2016 Progress Report: Mechanistic Air Quality Impact Models for Assessment of Multiple Pollutants at High Spatial ResolutionEPA Grant Number: R835873C001
Subproject: this is subproject number 001 , established and managed by the Center Director under grant R835873
(EPA does not fund or establish subprojects; EPA awards and manages the overall grant for this center).
Center: Center for Air, Climate, and Energy Solutions
Center Director: Robinson, Allen
Title: Mechanistic Air Quality Impact Models for Assessment of Multiple Pollutants at High Spatial Resolution
Investigators: Adams, Peter , Apte, Joshua S. , Coggins, Jay S. , Hill, Jason , Marshall, Julian D. , Millet, Dylan B , Pandis, Spyros N.
Institution: Carnegie Mellon University , The University of Texas at Austin , University of Minnesota
EPA Project Officer: Chung, Serena
Project Period: May 1, 2016 through April 30, 2021
Project Period Covered by this Report: May 1, 2016 through April 30,2017
RFA: Air, Climate And Energy (ACE) Centers: Science Supporting Solutions (2014) RFA Text | Recipients Lists
Research Category: Air , Climate Change
The primary objectives of Project 1 are to advance the state-of-the-art in spatially explicit multipollutant air quality impact assessments by evaluating the performance of chemical transport models (CTM) at high spatial resolution, developing and evaluating next-generation reduced complexity models (RCM), and performing multipollutant and regional source apportionment and intake fraction analyses. This project interacts with other center activities: we will evaluate CTM predictions of regional and intra-urban variability against high-resolution monitoring (Project 2), and tools developed here will evaluate future air quality scenarios (Project 4) and assessments of health outcomes (Project 5). We also will provide source and species-resolved PM2.5 information to the land use regression (LUR) models (Project 3).
Project 1 focused on the development, evaluation and application of mechanistic air quality models, both chemical transport models (CTMs) and reduced-complexity models (RCMs).
Development and Evaluation of Chemical Transport Models (CTMs)
Chemical transport modeling activities fall into two categories: 1) historical modeling of exposures to PM2.5 and related pollutants and 2) high-resolution (1 km) modeling of the present day.
Historical modeling of exposure to PM2.5 and related pollutants (1980-2015).
The goal of this task is to use a chemical transport model (CTM) to produce gridded estimates (at 36x36 km) of health-relevant air pollutants over the time period 1980-2015. Together with empirically based exposure estimates from Project 3, these will be provided to Project 5 for use in health studies. During this reporting period, a review has been performed of prior modeling efforts in this area. As a result of this, one key need was identified: the CTM requires an inventory of pollutant emissions over the relevant time period, and no suitable one currently exists.
An emissions dataset has been acquired that is based on EPA’s National Emissions Inventory (NEI) and covers part (1990 to 2010) of the desired time period. The emissions processing program (the Sparse Matrix Operator Kernel Emissions system [SMOKE]) that is necessary to convert these emissions into CTM-ready format has been benchmarked, and used to process 1990 emissions.
A procedure for scaling this inventory forward and backward in time has been developed. Because this procedure relies on changes in activity levels and emission factors, extensive literature review has been performed to identify those parameters with as much spatial resolution as possible. Records from 1980 to present of dozens of spatially-resolved activity indicators (e.g., vehicle-miles traveled, acres of agricultural tillage, BTUs of power plant output) have been compiled. Published and measured emission factors have been gathered and compared. Previous EPA modeling work from the 1980s has been reviewed, and used to constrain estimates.
A large dataset of air quality measurements (PM2.5, PM10, NO2, NOx, SO2, O3, EC/OC, SO4-, NO3-) has been acquired and processed for eventual use in evaluating the CTM outputs. This has been used to investigate regional trends in pollutant levels and to constrain emissions trends. Significant effort has been put into finding the few measurements of PM2.5 that were performed prior to 1992; these will be used for model evaluation.
As the emissions inventory is nearing completion, focus has turned towards the other major input needed: meteorological fields. The development of these inputs using the Weather Research and Forecasting (WRF) model is currently underway.
High-resolution (1 km) modeling of present-day air quality.
