (AMoN), which reports ground-level NH 3 concentrations measured using Radiello ® passive samplers, integrated over two-week sampling periods (NADP 2015). Predicted NH 3 concentrations were aggregated to match the time frame of AMoN observations at each station, as recommended by Schiferl et al (2016). Predicted NH 3 concentrations were converted from units of parts per million by volume (ppm v ) to μg m −3 using average surface temperature and pressure observed across the Midwest obtained from ISWS (2016). Error introduced by using this standard conversion factor was<5%, when accounting for the range in observed maximum and minimum temperature and pressure values during May 2011, at the Midwest observation sites (ISWS 2016). Further model evaluation was considered by comparing predictions of total PM 2.5 , and PM-NH 4 , PM-SO 4 and PM-NO 3 concentrations to ground-level, integrated, 24-hour observations obtained from the quality assured Air Quality System (AQS) Data Mart (US EPA 2014c). PM 2.5 predictions were similarly averaged to the 24-hr time resolution of AQS observations. Observation stations located in the boundary grid cells of the Midwest domain were removed from analysis to minimize impact of boundary conditions. Closure between CAMx predictions and observations was evaluated using statistical metrics including normalized mean bias (NMB), root mean square error (RMSE) and correlation coefficient (r) (table S4). CAMx performance was compared to findings from other studies and to benchmarks recommended by Emery et al (2017), who identified the 'goal' (i.e., best achievable CTM performance that the top one-third of studies have met) and 'criteria' (i.e., current CTM performance that two-third of studies have met) performance. Recommended benchmarks are available for total and speciated PM 2.5 as follows: Total PM 2.5 , PM-NH 4 ,

CAMx performance evaluation
Performance evaluation statistics for NH 3 , PM 2.5 , PM-NH 4 , PM-SO 4 , and PM-NO 3 are presented in table 2. Modeled results at the hourly scale were averaged to match the time scales of the measurements which are 14 days for NH 3 as reported by AMoN, and 24 h for PM species as reported in AQS. We focus on comparing differences resulting from alternative spatial representation of NH 3 emissions from fertilizer application   combined with DNDC predicted NH 3 emission peak timing and temporal evolution, and model grid resolution. Evaluation statistics from this study are compared with other studies in table 3.  of cropland in Iowa, 50% of cropland in Indiana (NASS 2014)), in comparison to emission maps presented in Balasubramanian et al (2015a). Predicted NH 3 concentrations are lower for the 12 km×12 km scenarios compared to the 4 km×4 km scenarios; indicating spatial homogenization of NH 3 emissions in the coarser grid due to averaging of sub-grid scale emission peaks. Thus, NH 3 emissions and CAMx predicted concentration trends were further investigated, at three AMoN sites (Bondville and Alhambra, rural agricultural sites; and Indianapolis, urban site), where observations were available.

