Final Report

Project 10-008

Factors Influencing Ozone-Precursor Response in Texas Attainment Modeling

Principal Investigator: Daniel Cohan, Rice University

Co-Principal Investigators: Greg Yarwood, ENVIRON International Bonyoung Koo, ENVIRON International

Project Scientists: Xue Xiao, Rice University

Antara Digar, Rice University

Submitted to:

Texas Air Quality Research Program

August 31, 2011

Executive Summary

This report presents the findings of Texas Air Quality Program Project 10-008, which

investigated the influence of input uncertainties on model predictions of pollutant responsiveness

to emission controls. This project characterized how various model formulations (structural

uncertainty) and input parameters (parametric uncertainty) influence predictions of ozone-

precursor response in Texas State Implementation Plan (SIP) modeling episodes. Both Bayesian

and non-Bayesian approaches were applied to compute probabilistic representations of the

sensitivity of ozone to changes in precursor emissions.

Base case modeling was taken from TCEQ’s CAMx simulations of ozone during two month-

long episodes in 2006. Structural scenarios were then developed by applying alternate options for

the biogenic emissions model, the deposition scheme, the chemical mechanism, the global model

for deriving boundary conditions, and satellite-based photolysis rates. Screening analysis of the

impacts of these options on ozone concentrations and sensitivities led to a focus on scenarios

involving alternate choices for biogenic emissions model and chemical mechanism. The base

model achieved very low bias during the June 2006 episode (NMB = -1.0% relative to ozone

monitors in the 12-km domain), so the structural scenarios provide plausible alternatives but

could not dramatically improve model performance.

For parametric uncertainties, screening analysis identified the specific emission rates, reaction

rate constants, and boundary conditions that most influence ozone concentrations and their

sensitivities to nitrogen oxide (NOx) and volatile organic compound (VOC) emissions. Some

parameters such as ozone boundary conditions were found to impact concentrations far more

strongly than sensitivities, whereas the converse was true for some other parameters such as

anthropogenic VOC emissions.

Bayesian Monte Carlo analysis was then applied to weight the relative likelihood of alternate

structural and parametric scenarios, based on model performance in simulating observed

concentrations within the Dallas-Fort Worth (DFW) region during the June 2006 episode. Metric

1 evaluated model performance on high-ozone days at three DFW monitors, while Metric 2

considered average 8-hour ozone concentrations across all DFW monitors on each episode day.

A non-Bayesian metric for assigning weights based on standard model performance statistics

(Metric 3) was also developed and was applied to produce alternative weightings of the Monte

Carlo scenarios.

The Bayesian and non-Bayesian analyses generated probabilistic representations of ozone

responses to changes in precursor emissions and of model input parameters. All of the results

confirmed the findings of the base model that 8-hour ozone in the DFW region during the June

2006 episode was predominately NOx-limited. However, the three metrics yielded conflicting

shifts in the probability distributions of ozone sensitivities. For example, results from Metric 1

tended to increase the predicted sensitivity of ozone to NOx, whereas Metric 2 indicated slightly

greater sensitivity to VOC than originally modeled (Figure ES-1). Non-Bayesian Metric 3

yielded a slight shift toward greater sensitivity to VOCs, but retained the primarily NOx-limited

conditions of the base model. Further work is needed to refine the metrics and incorporate

consideration of other measurements beyond ozone for evaluating model performance.

Nevertheless, the project has demonstrated how probabilistic analyses via an ensemble approach

can supplement deterministic estimates of ozone response and characterize the uncertainty of

those results.

O3 sensitivity to ANOX Ozone sensitivity to AVOC M

ET

RIC

1

ME

TR

IC 2

ME

TR

IC 3

Figure ES-1. Cumulative probability distribution functions of the sensitivity of ozone at the Denton monitor in June 2006 to DFW anthropogenic NOx (left) and VOC (right) emissions for Bayesian metrics 1 and 2, and non-Bayesian Metric 3. Green line shows deterministic (base case) results.

TABLE OF CONTENTS

1. Introduction and Motivation 1

2. Description of Model and Inputs 3

3. Screening Analysis for Structural Scenarios 16

4. Screening Analysis for Parametric Scenarios 31

5. Bayesian and non-Bayesian Monte Carlo Analysis: Methods and Pseudo

Case Testing 38

6. Bayesian and non-Bayesian Monte Carlo analysis: Results and Discussion 52

7. Conclusions 72

8. References 76

9. Appendix 1: CB-6 Mechanism 80

1

1. Introduction and Motivation

Developing control strategies to provide for attainment of ozone standards relies upon

photochemical modeling to predict the responses of pollutants to emission changes. Model

estimates of pollutant-emission responses or “sensitivities” help inform control strategy selection

and indicate the amount of emission reduction needed to attain air quality standards. Thus,

models for attainment planning must reliably predict not only pollutant concentrations, but also

their responsiveness to emission changes.

However, despite the abundance of methods to gauge model performance for pollutant

concentrations, the accuracy of sensitivity predictions cannot be directly gauged. Pollutant

concentrations are routinely observed at numerous monitors, but concentration-emission

sensitivity relationships cannot be directly measured in the ambient atmosphere. Dynamic

evaluation of how pollutant concentrations respond to emission changes over weekly (i.e.,

weekday vs weekend) or interannual (e.g., before and after nitrogen oxides (NOx) SIP Call) time

scales (Gilliland et al., 2008; Dennis et al., 2010; Pierce et al., 2010) can provide a proxy for

ground-truthing sensitivity estimates, but the accuracy and uncertainty of those estimates remain

poorly characterized.

Uncertainties in pollutant-emission response can arise from choices of model formulations such

as vertical mixing scheme and chemical mechanism (structural uncertainty), and of input

parameters such as reaction rate constants and deposition velocities (parametric uncertainty)

(Fine et al., 2003; Deguillaume et al., 2008; Pinder et al., 2009). Uncertainty can be especially

pronounced for pollutants such as ozone which form from nonlinear interactions of multiple

precursor compounds (Lin et al., 1988; Cohan et al., 2005).

Recent work has introduced efficient methods for characterizing structural and/or parametric

uncertainties in ozone response (Pinder et al., 2009; Digar and Cohan, 2010; Tian et al., 2010).

Of these, only Pinder et al. jointly considered structural and parametric uncertainties. Other

studies have shown that Bayesian Monte Carlo analysis can be applied to weight the relative

likelihood of each model formulation based on its performance in simulating observed

concentrations (Bergin and Milford, 2000; Deguillaume et al., 2007a). These weighted results

2

can yield probability distributions for predicting the actual values of pollutant-emission

sensitivities as well as model inputs such as emission rates.

This project merges some of the best practices of the recent uncertainty studies and applies them

for a reanalysis of photochemical modeling episodes in Texas. Specifically, the reduced form

model methods introduced by Tian et al. (2010) and Digar and Cohan (2010) are extended

beyond parametric uncertainties to also consider structural uncertainties in model inputs and

formulations. Whereas these previous applications of reduced form models assumed each

scenario to be equally likely, here we apply the Bayesian Monte Carlo method of Bergin and

Milford to weight each scenario based on its performance in simulating observed pollutant

concentrations in Texas. This work yields probabilistic representations of ozone responses to

emission reductions in the Dallas-Fort Worth region, and highlights model inputs that most

strongly influence those estimates.

3

2. Description of Model and Inputs

All analyses are conducted with the Comprehensive Air quality Model with extensions (CAMx)

version 5.32 (http://www.camx.com) for Texas Commission on Environmental Quality (TCEQ)-

developed episodes in order to optimize the relevance and transferability of the results to Texas

attainment planners. Specifically, an August 13-September 15, 2006 episode is considered for

Houston-Galveston-Brazoria (HGB) (http://www.tceq.texas.gov/airquality/airmod/data/hgb8h2),

and a May 31-July 2, 2006 episode for Dallas-Fort Worth (DFW) and other regions

(http://www.tceq.texas.gov/airquality/airmod/data/dfw8h2). These two episodes were identified

by TCEQ based on their prevalence of observed eight-hour (8-hr) daily maximum ozone

concentrations exceeding the 1997 8-hr ozone National Ambient Air Quality Standard (NAAQS)

(Chapter 3 of (TCEQ, 2009)). For example, there were ten days on which 8-hr ozone exceeded

84 ppb at one or more HGB monitors during the Aug/Sept episode (Chapter 3, TCEQ, 2009).

The two episodes encompassed a sufficient number of days to reflect a variety of meteorological

conditions that favor ozone production. The HGB episode occurred during the TexAQS II field

intensive period, which enhances opportunity for comparing model results with observations.

Figure 2.1. DFW (left) and HGB (right) CAMx modeling domain.

4

The horizontal modeling domain structure consists of a coarse-grid (36 km resolution) eastern

US domain and nested fine-grid subdomains: an East Texas subdomain (12 km resolution), and a

DFW subdomain (4 km resolution) or HGB subdomain (4 km resolution) (Figure 2.1). The 2km

HG subdomain was not considered. CAMx applies two-way nesting across the domains.

Vertically there are 28 layers for the 36/12/4 km domains for the DFW episode, and 17 layers for

the 36/12 km domains and 28 layers for the 4 km domain for the HGB episode.

The input meteorology, emissions, initial/boundary concentrations and overall model

configuration were obtained from the TCEQ base case for the two episodes. Development details

of the base case inputs can be found elsewhere (TCEQ, 2009, Chapter 3). MM5 Version 3.7.3

was used to generate meteorological inputs to CAMx including wind speed, wind direction,

temperature, humidity, etc. The four model parameters have been evaluated against

meteorological observations by a statistical package developed by TCEQ and shown good

performance. Day-specific, gridded, speciated and temporally (hourly) allocated emission

inventories were created by the emissions modeling processors, version 3 of the Emissions

Processing System (EPS3) (for point, area, and mobile sources) and the Global Biosphere

Emissions and Interactions System (GloBEIS) biogenics emissions model (for biogenic sources).

The global air quality model MOZART was used to derive the boundary conditions. The base

case model uses the Carbon Bond Mechanism (CB-05) and the Regional Acid Deposition Model

(RADM) dry deposition scheme.

A primary aim of this work is to assess how structural changes to the photochemical model and

its inputs may influence predictions of ozone sensitivity to precursor emissions. Structural

uncertainty analysis thus requires obtaining a variety of alternate CAMx inputs that differ from

those used for the original TCEQ SIP modeling, focusing on the factors of emission inventories,

chemical mechanism, photolysis rates, boundary conditions, and dry deposition scheme Plausible

scenarios of model formulations and input parameter settings were then developed based on

combinations of these inputs.

5

2.1. Alternate Boundary Conditions

The original modeling provided by TCEQ used boundary conditions (BCs) generated by the

MOZART model. ENVIRON obtained the 2006 annual GEOS-CHEM global model simulation

outputs from NASA. BCs for both episodes on the national RPO 36-km modeling domain were

extracted from the GEOS-CHEM outputs using the GEOS2CMAQ processor (version 3.0). BCs

for the TCEQ 36-km modeling grid were generated using CAMx simulations with 2-way nesting

as described in reports to TCEQ1. Table 2.1 shows how the GEOS-CHEM model species are

mapped to CAMx CB05 species. The 3-hourly GEOS-CHEM output data were linearly

interpolated to hourly CAMx BCs.

Table 2.1. Mapping of GEOS-CHEM to CAMx CB05 species. CAMx GEOS-CHEM

NO2 NOx O3 Ox - NOx CO CO NXOY 2 N2O5 HNO3 HNO3 PNA HNO4 H2O2 H2O2 NTR R4N2 FORM CH2O ALD2 0.5 ALD2 ALDX RCHO PAR 0.333 PRPE + ALK4 +0.5 C3H8 + ACET + MEK + RCHO OLE 0.333 PRPE ETHA 0.5 C2H6 MEPX MP PAN PAN PANX PPN + PMN ISOP 0.2 ISOP ISPD MACR + MVK SO2 SO2 + DMS NH3 NH3 ISP 0.2 ISOP PSO4 SO4 + MSA

1 Reports are available at http://www.tceq.texas.gov/assets/public/implementation/air/am/contracts/reports/pm/5820784005FY0810-

20080831-environ-bcic_final_report.pdf http://www.tceq.state.tx.us/assets/public/implementation/air/am/contracts/reports/pm/5820784005FY0916-

20090730-environ-updated_bc.pdf

6

Figure 2.2 compares vertical profiles of average ozone boundary conditions from the MOZART

(base case) and GEOS-CHEM (alternate case) global model simulations. GEOS-CHEM BCs

exhibit higher ozone concentrations than MOZART BCs at all model layers and the differences

range from 0.7 ppb (west boundary) to 8 ppb (north boundary).

7

(a) West Boundary

(b) East Boundary

(c) South Boundary

(d) North Boundary

Figure 2.2. Average ozone boundary conditions (ppb) by model layer estimated from the

MOZART and GEOS-CHEM global model simulations (Aug 13 – Sep 15 episode).

8

2.2. Alternate Biogenic Emissions

Alternate biogenic emissions inputs were prepared for each modeling grid using the Model of

Emissions of Gases and Aerosols from Nature (MEGAN) biogenic emission model (version

2.03a) which employs updated land cover data with 1-km of spatial resolution based on satellite

and ground observations (Guenther et al., 2006).

Figure 2.3 shows spatial distributions of average daily total biogenic NOx, non-methane volatile

organic compound (NMVOC) and carbon monoxide (CO) emissions from the GloBEIS (base

case) and MEGAN (alternate case) biogenic emission models. Although both models produced

similar spatial patterns of biogenic emissions, MEGAN estimated lower NOx emissions and

higher NMVOC emissions than GloBEIS. CO emissions produced by MEGAN are higher than

those from GloBEIS. Table 2.2 presents domain total biogenic emissions from the two models

for the 36- and 12-km grids. Strong differences between biogenic emission estimates from BEIS

and MEGAN have been documented in previous studies (Carlton and Baker, 2011), though it is

unclear the extent to which they arise from different model formulations or different inputs such

as land cover data.

9

(a) NOx (GloBEIS)

(b) NOx (MEGAN)

(c) NMVOC (GloBEIS)

(d) NMVOC (MEGAN)

(e) CO (GloBEIS)

(f) CO (MEGAN)

Figure 2.3. Average daily total biogenic NOx, NMVOC and CO emissions over the 12-km

modeling domain estimated from the GloBEIS and MEGAN biogenic emission models (Aug 13

– Sep 15 episode)

10

Table 2.2. Domain total daily NOx, NMVOC and CO emissions (TPD) estimated from the

GloBEIS and MEGAN biogenic emission models (Aug 13 – Sep 15 episode)

Species 36-km Domain 12-km Domain GloBEIS MEGAN GloBEIS MEGAN

NOx 6,932 2,168 1,123 590 NMVOC 159,943 185,059 48,176 59,527 CO 16,622 19,684 4,457 5,723

2.3. Land Use Inputs for the Zhang Dry Deposition Scheme

CAMx version 5.3 (ENVIRON, 2010) offers two dry deposition options: the original approach

based on the work of Wesely (1989) and Slinn and Slinn (1980); and an updated approach based

on the work of (Zhang et al., 2001; 2003). The new Zhang scheme incorporates vegetation

density effects via leaf area index (LAI), possesses an updated representation of non-stomatal

deposition pathways including a better snow cover treatment, and has been tested extensively

through its use in daily air quality forecasting. The original Wesely/Slinn model is formulated

for 11 land use categories, while the Zhang model uses 26 land use categories (Table 2.3). A

new land use input file format is introduced that supports both land use categorizations as well as

an optional LAI data field. ENVIRON prepared the new land use inputs for the Zhang scheme

for each modeling grid.

11

Table 2.3. CAMx land use categories for the Wesely/Slinn model and the Zhang model.

Wesely/Slinn Model Zhang Model Category Number Land Cover Category Category

Number Land Cover Category

1 Urban 1 Water 2 Agricultural 2 Ice 3 Rangeland 3 Inland lake 4 Deciduous forest 4 Evergreen needleleaf trees 5 Coniferous forest, wetland 5 Evergreen broadleaf trees 6 Mixed forest 6 Deciduous needleleaf trees 7 Water 7 Deciduous broadleaf trees 8 Barren land 8 Tropical broadleaf trees 9 Non-forested wetlands 9 Drought deciduous trees 10 Mixed agricultural/range 10 Evergreen broadleaf shrubs 11 Rocky (with low shrubs) 11 Deciduous shrubs 12 Thorn shrubs 13 Short grass and forbs 14 Long grass 15 Crops 16 Rice 17 Sugar 18 Maize 19 Cotton 20 Irrigated crops 21 Urban 22 Tundra 23 Swamp 24 Desert 25 Mixed wood forest 26 Transitional forest

2.4. Chemistry Mechanism

2.4.1. CB6 Chemistry Mechanism

CB6 is the 6th version of the Carbon Bond mechanism (Yarwood et al., 2010). Several organic

compounds that are long-lived and relatively abundant, namely propane, acetone, benzene and

ethyne (acetylene), are added explicitly in CB6 so as to improve oxidant formation from these

compounds as they are oxidized slowly at the regional scale. CB6 also includes several updates

for organic and inorganic aerosol chemistry. The core inorganic chemistry mechanism for CB6

is based on evaluated data from the International Union of Pure and Applied Chemistry (IUPAC)

12

tropospheric chemistry panel as of January, 2010 (Atkinson et al., 2010). IUPAC also is the

primary source for photolysis data in CB6 with some data from the 2006 NASA/JPL data

evaluation (http://hdl.handle.net/2014/41648) or other sources for photolysis of some organic

compounds. There are changes to the organic chemistry for alkanes, alkenes, aromatics and

oxygenates. The most extensive changes are for aromatics and isoprene. Chemistry updates for

aromatics were based on the updated toluene mechanism (CB05-TU) developed by Whitten et al.

(2010) extended to benzene and xylenes. The isoprene mechanism was revised based on several

recently published studies (Paulot et al., 2009a; Paulot et al., 2009b; Peeters et al., 2009).

Compared to the CB05 mechanism, CB6 increases the number of model species from 51 to 76

and the number of reactions from 156 to 218. A listing of reactions in the CB6 mechanism is

provided in Appendix 1. Note that for CB6 (as well its modified form presented in Section 2.2),

the rate constant for Reaction 45 (OH+NO2) was adjusted as shown in Appendix 1 to reflect the

findings of Mollner et al. (2010). However, there remains considerable uncertainty in this rate

constant, as a subsequent chemical kinetics report from NASA JPL in 2011 chose not to update

this value (http://jpldataeval.jpl.nasa.gov).

2.4.2. Alternate Chemistry Mechanism

ENVIRON developed an alternate CB6 isoprene mechanism that produces more OH radicals at

low NOx conditions without breaking the mechanism evaluation against chamber experiments.

The motivation for this alternate mechanism is to explore a potential approach for addressing

reported underpredictions of OH by most chemical mechanisms in isoprene-rich, low-NOx

conditions (Lelieveld et al., 2008). Potential approaches to adjusting isoprene oxidation

mechanisms to influence HOx levels were discussed by (Archibald et al., 2010). Table 2.4 shows

the adjustments to the base CB6 mechanism that were adopted for the alternate CB6 mechanism,

which is termed CB6MOD4 in the sensitivity analyses in subsequent sections of this report.

