in support of US air quality management application to the GEOSChem model Daniel Jacob with Kelvin Bates Xuan Wang 1 Viral Shah Lu Shen Makoto Kelp Bob Yantosca An EPA STAR Chemical Mechanisms for Air Quality Modeling Project ID: 1048457
Download Presentation The PPT/PDF document "Improving chemical mechanisms for region..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
1. Improving chemical mechanisms for regional/global models in support of US air quality management: application to the GEOS-Chem modelDaniel Jacobwith Kelvin Bates, Xuan Wang1, Viral Shah, Lu Shen, Makoto Kelp, Bob Yantosca An EPA STAR Chemical Mechanisms for Air Quality Modeling Project1now at City University of Hong Kong
2. GEOS-Chem as a modular platform for mechanism development(x, y)C(x, t)C(x, t + Δt)dC/dt = P – L + E - Dchemistry (FlexChem-KPP)emissionsdepositionGEOS-Chem chemical module:transportoff-line (archived winds)GC-Classic [Bey et al.,JGR 2001]GCHP [Eastham et al., GMD 2018]Stretched-grid GCHP [Bindle et al., GMD 2020]NASA GEOS [Hu et al., GMD 2018], GEOS forecasts [Keller et al., JAMES 2020]WRF [Lin et al., GMD 2019]BCC-ESM [Lu et al., GMD 2020]NCAR CESM (in progress)on-line in meteorological model (solves for chemical transport)
3. Ongoing workNew reduced mechanism for aromatic VOC chemistry (Kelvin Bates)New mechanism for tropospheric halogen (Cl-Br-I) chemistry (Xuan Wang) New mechanism for mercury redox chemistry (Viral Shah)Adaptive reduction of chemical mechanisms for global models (Lu Shen)Application of machine learning to chemical solvers (Makoto Kelp)Unification of GEOS-Chem chemical mechanisms under KPP (Bob Yantosca)
4. New reduced mechanism for aromatic VOCsMechanism Species ReactionsMCM v3.1 1271 3788MECCA 229 666SAPRC-11 55 374CRI v2-R5 56 128RACM2 34 115MOZART-T1 13 43This work 18 46First stage:Second stage:Bates et al., in prep.Motivation:Simulation of ozone, PAN, SOAUse of glyoxal/formaldehyde satellite dataphenoxy-phenylperoxy cycle:catalytic ozone lossBenzaldehyde
5. Mechanisms intercomparison in box modelLarge production of glyoxal and methylglyoxal following ring-breakingSource of HOx from photolysis of dicarbonylsOzone sink at low NOx from phenoxy-phenylperoxy cyclingBates et al., in prep.This workThis work
6. Seasonally varying effects on OH concentrationsGEOS-Chem simulation with vs. without aromatic VOCs: change in boundary layer OHLarge HOx source in winter from photolysis of HCHO and dicarbonyls drives winter-spring spread of ozone pollution in northern ChinaBates et al., in prep.
7. New mechanism for tropospheric Cl-Br-I chemistryGEOS-Chem global masses in Gg [ppt in brackets], rates in Gg a-1Wang et al., in prepCl radical production from SSA acid displacement and heterogeneous chemistryClNO2 chemistry as ozone source in high-NOx regionsBrOx radical production from HOBr + Cl-Cl atom accounts for only 1% of global methane loss
8. Wang et al., in prepGlobal daytime BrO 0.5 ppt, drives ozone lossBr atoms provide dominant oxidant for Hg(0)GEOS-Chem global masses in Gg [ppt in brackets], rates in Gg a-1New mechanism for tropospheric Cl-Br-I chemistry
9. Wang et al., in prepCatalytic ozone loss and OH production through HOIGEOS-Chem global masses in Gg [ppt in brackets], rates in Gg a-1New mechanism for tropospheric Cl-Br-I chemistry
10. Global annual halogen-driven changes in oxidantsWang et al., in prepsurfacetropopauseOH and NOx increases over land mostly from ClNO2 chemistryOzone loss mostly from BrOx chemistry
11. New mechanism for Hg(0)/Hg(II) redox chemistryShah et al., in progressFirst-generation Hg(II) products have very short lifetimes against photolysisPush Hg(II) to stable forms: HgBrOH, Hg-Org, HgCl2
12. Observational constraints on Hg(0)/Hg(II) redox chemistryGlobal Hg(0) distribution: lifetime 6-9 months against deposition (here 9 months)Hg (II) wet deposition fluxesHg(0) seasonal variation at northern and southern mid-latitudesFast depletion of Hg(0) above tropopauseShah et al., in progress
13. Evaluation with MDN observations over North AmericaWe do not capture Southeast US maximum associated with deep convectionNeed to increase Hg(II) pool in free troposphereShah et al., in progressGlobal Hg(II) poolscavenging
14. Adaptive reduction of chemical mechanisms in global modelsGEOS-Chem mechanism (262 species, 850 reactions) covers all bases in troposphere and stratosphere but is often a severe overkill:reactive VOCs may be unimportant outside source regionsday and night, troposphere and stratosphere involve different species% of ‘fast’ species in mechanism with P or L larger than threshold δ = 102 cm-3 s-1Santillana et al., AE 2010; Shen et al., GMD 2019Can we reduce the mechanism locally to include only ‘fast’ species in the coupled chemical solver?Problem is computational overhead in diagnosing/constructing local mechanism and Jacobian matrixSolution: have library of pre-constructed sub-mechanisms from which model picks locally and on the flya July day at 0 GMT
15. Selection of 20 sub-mechanisms to cover all atmospheric regimesApply machine learning algorithm to GEOS-Chem gridboxes globally, diagnosing species with max(P,L) < δ and accounting for chemical connections13 species groups (blocks)20 submechanismsoxidants inorganichalogensanthro VOCsaerosolsbio VOCSand SOAorganic halogensShen et al., submittedAvoid NMVOC chemistry in half of gridboxes
16. Adaptive mechanism reduction in GEOS-Chem:70% decrease in # coupled species, 50% gain in CPU time, <1% errorsimulation month% error over 3-year simulationShen et al., submittedδ = 500 cm-3 s-1Advantages:Simplify only where warrantedKeep full diagnostic infoAllow mechanism updates
17. Application of machine learning to chemical solversWhy: chemical calculation is expensive, highly repetitive, fully deterministicWhy not: high dimensionality, error growthRecursive neural network algorithm with encoder/decoder:Reduces dimension along dominant manifoldsRecursive fit allows capture of longer modes to reduce error growthKelp et al. [JGR 2020]
18. Application of NN solver to SuperFast mechanism in GEOS-ChemGlobal chemical mechanism with only 15 species [Brown-Steiner et al., GMD 2018]Error in simulating surface ozone in 30-day simulationRandom-forest algorithm [Keller and Evans, GMD 2019]On-line trained NN algorithm (this work)Kelp et al., in progressBeijingoceanNeed to check other species, error growthWhy is on-line training so important?
19. Plans for coming yearUnify all GEOS-Chem chemical mechanisms under KPPComplete/submit aromatic VOC mechanism paper, develop SOA mechanismComplete/submit halogen mechanism paperComplete/submit mercury mechanism paperIntercompare isoprene mechanisms, develop SOA mechanismImplement adaptive mechanism reduction in standard GEOS-ChemComplete/submit implementation of ML solver in SuperFast mechanism, apply to full GEOS-Chem mechanism