Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from:...

22
Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow Fellowship Source evaluation of aerosol precursors with the adjoint of GEOS-Chem
  • date post

    22-Dec-2015
  • Category

    Documents

  • view

    216
  • download

    0

Transcript of Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from:...

Page 1: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Daven K. Henzewith Amir Hakami and John H. Seinfeld

Caltech, Chemical Engineering

Support from:

NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow Fellowship

Source evaluation of aerosol precursors with the adjoint of GEOS-Chem

Page 2: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Forward sensitivity

Page 3: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Adjoint sensitivity

Page 4: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Adjoint method

Depending on “model response,” can be used for:

Sensitivity analysis: quantifying influence of uncertain model parameters (emissions, reaction rates, …)

Attainment studies: assessing the effectiveness of emissions abatement

Inverse modeling: using large data sets, optimizing parameters on resolution commensurate with forward model.

Page 5: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Model Description, v6-02-05 (Bey et al., 2001; Park et al., 2004)• GEOS-3 Assimilated meteorology

• 4°x5°(Global) resolution, 30 vertical levels• HOx - NOx - HC full gas-phase chemistry• Aerosols

- Secondary inorganic- Carbonaceous (primary) aerosol- Sea salt- Dust

Forward Model: GEOS-CHEM

Gas-phase emissions

SO2, NOx, NH3

Aerosol SO4

2-, NO3-, NH4

+

Gas-phase chemistryCloud processingAerosol thermo

Page 6: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Discrete (adjoint of algorithm)KPP (Damian et al., 2002; Sandu et al., 2003; Daescu et al., 2003)

- chemistryTAMC (Giering & Kaminski, 1998) and manual

- aerosol thermo- cloud processing- convection

Manual- turbulent mixing- deposition- heterogeneous chemistry

The adjoint of GEOS-Chem: hybrid

Continuous (adjoint of equation) - advection (Vukicevic et al., 2001; Thuburn and Haine, 2001; Liu and Sandu, 2006; Hakami et al., 2006; Singh et al., 2006)

Henze et. al, 2007

Resources- CPU: tadj ~ 1.5 tfwd As || as the fwd model- HD: 45 GB for 1 month (4x5)

Page 7: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Testing the Adjoint Model: Gradient Check

Check gradient using finite difference calculation

Component-wise analysis affords domain wide points-of-comparison

cost function

control parameter

adjoint sensitivity

Page 8: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Testing the Adjoint: single processes, 1 week

(thermo only)

Page 9: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Testing the Adjoint: single processes, 1 week

(thermo only)

(chem only)

Page 10: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

GEOS-Chem Adjoint: full chemistry

Initial Conditions (all species and tracers)

Emissions sectors- NOx (lightning, anthro)- SOx (anthro, bioburn, biofuel, ships)- NH3 (anthro, bioburn, biofuel, natural)- OC/BC (anthro, bioburn, biofuel)- others are easy to add

Reaction rate constants- All reactions- gas-phase emissions (NO, ISOP, ACET, etc.)- dry deposition

Page 11: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Pollution

= non-attainment of NAAQS for PM2.5 of 15 µg/m3 (annual ave)

www.epa.gov

Page 12: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Adjoint method

Depending on “model response,” can be used for:

Sensitivity analysis: quantifying influence of uncertain model parameters (emissions, reaction rates, …)

Attainment studies: assessing the effectiveness of emissions abatement

Inverse modeling: using large data sets, optimizing parameters on resolution commensurate with forward model.

Page 13: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Attainment -- Aerosols

Define cost function ~ non-attainment for PM2.5

July 2001

NH4+ non-attainment

Page 14: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Attainment -- Aerosols

Define cost function ~ non-attainment for PM2.5

NH4+ non-attainment

Emissions (normalized) Sensitivities (normalized)

anth NH3

Responsibility Effectiveness

Benefit(Hakami et al., 2006)

Page 15: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Attainment -- Aerosols

Seasonal variability

July

Emissions Sensitivitiesanth NH3 stack SOx

April

NH3 controls effective in spring, SO2 in summer. Also consider $$ (Pinder et. al, 2007)

Page 16: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Attainment -- Aerosols

Long range transport

Emissions Sensitivities w.r.t. surface SOx

Influences concentrations, not AQ attainmentFuture emissions scenarios? Climate change? Cost?

Page 17: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Adjoint method

Depending on “model response,” can be used for:

Sensitivity analysis: quantifying influence of uncertain model parameters (emissions, reaction rates, …)

Attainment studies: assessing the effectiveness of emissions abatement

Inverse modeling (Data Assimilation): using large data sets, optimizing parameters on resolution commensurate with forward model.

Page 18: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Observed Aerosol (IMPROVE): January 2002

Observed:

NIT SO4

Model:

Diff:MOD - OBS

Page 19: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Inverse Model

Parameter Estimate

PredictionsAdjoint Forcing

Gradients(sensitivities)

Optimization

Forward Model Adjoint Model

Observations

Improved Estimate

-

t0 tf tf t0

Inverse Modeling using Adjoint Model

Page 20: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

Emissions Scaling Factors

1

10

. . .

DIFF(GC-IMPRV)

NIT

Domain wide NH3 adjustmentssimilar to inverse modelingstudy by Gilliland et al.,2006.

Optimized Anth NH3

1

10

[kg/box/s]. . .

scaling f = ln( e10/e1)

Page 21: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

y = 1.33x + 0.08

R2 = 0.77

y = 1.09x + 0.14

R2 = 0.780

1

2

3

0 1 2 3

Emissions Scaling Factors

NH4+

CASTNet

Optimized Anth NH3

1

10

[kg/box/s]. . .

scaling f = ln( e10/e1)

Page 22: Daven K. Henze with Amir Hakami and John H. Seinfeld Caltech, Chemical Engineering Support from: NSF, EPA, TeraGrid and JPL Supercomp., W. & S. Davidow.

The End

Thanks!