13

Table 2.4. Changes between the base and alternate CB6 mechanisms

Rxn #

Base Mechanism Alternate Mechanism

116 GLY + OH = 1.7 CO + 0.3 XO2 + 0.3 RO2 + HO2

GLY + OH = 1.7 CO + 0.3 HO2 + 0.3 OH

118 GLY + NO3 = HNO3 + CO + HO2 + XO2 + RO2

GLY + NO3 = HNO3 + 1.7 CO + 0.3 HO2 + 0.3 OH

151 ISO2 + HO2 = 0.88 ISPX + 0.12 OH + 0.12 HO2 + 0.12 FORM + 0.12 ISPD

ISO2 + HO2 = 0.88 EPOX + 0.12 OH + 0.12 HO2 + 0.12 FORM + 0.12 ISPD

154 ISO2 = 0.8 HO2 + 0.04 OH + 0.04 FORM + 0.8 ISPD k = 1.0 s-1

ISO2 = 0.95 HO2 + 0.05 OH + 0.05 FORM + 0.95 ISPX k = 0.1 s-1

161 ISPX + OH = 0.904 EPOX + 0.933 OH + 0.067 ISO2 + 0.067 RO2 + 0.029 IOLE + 0.029 ALDX

ISPX = OH + CXO3 j(ISPX) = j(ALDX) × 60.0

ENVIRON evaluated the alternate CB6 mechanism by 6 environmental chamber experiments

using isoprene performed by University of California (UC) Riverside:

• Xenon arc Teflon Chamber (XTC) experiment 093

• Evacuable Chamber (EC) experiment 520

• Outdoor Teflon Chamber (OTC) experiments 309A, 309B, 316A & 316B

The box model performance was evaluated using model-experiment errors in the following

quantities:

• Maximum ozone concentration (Max(O3))

• Maximum D(O3 – NO)

− D(O3 – NO) = ([O3] – [NO])t=t – ([O3] – [NO])t=0 (quantifies the amount of O3

formed and NO oxidized during the experiment)

• NOx crossover time

− The time when [NO2] = [NO] (provides information on the rate of NO oxidation

into NO2)

14

Table 2.5 shows averages and standard deviations of model errors with the base and alternate

mechanisms. Performance differences between the base and alternate mechanisms are within

experimental uncertainty range. Figure 2.4 compares time-series plots of the box model

simulations and chamber experiments for ozone, NO and isoprene.

Table 2.5. Box model performance statistics for the base and alternate CB6 mechanisms

Model error (%)

in Max(O3)

Model error (%)

in Max(D(O3 – NO))

Model error (min)

in NOx crossover time

Base Alternate Base Alternate Base Alternate Average 2 5 1 3 7 6 Std. dev. 18 17 13 13 5 5

15

XTC093 EC520 OTC309A OTC309B OTC316A OTC316B

Figure 2.4. Time-series plots of the box model simulations and chamber experiments for ozone, NO and isoprene.

16

3. Screening Analysis for Structural Scenarios

Initial screening with Decoupled Direct Method in CAMx (CAMx-DDM) was applied to identify

the structural input choices that most significantly influence ozone-precursor sensitivity results.

Structural scenarios were constructed from combinations of discrete choices of four structural

factors: dry deposition (RADM/Zhang), chemical mechanism (CB05/CB-6), boundary

conditions (MOZART/GEOS-CHEM), and biogenic emissions models (GloBEIS/MEGAN),

plus one case using the alternate CB-6 mechanism (CB6MOD4) and one case using GOES-based

photolysis rates (SATTR) provided by the University of Alabama-Huntsville. CAMx-DDM

simulations were conducted on the 36/12 km domains for all 16 possible combinations of the

four main factors, plus two additional cases for CB6MOD4 and SATTR. Outputs were generated

to assess ozone (and other species) concentrations and their sensitivities to anthropogenic NOx

(ANOx) and VOC (AVOC) emissions from each of five regions (DFW, Austin, San Antonio,

HGB, and the rest of 12 km domain) for the June 2006 DFW episode. Analysis are mainly

focused on the DFW region (assumed control scenario) while the methodology can be applied to

other regions. Note that the base case uses the original RADM dry deposition scheme, the CB-05

chemical mechanism, boundary conditions from the MOZART global model, and GloBEIS-

generated biogenic emissions. The perturbation cases switch one or more of the structural factors

to its alternate setting.

3.1. Comparison between perturbation cases and the base case

We compare each perturbation case to the base case, focusing on results averaged across the

episode for an 8-hr window (10:00-18:00) each day. A fixed time window rather than daily 8-

hour maximum is chosen so that the same hours are compared in each case, making conditions

such as plume locations more comparable. The statistical measures that serve as the basis for the

comparisons are presented in Table 3.1.

17

Table 3.1. Statistical measures used to compare the structural cases.

Statistical measures Formula*

Correlation Coefficient (R2) ∑∑

∑

−−

−−=

NN

jj

N

jj

1

2

1

2

12

)basebase()casecase(

)basebase)(casecase(R

Root-Mean-Squares (RMS) (ppb) N

N

j∑ −= 1

2)basecase(RMS

Mean Bias (BIAS) (ppb) ∑ −=N

jN 1)basecase(1BIAS

Normalized Mean Bias (NMB) (percent) %100base

)basecase(NMB

1

1 ⋅−

=

∑

∑N

N

j

Normalized Mean Error (NME) (percent) %100base

|basecase|NME

1

1 ⋅−

=

∑

∑N

N

j

Mean Normalized Bias (MNB) (percent) %100base

base)(case1MNB1

⋅−

= ∑N

j

N

Mean Normalized Gross Error (MNGE)

(percent) %100

base|basecase|1MNGE

1⋅

−= ∑

Nj

N

* ‘base’ denotes concentrations or sensitivities from base case CAMx simulations; casej denotes

each perturbation case; jcase represents the average value of the N data points in that case.

18

Results are compared for ozone concentrations at US Environmental Protection Agency (EPA)

monitors within DFW (Table 3.2) and their sensitivities to anthropogenic NOx (ANOx) (Table

3.4) and VOC (AVOC) (Table 3.5) emissions from the DFW non-attainment region. The 8-hr

ozone concentrations from each structural case are also compared to ambient ozone observations

from EPA Air Quality System (AQS) database (http://www.epa.gov/air/data/aqsdb.html) (Table

3.3).

Among the individual choices in structural inputs, ozone concentrations are most influenced by

the chemical mechanism (CB-6), satellite-based photolysis rates (SATTR), and deposition

scheme (Zhang), based on the RMS results in Table 3.2. The boundary conditions, biogenic

emissions inventory, and modified form of CB-6 all exert much smaller influences on

concentrations. As expected, perturbation cases that involve combinations of structural changes

typically yield larger changes in ozone concentrations. Statistical evaluation of the scenarios

against ambient ozone observations shows that none of the perturbation scenarios dramatically

improve model performance in terms of bias or error (Table 3.3). This largely reflects the fact

that the base model achieved very low bias during the June 2006 episode, so perturbations can

impair that performance.

For informing the prioritization of control strategies, the sensitivities of ozone to changes in

emission rates are of paramount importance. Ozone sensitivities both to DFW ANOx (Table 3.4)

and to DFW AVOC (Table 3.5) are most influenced by the choices of photolysis rates (SATTR),

biogenic emissions (MEGAN), and chemical mechanism (CB-6). The other two structural input

choices (boundary conditions and deposition scheme) yield far smaller influences on ozone

sensitivities. The small impact of boundary conditions was to be expected, given the large size of

the 36-km domain (Figure 2.1) and the small impact of boundary conditions on ozone

concentrations (Table 3.2). However, the lack of influence of deposition scheme on sensitivities

is surprising, given its strong influence on concentrations. Switching between CB6 and

CB6MOD4 significantly changes ozone sensitivities to ANOx emissions, but has little impact on

sensitivity to AVOC emissions.

19

Table 3.2. Comparison of each perturbation case to base case simulations of 8-hour (10:00-

18:00) ozone concentrations at all regulatory monitors within the 12 km domain for the June

2006 episode.

Cases R2 RMS ppb

BIAS ppb

NMB %

NME %

MNB %

MNGE %

Base - - - - - - - Zhang Dep. (Z) 0.991 3.18 2.44 4.23 4.54 3.76 4.28 CB-6 (C) 0.964 6.69 5.22 9.07 9.99 8.19 9.64 GEOSCHEM_BC (G) 0.996 1.06 0.55 0.95 1.16 1.27 1.51 MEGAN EI (M) 0.987 1.77 -0.46 -0.80 2.29 -0.91 2.30 Z + C 0.972 11.69 10.37 18.00 18.07 16.88 17.03 Z + G 0.989 3.55 2.98 5.18 5.37 5.01 5.31 Z + M 0.988 3.04 1.86 3.22 4.05 2.65 3.82 C + G 0.961 6.84 5.55 9.64 10.36 9.06 10.07 C + M 0.972 6.59 4.96 8.62 9.54 7.54 9.00 G + M 0.983 1.90 0.13 0.23 2.53 0.44 2.78 Z + C + G 0.971 11.82 10.68 18.55 18.61 17.76 17.86 Z + C + M 0.977 11.47 9.94 17.25 17.36 15.95 16.16 Z + G + M 0.985 3.36 2.45 4.25 4.73 3.98 4.64 C + G + M 0.968 6.79 5.37 9.33 10.01 8.54 9.51 Z + C + G + M 0.975 11.66 10.34 17.95 18.04 16.97 17.12 CB6MOD4 0.971 6.67 5.34 9.27 9.88 8.44 9.37 CB6MOD4 vs CB6 0.995 1.24 0.12 0.19 1.51 0.31 1.70 SATTR* 0.910 3.87 1.76 2.80 3.04 3.32 3.55

* Results for SATTR are available only for 5/31 - 6/15.

20

Table 3.3. Comparison for daily eight-hour averaged ozone concentrations (8-hr [O3]) from the

structural cases, evaluated against observations from regulatory ozone monitors within the 12-km

domain for the June 2006 episode.

Cases R2 RMS ppb

BIAS ppb

NMB %

NME %

MNB %

MNGE %

Base 0.557 13.01 -0.61 -1.04 17.79 6.42 22.29 Zhang Dep. (Z) 0.576 12.95 1.88 3.22 17.08 10.17 22.13 CB-6 (C) 0.614 13.21 4.59 7.83 16.88 14.01 22.02 GEOSCHEM_BC (G) 0.555 13.01 0.02 0.04 17.76 8.03 22.77 MEGAN EI (M) 0.518 13.63 -1.06 -1.82 18.85 5.71 23.24 Z + C 0.556 17.15 9.96 17.00 22.20 24.22 29.03 Z + G 0.576 13.00 2.51 4.29 17.11 11.75 22.63 Z + M 0.542 13.47 1.30 2.22 17.99 9.26 22.91 C + G 0.613 13.28 4.97 8.49 16.99 15.14 22.52 C + M 0.587 13.72 4.42 7.54 17.58 13.69 22.66 G + M 0.517 13.59 -0.38 -0.65 18.79 7.40 23.71 Z + C + G 0.557 17.22 10.33 17.62 22.32 25.35 29.59 Z + C + M 0.532 17.49 9.61 16.40 22.45 23.61 29.21 Z + G + M 0.542 13.49 1.98 3.39 17.98 10.93 23.39 C + G + M 0.588 13.78 4.89 8.35 17.68 14.96 23.16 Z + C + G + M 0.533 17.58 10.08 17.20 22.61 24.89 29.81 CB6MOD4 0.597 13.56 4.78 8.16 17.39 14.66 22.78 SATTR* 0.361 10.50 -2.96 -4.23 12.16 -2.73 12.48

* Results for SATTR are available only for 5/31 - 6/15.

21

Table 3.4. Comparison of each perturbation case to base case for 8-hr sensitivities of ozone at

DFW monitors to ANOx emissions from the DFW region for the June 2006 episode.

Cases R2 RMS ppb

BIAS ppb

NMB %

NME %

MNB %

MNGE %

Base - - - - - - - Zhang Dep. (Z) 0.999 0.12 -0.01 -0.25 2.13 -0.87 1.93 CB-6 (C) 0.983 0.79 -0.40 -12.07 15.35 -2.32 4.77 GEOSCHEM_BC (G) 0.999 0.16 -0.09 -2.81 2.83 -2.30 2.29 MEGAN EI (M) 0.967 1.37 0.75 22.81 26.08 11.00 11.15 Z + C 0.983 0.77 -0.39 -11.79 15.34 -3.11 5.68 Z + G 0.999 0.19 -0.10 -3.12 3.76 -3.12 3.43 Z + M 0.967 1.38 0.74 22.44 25.97 9.97 10.84 C + G 0.978 0.92 -0.51 -15.42 18.01 -5.02 7.13 C + M 0.974 1.11 0.54 16.46 22.38 18.81 17.06 G + M 0.969 1.28 0.66 20.03 24.02 8.53 9.04 Z + C + G 0.978 0.91 -0.51 -15.28 18.07 -5.82 8.07 Z + C + M 0.973 1.14 0.56 16.84 22.77 18.47 18.03 Z + G + M 0.968 1.28 0.65 19.60 23.85 7.55 8.71 C + G + M 0.974 1.02 0.43 13.08 20.44 16.20 12.10 Z + C + G + M 0.974 1.04 0.44 13.33 20.71 15.83 13.10 CB6MOD4 0.971 0.82 -0.04 -1.06 16.11 -9.25 0.39 CB6MOD4 vs CB6 0.985 0.67 0.36 12.51 16.16 4.72 7.81 SATTR* 0.857 2.06 0.70 19.48 21.42 10.12 11.58

* Results for SATTR are available only for 5/31 - 6/15.

22

Table 3.5. Comparison of each perturbation case to base case for 8-hr ozone sensitivities at

DFW monitors to AVOC emissions from the DFW region for the June 2006 episode.

Cases R2 RMS ppb

BIAS ppb

NMB %

NME %

MNB %

MNGE %

Base - - - - - - - Zhang Dep. (Z) 0.999 0.02 -0.00 -0.80 2.04 -4.22 4.89 CB-6 (C) 0.964 0.44 0.26 63.35 63.45 122.38 118.80 GEOSCHEM_BC (G) 0.999 0.02 0.01 1.88 2.14 3.65 4.12 MEGAN EI (M) 0.948 0.17 -0.08 -19.33 21.90 -28.04 34.35 Z + C 0.970 0.43 0.25 61.86 61.97 115.51 107.52 Z + G 0.998 0.03 0.01 1.12 2.78 -0.13 3.19 Z + M 0.944 0.18 -0.08 -19.91 22.61 -31.06 39.45 C + G 0.959 0.47 0.27 67.12 67.19 131 128.21 C + M 0.970 0.25 0.12 30.22 32.31 65.66 62.58 G + M 0.952 0.16 -0.07 -18.07 20.84 -25.63 34.69 Z + C + G 0.963 0.46 0.27 65.82 65.91 124.56 121.33 Z + C + M 0.970 0.25 0.12 28.96 31.22 59.96 50.76 Z + G + M 0.948 0.17 -0.08 -18.62 21.53 -28.50 35.91 C + G + M 0.970 0.26 0.13 33.08 34.58 71.97 73.75 Z + C + G + M 0.969 0.26 0.13 31.98 33.62 66.60 64.22 CB6MOD4 0.964 0.43 0.24 59.80 60.26 69.50 76.76 CB6MOD4 vs CB6 0.991 0.09 -0.01 -2.17 7.42 -14.04 19.49 SATTR* 0.850 0.38 0.08 21.31 22.14 10.66 19.83

* Results for SATTR are available only for 5/31 - 6/15.

Figure 3.1 shows the diurnal profile of ozone sensitivities to DFW ANOx and AVOC emissions

for the structural cases, averaged over the June episode and the DFW monitors. Afternoon ozone

in DFW is primarily NOx-limited in all of the structural cases, with ozone about an order of

magnitude more sensitive to ANOx than AVOC. In general, the MEGAN case increases (relative

to the base case) ozone-ANOx sensitivities and decreases ozone-AVOC sensitivities during

daytime because of its stronger biogenic VOC emissions. The CB-6 case also affected daytime

ozone sensitivities but in the opposite direction, yielding strong sensitivities to AVOC. The new

deposition scheme (Zhang) mainly affected nighttime ozone sensitivities and exerted little

influence the 8-hr ozone sensitivities very much. Shallow planetary boundary layer conditions at

night magnify the influence of deposition on conditions near the ground.

23

Figure 3.1. Diurnal profile of ozone sensitivities to DFW ANOx and AVOC emissions for the

indicated structural cases, averaged over the June episode and the DFW region.

0 5 10 15 20 25-8

-6

-4

-2

0

2

4

6

Time (hr)

∂[O

3] / ∂

(ED

FW A

NO

x) (pp

b)

Sens of Region DFW to EDFW ANOx

baseZhang(Z)CB6(C)GEOS(G)MEGAN(M)

0 5 10 15 20 250

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Time (hr)

∂[O

3] / ∂

(ED

FW A

VOC

) (pp

b)

Sens of Region DFW to EDFW AVOC

baseZhang(Z)CB6(C)GEOS(G)MEGAN(M)

24

The results of the initial screening shown in the figure and tables above led us to retain biogenic

emissions and chemical mechanism as the primary structural factors for further full analysis,

including joint consideration with parametric uncertainties on the finer 4 km domain. Thus, four

structural cases — base case, CB-6, MEGAN, and CB-6+MEGAN — are targeted. Two other

structural factors, deposition scheme and boundary conditions, are excluded due to their much

smaller impacts on sensitivities. The satellite-based photolysis rates did show a strong influence

on concentrations and sensitivities, but were not used in this phase of the analysis because inputs

for the full episode are not yet available.

Figures 3.2 and 3.3 show the diurnal profile of ozone sensitivities to HGB and DFW ANOx and

AVOC emissions for each of the four targeted structural cases. Consistent with the findings from

the June DFW episode, the MEGAN case increased (relative to the base case) ozone-ANOx

sensitivities and decreased ozone-AVOC sensitivities during daytime because of the stronger

magnitude of biogenic VOC emissions. The CB-6 case also affected daytime ozone sensitivities

but in the opposite direction, again with strong increases in sensitivity to AVOC. The CB-

6+MEGAN case showed the combined effects of the two structural factors.

25

Figure 3.2. Episode (Aug/Sept) and HGB region averaged diurnal profile of ozone sensitivities

to HGB ANOx and AVOC emissions under the four targeted structural cases.

0 5 10 15 20 25-8

-6

-4

-2

0

2

4

6

8

10

Time (hr)

∂[O

3] / ∂

(EH

GB

AN

Ox) (

ppb)

Sens of Region HGB to EHGB ANOx

baseCB6(C)MEGAN(M)C+M

0 5 10 15 20 250

0.5

1

1.5

2

2.5

3

3.5

Time (hr)

∂[O

3] / ∂

(EH

GB

AVO

C) (

ppb)

Sens of Region HGB to EHGB AVOC

baseCB6(C)MEGAN(M)C+M

26

Figure 3.3. Episode (June) and DFW region averaged diurnal profile of ozone sensitivities to

DFW ANOx and AVOC emissions under the four targeted structural cases.

