Adjoint clinic Introduction for new users

Transcription

1 Adjoint clinic Introduction for new users Daven Henze (U. Colorado) Monika Kopacz (Harvard) Kumaresh Singh (Virginia Tech.) April 9, 2009

2 Agenda What is an adjoint good for? and what it is not good for (Kopacz) What has been done with an adjoint so far (quick 1 sentence summary of all projects) (Kopacz) Overview of obtaining code, users guides, and benchmark simulations (Henze) Model setups (parameters, operators, active variables) - for a PM2.5 sensitivity problem (Henze) - for a CO optimization problem (Kopacz) - for a v7 O3 optimization problem (Singh) Soliciting feedback from new users - potential future applications? - what capabilities will need to be added / updated / improved for your application?

3 What is an adjoint (technically) Adjoint = forward + reverse run (~2.5 x GEOS-Chem time) Adjoint inversion = and iterations if performing optimization, each run at the full resolution of GEOS-Chem for the whole period of time

4 What is an adjoint good and not good for? ie. what efficiency do we gain by using an adjoint Sensitivity of 1 parameter (scalar) to many parameters (vector, matrix), e.g. surface station: adjoint is most efficient 1. Sensitivity studies Sensitivity of many parameters to 1 or many parameters (computing full Jacobian): adjoint not more efficient Adjoint Model (receptor-oriented) Forward Model (source-oriented) Concentration at the receptor t n Forward GC Perturbation at source region t 0 adjoint area of possible origin t 0 Changes of concentratio n t n

5 What is an adjoint good and not good for? ie. what efficiency do we gain by using an adjoint Many observations and high resolution estimates possible (eg. CO sources) use an adjoint 2. Parameter estimation Few observations, can only constrain a few regions (e.g. mercury sources) adjoint not efficient satellite obs. A posteriori emissions * Except for calculating gradients for analytical inversion (see Kopacz et al. [2009], Kaminski et al. 1999)

6 Limitations and cautions Nonlinearities (cf. Henze s talk) Sensitivity!= Source apportionment Sensitivities: cheap, inversion: expensive Each application requires extra code development Memory intensive Everything running together (observational operators etc.)

7 What has GEOS-Chem adjoint been applied to so far? Aerosol source estimation (Henze) CO source estimation: regional and global, using MOPITT, AIRS, SCIAMACHY (Kopacz) Sensitivity of Air Quality attainment metric to ozone precursors (Henze) Chemical sensitivity of ozone concentrations to precursor emissions (Singh) Ozone transport/production sensitivity study (Zhang) NOx source estimates using SCIAMACHY data (Shim) Sensitivity of Arctic O 3 to precursor emissions (Walker)

8 Obtaining code Model setup: PM2.5 sensitivity

9 Obtaining the adjoint codes Step 1: Sign up on the adjoint group mailing list: geos chem stay informed of bug patches new releases (many coming out soon after this meeting) Step 2: Peruse the adjoint wiki: Chem_Adjoint find out which adjoint version you need read the manual (linked from the wiki page) Step 3: Get code from CVS server for an account CVS server is moving from Caltech to CU Boulder very soon

10 Obtaining the adjoint codes...step 3: CVS is good for: tracking changes you ve made to the code upgrading to changes we ve made to the code e.g., adding 2x2.5 support: 106,107c102 < REAL*4, ALLOCATABLE :: QC_SO2_CHEMT(:,:,:,:) < REAL*4, ALLOCATABLE :: QC_SO2_DYNT(:,:,:,:) > REAL*4, ALLOCATABLE :: QC_SO2_CHK(:,:,:,:) 158a154 >! (11) Now completely split dynamic from chemical time step checkpoints! (dkh, 02/01/09) 354,365c350,362

11 Caveats about the code Code is distributed following GEOS Chem grass roots policy open access get code from the source share what you ve done too give proper credit

12 Running the code Benchmark simulations following links from wiki, download benchmark inputs / outputs try it on your computer! Supporting files