Another task of Project 1 is to develop and evaluate capabilities to simulate pollution exposures at the intra-urban scale using a 1 km CTM grid. These high-resolution simulations are focused on the three target cities (Pittsburgh, Austin, and Oakland CA) where Project 2 is making detailed observations. These new data will be used to evaluate model performance, specifically how well the models can predict observed intra-urban pollution gradients. These simulations require quality emissions inventories at the 1 km scale. We are using SMOKE for this purpose and are currently focusing on Pittsburgh, where the first set of measurements are being made. While we will use default SMOKE spatial surrogates for many emissions categories, we are developing Pittsburgh-specific inventories for two key sectors thought to drive much of the intra-urban variability: traffic and restaurants. For traffic, we are collaborating with CMU colleagues (Prof. Sean Qian and his group) who are expert traffic modelers to generate high spatially and temporally resolved traffic activity data.
Results from Project 2 as well as literature data indicate that commercial cooking activities contribute to the organic aerosol mass in urban areas. Currently the SMOKE modeling platform v6.3 uses population density as the spatial surrogates to distribute county-level estimates into modeling grids. We have obtained restaurant location data from the Google Places API for the Pittsburgh Combined Statistical Area (CSA) and surrounding counties. At the 12 km resolution, population density is very well correlated with restaurant density with an R2 of 0.88, but as the resolution increases the correlation breaks down with R2=0.55 for 4 km and R2=0.2 at 1 km. Therefore population-based spatial surrogates are not robust at the high resolutions needed to resolve emissions in a neighborhood scale. Essentially, at 12 km, population density does a good job since it reflects the fact that there are more restaurants in urban areas than rural ones. However, within Pittsburgh as with most cities, clusters of restaurants exist in specific commercial areas. Therefore, we have constructed a draft 1x1 km cooking emission inventory using restaurant location as the spatial surrogate for cooking emissions. In the next reporting period, these emissions will be used to model air quality using PMCAMx at 1 km resolution and these modeling results will be evaluated using the field data collected in the Center for Air, Climate, and Energy Solutions (CACES) Project 2.
An area of special focus for our 1 km modeling is the chemical transport modeling of ultrafine particulate matter (PM0.1) over Pittsburgh. In addition to associations with negative health outcomes such as ischemic heart disease and respiratory illness, currently unregulated PM0.1 has high spatial variability compared to PM2.5. Using Particulate Matter CAMx Ultrafine (PMCAMx-UF), a specialized version of the CAMx air quality model, and appropriate high resolution emission inventory inputs, we will simulate PM0.1 number concentrations over Pittsburgh across seasons in 2016 and 2017 and will compare results against Project 2 observations. During this past period, we developed the computational machinery to run PMCAMx-UF in a nested fashion starting with a 36 km resolution domain that covers the entire United States, and zooming into a 1 km domain over Pittsburgh. We performed test runs using available emission and meteorology input from 2001 while we develop more updated, higher resolution inputs.
Development and Evaluation of Reduced-Complexity Models (RCMs)
CACES employs three reduced-complexity models: AP2, EASIUR, and InMAP. A central activity of the past reporting period has been the drafting of an intercomparison paper that evaluates social cost estimates from these three models, described in more detail below. Additionally, each of the individual RCMs has undergone some enhancements.
Intercomparison of social cost estimates from three RCMs
Reliable estimates of externality costs—such as the costs of premature mortality from exposure to air pollution—are critical for policy analysis. To estimate these costs, ‘state of the science’ air quality models are often used to predict ambient concentrations and exposures from a given emission scenario. Full physical models, however, require expertise, time, and dedicated computational resources to operate, greatly limiting their application. To facilitate broader analysis, several datasets of social costs have been produced by a set of models known as reduced complexity models (RCMs). We compared per-metric ton marginal social costs of premature mortality from ambient PM2.5 exposure as estimated by three RCMs: Air Pollution Emission Experiments and Policy (APEEP/AP2), Estimating Air pollution Social Impacts Using Regression (EASIUR), and the Intervention Model for Air Pollution (InMAP). We compared the costs of ground level emissions in each county of SO2, NOx, NH3, and inert primary PM2.5 among the three models. Despite fundamental structural differences among the three models, predicted marginal social costs are generally within the same order of magnitude and usually within a factor of 2 or 3. The agreement varies with the complexity of the chemistry that links the emissions to their equivalent ambient PM2.5 concentrations; predictions are most similar for primary PM2.5 and most different for NOx and NH3. Even with these differences, the three models generate robust rankings of national-level air quality policies based on social costs and benefits summed across pollutants and geographical locations. For example, national emission-weighted damages for the United States varied across the three models by only 25%, 23%, 33%, and 12% for inert primary PM2.5, SO2, NOx, and NH3 respectively. However, differences in model predictions may matter for applications that are more source- or location-specific.