0 5 10 15 20 25-6

-4

-2

0

2

4

6

Time (hr)

∂[O

3] / ∂

(ED

FW A

NO

x) (pp

b)Sens of Region DFW to EDFW ANOx

baseCB6(C)MEGAN(M)C+M

0 5 10 15 20 250

0.2

0.4

0.6

0.8

1

1.2

1.4

Time (hr)

∂[O

3] / ∂

(ED

FW A

VOC

) (pp

b)

Sens of Region DFW to EDFW AVOC

baseCB6(C)MEGAN(M)C+M

27

3.2. Comparison between CB-6 and CB6MOD4 in CAMx

ENVIRON developed the alternate form of CB-6, referred to here as CB6MOD4, specifically to

test the importance of altered reaction rates that would boost the production of OH radicals under

low-NOx conditions without undermining performance relative to chamber experiments.

Considerable scientific attention has been devoted in recent years to the inability of most

chemical mechanisms to predict sufficient levels of OH in isoprene-rich, low-NOx conditions

(Lelieveld et al., 2008). Since the modified chemical mechanism allows the impact of isoprene

oxidation reactions to be isolated, we devote attention in the following subsections to comparing

the CB6 and CB6MOD4 results.

3.2.1. Comparison for OH concentrations between CB6 and CB6MOD4

Figure 3.4 shows the simulated OH fields using the CB6 and CB6MOD4 mechanisms at local

noontime on a sample high ozone day, June 30.

Figure 3.4. OH concentrations predicted by CAMx with the CB6 (left) and CB6MOD4 (middle)

chemical mechanisms on June 30 at 12:00 CST, and the difference (right).

OH concentrations are increased by the CB6MOD4 mechanism, not only in low NOx regions,

but also in the urban area. However, on a percentage basis, the greatest increases in OH

concentrations are simulated in rural low NOx areas. Thus, the modified chemical mechanism did

in fact enhance rural OH levels, though it is beyond the scope of this study to explore whether

28

the increase in OH levels would be sufficient to address the underpredictions noted in the

literature.

3.2.2. Comparison for ozone concentrations between CB6 and CB6MOD4

Figure 3.5 shows the simulated fields of ozone concentrations using the CB6 and CB6MOD4

mechanisms and their differences on a June 30 at 15:00 CST. The major high ozone plumes are

captured by both mechanisms, and the differences in concentrations are relatively small.

Figure 3.5. Ozone concentrations predicted by CAMx with the CB6 (left) and CB6MOD4

(middle) chemical mechanisms on June 30 at 12:00 CST, and the difference (right).

Averaged over the EPA sites in the 12-km domain over all episode days, 8-hr ozone

concentrations are only 0.12 ppb (0.19%) higher in CB6MOD4 than in CB-6 (Table 3.2). The

slight increase in 8-hr ozone concentrations by CB6MOD4 likely resulted from the increase in

OH concentrations noted in the previous subsection.

3.2.3. Comparison for ozone sensitivities between CB6 and CB6MOD4

Despite the small average change in ozone concentrations, were the sensitivities of ozone to

emissions affected by the modifications to the chemical mechanism? Figure 3.6 shows the

simulated ozone sensitivities to DFW ANOx emissions using the CB6 and CB6MOD4

mechanisms and their differences on several days during the June episode.

29

On June 10, 2006

On June 14, 2006

On June 30, 2006

Figure 3.6. Sensitivities of ozone to DFW ANOx emissions under the CB6 (left) and CB6MOD4

(middle) chemical mechanisms on three afternoons, and the difference (right).

Sensitivities of ozone in the DFW region to DFW ANOx emissions are usually negative during

early morning rush hours, and then turn positive during daytime hours (NOx-limited regime),

30

with smaller positive sensitivities simulated downwind. The plots for selected days show larger

simulated ozone sensitivities in the DFW region with CB6MOD4 than CB6 during the afternoon.

Statistically, CB6MOD4 also indicated larger sensitivities of 8-hr ozone concentrations to DFW

ANOx emissions than CB6 when considering the whole episode (0.36 ppb and 12.51% higher,

Table 3.3). With OH increased by CB6MOD4, daytime photochemistry became faster. More OH

removed NO2 (via OH+NO2HNO3) and increased the ozone production efficiency of each

NOx molecule, making ozone formation more sensitive to NOx emissions.

Although the differences caused by the modified mechanism were small, it is interesting to note

the much larger percentage impacts on ozone sensitivities than on ozone concentrations. This is

consistent with an earlier study which showed that changes in uncertain reaction rate constants

can exert much larger percentage influences on sensitivities than on concentrations (Cohan et al.,

2010).

31

4. Screening Analysis for Parametric Scenarios

A key feature of this project is the joint consideration of structural and parametric uncertainties

influencing model results. Following upon the analysis of the structural scenarios in the previous

section, parametric uncertainties are considered here.

Screening was conducted on a list of input parameters that were hypothesized to potentially

impact ozone concentrations and sensitivities. The list includes most of the parameters that have

been identified by previous studies that ranked the relative importance of various input

parameters in generating uncertainties in model outputs for ozone (Gao et al., 1995; Bergin et al.,

1999; Hanna et al., 2001b; Cohan et al., 2010; Digar and Cohan, 2010). Specifically, we focused

on the emission rates, reaction rate constants, and boundary conditions listed in Table 4.1. Each

parameter was assumed to have a lognormal probability distribution, characterized by the sigma

value reported in Table 4.1. Note that the “reported uncertainty” column of Table 4.1 reflects

how uncertainty for that parameter was reported in the literature; we then computed sigma from

the associated factors of uncertainty by the equation: sigma=ln(Factor).

32

Table 4.1. Uncertain CAMx input parameters considered in the initial screening.

Parameter Reported Uncertainty*

Factor of Uncertainty@ 1σ# Reference

Emission Rates: Domain-wide NOX ± 40% (1σ) 1.40 0.336 (Deguillaume

et al., 2007b)

Domain-wide Anthropogenic VOC ± 40% (1σ) 1.40 0.336 (Deguillaume et al., 2007b)

Domain-wide Biogenic VOC ± 50% (1σ) 1.50 0.405 (Deguillaume et al., 2007b)

Reaction Rate Constants: All Photolysis Frequencies Factor of 2 (2σ) 1.41 0.347 (Hanna et al.,

2001a)

R(All VOCs+OH) ± 10% (1σ) 1.10 0.095

(Hanna et al., 2001a),

(Deguillaume et al., 2007b)

R(OH+NO2) ± 30% (2σ) 1.14 0.131 (Sander S P, 2006)

R(NO+O3) ± 10% (1σ) 1.10 0.095 (Hanna et al., 2001a)

Boundary Conditions: Boundary Cond. O3 ± 50% (2σ) 1.23 0.203 (Deguillaume

et al., 2007b)

Boundary Cond. NOX Factor of 3 (2σ) 1.73 0.549 (Deguillaume et al., 2007b)

Boundary Cond. HNO3 Factor of 3 (2σ) 1.73 0.549 (Deguillaume et al., 2007b)

Boundary Cond. PAN Factor of 3 (2σ) 1.73 0.549 (Deguillaume et al., 2007b)

Boundary Cond. HONO Factor of 3 (2σ) 1.73 0.549 (Deguillaume et al., 2007b)

Boundary Cond. N2O5 Factor of 3 (2σ) 1.73 0.549 (Deguillaume et al., 2007b)

*Uncertainty as reported in literature, all of which assumed lognormal distribution;@Factor by which base value is multiplied or divided for a ±1σ range lognormal distribution; # sigma=ln(Factor)

To screen parameters that affect O3 concentrations and responses to emission, relevant impact

factors were evaluated by computing first-order sensitivity of O3 to source controls (DFW

anthropogenic NOX and VOCs) and its cross-sensitivity with each uncertain parameter, for a 2

week sub-episode spanning from June 6 - 20, 2006 for the base-case simulation. The impact

factors take into account both the uncertainty in the input parameter itself, and the rate at which a

33

change in that parameter leads to a change in model output. Specifically, impact factors for the

influence of a parameter on concentrations are calculated as

σjSi,j(1)/Ci

and impact factors for the influence of a parameter on sensitivities are calculated as

σjSi,j,k(2)/Si,k

(1)

In the equations above, σj denotes the 1σ uncertainty in parameter j (taken from column 4 of

Table 4.1); Si,j(1) denotes the first-order sensitivity coefficient of concentrations Ci to parameter j;

and Si,j,k(2) denotes the cross-sensitivity of Ci to parameters j and k. An impact factor of 0.1 would

mean that a one sigma increase in the input parameter would cause an approximately 10%

change in that concentration or first-order sensitivity.

CAMx-HDDM modeling was conducted on the 36/12/4 km domains, and the results within the

12-km domain were used to select key parameters needed for our final Bayesian analysis. Tables

4.2-4.4 show the results for these screening tests. The parameters with impact factors greater

than 0.010 (next to last column) were selected in each case (final column).

34

Table 4.2: Results of the screening test for the selection of uncertain input parameters

influencing O3 concentrations.

Parameter Uncertainty in parameter (1σ)

1st order sensitivity# (ppb) Impact Factor* Selection

(Y/N)

Emission Rates Domain-wide NOX 0.336 19.74 0.105 Y Domain-wide biogenic VOC 0.405 4.08 0.026 Y

Domain-wide anthropogenic VOC 0.336 1.07 0.006 N

Reaction Rates All photolysis rates 0.347 16.61 0.091 Y R(NO2+OH) 0.131 -8.02 -0.017 Y R(NO+O3) 0.095 -15.35 -0.023 Y R(all VOCs+OH) 0.095 1.75 0.003 N Boundary Conditions BC(O3) 0.203 11.23 0.036 Y BC(NOX) 0.549 0.20 0.002 N BC(HNO3) 0.549 0.09 0.001 N BC(PAN) 0.549 0.98 0.008 N BC(HONO) 0.549 0.00 0.000 N BC(N2O5) 0.549 0.00 0.000 N #First-order sensitivity of O3 to each uncertain parameter at time of the maximum daily 8-h average O3, averaged over all the 12-km grid-cells corresponding to the regulatory monitors within DFW and over a 2 week period in summer spanning from June 6-20, 2006. *Impact factor: The fractional change in first-order sensitivity of ozone to emissions, due to a 1σ change in an input parameter. Computed as Impact Factor = σSj

(1)/C where Sj(1) is the first-order sensitivity of O3 to an uncertain

parameter and C is the concentration of O3 (63.33 ppb).

35

Table 4.3: Results of the screening test for the selection of uncertain input parameters

influencing O3 sensitivity to ANOx.

Parameter Uncertainty in parameter (1σ)

Cross-sensitivity# (ppb) Impact Factor*

Selection (Y/N)

Emission Rates Domain-wide NOX 0.336 -7.444 -0.463 Y Domain-wide biogenic VOC 0.405 2.881 0.216 Y

Domain-wide anthropogenic VOC 0.336 1.172 0.073 Y

Reaction Rates All photolysis rates 0.347 6.244 0.401 Y R(NO2+OH) 0.131 -2.346 -0.057 Y R(NO+O3) 0.095 -3.276 -0.058 Y R(all VOCs+OH) 0.095 1.182 0.021 Y Boundary Conditions BC(O3) 0.203 0.153 0.006 N BC(NOX) 0.549 -0.006 -0.001 N BC(HNO3) 0.549 -0.003 -0.000 N BC(PAN) 0.549 -0.032 -0.003 N BC(HONO) 0.549 -0.000 -0.000 N BC(N2O5) 0.549 -0.000 -0.000 N #Cross-sensitivity of O3 to DFW anthropogenic NOX (ANOX) emissions and each uncertain parameter at time of the maximum daily 8-h average O3, averaged over all the 12-km grid-cells corresponding to the regulatory monitors within DFW and over a 2 week period in summer spanning from June 6-20, 2006. *Impact factor: The fractional change in first-order sensitivity of ozone to emissions, due to a 1σ change in an input parameter. Computed as Impact Factor = σSj,k

(2)/Sj(1) where Sj

(1) is the first-order sensitivity of O3 to DFW ANOX (5.40 ppb) and Sj,k

(2) is the cross sensitivity of Sj(1) with an uncertain parameter.

36

Table 4.4: Results of the screening test for the selection of uncertain input parameters

influencing O3 sensitivity to AVOC.

Parameter Uncertainty in parameter (1σ)

Cross-sensitivity# (ppb) Impact Factor* Selection

(Y/N)

Emission Rates Domain-wide NOX 0.336 0.925 0.496 Y Domain-wide biogenic VOC 0.405 -0.493 -0.319 Y

Domain-wide anthropogenic VOC 0.336 -0.279 -0.150 Y

Reaction Rates All photolysis rates 0.347 0.164 0.091 Y R(NO2+OH) 0.131 0.140 0.029 Y R(NO+O3) 0.095 -0.161 -0.024 Y R(all VOCs+OH) 0.095 0.092 0.014 Y Boundary Conditions BC(O3) 0.203 -0.131 -0.042 Y BC(NOX) 0.549 -0.001 -0.001 N BC(HNO3) 0.549 -0.000 -0.000 N BC(PAN) 0.549 -0.002 -0.002 N BC(HONO) 0.549 -0.000 -0.000 N BC(N2O5) 0.549 0.000 0.000 N #Cross-sensitivity of O3 to DFW anthropogenic VOC (AVOC) emissions and each uncertain parameter at time of the maximum daily 8-h average O3, averaged over all the 12-km grid-cells corresponding to the regulatory monitors within DFW and over a 2 week period in summer spanning from June 6-20, 2006. *Impact factor: The fractional change in first-order sensitivity of ozone to emissions, due to a 1σ change in an input parameter. Computed as Impact Factor = σSj,k

(2)/Sj(1) where Sj

(1) is the first-order sensitivity of O3 to DFW AVOC (0.626 ppb) and Sj,k

(2) is the cross sensitivity of Sj(1) with an uncertain parameter.

Although there was considerable overlap in the selected parameters, there were also some

differences in those found to most influence concentrations and the two sensitivities. Domain-

wide NOx and biogenic VOC emissions, photolysis rates, and the reaction rate constants

R(NO2+OH) and R(NO+O3) significantly impacted all three categories. Meanwhile, boundary

conditions of all of the NOy compounds were not major influences on any of the results.

However, the BC(O3) parameter significantly impacted concentrations but not sensitivity to NOx,

whereas anthropogenic VOC emissions impacted sensitivities but not concentrations.

37

Targeting the parameters separately for the concentrations and the two sensitivity cases allows

the scenarios to be modeled more efficiently by CAMx-HDDM. As will be explained in Chapter

5.2.2, reduced form model (RFM) calculations of ozone concentrations require first- and second-

order sensitivity results for each selected input parameter, along with cross-sensitivities between

each parameter; RFM calculations of ozone sensitivities require first-order and cross-sensitivities

between each targeted parameter and the control scenario (i.e., DFW ANOx or AVOC). The

Monte Carlo sampling in Chapters 5 and 6 selects perturbation factors for all of the parameters

that significantly impact concentrations and/or sensitivities. However, due to the requirements of

the RFM, for computational efficiency Bayesian comparisons of modeled and observed

concentrations adjust model results only by the parameters that influence concentrations,

whereas sensitivities are adjusted only by the parameters that influence that sensitivity.

38

5. Bayesian and non-Bayesian Monte Carlo Analysis: Methods and Pseudo Case Testing

A Bayesian inference approach (Bergin and Milford, 2000; Deguillaume et al., 2007) is applied

to construct probabilistic representations of ozone-precursor response based on the relative

performance of the model under various structural and parametric settings in simulating observed

ozone and precursor concentrations. Figure 5.1 shows the concept of the Bayesian Monte Carlo

analysis as an extension to the standard Monte Carlo method. Both analyses involve generating

hundreds or thousands of Monte Carlo simulations with different model formulations and input

parameter settings randomly selected from predefined probability density functions. The

standard Monte Carlo then develops a priori estimates of the probabilistic distributions of model

outputs (ozone concentrations and response to emissions), assuming the simulations are equally

likely. The Bayesian Monte Carlo analyses adjust the probability of each simulation by taking

into account its model performance relative to the observations, leading to the a posteriori

estimates of the probabilistic distributions.

39

Figure 5.1. Conceptual diagram of Bayesian Monte Carlo analysis, adapted from Deguillaume et

al., 2007.

The Monte Carlo method of randomly sampling inputs has often been used to explore how

various input settings influence model outputs. Most previous applications of Monte Carlo to

characterize photochemical model uncertainty have assumed that each of the input scenarios is

equally likely to reflect “true” conditions (Bergin et al., 1999; Digar and Cohan, 2010; Tian et

al., 2010). However, some combinations of input settings may yield model results that perform

poorly relative to observations. Bayesian inference methods allow the relative likelihood of each

model scenario to be weighted based on model performance. Initial applications of Bayesian

Monte Carlo to parametric uncertainty analysis of photochemical models have been

demonstrated by a few studies (Bergin and Milford, 2000; Beekman and Derognat, 2003;

Deguillaume et al., 2007a) but remains an area of emerging interest.

40

5.1. Pseudo case tests Although the ultimate goal of Bayesian analysis is to use observations to assess the relative

likelihood of each model scenario, there are many choices to be made regarding the metric(s) for

comparing modeling results with observations and the computation of the likelihood function.

For example, should model results be compared against observations for each monitor-day, or

should results be aggregated spatially and/or temporally? Should the model be evaluated on high

ozone days that drive nonattainment, or on all days? To address these and other questions in a

controlled manner, pseudo case experiments were conducted.

We first conducted pseudo case tests to evaluate how different metrics would influence the

relative likelihoods assigned to different scenarios, following the Bayesian Monte Carlo method

of Bergin and Milford (2000). The pseudo case modeling was designed to provide insights into

how each metric would perform in assigning relative likelihoods to various model results of

specified performance.

To design the pseudo cases, we use model base case ozone concentrations kbY , (“b” for base; “k”

for each data point) as pseudo “observations” kO , i.e.,

kbk YO ,=

with Gaussian errors

)%25,20max( kk O⋅=σ

Bergin and Milford (2000) used a 30% standard deviation for the observations. Due to lack of

information a Gaussian error with the standard deviation of )%25,20max( kk O⋅=σ (the

maximum between 20 ppb and kO⋅%25 ) is used in pseudo case tests for the hourly pseudo

ozone observations. As discussed later in this report, subsequent analysis led us to choose a

lower value of 𝜎𝑘 for Bayesian analysis of the actual model scenarios.

41

Then we generate pseudo “model” cases ( kjY , ) by applying different systematic errors ( jBias in

ppb) and Gaussian random errors ( jζ ) to the base case ozone concentration fields (thus we know

how “bad” each model cases is). For each case j :

)1 ,0(random**Bias ,,, jkbjkbkj YYY ζ++= (5.1)

where random(0,1) represents a set of Gaussian random values with zero mean and unity

standard deviation.

For evaluating the model performance against observations under each scenario, a Gaussian

likelihood function )|( OYL j for simulation j given the observations O is used (as defined by

Bergin and Milford (2000), assuming that the errors in the interpolated concentrations at all

monitor/days are independent and normally distributed with mean of zero) and extended as

2,

21

1

( )1 1( | ) exp2( 2 )

Nk j k

j NN k k

kk

O YL Y O

σπ σ =

=

− = −

∑∏

(5.2)

Where kjY , and kO represent each ozone concentration from model simulation j and

observation, respectively. An observation standard error of kσ is considered for any

measurement kO (k = 1, 2, …., N where N = total number of data points for computing

likelihood). Note that Equation 5.2 in effect multiplies together the likelihoods that would be

computed for each of the N measurements; this tends to accentuate the differences between the

likelihoods assigned to the model cases as N grows larger.