NH 3 concentration predictions
The hourly modelled emission fluxes and predicted NH 3 concentrations at these sites are shown in figure 4 for all emission scenarios. As anticipated, rural sites exhibit higher emission fluxes (figures 4(a) and (b)) than urban sites (figure 4(c)) due to presence of agricultural activities. The impact of spatial homogenization at coarser grid scale is evident as the 4 km×4 km emission scenarios exhibit higher NH 3 emission flux in comparison to the 12 km×12 km scenarios. In terms of timing, localized emission peaks captured in the ISS-DNDC scenarios (predicted in the third week of May due to fertilizer application during active planting dates for corn) are higher in magnitude compared to the baseline approach for both 4 km×4 km and 12 km×12 km scenarios that represent seasonally averaged emissions. However, trends in diurnal patterns in emissions are similar both baseline and ISS-DNDC, as they are driven by the same hourly temporal profile (Balasubramanian et al 2015a) obtained from the NEIv6.2 modeling platform. Since NH 3 has a short lifetime (order of hours) (Adams et al 1999, Aneja 2001, trends in predicted NH 3 concentrations were anticipated to follow spatial and diurnal patterns in NH 3 emissions (Balasubramanian et al 2015a). However, the modeled spatial patterns are not identical across the domain. We hypothesize that NH 3 emissions in CAMx were potentially transported and consumed faster within the model domain compared to the ambient environment, but this hypothesis requires further analysis. We also examined hourly variations in NH 3 concentrations at the three chosen sites (figures 4(d)-(f)). The data show that: (1) while NH 3 emissions (figures 4(a)-(c)) peak around noon, predicted NH 3 concentrations (figures 4(d)-(f)) show maxima in late afternoon, and (2) peak modeled daily NH 3 concentrations do not follow trends in daily NH 3 emission peaks at site scale. Our findings are qualitatively similar to measurement and modeling results presented by Personne et al (2009). Our findings indicate that while the ISS-DNDC approach can capture localized and episodic NH 3 emissions in comparison to the baseline approach, errors due to uncertainties in emission estimates still result in errors in CAMx predictions of NH 3 concentrations.

PM 2.5 concentration predictions
Comparison of predicted (figures 5(a)-(d)) and observed total PM 2.5 (figure 5(e)) concentrations in the Midwest, showed an underestimate for all emissions scenarios with lowest NMB for both baseline scenarios (−27% for baseline-4 km, −25% for the baseline-12 km), and comparable correlations, with the two ISS-DNDC scenarios producing slightly higher r (0.55 for ISS-DNDC-4 km, 0.57 for ISS-DNDC-12 km). The extent of PM 2.5 under-predictions was much smaller, compared to NH 3 underpredictions. Spatial distribution of predictions of total PM 2.5 concentrations in the Midwest under four emission scenarios are shown in figures 5(a) and (b) (baseline-4 km and ISS-DNDC-4 km scenarios, respectively) and figures 5(c) and (d) (baseline-12 km  higher emission fluxes result in an increase of free ammonia over short periods of time, that is potentially advected or deposited before chemical transformation; thereby resulting in smaller magnitude of speciated PM 2.5 predictions. Spatial distribution of average predictions of PM-NH 4 concentrations under the four emissions scenarios are shown in figures 6(a) and (b) (baseline-4 km and ISS-DNDC-4 km scenarios, respectively) and figures 6(c) and (d) (baseline-12 km and ISS-DNDC-12 km12 km scenarios, respectively). Average predicted PM-NH 4 concentrations ranged between 0.1 and 3.6 μg m −3 across the Midwest. Baseline-4 km and baseline-12 km scenarios exhibited larger spatial variability and predicted higher concentrations (>2.2 μg m −3 ) in the eastern part of the Midwest not seen in the corresponding ISS-DNDC scenarios.
Similar grid related differences were observed for PM-SO 4 , with the 4 km scenarios (figures 7(a) and (b)) similarly underpredicting (NMBs=−40% and −41%) and the 12 km scenarios (figures 7(c) and (d)) similarly overpredicting (NMBs=25% and 23%). (>4 μg m −3 ) in the eastern part of the domain. Among all speciated PM species, PM-SO 4 is relatively well-mixed in the atmosphere and exhibited larger spatial homogeneity, because of the preferential uptake of sulfuric acid as compared to nitric acid by free NH 3 , and slow rates of formation and removal (Park et al 2006).