A manuscript entitled “Air quality social cost estimates from reduced-complexity models: critical review and guide for users” is in the final stages of internal review and will be submitted to Risk Analysis in the next 1-2 months.
Dr. Muller has begun to update the module of AP2 that connects emissions of nitrogen oxides (NOx) to ammonium nitrate, an important constituent of ambient fine particulate matter (PM2.5). The updates entail using regression analysis applied to output from the PMCAMx photochemical air quality model to characterize or estimate the relationships between ambient sulfate, ammonium, and ammonium nitrate. This is done using ordinary least squares regression. The concentration data provided by PMCAMx are hourly. Temperature and atmospheric pressure data allow the regression models to control for these important confounders.
Preliminary results suggest that this method produces baseline ambient concentration estimates of ammonium nitrate that are strongly correlated with the available ambient monitoring data (provided by USEPA Air Quality System [AQS]). Additional tests will compare the ammonium nitrate concentrations produced by AP2 to those predicted by the PMCAMx model. This new module for AP2 will also be used to estimate marginal ($/ton) damage estimates for NOx emissions.
The effort to update nitrate modeling in AP2 is part of a larger update to the 2014 model year. Major components of this update include U.S. Census data for human populations. Specifically, these data consist of county-level population estimates for each of 19 age groups. Another important data update is to baseline mortality rates provided by the CDC. These data are also provided at the county-level by age group. Finally, AP2 is being updated to include the USEPA’s 2014 National Emission Inventory (NEI). These data consist of county-level estimates of emissions for the five pollutant species that AP2 tracks: NOx, PM2.5, sulfur dioxide (SO2), ammonia (NH3), and volatile organic compounds (VOC). The NEI also provides point source emissions data for these species which will be imported into AP2’s current structure.
The original version of the EASIUR model was source-oriented. That is, it provided per ton social costs for emissions from a given location, summing damages over all downwind receptors. In this reporting period, we developed a source-receptor version of EASIUR, known as the APSCA (Air Pollution Social Cost Accounting) model. APSCA disaggregates the downwind damages, allowing for a full source-receptor treatment of PM2.5 pollution in the United States. This extends EASIUR beyond simply estimating aggregate downwind damages to allow us to estimate where those damages occur. Conversely, we can take a given receptor location and determine and quantify the source locations responsible for PM2.5 health effects at that receptor. A manuscript describing this work, entitled “Public health costs accounting of inorganic PM2.5 pollution in metropolitan areas of the United States using a risk-based source-receptor model” is now published in Environment International. The abstract is reproduced below.
In order to design effective strategies to reduce the public health burden of ambient fine particulate matter (PM2.5) in an area, it is necessary to identify the emissions sources affecting that location and quantify their contributions. However, it is challenging because PM2.5 travels long distances and most constituents are the result of complex chemical processes. We developed a reduced-form source-receptor model for estimating locations and magnitudes of downwind health costs from a source or, conversely, the upwind sources that contribute to health costs at a receptor location. Built upon outputs from a state-of-the-science air quality model, our model produces comprehensive risk-based source apportionment results with trivial computational costs. Using the model, we analyzed all the sources contributing to the inorganic PM2.5 health burden in 14 metropolitan statistical areas (MSAs) in the United States. Our analysis for 12 source categories shows that 80–90% of the burden borne by these areas originates from emissions sources outside of the area and that emissions sources up to 800 km away need to be included to account for 80% of the burden. Conversely, 60–80% of the impacts of an MSA's emissions occurs outside of that MSA. The results demonstrate the importance of regionally coordinated measures to improve air quality in metropolitan areas.
InMAP (Intervention Model for Air Pollution) offers an alternative to comprehensive air quality models for estimating the air pollution health impacts of emission reductions and other potential interventions. InMAP estimates annual-average changes in primary and secondary fine particle (PM2.5) concentrations—the air pollution outcome generally causing the largest monetized health damages—attributable to annual changes in emissions. InMAP leverages pre-processed physical and chemical information from the output of a state-of-the-science chemical transport model and a variable spatial resolution computational grid to perform simulations that are several orders of magnitude less computationally intensive than comprehensive model simulations.