Bayes’ theorem is then applied to compute the a posteriori probability (or Bayesian weight) of

the relative likelihood of model output for the j th simulation ( j = 1, 2, …. M , where M =

total number of Monte Carlo simulations) as follows,

1

( | ) ( )'( | )

( | ) ( )

j jj M

j jj

L Y O p Yp Y O

L Y O p Y=

=

∑ (5.3)

42

where '( | )jp Y O and )( jYp represent a posteriori and a priori probability of the model output,

respectively.

For representation of kjY , and kO , metrics have been designed mainly regarding the high

concentrations of ozone at the targeted regulatory monitors

(http://www.epa.gov/air/data/aqsdb.html). For the June episode three ozone monitors are chosen

within the DFW region: Denton, Eagle Mountain Lake, and Keller. The 2006 8-hr ozone design

values at these three sites are 95, 96, and 94 ppb, respectively, which are among the highest of

the monitors within the DFW region (dfw8h_o3_dv_20100628.xlsx,

ftp://amdaftp.tceq.texas.gov/pub/DFW8H2/data/). Note that for the pseudo-data testing we have

run each structural scenario for May 31 – July 2, 2006. When applying the metrics for the

pseudo-data testing, the first five days were discarded as spin up days. On each episode day,

hourly ozone concentrations were averaged within a fixed window of 10-18 hr local time at each

target monitor. We chose the fixed 8-hr window instead of the running 8-hr for comparison

between different structural cases. The 8-hr averaged concentrations (from both pseudo model

cases and pseudo observations) are calculated at these three target monitors, and the standard

deviations for the 8-hr averaged pseudo observations are derived from those for the hourly

concentrations. Then the 8-hr ozone results for both the model and the observations are further

aggregated or selected in the following metrics in Table 5.1 (N = number of data points selected;

Mo = number of monitors (3); D = number of days (28); and R = number of regions (DFW and

HGB)).

43

Table 5.1. Metrics to select/aggregate model and observation data points in model performance

evaluation for the pseudo-data tests.

Metric Description

A Daily 8-hr ozone at each target monitor on each day, considering only monitor-days when obs > 70 ppb (N <= Mo*D)

B Average of unpaired 3 highest 8-hr ozone days at each target monitor (N = Mo)

C Rank-order the 8-hr results from all episode days at each target monitor, and then compare only 4 cut-points (95th percentile, 75th percentile, median, and 25th percentile) (N = Mo*4). The cut-points can be unpaired in time.

D Unpaired 8-hr peaks within the DFW region on each day, considering only days when obs > 70 ppb (N <= R*D). The 8-hr peaks are picked among the monitors in DFW for the observations, and the grid cells containing the monitors for the model simulation.

We use the following statistics calculated using the simulated and observed metrics kY (j

omitted) and kO to evaluate the model-observation performance.

Normalized Mean Bias (percent):

%100)(

B%

1

1 ⋅−

=

∑

∑

=

=N

kk

N

kkk

O

OY

Weighted root-mean-squared error (WRMSE):

E: ∑∑==

−=

N

k k

N

k k

kk OY1

21

2

2 1)(WRMSEσσ

Sum of the weighted squared errors (Sum):

∑=

−=

N

k k

kk OY1

2

2)(Sum

σ

44

Product of the standard deviations:

∏=

=N

kk

NP1

)2( σπ

Relative likelihood:

)2/Sumexp()(

21expL

12

2* −=

−−= ∑

=

N

k k

kk OYσ

The following two tables show how the four metrics compare in evaluating the performance of

various pseudo model cases in predicting the pseudo observations. Recall that the pseudo model

cases were developed by applying Equation 5.1 with pre-assigned levels of model bias and error

( jBias , jζ ) to the base ozone field, and the pseudo observations were developed by applying

random Gaussian errors of )%25,20max( kk O⋅=σ to that same field. Ideally, a metric would

assign greater likelihoods (L*) to pseudo models with lower bias and error, though it is unclear

what level of spread in likelihoods is optimal. It would also be hoped that a metric would assign

equal weights to two models assigned to have the same error and biases of equal magnitude but

opposite sign.

As can be seen, the larger jBias and/or jζ for a case, the larger percentage bias, weighted root-

mean-squared error, and sum of the weighted squared errors between the pseudo case

concentrations and pseudo observations, and the smaller relative likelihood of the case.

Comparing the different metrics, we see that the fewer data points (N) used by a metric, the

narrower the spread of likelihoods assigned to the model cases (see Column L* in Tables 5.2 and

5.3). For example, metrics A and D indicate that the relative likelihoods of the model cases range

by more than 7 orders of magnitude, whereas metric B shows a range of only 1 order of

magnitude. This occurs because the more data points that are considered, the greater confidence

that Bayesian analysis will place on a better performing model being the “correct” one.

45

Table 5.2. Statistical measures and likelihoods for pseudo model cases using metrics A and B.

Metric A (N=31) Metric B (N=3) jBias , jζ B% E Sum L* B% E Sum L*

-20%, 0% -25.7 20.0 29.49 3.94E-7 -23.8 20.0 2.61 0.27 -10%, 0% -12.8 10.0 7.37 2.51E-2 -11.9 10.0 0.65 0.72 -5%, 0% -6.4 5.0 1.84 0.40 -6.0 5.0 0.16 0.92 5%, 0% 6.4 5.0 1.84 0.40 6.0 5.0 0.16 0.92 10%, 0% 12.8 10.0 7.37 2.51E-2 11.9 10.0 0.65 0.72 20%, 0% 25.7 20.0 29.49 3.94E-7 23.8 20.0 2.61 0.27 0%, 5% -0.5 1.4 0.14 0.93 0.1 0.3 5.13E-4 1.00 0%, 10% -0.9 2.8 0.58 0.75 0.3 0.6 2.05E-3 1.00 0%, 20% -1.8 5.6 2.31 0.31 1.2 2.0 2.67E-2 0.99 0%, 30% -2.8 8.4 5.21 7.41E-2 2.3 3.6 8.52E-2 0.96 0%, 40% -3.7 11.2 9.25 9.78E-3 3.8 5.4 0.19 0.91 0%, 50% -4.6 14.0 14.46 7.25E-4 6.2 7.2 0.34 0.84 10%, 30% 10.1 11.2 9.26 9.73E-3 14.2 12.3 0.99 0.61 10%, 40% 9.1 12.9 12.21 2.23E-3 15.7 13.9 1.26 0.53 10%, 50% 8.2 14.9 16.31 2.88E-4 18.1 16.0 1.67 0.43 20%, 30% 22.9 19.5 28.07 8.03E-7 26.1 22.1 3.19 0.20 20%, 40% 22.0 20.1 29.91 3.20E-7 27.7 23.6 3.64 0.16 20%, 50% 21.1 21.1 32.90 7.16E-8 30.0 25.7 4.31 0.12 -20%, 20% -27.5 22.2 36.23 1.36E-8 -22.6 19.1 2.38 0.30 -10%, 20% -14.7 12.7 11.90 2.61E-3 -10.7 9.2 0.55 0.76 -5%, 20% -8.3 8.4 5.26 7.20E-2 -4.8 4.4 0.13 0.94 5%, 20% 4.6 6.4 3.05 0.22 7.1 6.2 0.25 0.88 10%, 20% 11.0 10.1 7.48 2.38E-2 13.1 11.1 0.81 0.67 20%, 20% 23.8 19.3 27.39 1.13E-6 25.0 21.1 2.90 0.23

46

Table 5.3. Statistical measures and likelihoods for pseudo model cases using metrics C and D.

Metric C (N=12) Metric D (N=34) jBias , jζ B% E Sum L* B% E Sum L*

-20, 0% -29.5 20.0 11.49 3.20E-3 -23.9 20.0 29.31 4.32E-7 -10, 0% -14.8 10.0 2.87 0.24 -11.9 10.0 7.33 2.56E-2 -5, 0% -7.4 5.0 0.72 0.70 -6.0 5.0 1.83 0.40 5, 0% 7.4 5.0 0.72 0.70 6.0 5.0 1.83 0.40 10, 0% 14.8 10.0 2.87 0.24 11.9 10.0 7.33 2.56E-2 20, 0% 29.5 20.0 11.49 3.20E-3 23.9 20.0 29.31 4.32E-7 0, 5% -4.3 1.0 3.01E-2 0.99 0.7 1.5 0.16 0.92 0, 10% 0.4 1.38 5.45E-2 0.97 2.1 3.2 0.74 0.69 0, 20% 0.9 2.1 0.13 0.94 6.7 7.1 3.73 0.15 0, 30% 1.8 3.3 0.31 0.86 12.3 11.6 9.87 7.17E-3 0, 40% 3.4 4.7 0.64 0.73 18.0 16.3 19.58 5.61E-5 0, 50% 4.8 6.8 1.32 0.52 23.8 21.2 32.95 6.99E-8 10, 30% 16.6 11.5 3.79 0.15 24.2 20.8 31.57 1.40E-7 10, 40% 18.1 12.8 4.67 9.67E-2 29.9 25.6 48.05 3.68E-11 10, 50% 19.5 14.3 5.83 5.41E-2 35.7 30.5 68.28 1.49E-15 20, 30% 31.3 21.3 13.02 1.49E-3 36.1 30.4 67.91 1.79E-15 20, 40% 32.9 22.4 14.45 7.27E-4 41.8 35.3 91.17 1.59E-20 20, 50% 34.3 23.7 16.09 3.20E-4 47.6 40.2 118.27 2.08E-26 -20, 20% -28.7 19.6 11.01 4.06E-3 -17.2 15.5 17.52 1.57E-4 -10, 20% -13.9 9.7 2.70 0.26 -5.2 6.7 3.30 0.19 -5, 20% -6.5 4.9 0.70 0.71 0.7 4.8 1.69 0.43 5, 20% 8.3 5.9 1.00 0.61 12.6 11.4 9.45 8.89E-3 10, 20% 15.6 10.7 3.30 0.19 18.6 16.0 18.82 8.19E-5 20, 20% 30.4 20.6 12.22 2.22E-3 30.5 25.7 48.56 2.85E-11

The results for Metric A are especially important to consider, since it most closely approximates

the manner in which SIP models are often evaluated. EPA-recommended SIP modeling

methodology focuses on results on days in which ozone at a monitor was observed to be above a

particular concentration threshold. The use of a threshold is motivated by the fact that the most

polluted days drive attainment status for ozone, which is regulated based on the annual fourth-

highest concentration. However, Table 5.2 shows that, because threshold-based metrics like

Metric A target the days that were observed, but not necessarily modeled, to have the highest

ozone concentrations, they will tend to favor models with a positive bias.

47

How can a threshold-based metric be maintained in order to maximize policy relevance, without

leading to skewed preferences toward positively biased models? Extensive discussions were

undertaken with statisticians to explore methods for overcoming the inherent positive bias in

some threshold-based metrics. As discussed in the following section, the truncated likelihood

function emerged as an effective approach for use in the Bayesian analysis.

5.2. Bayesian and non-Bayesian methods for full ensemble

5.2.1. Metrics for BMC Analysis

The Bayesian Monte Carlo analysis of the actual structural and parametric scenarios followed

much of the methodology described above for the pseudo case testing, but with important

adjustments due to the insights gained from those tests. Upon further consideration, it was

recognized that Pseudo-case Metrics B, C, and D from the pseudo case testing (Table 5.1) are

problematic, because their use of unpaired data means that model results and observations are

not being compared at the same places and/or times. However, the pseudo case testing also

demonstrated that for metrics such as Pseudo-case Metric 1 (Table 5.1) that apply thresholds to

screen observational data, normal likelihood functions will skew the weightings in favor of

biased models (Table 5.2). Thus, alternate metrics and/or alternate likelihood functions were

needed.

To retain the use of Metric 1 (daily 8-hr O3 at targeted monitors when observed O3>70ppb),

which well captures model performance for polluted monitor-days, a truncated normal

distribution function was adopted to avoid the bias noted above. The truncated likelihood

function computes the likelihood of ozone prediction ( kjY , ) given that observation ( kO ) exceeds

a threshold concentration (a) by the following equation:

𝐿�𝑌𝑗|𝑂 > 𝑎� = 𝑃(𝑂 > 𝑎|𝑌) =

1

�√2𝜋�𝑁∏ 𝜎𝑘𝑁

𝑘=1exp �− 1

2∑ �𝑂𝑘−𝑌𝑗,𝑘

𝜎𝑘�2

𝑁𝑘=1 �

∑ �12�1 − 𝑒𝑟𝑓 �𝑎−𝑌𝑗,𝑘

�2𝜎𝑘2���𝑁

𝑘=1

� (5.4)

48

In addition, a new Metric 2 was adopted in order to consider average conditions across the DFW

region. This metric does not include a threshold, and thus the original normal likelihood function

remains applicable. Thus, in sum, the following two metrics are considered for the BMC analysis

of the full ensemble comprising structural and parametric scenarios:

Metric 1: Daily 8-hr ozone at each target monitor on each day, considering only monitor-days

when obs > 70 ppb (N = 48) (using truncated normal function).

Metric 2: Daily 8-hr ozone averaged over all sites within DFW (N = 30) (normal likelihood

function).

For each metric, it is necessary to identify a value of σ that characterizes the amount of

uncertainty in each observation used to evaluate model results. Measurements of ozone are

conducted by well established techniques, and thus instrumental error is relatively small.

Additional uncertainty is caused by the use of a point observation to represent a grid-cell average

concentration. One indicator of this uncertainty can be obtained by examining the variability

between ozone concentrations measured by multiple monitors within the same grid cell. The 12-

km modeling domain has 5 grid cells that contain 2 or more ozone monitors. Analysis by Dr.

Kristen Foley at US EPA showed that the standard deviation between observed 8hr ozone values

at these same-grid-cell sites ranges from 3.0 to 10.5ppb (Figure 5.2). Thus, we chose σ = 8 ppb

for Metric 1 to be near the midpoint of this range. Metric 2 should have less uncertainty due to its

averaging across sites, and thus σ = 5 ppb was chosen for this metric. Note that both values of σ

are much smaller than those used in the pseudo data testing, which had been based on the earlier

literature rather than on the domain-specific conditions considered here.

49

Figure 5.2. Scatter plot showing the relationship between 8-hr ozone concentrations observed at

pairs of monitors within the same grid cell of the 12-km domain. The different colors for the dots

represent different grid cells with more than one monitor. Figure courtesy of Dr. Kristen Foley,

US EPA.

Most other aspects of the BMC methodology remained unchanged, except that 3 initialization

days were used rather than 5 in order to retain more days for consideration. Analysis focused on

the sensitivity of DFW ozone concentrations to DFW emissions during the June 2006 episode.

5.2.2. Metric for non-Bayesian Analysis

The Bayesian metrics are constrained to follow Bayes’ Theorem and the associated likelihood

functions (Eqs. 5.2-5.4). Under those metrics, the more presumably independent observations

that are used, the greater the number of multiplications involved in the likelihood functions,

50

leading to very large spreads in relative likelihoods. However, the more general goal of

weighting ensemble cases based on performance against observations can be accomplished

through alternate metrics that do not invoke Bayes’ Theorem. By “non-Bayesian” approaches,

we refer to any other observation-based effort to assign relative weights to ensemble cases

without directly invoking Bayesian approaches.

Numerous metrics could be postulated for assigning weights depending on the performance

evaluation statistics of interest. For example, Mallet and Sportisse (2006) used model evaluation

statistics as the basis for assigning weights to their original ensemble to better predict the

observed data. For the sake of analysis, a single new metric (Metric 3) was created based upon

three model evaluation statistics recommended by EPA for screening the adequacy of ozone SIP

models (US-EPA, 2007):

(1) Mean Normalized Bias (MNB) n

i i

i 1 i

y o1MNBn o=

−= ∑

(2) Mean Normalized Gross Error (MNGE)

ni i

i 1 i

y o1MNGEn o=

−= ∑

(3) Unpaired Peak Accuracy (UPA)

y oUPA

omax max

max

−=

To ensure meaningful results, MNB and MNGE were computed for model results (yi) when O3

observations (oi) were greater than the recommended threshold of 60 ppb (US-EPA, 2007). This

resulted in 356 data points being considered in DFW during the June 2006 episode. Although

weights could be assigned based on these statistics in any number of ways, for the sake of

analysis weights for Metric 3 were computed as the inverse of the sum of these 3 measures

(neglecting the signs for bias and accuracy), as shown in the equation below:

Non-Bayesian Weight = 1|𝑀𝑁𝐵|+𝑀𝑁𝐺𝐸+|𝑈𝑃𝐴| (Metric 3)

51

Weights were then normalized to sum to 100% as in the Bayesian analyses.

5.2.3. Reduced Form Models for parametric uncertainties

The approach to selecting input parameters for BMC analysis was described in Chapter 4, and

resulted in the targeting of the parameters highlighted in Tables 4.2-4.4. Adjusted O3

concentrations based on the uncertainties in selected input parameters can be determined using

the relationship given by Cohan et al. 2005:

𝐶𝑗+𝑘∗ = 𝐶0 + ∑ 𝜑𝑗𝑆𝑗(1) +𝑗 ∑ 𝜑𝑘𝑆𝑘

(1) +𝑘12∑ 𝜑𝑗2𝑆𝑗,𝑗

(2) +𝑗12∑ 𝜑𝑘2𝑆𝑘,𝑘

(2) +𝑘 ∑ 𝜑𝑗𝜑𝑘𝑆𝑗,𝑘(2)

𝑗,𝑘 (5.5)

where, C0 is the modeled concentration, ϕj and ϕk are the perturbations in parameters j and k

respectively, and S(1) and S(2) denotes first- and high-order sensitivities to parameters given in

suffix. Note that in the calculation of adjusted concentrations, Sj,k(2) denotes cross-sensitivity

between two input parameters.

We conduct Bayesian Monte Carlo simulations (sample size 1000) for selected O3 metrics within

the structural ensemble composed of the 4 selected members (Base, CB-6, MEGAN, and CB-

6+MEGAN) selected in Chapter 3. This enabled characterization of C* for various perturbations

ϕ in the input parameters. Thus, the total number of Monte Carlo cases was 4000 (M = 1000 x

4). Initially, we consider that each simulation within any given structural case to be equally

likely (prior probability, p(C*) = 1/M). Then we use eqs 1 - 3 to compute the posterior

probabilities, p’(C*) for each of the two ozone metrics. Testing with a parametric sample size of

10,000 showed that the larger number of samples did not substantially influence the posterior

distributions.

Finally, we assume that the posterior probabilities developed from the model output for ozone

concentration can also be applied to obtain Bayesian estimates of ozone responses to emission

changes (sensitivity results) and of input parameter values as well. To characterize adjusted O3

sensitivity due to uncertainties in input parameter j, we use the Reduced Form Model (RFM)

given by Digar and Cohan (2008),

52

𝑆𝑗(1)∗ = �1 + 𝜑𝑗��𝑆𝑗

(1) + 𝜑𝑗𝑆𝑗(2) + ∑ 𝜑𝑘𝑆𝑗,𝑘

(2)𝑘 � (5.6)

where, ϕj denotes the perturbation (factor of uncertainty) in parameter j. Note that in the

calculation of adjusted sensitivities, Sj,k(2) denotes cross-sensitivity between an input parameter

and the control scenario (DFW ANOx or DFW AVOC). In the RFMs for both concentrations and

sensitivities, the value of ϕj is restricted to within a 2σ range, to avoid extreme values of input

parameters which would extend beyond the reliability of the RFM equations.