13 Adjoint model setup Define your active variables: (CMN_ADJ)!============================================================! **** ACTIVE VARIABLE SELECTION ****! Set type of inverse problem to solve. Only uncomment one of! the following sections.!============================================================! INITIAL CONDITIONS! CHARACTER(LEN=10) :: ACTIVE_VARS = 'TRACERS'! INTEGER, PARAMETER :: NOPT = IIPAR * JJPAR * LLPAR * NADJ! ! EMISSION SCALE FACTORS! CHARACTER(LEN=10) :: ACTIVE_VARS = 'EMISSIONS'! INTEGER, PARAMETER :: NOPT = IIPAR * JJPAR * MMSCL * NNEMS! ! FINITE DIFFERENCE TEST CHARACTER(LEN=10) :: ACTIVE_VARS = 'FDTEST' INTEGER, PARAMETER :: NOPT = IIPAR * JJPAR * MMSCL * NNEMS

14 Adjoint model setup Pick a type of cost function (cpp directives in CMN_ADJ):!#define JACOBIAN 'JACOBIAN' #define PM_ATTAINMENT 'PM_ATTAINMENT'!#define O3_ATTAINMENT 'O3_ATTAINMENT'!#define NO2_SAT_OBS 'NO2_SAT_OBS'!#define IMPROVE_OBS 'IMPROVE_OBS'!#define CASTNET_OBS 'CASTNET_OBS and some other options: #define LOG_OPT 'LOG_OPT'!#define FD_GLOB 'FD_GLOB'

15 Adjoint model setup Set the number of iterations and folder locations in the run script: # Set the start (or current ) iteration number X=1 # Set the stopping iteration number XSTOP=1 # Give every run a unique name (default is $PBS_JOBID) RNAME=ADJv29d_f04C0 Set some classic geos chem input files (input.geos, input.ctm) TIMING VARIABLES--4 x 5rRUN NYMDb NHMSb NYMDe NHMSe NDT NTDT NDIAGTIME GEOS-CHEM FLAGS T T T T F LEMIS LDRYD LCHEM LTRAN LTPFV T T T F T LTURB LCONV LWETD LDBUG LMONOT F T T F F LWAIT LBBSEA LUNZIP LSVGLB LTOMSAI F T F F T LMFCT LFILL LSTDRUN LDEAD LSHIPSO2 T T T T F LSULF LCARB LDUST LSSALT LATEQ

16 Results in the gctm.gdt.01 file: IDL> gamap, file='gctm.gdt.01' CATEGORY ILUN TRCNAME TRC UNIT TAU0(DATE) DIMENSIONS 1 : IJ GDE $ 21 djdsox none ( ) : IJ GDE $ 21 djdsox none ( ) : IJ GDE $ 21 djdso2s 9803 none ( ) : IJ GDE $ 21 djdso2b 9804 none ( ) : IJ GDE $ 21 djdso2b 9805 none ( ) : IJ GDE $ 21 djdnh3b 9806 none ( ) : IJ GDE $ 21 djdnh3b 9807 none ( ) : IJ GDE $ 21 djdnh3a 9808 none ( ) : IJ GDE $ 21 djdnh3n 9809 none ( ) note: you will need the tracerinfo.dat and diaginfo.dat files provided in the tools package

17 Developing new capabilities Overall, adjoint models calculate gradients For your application, what is J? Does GEOS Chem alreadly calculate J? What is p? Name the GEOS Chem specifically variable. e.g., I want sensitivity w.r.t NOx emissions. OK. which NOx emissions? REMIS RRATE GEMISNOX2 NOXTOT EMIS_BL EMISRRN...

18 Adding your own code So you added a new (tracer, process, emission, reaction) to the adjoint. GREAT! But did you check the adjoint with finite difference comparison? try your inverse model with pseudo observations?

19 Validating code (IDL/matlab scripts provided)