In our model comparisons, InMAP recreates comprehensive model predictions of changes in total PM2.5 concentrations with population-weighted mean fractional bias (MFB) of −17% and population-weighted R2 = 0.90. Although InMAP is not specifically designed to reproduce total observed concentrations, it is able to do so within published air quality model performance criteria for total PM2.5.
We are currently working on further enhancements to the model:
We are currently developing a neural-network based chemical mechanism emulator to improve InMAP’s ability to represent nonlinear atmospheric chemistry without greatly increasing the computational requirements of the model. This will allow the model to estimate concentrations and health impacts of multipollutant mixtures more accurately.
We are using a box model version of the MOSAIC (Model for Stimulating Aerosol Interactions and Chemistry) chemical mechanism (Zaveri, et al., 2008) to train a neural network to predict, given initial concentrations of approximately 100 gas and aerosol chemical species, what the changes in concentration of those same 100 chemical species will be after one model time step. Overall agreement is good with an average R2 among species of 0.989, but for a small number of species the agreement is currently lower. In continuing work, we will focus on further refining and testing this method and incorporating it in the InMAP model.
Over the past year, we have added, deleted, or moved approximately 150,000 lines of the code that comprises the InMAP model, focusing on making the model more flexible, expandable, and easy to use. Significant improvements include:
- Rewriting parts of the preprocessor that converts outputs from chemical transport models (e.g., CMAQ, CAMx, or WRF-Chem) to InMAP data inputs to make it easier to add the capability to InMAP to ingest the output of different models.
- Allowing users to customize model output using mathematical expressions in a configuration file rather than having the outputs hard-coded into the model. This feature is important because it greatly simplifies the process exploring alternative health impact assessment methods, for instance using different concentration-response functions or calculating differences in exposure or health impacts among demographic groups.
We plan to continue this focus on usability in InMAP by adding a graphical user interface and a web-based version of the model.
The effect of spatial resolution on health impacts and disparities
While previous research has investigated the effects of horizontal spatial grid resolution on modeled air pollution concentrations and overall health impacts at various geographic scales, there is little knowledge of how air quality model resolution affects estimates of disparities in air pollution exposure among population sub-groups. The ability to identify potential changes in air pollution exposure disparities caused by future emissions changes is critical to conducting environmental justice evaluations of policy proposals.
We used InMAP to quantify how estimates of fine particulate matter exposure in the United States vary with grid resolution. We tested six variable resolution grids with progressively decreasing grid cell size, with population-weighted average grid cell length ranging from 80 to 3 km.
Figure 1.1 shows estimated population-weighted annual average PM2.5 concentrations—a surrogate for pollution exposure—as a function of spatial resolution. Coarse grids result in lower exposure estimates than fine grids. Exposure disparities between white and nonwhite populations grow as grid resolution increases. Among nonwhite populations, blacks see the largest increase in exposure disparities relative to whites from the coarsest to finest grids, 1.3 μg m-3 (a 65% increase). At the finest grid resolution, exposure disparities between whites and nonwhites persist across all 16 U.S. Census income groups. Exposure disparities by income alone remain relatively small, irrespective of grid resolution.
These results suggest that fine grid resolution is important in enabling air quality models to resolve air pollution concentration differences over the small spatial scales that are relevant for environmental justice analyses.
Figure 1.1: Population-weighted PM2.5 concentrations—a surrogate for pollution exposure—for different population groups (left panel) and exposures for non-white groups relative to the white group (right panel). Latino refers to individuals identified as Hispanic or Latino in the U.S. Census; Asian, Black, and White refer to individuals identified as the corresponding racial groups in the U.S. Census but not as Hispanic or Latino.