53

6. Bayesian and non-Bayesian Monte Carlo analysis: Results and Discussion

This chapter presents results for the Bayesian and non-Bayesian Monte Carlo analyses of the

structural ensemble and of the final ‘full ensemble’. The full ensemble allows the parameters

targeted in Tables 4.2-4.4 to vary within the four selected structural scenarios – Base Case (B),

CB-6 (C), MEGAN (M) and CB-6 with MEGAN (C+M). Monte Carlo randomly samples 1000

sets of input parameter values from their a priori lognormal probability distributions, with σ

taken from Table 4.1 and truncation applied at ±2σ to avoid extreme values. Each set is paired

with each of the four structural scenarios, resulting in 4000 cases that are originally assumed to

be equally likely. The Bayesian analysis then “weights” each of the 4000 cases based on its

performance in simulating observed O3 concentrations, yielding a posteriori probability

distributions not only for the targeted input parameters but also for the simulated O3-emission

sensitivities. Differences between the a posteriori and a priori probability distributions for input

parameters may highlight potential changes to inputs that could be investigated in further

research. Meanwhile, the a posteriori sensitivity results will provide observation-adjusted

expectations for the amount of air quality improvement that would result from emission controls.

Due to shortcomings observed in the Bayesian results, results for non-Bayesian Metric 3 are also

presented.

6.1. Bayesian and non-Bayesian probability distributions of input scenarios and parameters

Section 5.2.2 discussed the selection of two Bayesian metrics and one non-Bayesian metric for

weighting each model case based on its performance in simulating observed O3 concentrations.

To recap, Metric 1 compares daily 8-hr O3 at each of three targeted DFW monitors when

observed O3>70ppb (N = 48), whereas Metric 2 considers average 8-hr O3 concentrations across

all sites on each day (N = 30). Non-Bayesian Metric 3 considers performance statistics for 8-hr

O3 concentrations across all DFW monitors and days (N=356) based on aggregate statistics.

Figure 6.1 shows the performance of each of the structural scenarios, with unperturbed input

parameters, in simulating each of the Bayesian observational metrics. Three of the four scenarios

54

tend to underpredict Metric 1 on most monitor-days, whereas the C+M scenario yields unbiased

predictions but with considerable scatter. Underpredictions of Metric 1 may in part reflect the 70

ppb threshold that is applied to observations in this metric. For Metric 2, the two cases with CB-

6 chemistry (C and C+M) yield the least biased predictions, whereas the other two cases tend to

underpredict O3 in the DFW region.

Figure 6.1: Boxplots showing ozone differences for the 4 structural members based on

(A) Metric 1 (B) Metric 2, with input parameters set at default values.

Bayesian analysis is first conducted for the structural-only ensemble (i.e., 4 structural cases with

unperturbed input parameters) under the two metrics. As discussed in Chapter 5, the truncated

likelihood function (Eq. 5.4) is applied to assign weights to Metric 1, in order avert the positive

bias that can arise due to the metric having a threshold (a=70 ppb). The normal likelihood

function (Eq. 5.2) is applied to Metric 2. Since the CB-6 cases simulate higher O3 levels in DFW

that better matched the metrics, these cases dominate the weightings (Table 6.1, “w/o

parametric” results). Metric 1 shows that pairing MEGAN biogenics with CB-6 yields the best

results, whereas Metric 2 favors the original biogenic inputs. The strong differences between

probabilities arise from the fact that Eqs. 5.2 and 5.4 essentially multiply together likelihoods

evaluated against each of the N observations.

Incorporating parametric uncertainties dramatically shifts the a posteriori weightings among the

structural scenarios (Table 6.1, “w/ parametric”). For Metric 1, the rankings are flipped, with the

(A) (B)

55

CB-05 cases (M and B) now preferred over CB-6. For Metric 2, CB-6 chemistry remains

preferred, but with MEGAN rather than GloBEIS biogenic emissions. Again, since Bayesian

Eqs. 5.2 and 5.4 essentially multiply together N likelihoods, there are enormous spreads in the

weightings of the 4000 cases under Metrics 1 and 2, with most of the weight placed on a handful

of cases (Figure 6.2). Non-Bayesian Metric 3 yields more even spread among the cases (Figure

6.2) and thus a flatter spread among the structural scenarios (Table 6.1).

Table 6.1: Posterior probability of the structural ensemble. The “structural only” results consider only the four cases with their default inputs; the “with parametric” results consider the full ensemble of 4000 cases (4 structural * 1000 Monte Carlo samplings of parameters).

a posteriori probabilities BASE CB6 (C) MEGAN (M) C+M

Metric 1 (N = 48)

structural only 0.00% 16.34% 0.00% 83.66%

w/ parametric 14.91% 5.26% 65.01% 14.82%

Metric 2 (N = 30)

structural only 0.19% 80.06% 0.16% 19.58%

w/ parametric 0.00% 25.32% 0.00% 74.68%

Metric 3 (non-Bayesian, N=356) w/ parametric 21.63% 29.57% 21.42% 27.39%

56

Figure 6.2. Weights assigned to the 4000 members of the full ensemble under Bayesian Metrics 1 and 2 and non-Bayesian Metric 3. Note for Metric 1 and 2 a handful of cases receive most of the weight, and most cases receive near zero weight; weightings are more dispersed in Metric 3. (Kristen Foley, US EPA, assisted with this image).

How could the rankings of structural scenarios in Table 6.1 differ so radically with the inclusion

of parametric variability? This could only occur if there are perturbation values of the input

parameters which, when paired with a structural scenario, help it far better match the observed

data. We closely examine the prior and posterior input distributions within each of the structural

scenarios to investigate the cause of the flip in Bayesian weights (Figures 6.3 and 6.4).

The blue lines in Figures 6.3 and 6.4 depict the probability densities for the 1000 Monte Carlo

cases randomly sampled from the truncated lognormal a priori probability distributions of each

input parameter. The dashed lines show the a posteriori probability distributions from Bayesian

weightings within each of the structural scenarios, and the solid red lines show the final a

posteriori distributions resulting from joint consideration of the full 4000 case ensemble. For

Metric 1, Bayesian weightings under all the scenarios tended to prefer higher levels of BC(O3)

and photolysis rates (i.e., 1+φ>1.0), and lower levels of R(OH+NO2) (i.e., 1+φ<1.0) (Table 6.2

and Figure 6.3). All of these changes tend to favor higher O3 concentrations. Results were

57

ambiguous across the structural scenarios for scaling biogenic VOC emissions and R(NO+O3).

However, the Bayesian analysis favored scaling down the NOx emission inventory for the CB-6

cases (C and C+M), but scaling it up for the CB-05 cases. The boost in E(NOx) for the CB-05

cases allowed them to overcome their negative bias that had been documented in Figure 6.1, and

to actually outperform the CB-6 cases.

58

Figure 6.3: Prior and posterior distributions of input parameters under Metric 1.

ENOX EBVOC

R(photo) R(NO2+OH)

R(NO+O3) BC(O3)

59

Figure 6.4: Prior and posterior distributions of input parameters under Metric 2.

ENOX EBVOC

R(photo) R(NO2+OH)

R(NO+O3) BC(O3)

60

For Metric 2, scaling up the ozone boundary conditions and photolysis rates was again preferred

(Figure 6.4 and Table 6.2), which tends to raise O3 concentrations. However, scaling up

R(NO2+OH) became preferred under Metric 2, yielding an opposite tendency. Again, the CB-05

cases tended to prefer scaling up the NOx emissions inventory to overcome their initial low bias

in O3 predictions, whereas the CB-6 cases performed better with NOx scaled down (for CB-6

alone) or held near its base levels (for CB-6+MEGAN). The MEGAN cases (M and C+M)

tended to prefer scaling down the BVOC emissions inventory under Metric 2, counteracting the

larger BVOC inventory of MEGAN compared to the default GloBEIS (see Figure 2.3 and Table

2.2).

Even small changes in the input parameter distributions can cause major changes in the

likelihood weightings assigned to the cases, due to the nature of the likelihood functions applied

here. That is because, as noted earlier, the functions essentially multiply the likelihoods for each

of the observation points, accentuating differences. This is illustrated in Figure 6.5, which

depicts the performance of each structural case against observations for Metric 2. In the top plot,

note that under default parameter settings, the CB-6 and C+M cases do perform better than the

other cases, especially during the first two weeks of the episode. However, it is questionable

whether the outperformance in Figure 6.5(top) merits over 99% of the Bayesian weighting being

placed on the two cases as shown in Table 6.1, or that the CB-6 case is 4 times more likely than

the very similar C+M results. Metric 3 yielded much flatter weightings.

The bottom plot of Figure 6.5 shows an initial visualization of how the parametric adjustments

affected performance of the CB-6 and C+M cases under Metric 2. It applies the posterior mean

of the input parameters that were derived in each of those cases, to provide a rough

approximation of the performance of these two scenarios within the full ensemble. Close

inspection comparing the bottom and top plots shows that the parametric adjustments do indeed

slightly improve how well the CB-6 and C+M cases match the observations. Again, however, it

appears unlikely that the difference between the cases really merits the three times greater

likelihood placed on C+M in the full ensemble (Table 6.1). Alternate approaches to the metrics

and likelihood functions will need to be considered in further research.

61

Figure 6.5: Daily 8-hr ozone concentration observations averaged over all 20 monitors in DFW (i.e., Metric 2; dash-dotted line), compared to each of the structural cases under default parameters (top plot) and to the C and C+M cases under their mean settings of input parameters resulting from the Metric 2 Bayesian analysis (bottom plot).

Table 6.2 summarizes the a posteriori scaling factors for the input parameters derived from the

full ensemble under the three metrics. Under both Bayesian metrics, the weightings tended to

prefer higher photolysis rates and ozone boundary conditions in order to avert the ozone

underpredictions that occurred under default parameters. However, the results for other reaction

30

40

50

60

70

80

90

6/2/2006 6/7/2006 6/12/2006 6/17/2006 6/22/2006 6/27/2006 7/2/2006

8-h

Ozo

ne C

once

ntra

tions

, ppb

OBS BASE CB-6 MEGAN C+M

30

40

50

60

70

80

90

6/2/2006 6/7/2006 6/12/2006 6/17/2006 6/22/2006 6/27/2006 7/2/2006

8-h

Ozo

ne C

once

ntra

tions

, ppb

CB-6_w_posterior_mean C+M_w_posterior_mean OBS

62

rates and for NOx emissions are ambiguous. The parametric scaling factors for BVOC in Table

6.2 cannot be reliably interpreted, since they aggregate across scenarios that used different

biogenic emissions models; while the scenario-specific results (dashed lines in Figures 6.3 and

6.4) avoid that problem, they yielded conflicting signals between the two metrics.

Non-Bayesian Metric 3 placed less varied weights on the cases (Figure 6.2), and thus did not

show such wide perturbations in the input parameters, with all of the mean values within 1σ of

the original values (Table 6.2). Metric 3 did tend to scale up NOx emissions, to compensate for

the slight underprediction of ozone by the base model.

Table 6.2: Comparison of weighted distributions of input parameter scaling factors for the full

ensemble of cases based on the 3 metrics.

Input Parameters a posteriori mean ± 1σ

Metric 1 Metric 2 Metric 3 ENOX 1.05 ± 0.16 0.90 ± 0.07 1.06 ± 0.25 EBVOC 1.07 ± 0.27 0.85 ± 0.14 1.03 ± 0.25 R(photolysis) 1.06 ± 0.07 1.11 ± 0.07 1.01 ± 0.08 R(NO2+OH) 0.89 ± 0.24 1.45 ± 0.16 1.03 ± 0.27 R(NO+O3) 0.97 ± 0.06 1.04 ± 0.08 1.00 ± 0.08 BC(O3) 1.23 ± 0.13 1.30 ± 0.05 1.02 ± 0.16

6.2. Ensemble Evaluation

Do the weighted ensembles outperform simple equal weighting of cases in representing DFW

ozone observations? Ensemble accuracy is tested by evaluating the root mean squared error

(RMS), normalized mean bias (NMB) and the correlation of the ensemble mean with the

observations. These statistical measures are represented in the equations below:

N2

k kk 1

y oRMSE

N

( )=

−=∑

63

N

k kk 1

N

kk 1

y oNMB

o

( )=

=

−=∑

∑

( ) ( )

( ) ( )

N

k kk 1

N N2

k kk 1 k 1

y y o ocorrelation

y y o o

=

= =

− −=

− −

∑

∑ ∑

where N is the number of observation data (site/days); ky represents the ensemble mean

prediction for kth observation ko (k = 1, 2, 3, …, N); y denotes the average value for all the N

mean ensemble predictions; and o is the mean observation over all site/days.

The accuracy test statistics, evaluated based on 8-hour ozone concentrations for all monitors

within the DFW regions and all days of the June 2006 episode (excluding initialization), are

provided in Table 6.3 and Figure 6.6. Even before application of weights, the equal-weighted

ensemble already outperforms the deterministic base case, in part because it includes runs with

the CB-6 mechanism that help correct for the slight low bias of the base model within the DFW

region. Note in Table 6.3 that Metric 1 slightly overcorrects the original underprediction of

ozone, and thus does not improve overall accuracy. This may reflect the fact that Metric 1

considered only monitor-days above the 70 ppb threshold, and thus placed large weights on cases

that overpredict low ozone days. Although its median result most closely matched observation,

the bias arises from the high cases (Figure 6.6). Metric 2, which considered average 8-hour

ozone across DFW monitors on all episode days, led to better ensemble accuracy, reducing RMS

by 10% and improving the bias and correlation. Metric 3 placed greater weights on CB-6 and

high ENOx cases to correct for the negative bias of ozone predictions in the DFW region, but did

not improve performance in terms of RMS or correlation (Table 6.3).

64

Table 6.3. Statistical evaluation of the original and weighted ensembles (i.e, the 4000 cases),

evaluated against 8-hour ozone at all sites/days within DFW during the June 2006 episode.

Statistics Base Case (deterministic)

Equal weighted full ensemble

Bayesian (Metric-1)

Bayesian (Metric-2)

Non-Bayesian

(Metric-3) RMS (ppb) 12.62 11.50 11.58 10.23 11.45 NMB (%) -7.81 -2.44 3.21 -0.31 -0.46

Correlation 0.759 0.770 0.769 0.819 0.770

Figure 6.6. Boxplot evaluating model performance against 8-hr ozone at all site-days within

DFW for the June 2006 episode.

The Talagrand diagram, popularly known as the rank histogram (Talagrand et al., 1997), is a

statistical tool to assess the measure of differences in the ensemble predictions (spread). The

ensemble is distributed into (B + 1) bins, where B = number of ensemble predictions (in our case

B = 4000). For each of the N observations (site/days), the ensemble predictions are ranked along

with the observed value to find out the bin in which the observation is falling. A rank histogram

65

is then constructed by tallying over these N site/days and plotting the frequency of the rank of

the observation. A rank histogram therefore evaluates whether the model-ensemble is able to

predict the actual observations such that the occurrence of the observation within each bin is

equally likely, and a flat rank histogram would indicate that the ensemble has the correct spread

(rank uniformity).

For the prior full ensemble with equal weights (Figure 6.7, top plot), the rank histogram shows

an underforecasting bias, reflected in the preponderance of observations that fall on the right of

the histogram, above the majority of the model cases. The rank histogram also shows the prior

ensemble spread to be too narrow (underdispersive), as reflected in the U shape. Note both the

large first bin, which shows that many observations fall below most or all of the model cases (see

Figure 6.5, which showed episode days around June 21 and July 2 to have observed ozone lower

than model results), and the large bins toward the right.

For the a posteriori ensembles, the U shape of the rank histograms (over-confidence) becomes

even more pronounced (Figure 6.7, middle and bottom plots). That is because, as had been noted

in Figure 6.2, the Bayesian analysis for each metric placed most of the weight on a handful of

cases. Thus, the bulk of each weighted ensembles lie above some observations (leading to the

large left-most column), and above other observations (leading to the large right-most column).

Metric 1 (M1) resulted in a slight positive bias (reflected by the larger leftmost bin than

rightmost bin), despite the use of the truncated normal function (Eq. 5.4) to counteract its

threshold. Metric 2 is essentially unbiased, but is even more over-confident, since it places more

than half of its weight on 2 of the 4000 cases (see the two outlier points in Figure 6.2).

Metric 3 applied a relatively narrow spread of weights to the ensemble cases (Figure 6.2). Thus,

its rank histogram (Figure 6.7, bottom) retains much of the structure of the prior (equal-

weighted) ensemble. Its distribution is essentially unbiased but is still somewhat over-confident

(under-dispersive), though not as strongly so as Metrics 1 and 2.

66

Figure 6.7. Rank histogram for the full ensemble with equal weights (top) and weighted by Metrics 1, 2, and 3, evaluated based on 8-hr ozone observations from DFW monitors during the June 2006 episode. (Method for images courtesy of K. Foley, US EPA).

67

Figure 6.7 (cont.). Rank histogram for the full ensemble with equal weights (top) and weighted by Metrics 1, 2, and 3, evaluated based on 8-hr ozone observations from DFW monitors during the June 2006 episode. (Method for images courtesy of K. Foley, US EPA).

6.3. Bayesian and non-Bayesian estimation of ozone sensitivities

We now turn to considering how the Bayesian and non-Bayesian weights affect predictions of

ozone sensitivity to emissions within the DFW region during the June 2006 episode. As

discussed in Chapter 3, afternoon ozone in DFW is primarily NOx-limited under base conditions,

with ozone about an order of magnitude more sensitive to DFW ANOx than to DFW AVOC

(Figure 3.1). As shown in Figure 6.8, NOx-limited conditions persist regardless which structural

scenario is assumed. As noted earlier, MEGAN tends to enhance sensitivities to NOx because it

predicts more biogenic VOC and less biogenic NOx. CB-6 favors sensitivity to VOC, though

conditions remain predominantly NOx-limited under either chemical mechanism.

68

S(1) ANOX (ppm) S(1) AVOC (ppm)

BA

SE

ME

GA

N

CB

-6

Figure 6.8: Ozone sensitivity to NOx and VOC emission from DFW for different structural model scenarios under default settings of input parameters. Episode average results are shown for a 4-km grid resolution for the region near DFW.

How do the relative sensitivities of ozone to NOx and VOC change as Bayesian and non-

Bayesian analyses are applied to the full ensemble? The Bayesian and non-Bayesian weights

from the full parametric-structural ensemble are used to characterize the final a posteriori

distribution for O3 sensitivity to NOX and VOC emissions from DFW. Results are probed for

three grid-cells corresponding to three EPA sites in DFW - two having the highest O3 historical

DVs (Denton and Eagle Mountain Lake) and one with the lowest O3 concentrations (Kaufman)

69

among all the 20 DFW sites (Figure 6.9). Daily 8-hr O3 sensitivities were averaged over the

episode to have a representative value for each site.