Comparison of PM-NO 3 predicted (figures 8(a)-(d)) and observed (figure 8(e)) values showed poor performance of the 4 km×4 km scenarios with r=0.12 and 0.1 for baseline-4 km and ISS-DNDC-4 km, respectively, with these two scenarios underpredicting (NMBs=−79% and −84%). Results for the 12 km×12 km scenarios are confounding, with higher correlations (r=0.64 and 0.52 for baseline-12 km and ISS-DNDC-12 km) but baseline-12 km overpredicted (NMB=38%) while ISS-DNDC-12 km underpredicted (NMB=−54%) PM-NO 3 . Spatial distribution of average predictions of PM-NO 3 concentrations range between 0.07 and 3.8 μg m −3 across the Midwest domain under the four emissions scenarios that are shown in figures 8(a) and (b) (baseline-4 km and ISS-DNDC-4 km scenarios, respectively) and figures 8(c) and (d) (baseline-12 km and ISS-DNDC-12 km scenarios, respectively). Observed PM-NO 3 concentrations at AQS stations ranged between 0.09 and 1.6 μg m −3 , in comparison to the corresponding predicted concentrations (0.13 to 4 μg m −3 ). Compared to the ISS-DNDC scenarios, baseline-4 km and baseline-12 km scenarios predicted higher concentrations (>1.0 μg m −3 ) in the north-east part of the Midwest, where colder temperatures were observed in comparison with the rest of the Midwest (ISWS 2016). Further research is warranted to investigate these trends and establish how speciated PM formation is impacted by availability of NH 3 in the atmosphere and in consideration of meteorological parameters such as temperature and atmospheric stability.

Discussion
Evaluation of NH 3 predictions is limited by the lack of NH 3 observational data. Currently the AMoN network provides the best ground-based spatial coverage of NH 3 observations in the US, available since 2007, with coarse time resolution averaged over 14 days. While satellite monitoring and aerial measurements provide additional information, such information is not readily available because the effort currently requires complex data assimilation and processing, and is not directly comparable with ground level measurements because it is integrated over the atmospheric column (Clarisse et al 2009, Schiferl et al 2016, Warner et al 2016. In this study, due to the sparse coverage of AMoN stations, there were only 5 prediction-observation pairs available for statistical evaluation, over our focal region of the US Corn Belt that has high agricultural NH 3 emissions from fertilizers, in May 2011. Therefore, the reported evaluation is indicative of model performance rather than conclusive. However, our findings reassert the underpredictions reported in the literature, irrespective of use of ground-based observations (table 2) (Schiferl et al 2016). The largest underpredictions were identified near agricultural sites that were characterized by high ambient NH 3 concentrations (Kruit et al 2012, Butler et al 2015, Vogt et al 2013. The magnitude of our underpredictions are higher and similar to Bray et al (2017), where CMAQ underpredicted NH 3 concentrations by a factor of 4.5 in comparison with satellite observations. One explanation is the smaller source contributions of NH 3 from chemical fertilizers in cited studies, as compared to this study. In studies over Eastern US and Colorado, fertilizer usage accounted for 20% or less of anthropogenic NH 3 emissions, whereas in the Midwest region we examine, the contribution is as high as 55% (US EPA 2015). This indicates the need for regionally specific modeling of NH 3 emissions from fertilizer application; because even though this source contributes half or less of total NH 3 emissions, it can introduce large biases in ambient NH 3 and PM 2.5 . Identified underpredictions in NH 3 concentrations could result from potential errors in total NH 3 emissions developed using the emission-factor approach for the Midwest, a case supported by comparisons with satellite measurements (Clarisse et al 2009) and inverse modeling efforts (Paulot et al 2014). Such uncertainties include uncertainties in planting and fertilization dates and use of generalized regional fertilizer application practices, as farm scale practices are not available. There is a further need for development and evaluation of alternative methods to estimate total NH 3 emissions used as inputs to CTMs, supported by a denser monitoring network for both NH 3 fluxes and concentrations. While this study indicates an advancement in this direction, reported findings indicate a need to further explore the use of satellite data and process models to capture variations in NH 3 emissions based on complex crop, soil and weather interactions. Examples in this direction include US EPA's use of EPIC (Williams et al 1983(Williams et al , 1984  Another source of differences in NH 3 concentration predictions between baseline and ISS-DNDC scenarios could result from the representation of NH 3 deposition within CAMx. CAMx currently does not model bidirectional exchange of NH 3 , unlike other CTMs that account for canopy compensation points to predict ambient NH 3 concentrations (Flechard et al 2013). An example is integration of the EPIC model with bidirectional exchange in CMAQ that increased ambient NH 3 concentrations by 14% across the US . However, this may vary across CTMs. A similar implementation in the GEOS-Chem model resulted in larger underestimates (factor 2-5) in predicted NH 3 concentrations (Zhu et al 2015a). Subsequently, Schiferl et al (2016) did not include the bidirectional scheme, and recommended an increase in model grid resolution. In comparison, introduction of bidirectional exchange in the LOTOS-EUROS model increased predictions of NH 3 concentrations by 30%-40% over agricultural areas due to increase in NH 3 lifetime and transport distance (Kruit et al 2012). Hence, implementation of a bidirectional exchange scheme within CAMx, at high spatial resolutions ( 12 km×12 km), is pertinent, as it could reduce identified underpredictions by extending the time period of NH 3 volatilization and lifetime in the atmosphere. Still all studies suggest that the current magnitude of NH 3 emissions is underestimated and require further development.