Source receptor matrix (ISRM)
We have used InMAP to create a source receptor matrix we call ISRM (InMAP Source Receptor Matrix). A source-receptor matrix is a set of relationships between emissions at source locations (PM2.5, SO2, NOx, NH3, and VOC emissions at 52,000 locations in the United States at three plume heights in this case) and receptor locations (52,000 locations in the United States in this case). This source-receptor matrix can be used to estimate the impacts of emissions almost instantaneously and with high spatial resolution in urban areas. We use ISRM to estimate marginal damages of emissions at each of the modeled pollution source locations. Our work reveals the enormous effect that emission location has on loss of life attributable to atmospheric emissions. The mean marginal damages of primary PM2.5 is $94,000 per metric ton (t-1), with large spatial variability among source locations (10th-90th percentile range: $11,000-$177,000 t-1). Spatial variability in marginal damages from PM2.5 precursors of VOC, NH3, SO2 and NOx is similarly large. Within individual counties, marginal damages can vary by over an order of magnitude, depending on emission location. Averaging across all emissions, one-third of damages occur within 8 km of the sources, and another one-quarter occur more than 256 km away. Thus, devising and prioritizing strategies for PM2.5 mitigation requires understanding spatial variability in marginal damages and including both near-to- and far-from-source impacts.
Source apportionment of PM2.5 exposure and exposure disparities
Efficiently reducing overall exposure to air pollution and disparities in exposure among demographic groups requires an understanding of which sources of emissions are contributing most to population exposure and exposure disparities, both in absolute terms and per unit of marginal emissions changes. To further the quantitative understanding of this issue, we are using the InMAP source receptor matrix (ISRM) to quantify the impacts to population exposure of both total and marginal emissions corresponding to each of the approximately 6,000 source classification codes (SCCs) in the 2014 U.S. National Emissions Inventory. We expect the future results of this work to suggest national and regional strategies for source-specific emissions control to reduce PM2.5 exposure and exposure disparities.
Coupling InMAP with life cycle emissions inventory models
In support of activities in Project 4, we have coupled the InMAP source receptor matrix (ISRM) with the GREET-cst (Tessum, et al., 2012) and Economic Input Output (EIO) life cycle inventory models. Having these models coupled will make it easier for Project 4 researchers to investigate the life cycle air quality impacts of technology or behavior scenarios. We additionally plan to couple the AP2 and EASIUR models to these emissions models to allow ensemble predictions of air quality impacts.
In the next reporting period, we will complete and evaluate historical modeling of criteria pollutants for the 1980 to 2015 time period. Model output will be evaluated against Inhalable Particulate Network (IPN), chemical speciation network (CSN), and IMPROVE data as well as other measurements. The results will be passed to Projects 3 and 5.
We also will complete the 1 km emissions inventories and modeling for Pittsburgh, evaluating the ability of CTMs to predict intraurban pollution variability against observations collected as part of Project 2. We then will use the validated model results to characterize the extent of human exposure to PM0.1 and sources across a typical urban area. Future analyses also will examine importance of model resolution, i.e., whether the computational expense of finer spatial resolution yields more information about human exposure gradients than coarse spatial resolution with comparable model inputs. Future applications of results to environmental justice also will yield more detailed information on whether certain demographic groups face disproportionate exposure to PM0.1.
We will extend EASIUR to include social cost estimates for VOCs using the volatility basis set framework. The enhancements to the treatment of difficult species like nitrate PM will be completed for AP2 and InMAP.
Journal Articles on this Report : 2 Displayed | Download in RIS Format
|Other subproject views:||All 11 publications||2 publications in selected types||All 2 journal articles|
|Other center views:||All 20 publications||8 publications in selected types||All 8 journal articles|
||Heo J, Adams PJ, Gao HO. Public health costs accounting of inorganic PM2.5 pollution in metropolitan areas of the United States using a risk-based source-receptor model. Environment International 2017;106:119-126.||
||Tessum CW, Hill JD, Marshall JD. InMAP: a model for air pollution interventions. PloS ONE 2017;12(4):e0176131 (26 pp.).||
Supplemental Keywords:reduced form, evaluation, mechanistic air quality modeling
Progress and Final Reports:Original Abstract
Main Center Abstract and Reports:R835873 Center for Air, Climate, and Energy Solutions
Subprojects under this Center: (EPA does not fund or establish subprojects; EPA awards and manages the overall grant for this center).
R835873C001 Mechanistic Air Quality Impact Models for Assessment of Multiple Pollutants at High Spatial Resolution
R835873C002 Air Quality Observatory
R835873C003 Next Generation LUR Models: Development of Nationwide Modeling Tools for Exposure Assessment and Epidemiology
R835873C004 Air Pollutant Control Strategies in a Changing World
R835873C005 Health Effects of Air Pollution and Mitigation Scenarios