Figure 6.9: Map showing locations of location of EPA Ozone Monitors (results are probed for the sites marked as Yellow).

The mean (𝜇′) and standard deviation (𝜎′) of the resulting posterior distribution of values 𝑌𝑗 can

be computed by

𝜇′ = ∑ (𝑌𝑗𝑀𝑗=1 × 𝑝𝑗′) (6)

𝜎′ = �∑ �𝑌𝑗 − 𝜇′�2∙ 𝑝𝑗′𝑀

𝑗=1 (7)

where 𝑝𝑗′ denotes the posterior probability for the jth iteration and M = 4000.

Table 6.4 summarizes the statistics for the a priori and a posteriori distributions of O3 sensitivity

to ANOX and AVOC emission from DFW. We find that at all three sites, for Metric 1, the

posterior mean for O3 sensitivity to ANOx always exceeds the prior mean; whereas, for Metric 2,

it is just the opposite. In other words, under Metric 1, Bayesian weighting enhances the relative

importance of anthropogenic NOx emission controls compared to VOC controls. On the contrary,

Metric 2 makes O3 slightly more sensitive to VOC and less sensitive to NOx. Under each metric,

Eagle Mt. Lake

Kaufman

Denton

70

the uncertainty of sensitivities (1σ standard deviation) is significantly reduced for all of the

posterior estimates, as the Bayesian weightings narrow the spread in the sensitivity results.

Table 6.4: Comparison of prior and posterior ozone sensitivities.

SANOx (ppb) SAVOC (ppb)

a priori

mean ± 1σ a posteriori mean ± 1σ

a priori mean ± 1σ

a posteriori mean ± 1σ

Denton Metric 1

6.79 ± 2.59 9.23 ± 1.64

1.10 ± 0.82 0.73 ± 0.29

Metric 2 6.21 ± 1.01 1.12 ± 0.30 Eagle Mt.

Lake Metric 1

6.79 ± 2.46 9.17 ± 1.58

0.95 ± 0.76 0.60 ± 0.28

Metric 2 6.28 ± 0.94 0.98 ± 0.27

Kaufman Metric 1

3.39 ± 0.57 3.92 ± 0.34

0.05 ± 0.06 0.01 ± 0.02

Metric 2 3.22 ± 0.17 0.06 ± 0.02

To explain the contrasting impacts of Bayesian analysis on sensitivity outputs, we must refer

back to the impacts that it had on adjusting the structural and parametric inputs (Table 6.1 and

Figures 6.3-6.4). For Metric 1, the Bayesian weights were predominantly driven by structural

case MEGAN, which had maximum O3 sensitivity to NOx and minimal sensitivity to VOC

among all the other structural members of the ensemble (Figure 3.1). This reflects the higher

biogenic VOC emissions and lower NOx emissions from the MEGAN model (Figure 2.3 and

Table 2.2). The parametric weightings (Figure 6.3) tended to scale up E(BVOC) even further,

counteracting the scaling up of E(NOx). At the same time, the parametric weightings favored

increasing photolysis rates and decreasing the rate of R(OH+NO2), in each case increasing

sensitivity to NOx (demonstrated by the cross-sensitivity coefficients in Tables 4.3-4.4, and by

(Cohan et al., 2010)).

By contrast, Metric 2 gives maximal weight to structural case C+M (Table 6.1). That case has

relatively high O3 sensitivity to VOC emissions (see Figure 3.3), as the tendency of the CB-6

chemical mechanism to favor VOC sensitivity outweighs the MEGAN biogenics. The parametric

weightings under Metric 2 (Figure 6.4) favored scaling up the rate of R(OH+NO2), thereby

71

reducing sensitivity to NOx. The weightings also favored scaling down E(BVOC), counteracting

the high BVOC from the MEGAN inventory and further favoring sensitivities to VOC rather

than NOx. Taken together, these Bayesian weightings led to the opposite influence on ozone

sensitivities in Metric 2 compared to Metric 1.

Since all the three sites exhibited similar patterns for O3 sensitivity estimates, we limit our study

results for the subsequent analyses only for Denton. This site is historically the ‘worst’ O3 non-

attainment site in the O3 SIP modeling with a 3-yr mean O3 design value of 93.33 ppb. The

deterministic (base case) model showed average first-order sensitivity coefficient of ozone at this

site to DFW anthropogenic NOx to be 6.99 ppb, and 0.84 ppb for DFW anthropogenic VOC.

The cumulative probability density functions (CDFs) of O3 sensitivities at Denton are presented

in Figure 6.10. We notice a significant positive shift in the posterior CDF of O3 sensitivity to

NOx for Metric 1, consistent with the results shown in Table 6.2. This indicates O3

responsiveness to NOx increases when model estimates are weighted by their relative

performance in predicting actual O3 measurements. Bayesian weightings under Metric 2 have the

opposite effect, shifting sensitivities to NOx slightly lower. For example, the prior probability

that the sensitivity of O3 to NOx exceeds 7 ppb is approximately 52% whereas the posterior

probability for this case gets as high as 90% for Metric 1 and as low as 13% for Metric 2.

Metric 3 retains most of the shape of the unweighted distribution, due to its relatively flat

weighting of cases. However, Metric 3 does tend to increase predicted sensitivity of ozone to

VOC, since its weightings favor the CB-6 mechanism (Table 6.1) and a larger NOx emissions

inventory (Table 6.2).

72

O3 sensitivity to ANOX Ozone sensitivity to AVOC M

ET

RIC

1

ME

TR

IC 2

ME

TR

IC 3

Figure 6.10. Cumulative probability distribution functions of the sensitivity of ozone at the Denton monitor in June 2006 to DFW anthropogenic NOx (left) and VOC (right) emissions for Bayesian metrics 1 and 2, and non-Bayesian Metric 3. Green line shows deterministic (base case) results.

73

7. Conclusions

This report has provided an initial case study demonstration of how Bayesian Monte Carlo

analysis could be applied to assess the uncertainties in pollutant-emission sensitivities that arise

from various choices of uncertain structural and parametric inputs for photochemical air quality

models. Modeling episodes used in the development of SIPs for the 1997 8-hr ozone standard

have provided the basis for this analysis.

The extensive photochemical modeling conducted by TCEQ and its contractors in support of

recent SIP development provided a wealth of structural scenarios for consideration. Meanwhile

the recently introduced Reduced Form Model method utilizing HDDM sensitivity coefficients

enabled efficient analysis of a large number of input parameters. Screening analysis of these

input choices highlighted both structural and parametric input choices that strongly influence

predictions of ozone concentrations and their sensitivities to anthropogenic emissions.

Among the structural options considered here, the choice of chemical mechanism and biogenic

emissions model emerged as leading contributors to uncertainty in sensitivity results. Deposition

scheme and boundary condition models exerted some influence on afternoon ozone

concentrations, but generally had a much smaller effect on sensitivities to emissions; only if

nighttime conditions were of interest would deposition scheme become an important influence

on sensitivities. Preliminary consideration of satellite photolysis rates showed that they could

have a major influence on results, but analysis was precluded by the availability of only one-half

month of data for the episode so far.

Among the input parameters considered, anthropogenic NOx emissions, biogenic VOC

emissions, and photolysis rates emerged as the biggest contributors to uncertainty in DFW

sensitivities, consistent with previous studies for other regions. Boundary conditions of various

NOy species generally had little impact on results. Some distinctions were found between the

parametric influences on concentrations and sensitivities, providing opportunities for slight

computational savings by targeting the relevant parameters influencing each output.

74

The underlying motivation for using Bayesian or non-Bayesian observation-based probabilistic

approaches to characterize pollutant-emission responsiveness is clear. Deterministic use of a

single model formulation with a single set of input parameters yields a false sense of precision

that the air quality impacts of emission controls are perfectly known. Traditional Monte Carlo

analysis of uncertain inputs or ensemble runs of models yield probabilistic estimates of model

outputs, but naively assume that each of the scenarios is equally likely.

Bayesian and non-Bayesian Monte Carlo approaches offer great potential to capitalize upon the

wealth of information contained in observational data to weight each of the cases based on

performance. As shown in this study, those weightings can be applied to generate posterior

probability distribution estimates for both the input parameters and for pollutant-emission

responsiveness. The input parameter distributions essentially provide an alternate approach to

inverse modeling. The responsiveness results address a core question of control strategy

modeling: how much air quality improvement will be obtained from each emission control

option.

However, the analysis of the full ensemble highlighted key challenges in applying Bayesian

methods to assess air quality model results. In particular, the BMC results proved to be quite

sensitive to the choice of observational metric used for the likelihood estimation. The two

Bayesian metrics considered here for the full ensemble analysis relied upon a large number of

data points. Multiplication of the associated likelihoods by the BMC likelihood function led to

huge discrepancies between the weights assigned to fairly similar model scenarios. This led to

posterior distributions that were heavily weighted toward model cases whose performance

appeared to be only slightly better than less favored cases. The non-Bayesian Metric 3 led to a

flatter distribution of weights. Furthermore, conflicting results arose across the three metrics,

even though all were based upon ozone observations from within the same region.

The underdispersive nature of the prior distribution (i.e., the fact that many observations fell

above or below the range of the model cases; see Figure 6.7) proved to be problematic, since

weighted ensemble approaches are generally poorly equipped to overcome the lack of spread.

Other approaches such as Bayesian Model Averaging (Raftery et al., 2005) are better suited for

improving ensemble spread and performance, but would not yield reliable results for

75

distributions of input parameters and sensitivities as desired here. Greater spread could also be

introduced within the current framework by allowing greater ranges in the inputs (Table 4.1), but

that may lead to unrealistic values for emissions or other parameters. Most likely, the lack of

spread reflects model error unaccounted for by the current ensemble; for example, if

meteorological conditions are poorly simulated on some days, then all of the model cases may

struggle to match observations. Additional structural cases such as alternate meteorological

inputs or satellite-based photolysis rates may lead to a more disperse ensemble (i.e., broader

ranges of predictions for ozone concentrations at each monitor-day) that could facilitate the

analysis approaches applied here.

7.1 Recommendations for further research

Methods applied to the DFW region for the June 2006 episode in Chapter 6 could readily be

extended to consider other regions and episodes. Results could also be computed for sensitivities

to various source sectors of interest.

Given the conflicting results across the three metrics, further research is needed to refine the

choice of observational metric. Papers that have applied Bayesian Monte Carlo analysis to

analyze ozone formation in Paris (Beekman and Derognat, 2003; Deguillaume et al., 2007a;

Deguillaume et al., 2008) used highly aggregated metrics such as concentrations averaged across

the region and episode. Such approaches yield a broader spread of weights, but fail to capitalize

upon the spatial and temporal resolution of the data. Metrics could also move beyond the ground-

level ozone data to consider precursor compounds as well as observations aloft. In particular, the

contrasting weightings applied to NOx emissions between the two Bayesian metrics highlights

the need for a NOx-based metric that would help constrain this crucial input.

As additional metrics are considered, ensemble evaluation statistics provide an objective basis

for assessing the posterior ensembles that result. The rank histograms presented in Chapter 6.2

helped visually show whether the weighted ensembles were appropriately dispersive and

accurate. Future work could compute statistics such as reliability, resolution, and Brier skill

scores to assess the weighted scenarios that result from Bayesian and non-Bayesian Monte Carlo

analyses. Data withholding could also be applied to test the robustness of results.

76

Future work could also expand upon the structural and parametric uncertainties considered here.

The satellite-based photolysis rates showed initial promise for substantially affecting results, and

should soon become available for the full episodes. Other structural changes in inputs, including

satellite-based NOx emission emissions and alternate meteorological inputs, could be considered

as well. The underdispersive nature of the ensemble considered here (i.e., many ozone

observations fell outside the range of model cases) adds impetus to expanding the ensemble.

In the end, Bayesian and non-Bayesian Monte Carlo analyses have been shown to be promising

approaches to probabilistic characterization of photochemical air quality models, but one that

will require further method development to enhance the reliability of its results. Our most crucial

next step will be to explore alternative metrics and/or likelihood functions that can capitalize

upon the wealth of available observations without yielding unrealistically stark discrepancies

between the weightings of cases. Ensemble statistics will help us judge the weightings that result

from these alternate metrics. The analysis of posterior distributions for input parameters and

sensitivity results in Chapter 6 provided a glimpse of the types of findings that can be gleaned

from Bayesian Monte Carlo analysis; further development of metrics and methods will enable

such results to be obtained in as reliable a manner as possible.

77

References

Archibald, A. T., M. C. Cooke, S. R. Utembe, D. E. Shallcross, R. G. Derwent and M. E. Jenkin (2010). "Impacts of mechanistic changes on HO(x) formation and recycling in the oxidation of isoprene." Atmospheric Chemistry and Physics 10(17): 8097-8118. doi: 10.5194/acp-10-8097-2010.

Atkinson, R. A., D. L. Baulch, R. A. Cox, J. N. Crowley, R. F. Hampson, R. G. Hynes, M. E. Jenkin, J. A. Kerr, M. J. Rossi and J. Troe (2010). Evaluated kinetic and photochemical data for atmospheric chemistry - IUPAC subcommittee on gas kinetic data evaluation for atmospheric chemistry.

Beekman, M. and C. Derognat (2003). "Monte Carlo uncertainty analysis of a regional-scale transport chemistry model constrained by measurements from the Atmospheric Pollution Over the Paris Area (ESQUIF) campaign." Journal of Geophysical Research 108. doi: 10.1029/2003JD003391.

Bergin, M. S. and J. B. Milford (2000). "Application of Bayesian Monte Carlo analysis to a Lagrangian photochemical air quality model." Atmospheric Environment 34(5): 781-792.

Bergin, M. S., G. S. Noblet, K. Petrini, J. R. Dhieux, J. B. Milford and R. A. Harley (1999). "Formal uncertainty analysis of a Lagrangian photochemical air pollution model." Environmental Science & Technology 33: 1116-1126.

Carlton, A. G. and K. R. Baker (2011). "Photochemical Modeling of the Ozark Isoprene Volcano: MEGAN, BEIS, and Their Impacts on Air Quality Predictions." Environmental Science & Technology 45(10): 4438-4445. doi: 10.1021/es200050x.

Cohan, D. S., A. Hakami, Y. T. Hu and A. G. Russell (2005). "Nonlinear response of ozone to emissions: Source apportionment and sensitivity analysis." Environmental Science & Technology 39(17): 6739-6748.

Cohan, D. S., B. Koo and G. Yarwood (2010). "Influence of uncertain reaction rates on ozone sensitivity to emissions." Atmospheric Environment 44(26): 3101-3109. doi: 10.1016/j.atmosenv.2010.05.034.

Deguillaume, L., M. Beekman and C. Derognat (2008). "Uncertainty evaluation of ozone production and its sensitivity to emission changes over the Ile-de-France region during summer periods." Journal of Geophysical Research 113. doi: 10.1029/2007JD009081.

Deguillaume, L., M. Beekman and L. Menut (2007a). "Bayesian Monte Carlo analysis applied to regional-scale inverse emission modeling for reactive trace gases." Journal of Geophysical Research 112. doi: 10.1029/2006JD007518.

Deguillaume, L., M. Beekmann and L. Menut (2007b). "Bayesian Monte Carlo analysis applied to regional-scale inverse emission modeling for reactive trace gases." J. Geophys. Res. 112. doi: 10.1029/2006jd007518.

Dennis, R., et al. (2010). "A framework for evaluating regional-scale numerical photochemical modeling systems." Environmental Fluid Mechanics 10(4): 471-489. doi: 10.1007/s10652-009-9163-2.

Digar, A. and D. S. Cohan (2010). "Efficient Characterization of Pollutant-Emission Response under Parametric Uncertainty." Environmental Science & Technology 44(17): 6724-6730. doi: 10.1021/es903743t.

78

ENVIRON (2010). User's Guide - Comprehensive Air Quality Model with Extensions Version 5.30.

Fine, J., L. Vuilleumier, S. Reynolds, P. Roth and N. Brown (2003). "Evaluating uncertainties in regional photochemical air quality modeling." Annual Review of Environment and Resources 28: 59-106.

Gao, D., W. R. Stockwell and J. B. Milford (1995). "First-order sensitivity and uncertainty analysis for a regional-scale gas-phase chemical mechanism." Journal of Geophysical Research 100(D11): 23153-23166.

Gilliland, A. B., C. Hogrefe, R. W. Pinder, J. M. Godowitch, K. L. Foley and S. T. Rao (2008). "Dynamic evaluation of regional air quality models: Assessing changes in O-3 stemming from changes in emissions and meteorology." Atmospheric Environment 42(20): 5110-5123. doi: 10.1016/j.atmosenv.2008.02.018.

Guenther, A., T. Karl, P. Harley, C. Wiedinmyer, P. I. Palmer and C. Geron (2006). "Estimates of global terrestrial isoprene emissions using MEGAN (Model of Emissions of Gases and Aerosols from Nature)." Atmospheric Chemistry and Physics 6: 3181-3210.

Hanna, S. R., Z. Lu, H. Christopher Frey, N. Wheeler, J. Vukovich, S. Arunachalam, M. Fernau and D. Alan Hansen (2001a). "Uncertainties in predicted ozone concentrations due to input uncertainties for the UAM-V photochemical grid model applied to the July 1995 OTAG domain." Atmospheric Environment 35(5): 891-903.

Hanna, S. R., Z. Lu, H. C. Frey, N. Wheeler, J. Vukovich, S. Arunachalam, M. Fernau and D. A. Hansen (2001b). "Uncertainties in predicted ozone concentrations due to input uncertainties for the UAM-V photochemical grid model applied to the July 1995 OTAG domain." Atmospheric Environment 35(5): 891-903.

Lelieveld, J., et al. (2008). "Atmospheric oxidation capacity sustained by a tropical forest." Nature 452(7188): 737-740. doi: 10.1038/nature06870.

Lin, X., M. Trainer and S. C. Liu (1988). "On the nonlinearity of the tropospheric ozone production." Journal of Geophysical Research 93(D12).

Mallet, V. and B. Sportisse (2006). "Uncertainty in a chemistry-transport model due to physical parameterizations and numerical approximations: An ensemble approach applied to ozone modeling." Journal of Geophysical Research-Atmospheres 111(D1): 15. doi: D01302

10.1029/2005jd006149. Mollner, A. K., et al. (2010). "Rate of Gas Phase Association of Hydroxyl Radical and Nitrogen

Dioxide." Science 330(6004): 646-649. doi: 10.1126/science.1193030. Paulot, F., J. D. Crounse, H. G. Kjaergaard, J. H. Kroll, J. H. Seinfeld and P. O. Wennberg

(2009a). "Isoprene photooxidation: new insights into the production of acids and organic nitrates." Atmospheric Chemistry and Physics 9(4): 1479-1501.

Paulot, F., J. D. Crounse, H. G. Kjaergaard, A. Kurten, J. M. St Clair, J. H. Seinfeld and P. O. Wennberg (2009b). "Unexpected Epoxide Formation in the Gas-Phase Photooxidation of Isoprene." Science 325(5941): 730-733. doi: 10.1126/science.1172910.

Peeters, J., T. L. Nguyen and L. Vereecken (2009). "HO(x) radical regeneration in the oxidation of isoprene." Physical Chemistry Chemical Physics 11(28): 5935-5939. doi: 10.1039/b908511d.