In comparing results of performance evaluation for PM 2.5 and its components PM-NH 4 , PM-SO 4 and PM-NO 3 , the effect of model grid resolution is more pronounced, than the method used to allocate fertilizer NH 3 emissions. The method of emission allocation has an effect on spatial trends in predicted concentrations and is therefore important in studies where CTMs like CAMx are applied over local scales. Overall, estimated NMB and r metrics for total and speciated PM 2.5 in this study are in the reported range from other CTM studies. PM-NH 4 exhibited improved performance compared to most CTM studies (table 3), especially for the ISS-DNDC-12 km scenario. Regarding previous CAMx performance results, in a comparative model performance study, CAMx exhibited improved performance in predicting PM 2.5 in the southeast US (r=0.31 to 0.7), as compared to CMAQ (r=0.22 to 0.5) (Zhang et al 2013). Thompson et al (2015) demonstrated that CAMx satisfactorily reproduced annual predictions of PM-NH 4 (NMB=−31%, r=0.71), PM-SO 4 (NMB=9%, r=0.78) and PM-NO 3 (NMB=57%, r=0.58) over the Rocky Mountain National Park in the US In evaluating CAMx, Rodriguez et al (2011) point to the effect of the extent of the spatial domain where they obtained improved correlation for PM 2.5 (r=0.40) for the entire US, in comparison with the Rocky Mountain National Park region (r=0.15), indicating strong dependence of predictions on regional emission sources. These findings support recommendations by Emery et al (2017) to limit evaluation of CTM predictions of PM 2.5 to domains not exceeding 1000 km×1000 km, to capture regional source contributions, and over temporal scales of a month (but no more than three months), to account for seasonal differences. Similarly, impact of timing of NH 3 emissions on ambient speciated PM 2.5 at the sub-daily scale deserves further study to resolve the identified poor correlation (r) indicated in table 3. While winter PM formation has not been addressed in this study, there is a need to better characterize variability in NH 3 emissions to reduce wintertime PM 2.5 overpredictions in the Midwest (Pitchford et al 2009, Kim et al 2014. By extension, observed low nighttime temperatures during May 2011 (4°C to 15°C, across the Midwest (ISWS 2016)) could result in enhancement of nighttime PM-NO 3 . Identified poor closure metrics for PM-NO 3 could be attributed to larger underpredictions during nighttime in the Midwest and deserves further investigation.