Pierce, T., C. Hogrefe, S. T. Rao, P. S. Porter and J. Y. Ku (2010). "Dynamic evaluation of a regional air quality model: Assessing the emissions-induced weekly ozone cycle." Atmospheric Environment 44(29): 3583-3596. doi: 10.1016/j.atmosenv.2010.05.046.

79

Pinder, R. W., R. C. Gilliam, K. W. Appel, S. L. Napelenok, K. M. Foley and A. B. Gilliland (2009). "Efficient Probabilistic Estimates of Surface Ozone Concentration Using an Ensemble of Model Configurations and Direct Sensitivity Calculations." Environmental Science & Technology 43(7): 2388-2393. doi: 10.1021/es8025402.

Raftery, A. E., T. Gneiting, F. Balabdaoui and M. Polakowski (2005). "Using Bayesian model averaging to calibrate forecast ensembles." Monthly Weather Review 133(5): 1155-1174. doi: 10.1175/mwr2906.1.

Sander S P, F. R. R., Ravishankara A R, Golden D M, Kolb C E, Kurylo M J, Molina M J, Moortgat G K, Rudek H K, Finlayson-Pitts B J, Wine P H, Huie R E (2006). Chemical kinetics and photochemical data for use in atmospheric studies, NASA JPL.

Slinn, S. A. and W. G. N. Slinn (1980). "Predictions for particle deposition on natural waters." Atmospheric Environment 14(9): 1013-1016. doi: 10.1016/0004-6981(80)90032-3.

Talagrand, O. R., R. Vautard and B. Strauss (1997). "Evaluation of probabilistic prediction systems." Proceedings, ECMWF Workshop on Predictability: 1-25.

TCEQ (2009). Houston-Galveston-Brazoria Attainment Demonstration State Implementation Plan Revision for the 1997 Eight-Hour Ozone Standard

Tian, D., M. S. Bergin, D. S. Cohan, S. L. Napelenok, Y. Hu, M. E. Chang and A. G. Russell (2010). "Uncertainty analysis of ozone formation and response to emission controls using higher-order sensitivities." Journal of the Air & Waste Management Association 60: 797-804.

US-EPA (2007). Guidance on the use of models and other analyses for demonstrating attainment of air quality goals for ozone, PM2.5, and regional haze.

Wesely, M. L. (1989). "Paramaterization of surface resistances to gaseous dry deposition in regional-scale numerical models." Atmospheric Environment 23(6): 1293-1304. doi: 10.1016/0004-6981(89)90153-4.

Whitten, G. Z., G. Heo, Y. Kimura, E. McDonald-Buller, D. T. Allen, W. P. L. Carter and G. Yarwood (2010). "A new condensed toluene mechanism for Carbon Bond CB05-TU." Atmospheric Environment 44(40): 5346-5355. doi: 10.1016/j.atmosenv.2009.12.029.

Zhang, L., J. R. Brook and R. Vet (2003). "A revised parameterization for gaseous dry deposition in air-quality models." Atmospheric Chemistry and Physics 3: 2067-2082.

Zhang, L. M., S. L. Gong, J. Padro and L. Barrie (2001). "A size-segregated particle dry deposition scheme for an atmospheric aerosol module." Atmospheric Environment 35(3): 549-560. doi: 10.1016/s1352-2310(00)00326-5.

80

Acknowledgment

The preparation of this report is based on work supported by the State of Texas through the Air

Quality Research Program administered by The University of Texas at Austin by means of a

Grant from the Texas Commission on Environmental Quality. We appreciate the support from

scientists at TCEQ who provided the input data for the modeling runs. Dr. Kristen Foley, a

statistician for US EPA, provided extensive consultations on the Bayesian method applications.

81

Appendix 1: CB6 Mechanism

Reactions and rate constants/parameters for the CB6 mechanism

Number Reactants and Products k298 Rate Parameters

A Ea B 1 NO2 = NO + O Photolysis 2 O + O2 + M = O3 + M 5.78E-34 5.68E-34 0.0 -2.60 3 O3 + NO = NO2 1.73E-14 1.40E-12 1310.0 0.00 4 O + NO + M = NO2 + M 1.01E-31 1.00E-31 0.0 -1.60 5 O + NO2 = NO 1.03E-11 5.50E-12 -188.0 0.00 6 O + NO2 = NO3 2.11E-12 Falloff, F=0.60 ,N=1.00 k0 1.30E-31 0.0 -1.50 k∞ 2.30E-11 0.0 0.24 7 O + O3 = 7.96E-15 8.00E-12 2060.0 0.00 8 O3 = O Photolysis 9 O3 = O1D Photolysis 10 O1D + M = O + M 3.28E-11 2.23E-11 -115.0 0.00 11 O1D + H2O = 2 OH 2.14E-10 2.14E-10 12 O3 + OH = HO2 7.25E-14 1.70E-12 940.0 0.00 13 O3 + HO2 = OH 2.01E-15 2.03E-16 -693.0 4.57 14 OH + O = HO2 3.47E-11 2.40E-11 -110.0 0.00 15 HO2 + O = OH 5.73E-11 2.70E-11 -224.0 0.00 16 OH + OH = O 1.48E-12 6.20E-14 -945.0 2.60 17 OH + OH = H2O2 5.25E-12 Falloff, F=0.50 ,N=1.13 k0 6.90E-31 0.0 -0.80 k∞ 2.60E-11 0.0 0.00 18 OH + HO2 = 1.11E-10 4.80E-11 -250.0 0.00 19 HO2 + HO2 = H2O2 2.90E-12 k = k1 + k2[M] k1 2.20E-13 -600.0 0.00 k2 1.90E-33 -980.0 0.00 20 HO2 + HO2 + H2O = H2O2 6.53E-30 k = k1 + k2[M] k1 3.08E-34 -2800.0 0.00 k2 2.66E-54 -3180.0 0.00 21 H2O2 = 2 OH Photolysis 22 H2O2 + OH = HO2 1.70E-12 2.90E-12 160.0 0.00 23 H2O2 + O = OH + HO2 1.70E-15 1.40E-12 2000.0 0.00 24 NO + NO + O2 = 2 NO2 1.95E-38 3.30E-39 -530.0 0.00 25 HO2 + NO = OH + NO2 8.54E-12 3.45E-12 -270.0 0.00 26 NO2 + O3 = NO3 3.52E-17 1.40E-13 2470.0 0.00 27 NO3 = NO2 + O Photolysis 28 NO3 = NO Photolysis 29 NO3 + NO = 2 NO2 2.60E-11 1.80E-11 -110.0 0.00 30 NO3 + NO2 = NO + NO2 6.56E-16 4.50E-14 1260.0 0.00 31 NO3 + O = NO2 1.70E-11 1.70E-11 32 NO3 + OH = HO2 + NO2 2.00E-11 2.00E-11 33 NO3 + HO2 = OH + NO2 4.00E-12 4.00E-12 34 NO3 + O3 = NO2 1.00E-17 1.00E-17 35 NO3 + NO3 = 2 NO2 2.28E-16 8.50E-13 2450.0 0.00

82

Number Reactants and Products k298 Rate Parameters

A Ea B 36 NO3 + NO2 = N2O5 1.24E-12 Falloff, F=0.35 ,N=1.33 k0 3.60E-30 0.0 -4.10 k∞ 1.90E-12 0.0 0.20 37 N2O5 = NO3 + NO2 4.46E-02 Falloff, F=0.35 ,N=1.33 k0 1.30E-03 11000.0 -3.50 k∞ 9.70E+14 11080.0 0.10 38 N2O5 = NO2 + NO3 Photolysis 39 N2O5 + H2O = 2 HNO3 1.00E-22 1.00E-22 40 NO + OH = HONO 9.77E-12 Falloff, F=0.81 ,N=0.87 k0 7.40E-31 0.0 -2.40 k∞ 3.30E-11 0.0 -0.30 41 NO + NO2 + H2O = 2 HONO 5.00E-40 5.00E-40 42 HONO + HONO = NO + NO2 1.00E-20 1.00E-20 43 HONO = NO + OH Photolysis 44 HONO + OH = NO2 5.98E-12 2.50E-12 -260.0 0.00 45 NO2 + OH = HNO3 9.25E-12 Falloff, F=0.60 ,N=1.00 k0 1.48E-30 0.0 -3.00 k∞ 2.58E-11 0.0 0.00 46 HNO3 + OH = NO3 1.54E-13 k = k1+k3M/(1+k3M/k2) k1 2.40E-14 -460.0 0.00 k2 2.70E-17 -2199.0 0.00 k3 6.50E-34 -1335.0 0.00 47 HNO3 = OH + NO2 Photolysis 48 HO2 + NO2 = PNA 1.38E-12 Falloff, F=0.60 ,N=1.00 k0 1.80E-31 0.0 -3.20 k∞ 4.70E-12 0.0 0.00 49 PNA = HO2 + NO2 8.31E-02 Falloff, F=0.60 ,N=1.00 k0 4.10E-05 10650.0 0.00 k∞ 4.80E+15 11170.0 0.00 50 PNA = 0.59 HO2 + 0.59 NO2 + 0.41 OH + 0.41 NO3 Photolysis

51 PNA + OH = NO2 3.24E-12 3.20E-13 -690.0 0.00 52 SO2 + OH = SULF + HO2 8.12E-13 Falloff, F=0.53 ,N=1.10 k0 4.50E-31 0.0 -3.90 k∞ 1.30E-12 0.0 -0.70 53 C2O3 + NO = NO2 + MEO2 + RO2 1.98E-11 7.50E-12 -290.0 0.00 54 C2O3 + NO2 = PAN 1.05E-11 Falloff, F=0.30 ,N=1.00 k0 2.70E-28 0.0 -7.10 k∞ 1.20E-11 0.0 -0.90 55 PAN = NO2 + C2O3 3.31E-04 Falloff, F=0.30 ,N=1.00 k0 4.90E-03 12100.0 0.00 k∞ 5.40E+16 13830.0 0.00 56 PAN = 0.6 NO2 + 0.6 C2O3 + 0.4 NO3 + 0.4 MEO2

+ 0.4 RO2 Photolysis

57 C2O3 + HO2 = 0.41 PACD + 0.15 AACD + 0.15 O3 + 0.44 MEO2 + 0.44 RO2 + 0.44 OH

1.39E-11 5.20E-13 -980.0 0.00

58 C2O3 + RO2 = C2O3 1.30E-11 8.90E-13 -800.0 0.00 59 C2O3 + C2O3 = 2 MEO2 + 2 RO2 1.55E-11 2.90E-12 -500.0 0.00

83

Number Reactants and Products k298 Rate Parameters

A Ea B 60 C2O3 + CXO3 = MEO2 + ALD2 + XO2H + 2 RO2 1.55E-11 2.90E-12 -500.0 0.00

61 CXO3 + NO = NO2 + ALD2 + XO2H + RO2 2.10E-11 6.70E-12 -340.0 0.00 62 CXO3 + NO2 = PANX 1.16E-11 Falloff, F=0.30 ,N=1.00 k0 3.00E-28 0.0 -7.10 k∞ 1.33E-11 0.0 -0.90 63 PANX = NO2 + CXO3 3.68E-04 Falloff, F=0.30 ,N=1.00 k0 1.70E-03 11280.0 0.00 k∞ 8.30E+16 13940.0 0.00 64 PANX = 0.6 NO2 +0.6 CXO3 + 0.4 NO3 + 0.4 ALD2

+ 0.4 XO2H + 0.4 RO2 Photolysis

65 CXO3 + HO2 = 0.41 PACD + 0.15 AACD + 0.15 O3 + 0.44 ALD2 + 0.44 XO2H + 0.44 RO2 + 0.44 OH

1.39E-11 5.20E-13 -980.0 0.00

66 CXO3 + RO2 = 0.8 ALD2 + 0.8 XO2H + 0.8 RO2 1.30E-11 8.90E-13 -800.0 0.00 67 CXO3 + CXO3 = 2 ALD2 + 2 XO2H + 2 RO2 1.71E-11 3.20E-12 -500.0 0.00 68 RO2 + NO = NO 8.03E-12 2.40E-12 -360.0 0.00 69 RO2 + HO2 = HO2 7.03E-12 4.80E-13 -800.0 0.00 70 RO2 + RO2 = 3.48E-13 6.50E-14 -500.0 0.00 71 MEO2 + NO = FORM + HO2 + NO2 7.70E-12 2.30E-12 -360.0 0.00 72 MEO2 + HO2 = 0.9 MEPX + 0.1 FORM 5.21E-12 3.80E-13 -780.0 0.00 73 MEO2 + C2O3 = FORM + 0.9 HO2 + 0.9 MEO2 +

0.1 AACD + 0.9 RO2 1.07E-11 2.00E-12 -500.0 0.00

74 MEO2 + RO2 = 0.685 FORM + 0.315 MEOH + 0.37 HO2 + RO2

3.48E-13 k = kref*K k(ref) ref = 70 K 1.00E+00 0.0 0.00 75 XO2H + NO = NO2 + HO2 9.04E-12 2.70E-12 -360.0 0.00 76 XO2H + HO2 = ROOH 9.96E-12 6.80E-13 -800.0 0.00 77 XO2H + C2O3 = 0.8 HO2 + 0.8 MEO2 + 0.2 AACD +

0.8 RO2 1.30E-11 k = kref*K

k(ref) ref = 58 K 1.00E+00 0.0 0.00 78 XO2H + RO2 = 0.6 HO2 + RO2 3.48E-13 k = kref*K k(ref) ref = 70 K 1.00E+00 0.0 0.00 79 XO2 + NO = NO2 9.04E-12 k = kref*K k(ref) ref = 75 K 1.00E+00 0.0 0.00 80 XO2 + HO2 = ROOH 9.96E-12 k = kref*K k(ref) ref = 76 K 1.00E+00 0.0 0.00 81 XO2 + C2O3 = 0.8 MEO2 + 0.2 AACD + 0.8 RO2 1.30E-11 k = kref*K

k(ref) ref = 58 K 1.00E+00 0.0 0.00 82 XO2 + RO2 = RO2 3.48E-13 k = kref*K k(ref) ref = 70 K 1.00E+00 0.0 0.00 83 XO2N + NO = NTR 9.04E-12 k = kref*K

84

Number Reactants and Products k298 Rate Parameters

A Ea B k(ref) ref = 75 K 1.00E+00 0.0 0.00 84 XO2N + HO2 = ROOH 9.96E-12 k = kref*K k(ref) ref = 76 K 1.00E+00 0.0 0.00 85 XO2N + C2O3 = 0.8 HO2 + 0.8 MEO2 + 0.2 AACD +

0.8 RO2 1.30E-11 k = kref*K

k(ref) ref = 58 K 1.00E+00 0.0 0.00 86 XO2N + RO2 = RO2 3.48E-13 k = kref*K k(ref) ref = 70 K 1.00E+00 0.0 0.00 87 MEPX + OH = 0.6 MEO2 + 0.6 RO2 + 0.4 FORM +

0.4 OH 1.00E-11 5.30E-12 -190.0 0.00

88 MEPX = MEO2 + RO2 + OH Photolysis 89 ROOH + OH = 0.54 XO2H + 0.06 XO2N + 0.6 RO2

+ 0.4 OH 6.05E-12 3.20E-12 -190.0 0.00

90 ROOH = HO2 + OH Photolysis 91 NTR + OH = HNO3 + XO2H + RO2 8.10E-13 8.10E-13 92 NTR = NO2 + XO2H + RO2 Photolysis 93 FACD + OH = HO2 4.50E-13 4.50E-13 94 AACD + OH = MEO2 + RO2 6.93E-13 4.00E-14 -850.0 0.00 95 PACD + OH = C2O3 6.93E-13 4.00E-14 -850.0 0.00 96 FORM + OH = HO2 + CO 8.49E-12 5.40E-12 -135.0 0.00 97 FORM = 2 HO2 + CO Photolysis 98 FORM = CO + H2 Photolysis 99 FORM + O = OH + HO2 + CO 1.58E-13 3.40E-11 1600.0 0.00 100 FORM + NO3 = HNO3 + HO2 + CO 5.50E-16 5.50E-16 101 FORM + HO2 = HCO3 7.90E-14 9.70E-15 -625.0 0.00 102 HCO3 = FORM + HO2 1.51E+02 2.40E+12 7000.0 0.00 103 HCO3 + NO = FACD + NO2 + HO2 5.60E-12 5.60E-12 104 HCO3 + HO2 = 0.5 MEPX + 0.5 FACD + 0.2 OH +

0.2 HO2 1.26E-11 5.60E-15 -2300.0 0.00

105 ALD2 + O = C2O3 + OH 4.49E-13 1.80E-11 1100.0 0.00 106 ALD2 + OH = C2O3 1.50E-11 4.70E-12 -345.0 0.00 107 ALD2 + NO3 = C2O3 + HNO3 2.73E-15 1.40E-12 1860.0 0.00 108 ALD2 = MEO2 + RO2 + CO + HO2 Photolysis 109 ALDX + O = CXO3 + OH 7.02E-13 1.30E-11 870.0 0.00 110 ALDX + OH = CXO3 1.91E-11 4.90E-12 -405.0 0.00 111 ALDX + NO3 = CXO3 + HNO3 6.30E-15 6.30E-15 112 ALDX = ALD2 + XO2H + RO2 + CO + HO2 Photolysis 113 GLYD + OH = 0.2 GLY + 0.2 HO2 + 0.8 C2O3 8.00E-12 8.00E-12 114 GLYD = 0.74 FORM + 0.89 CO + 1.4 HO2 + 0.15

MEOH + 0.19 OH + 0.11 GLY + 0.11 XO2H + 0.11 RO2

Photolysis

115 GLYD + NO3 = HNO3 + C2O3 2.73E-15 1.40E-12 1860.0 0.00 116 GLY + OH = 1.7 CO + 0.3 XO2 + 0.3 RO2 + HO2 9.70E-12 3.10E-12 -340.0 0.00 117 GLY = 2 HO2 + 2 CO Photolysis 118 GLY + NO3 = HNO3 + CO + HO2 + XO2 + RO2 2.73E-15 1.40E-12 1860.0 0.00

85

Number Reactants and Products k298 Rate Parameters

A Ea B 119 MGLY = C2O3 + HO2 + CO Photolysis 120 MGLY + NO3 = HNO3 + C2O3 + XO2 + RO2 2.73E-15 1.40E-12 1860.0 0.00 121 MGLY + OH = C2O3 + CO 1.31E-11 1.90E-12 -575.0 0.00 122 H2 + OH = HO2 6.70E-15 7.70E-12 2100.0 0.00 123 CO + OH = HO2 2.28E-13 k = k1 + k2[M] k1 1.44E-13 0.0 0.00 k2 3.43E-33 0.0 0.00 124 CH4 + OH = MEO2 + RO2 6.37E-15 1.85E-12 1690.0 0.00 125 ETHA + OH = 0.991 ALD2 + 0.991 XO2H + 0.009

XO2N + RO2 2.41E-13 6.90E-12 1000.0 0.00

126 MEOH + OH = FORM + HO2 8.95E-13 2.85E-12 345.0 0.00 127 ETOH + OH = 0.95 ALD2 + 0.9 HO2 + 0.1 XO2H +

0.1 RO2 + 0.078 FORM + 0.011 GLYD 3.21E-12 3.00E-12 -20.0 0.00

128 KET = 0.5 ALD2 + 0.5 C2O3 + 0.5 XO2H +0.5 CXO3 + 0.5 MEO2 + RO2 - 2.5 PAR

Photolysis

129 ACET = 0.38 CO + 1.38 MEO2 + 1.38 RO2 + 0.62 C2O3

Photolysis

130 ACET + OH = FORM + C2O3 + XO2 + RO2 1.76E-13 1.41E-12 620.6 0.00 131 PRPA + OH = 0.71 ACET + 0.26 ALDX + 0.26 PAR +

0.97 XO2H + 0.03 XO2N + RO2 1.07E-12 7.60E-12 585.0 0.00

132 PAR + OH = 0.11 ALDX + 0.76 ROR + 0.13 XO2N + 0.11 XO2H + 0.76 XO2 + RO2 - 0.11 PAR

8.10E-13 8.10E-13

133 ROR = 0.2 KET + 0.42 ACET + 0.74 ALD2 + 0.37 ALDX + 0.04 XO2N + 0.94 XO2H + 0.98 RO2 + 0.02 ROR - 2.7 PAR

2.15E+04 5.70E+12 5780.0 0.00

134 ROR + O2 = KET + HO2 3.78E+04 1.50E-14 200.0 0.00 135 ROR + NO2 = NTR 3.29E-11 8.60E-12 -400.0 0.00 136 ETHY + OH = 0.7 GLY + 0.7 OH + 0.3 FACD + 0.3 CO

+ 0.3 HO2 7.52E-13 Falloff, F=0.37 ,N=1.30

k0 5.00E-30 0.0 -1.50 k∞ 1.00E-12 0.0 0.00 137 ETH + O = FORM + HO2 + CO + 0.7 XO2H + 0.7

RO2 + 0.3 OH 7.29E-13 1.04E-11 792.0 0.00

138 ETH + OH = XO2H + RO2 + 1.56 FORM + 0.22 GLYD

7.84E-12 Falloff, F=0.48 ,N=1.15 k0 8.60E-29 0.0 -3.10 k∞ 9.00E-12 0.0 -0.85 139 ETH + O3 = FORM + 0.51 CO + 0.16 HO2 + 0.16 OH

+ 0.37 FACD 1.58E-18 9.10E-15 2580.0 0.00

140 ETH + NO3 = 0.5 NO2 + 0.5 NTR + 0.5 XO2H + 0.5 XO2 + RO2 + 1.125 FORM

2.10E-16 3.30E-12 2880.0 0.00

141 OLE + O = 0.2 ALD2 + 0.3 ALDX + 0.1 HO2 + 0.2 XO2H + 0.2 CO + 0.2 FORM + 0.01 XO2N + 0.21 RO2 + 0.2 PAR + 0.1 OH

3.91E-12 1.00E-11 280.0 0.00

142 OLE + OH = 0.781 FORM + 0.488 ALD2 + 0.488 ALDX + 0.976 XO2H + 0.195 XO2 + 0.024 XO2N + 1.17 RO2 - 0.73 PAR

2.86E-11 Falloff, F=0.50 ,N=1.13 k0 8.00E-27 0.0 -3.50 k∞ 3.00E-11 0.0 -1.00 143 OLE + O3 = 0.295 ALD2 + 0.555 FORM + 0.27

ALDX + 0.15 XO2H + 0.15 RO2 + 0.334 OH +0.08 HO2 + 0.378 CO + 0.075 GLY + 0.075 MGLY + 0.09 FACD + 0.13 AACD + 0.04 H2O2 - 0.79 PAR

1.00E-17 5.50E-15 1880.0 0.00

86

Number Reactants and Products k298 Rate Parameters

A Ea B 144 OLE + NO3 = 0.5 NO2 + 0.5 NTR + 0.48 XO2 + 0.48

XO2H + 0.04 XO2N + RO2 + 0.5 FORM + 0.25 ALD2 + 0.375 ALDX - PAR

9.54E-15 4.60E-13 1155.0 0.00

145 IOLE + O = 1.24 ALD2 + 0.66 ALDX + 0.1 XO2H + 0.1 RO2 + 0.1 CO + 0.1 PAR

2.30E-11 2.30E-11

146 IOLE + OH = 1.3 ALD2 + 0.7 ALDX + XO2H + RO2 5.99E-11 1.05E-11 -519.0 0.00 147 IOLE + O3 = 0.732 ALD2 + 0.442 ALDX + 0.128

FORM + 0.245 CO + 0.5 OH + 0.3 XO2H + 0.3 RO2 + 0.24 GLY + 0.06 MGLY + 0.29 PAR + 0.08 AACD + 0.08 H2O2

1.57E-16 4.70E-15 1013.0 0.00

148 IOLE + NO3 = 0.5 NO2 + 0.5 NTR + 0.48 XO2 + 0.48 XO2H + 0.04 XO2N + RO2 + 0.5 ALD2 + 0.625 ALDX + PAR

3.70E-13 3.70E-13

149 ISOP + OH = ISO2 + RO2 9.99E-11 2.70E-11 -390.0 0.00 150 ISO2 + NO = 0.117 INTR + 0.883 NO2 + 0.803 HO2

+ 0.66 FORM + 0.66 ISPD + 0.08 XO2H + 0.08 RO2 + 0.05 IOLE + 0.042 GLYD + 0.115 PAR + 0.038 GLY + 0.042 MGLY + 0.093 OLE + 0.117 ALDX

8.13E-12 2.39E-12 -365.0 0.00

151 ISO2 + HO2 = 0.88 ISPX + 0.12 OH + 0.12 HO2 + 0.12 FORM + 0.12 ISPD

7.78E-12 7.43E-13 -700.0 0.00

152 ISO2 + C2O3 = 0.709 HO2 + 0.583 FORM + 0.583 ISPD + 0.071 XO2H + 0.044 IOLE + 0.037 GLYD + 0.102 PAR + 0.034 GLY + 0.037 MGLY + 0.082 OLE + 0.103 ALDX + 0.8 MEO2 + 0.2 AACD + 0.871 RO2

1.30E-11 k = kref*K k(ref) ref = 58 K 1.00E+00 0.0 0.00

153

ISO2 + RO2 = 0.803 HO2 + 0.66 FORM + 0.66 ISPD + 0.08 XO2H + 0.05 IOLE + 0.042 GLYD + 0.115 PAR + 0.038 GLY + 0.042 MGLY + 0.093 OLE + 0.117 ALDX + 1.08 RO2

3.48E-13 k = kref*K

k(ref) ref = 70 K 1.00E+00 0.0 0.00 154 ISO2 = 0.8 HO2 + 0.04 OH + 0.04 FORM + 0.8 ISPD 1.00E+00 1.00E+00 155 ISOP + O3 = 0.6 FORM + 0.65 ISPD + 0.15 ALDX +

0.2 CXO3 + 0.35 PAR + 0.266 OH + 0.2 XO2 + 0.2 RO2 + 0.066 HO2 + 0.066 CO

1.27E-17 1.03E-14 1995.0 0.00

156 ISOP + NO3 = 0.35 NO2 + 0.65 INTR + 0.64 XO2H + 0.33 XO2 + 0.03 XO2N + RO2 + 0.35 FORM + 0.35 ISPD

6.74E-13 3.03E-12 448.0 0.00

157 ISPD + OH = 0.095 XO2N + 0.379 XO2 + 0.318 XO2H + 0.792 RO2 + 0.843 PAR + 0.379 C2O3 + 0.209 CXO3 + 0.379 GLYD + 0.24 MGLY + 0.24 FORM + 0.067 OLE + 0.079 CO + 0.028 ALDX

3.38E-11 6.31E-12 -500.0 0.00

158 ISPD + O3 = 0.02 ALD2 + 0.15 FORM + 0.225 CO + 0.85 MGLY + 0.36 PAR + 0.114 C2O3 + 0.064 XO2H + 0.064 RO2 + 0.268 OH + 0.09 HO2

7.10E-18 4.17E-15 1900.0 0.00

159 ISPD + NO3 = 0.643 CO + 0.282 FORM + 0.357 ALDX + 1.282 PAR + 0.85 HO2 + 0.075 CXO3 + 0.075 XO2H + 0.075 RO2 + 0.85 NTR + 0.15 HNO3

1.00E-15 1.00E-15

160 ISPD = 0.333 CO + 0.067 ALD2 + 0.9 FORM + 0.832 PAR + 0.333 HO2 + 0.7 XO2H + 0.7 RO2 + 0.967 C2O3

Photolysis

161 ISPX + OH = 0.904 EPOX + 0.933 OH + 0.067 ISO2 + 0.067 RO2 + 0.029 IOLE + 0.029 ALDX

7.77E-11 2.23E-11 -372.0 0.00

162 EPOX + OH = EPX2 + RO2 1.51E-11 5.78E-11 400.0 0.00

87

Number Reactants and Products k298 Rate Parameters

A Ea B 163 EPX2 + HO2 = 0.275 GLYD + 0.275 GLY + 0.275

MGLY + 1.125 OH + 0.825 HO2 + 0.375 FORM + 0.074 FACD + 0.251 CO + 2.175 PAR

7.78E-12 7.43E-13 -700.0 0.00

164 EPX2 + NO = 0.275 GLYD + 0.275 GLY + 0.275 MGLY + 0.125 OH + 0.825 HO2 + 0.375 FORM + NO2 + 0.251 CO + 2.175 PAR

8.13E-12 2.39E-12 -365.0 0.00

165 EPX2 + C2O3 = 0.22 GLYD + 0.22 GLY + 0.22 MGLY + 0.1 OH + 0.66 HO2 + 0.3 FORM + 0.2 CO + 1.74 PAR + 0.8 MEO2 + 0.2 AACD + 0.8 RO2

1.30E-11 k = kref*K k(ref) ref = 58 K 1.00E+00 0.0 0.00 166 EPX2 + RO2 = 0.275 GLYD + 0.275 GLY + 0.275

MGLY + 0.125 OH + 0.825 HO2 + 0.375 FORM + 0.251 CO + 2.175 PAR + RO2

3.48E-13 k = kref*K k(ref) ref = 70 K 1.00E+00 0.0 0.00 167 INTR + OH = 0.63 XO2 + 0.37 XO2H + RO2 + 0.444

NO2 + 0.185 NO3 + 0.104 INTR + 0.592 FORM + 0.331 GLYD + 0.185 FACD + 2.7 PAR + 0.098 OLE + 0.078 ALDX + 0.266 NTR

3.10E-11 3.10E-11

168 TERP + O = 0.15 ALDX + 5.12 PAR 3.60E-11 3.60E-11 169 TERP + OH = 0.75 XO2H + 0.5 XO2 + 0.25 XO2N +

1.5 RO2 + 0.28 FORM + 1.66 PAR + 0.47 ALDX 6.77E-11 1.50E-11 -449.0 0.00

170 TERP + O3 = 0.57 OH + 0.07 XO2H + 0.69 XO2 + 0.18 XO2N + 0.94 RO2 + 0.24 FORM + 0.001 CO + 7 PAR + 0.21 ALDX + 0.39 CXO3

7.63E-17 1.20E-15 821.0 0.00

171 TERP + NO3 = 0.47 NO2 + 0.28 XO2H + 0.75 XO2 + 0.25 XO2N + 1.28 RO2 + 0.47 ALDX + 0.53 NTR

6.66E-12 3.70E-12 -175.0 0.00

172 BENZ + OH = 0.53 CRES + 0.352 BZO2 + 0.352 RO2 + 0.118 OPEN + 0.118 OH + 0.53 HO2

1.22E-12 2.30E-12 190.0 0.00

173 BZO2 + NO = 0.918 NO2 + 0.082 NTR + 0.918 GLY + 0.918 OPEN + 0.918 HO2

9.04E-12 2.70E-12 -360.0 0.00

174 BZO2 + C2O3 = GLY + OPEN + HO2 + MEO2 + RO2 1.30E-11 k = kref*K k(ref) ref = 58 K 1.00E+00 0.0 0.00 175 BZO2 + HO2 = 1.49E-11 1.90E-13 -1300.0 0.00 176 BZO2 + RO2 = GLY + OPEN + HO2 + RO2 3.48E-13 k = kref*K k(ref) ref = 70 K 1.00E+00 0.0 0.00 177 TOL + OH = 0.18 CRES + 0.65 TO2 + 0.72 RO2 + 0.1

OPEN + 0.1 OH + 0.07 XO2H + 0.18 HO2 5.63E-12 1.80E-12 -340.0 0.00

178 TO2 + NO = 0.86 NO2 + 0.14 NTR + 0.417 GLY + 0.443 MGLY + 0.66 OPEN + 0.2 XOPN + 0.86 HO2

9.04E-12 2.70E-12 -360.0 0.00

179 TO2 + C2O3 = 0.48 GLY + 0.52 MGLY + 0.77 OPEN + 0.23 XOPN + HO2 + MEO2 + RO2

1.30E-11 k = kref*K k(ref) ref = 58 K 1.00E+00 0.0 0.00 180 TO2 + HO2 = 1.49E-11 1.90E-13 -1300.0 0.00 181 TO2 + RO2 = 0.48 GLY + 0.52 MGLY + 0.77 OPEN +

0.23 XOPN + HO2 + RO2 3.48E-13 k = kref*K

k(ref) ref = 70 K 1.00E+00 0.0 0.00 182 XYL + OH = 0.155 CRES + 0.544 XLO2 + 0.602 RO2

+ 0.244 XOPN + 0.244 OH + 0.058 XO2H + 0.155 HO2

1.85E-11 1.85E-11

183 XLO2 + NO = 0.86 NO2 + 0.14 NTR + 0.221 GLY + 0.675 MGLY + 0.3 OPEN + 0.56 XOPN + 0.86 HO2

9.04E-12 2.70E-12 -360.0 0.00

88

Number Reactants and Products k298 Rate Parameters

A Ea B 184 XLO2 + HO2 = 1.49E-11 1.90E-13 -1300.0 0.00 185 XLO2 + C2O3 = 0.26 GLY + 0.77 MGLY + 0.35

OPEN + 0.65 XOPN + HO2 + MEO2 + RO2 1.30E-11 k = kref*K

k(ref) ref = 58 K 1.00E+00 0.0 0.00 186 XLO2 + RO2 = 0.26 GLY + 0.77 MGLY + 0.35 OPEN

+ 0.65 XOPN + HO2 + RO2 3.48E-13 k = kref*K

k(ref) ref = 70 K 1.00E+00 0.0 0.00 187 CRES + OH = 0.06 CRO + 0.12 XO2H + HO2 + 0.13

OPEN + 0.732 CAT1 + 0.06 CO + 0.06 XO2N + 0.18 RO2 + 0.06 FORM

4.12E-11 1.70E-12 -950.0 0.00

188 CRES + NO3 = 0.3 CRO + HNO3 + 0.24 XO2 + 0.36 XO2H + 0.48 ALDX + 0.24 FORM + 0.24 MGLY + 0.12 OPEN + 0.1 XO2N + 0.7 RO2 + 0.24 CO

1.40E-11 1.40E-11

189 CRO + NO2 = CRON 2.10E-12 2.10E-12 190 CRO + HO2 = CRES 5.50E-12 5.50E-12 191 CRON + OH = CRNO 1.53E-12 1.53E-12 192 CRON + NO3 = CRNO + HNO3 3.80E-12 3.80E-12 193 CRNO + NO2 = 2 NTR 2.10E-12 2.10E-12 194 CRNO + O3 = CRN2 2.86E-13 2.86E-13 195 CRN2 + NO = CRNO + NO2 8.50E-12 2.54E-12 -360.0 0.00 196 CRN2 + HO2 = CRPX 1.88E-11 2.40E-13 -1300.0 0.00 197 CRPX = CRNO + OH Photolysis 198 CRPX + OH = CRN2 3.59E-12 1.90E-12 -190.0 0.00 199 XOPN = CAO2 + 0.7 HO2 + 0.7 CO + 0.3 C2O3 +

RO2 Photolysis

200 XOPN + OH = CAO2 + MGLY + XO2H + RO2 9.00E-11 9.00E-11 201 XOPN + O3 = 1.2 MGLY + 0.5 OH + 0.6 C2O3 + 0.1

ALD2 + 0.5 CO + 0.3 XO2H + 0.3 RO2 2.02E-17 1.08E-16 500.0 0.00

202 XOPN + NO3 = 0.5 NO2 + 0.5 NTR + 0.45 XO2H + 0.45 XO2 + 0.1 XO2N + RO2 + 0.25 OPEN + 0.25 MGLY

3.00E-12 3.00E-12

203 OPEN = OPO3 + HO2 + CO Photolysis 204 OPEN + OH = 0.6 OPO3 + 0.4 CAO2 + 0.4 RO2 4.40E-11 4.40E-11 205 OPEN + O3 = 1.4 GLY + 0.24 MGLY + 0.5 OH + 0.12

C2O3 + 0.08 FORM + 0.02 ALD2 + 1.98 CO + 0.56 HO2

1.01E-17 5.40E-17 500.0 0.00

206 OPEN + NO3 = OPO3 + HNO3 3.80E-12 3.80E-12 207 CAT1 + OH = CAO2 + RO2 7.00E-11 7.00E-11 208 CAT1 + NO3 = CRO + HNO3 1.70E-10 1.70E-10 209 CAO2 + NO = 0.86 NO2 + 0.14 NTR + 1.2 HO2 +

0.344 FORM + 0.344 CO 8.50E-12 2.54E-12 -360.0 0.00

210 CAO2 + HO2 = 1.88E-11 2.40E-13 -1300.0 0.00 211 CAO2 + C2O3 = HO2 + 0.4 GLY + MEO2 + RO2 1.30E-11 k = kref*K k(ref) ref = 58 K 1.00E+00 0.0 0.00 212 CAO2 + RO2 = HO2 + 0.4 GLY + RO2 3.48E-13 k = kref*K k(ref) ref = 70 K 1.00E+00 0.0 0.00 213 OPO3 + NO = NO2 + XO2H + RO2 + ALDX 1.00E-11 1.00E-11

89

Number Reactants and Products k298 Rate Parameters

A Ea B 214 OPO3 + NO2 = OPAN 1.16E-11 k = kref*K k(ref) ref = 62 K 1.00E+00 0.0 0.00 215 OPAN = OPO3 + NO2 9.92E-05 Falloff, F=0.30 ,N=1.00 k0 4.60E-04 11280.0 0.00 k∞ 2.24E+16 13940.0 0.00 216 OPO3 + HO2 = 0.41 PACD + 0.15 AACD + 0.15 O3

+ 0.44 ALDX + 0.44 XO2H + 0.44 RO2 + 0.44 OH 1.39E-11 k = kref*K

k(ref) ref = 57 K 1.00E+00 0.0 0.00 217 OPO3 + C2O3 = MEO2 + XO2 +ALDX + 2 RO2 1.55E-11 k = kref*K k(ref) ref = 59 K 1.00E+00 0.0 0.00 218 OPO3 + RO2 = 0.8 XO2H + 0.8 RO2 + 0.8 ALDX +

0.2 AACD 1.30E-11 k = kref*K

k(ref) ref = 58 K 1.00E+00 0.0 0.00 k298 is the rate constant at 298 K and 1 atmosphere using units molecules cm-3 and s-1