While it was anticipated that higher-spatial resolution emissions and meteorology inputs will improve predictions of PM 2.5 , implementation of the 4 km×4 km grid resolution did not improve model performance in comparison to the 12 km×12 km grid resolution. Similarly, a study over North Carolina concluded that increasing spatial resolution did not improve PM predictions . Slight improvements in PM 2.5 predictions using a 4 km×4 km grid over Texas (NMB=−22%) were observed as compared to 12 km×12 km (NMB=−33%) in a study by Misenis and Zhang (2010). While a multi-CTM evaluation in Europe identified significant improvement in PM 10 predictions at urban areas when reducing grid size from 50 km×50 km to 14 km×14 km, finer resolutions (<14 km×14 km) did not show further improvements (Schaap et al 2015). A comparison of PM 2.5 predictions at 36 km×36 km, 12 km×12 km and 4 km×4 km grid resolutions over Northeast US and western Europe indicated no statistically significant difference, except in winter for the finer 4 km×4 km grid resolution, which was due to improved emissions representation from urban areas (Fountoukis et al 2013). Similarly, no significant improvements were identified for health assessments as related to PM 2.5 when implementing finer 4 km×4 km grid spacing (Thompson et al 2014). Studies support that for current CTM formulations, 12 km×12 km is a suitable grid resolution for PM 2.5 predictions (Appel et al 2011, Schaap et al 2015 and results from this study seem to support that conclusion. It appears that the 12 km×12 km resolution represents an optimum trade-off between emissions representation within CTMs, and errors resulting from the higher-resolution emission inventory and meteorological inputs that can increase and propagate non-linearly within CTM algorithms (Valari andMenut 2008, Schaap et al 2015). Implementation of the 12 km×12 km grid resolution reduces computational time and resources required for generating predictions and output data management (Gan et al 2016). However, finer spatial resolution is desirable for reactive nitrogen deposition studies that are an important consequence of air pollution, with 1 km×1 km grid resolution identified as necessary for assessment of ecosystem damage (Dore et al 2012), especially near agricultural sources (Butler et al 2015). Therefore, the discussion about what is preferable spatial resolution given the current state of CTMs and what is desirable for health and ecosystem studies is far from closed.
Our findings come with two limitations. First, we consider a short modeling episode of one month (May 2011), which was an active planting month in the Corn Belt. Second, CAMx does not currently include bidirectional treatment of NH 3 fluxes. Given that exploring these limitations was out of scope for this study, we recommend two key directions for future research. First, a a higher-resolution monitoring network for NH 3 fluxes and concentrations is required to spatially increase coverage over both rural and urban locations, and increase time resolution in NH 3 flux measurements from fertilized fields. Such measurements will help evaluatete emissions estimated from process modeling and satellite-based approaches, and reduce uncertainties in total NH 3 emissions and also capture temporal occurrence of distinct emission peaks observed in the field . Second, increased specificity and regularity of data collection related to fertilizer type and management practices will also be required to support such efforts (US EPA 2011). With continued research to improve emission representation at higher spatial resolutions and CTM formulations, the impact of grid resolution on ambient NH 3 and PM 2.5 should be revisited.

Conclusion
In this paper we report findings from research aimed to understand the impact of NH 3 emissions from fertilizer application on CAMx predictions of ambient NH 3 and PM 2.5 concentrations. Based on our findings, we conclude that future studies should carefully account for tradeoffs in model prediction accuracy when increasing grid resolution for emission inventory and meteorological inputs. While total NH 3 emissions play a key role in determining ambient air pollutant concentrations, we highlight that how NH 3 emissions are distributed in space and time within chemical transport models (CTMs) also significantly impacts total magnitude and spatial distribution of predictions of ambient NH 3 and PM 2.5 concentrations, while keeping other inputs constant, including meteorological and model parameters, and emissions of other precursors and primary pollutants. Given that differences in spatial patterns of PM 2.5 species concentrations have implications for identifying geographical areas of concern for human health, environmental justice, and ecosystem services, further research is required to study abundance of ambient NH 3 predicted by CTMs at high-spatial and temporal scales and the impact on PM 2.5 formation. An approach, as we demonstrate, that uses process-based models to develop on-or off-line spatial and temporal factors can be further developed for preparing inputs to CTMs.