Minutes of the second Bredbeck workshop

============================================================================
==========                                                       ===========
==========   FORWARD MODEL WORKSHOP, Bredbeck 19-22 June, 2000   ===========
==========                                                       ===========
============================================================================

Chairmen:
---------
SB = Stefan Buehler
PE = Patrick Eriksson
JU = Joachim Urban
AE = Axel von Engeln
 
 
============================================================================

Participants:
-------------

SB  Stefan Buehler     IUP  

AE  Axel von Engeln    IUP  
PE  Patrick Eriksson   RSS  

DF  Dietrich Feist     IAP  

CJ  Carlos Jimenez     RSS  

YK  Yasuko Kasai       CRL                     
UK  Ulf Klein          IUP  
AK  Armin Kleinboehl   IUP     
GK  Gerhard Kopp       IMK                 
TK  Thomas Kuhn        IUP  
KK  Klaus Kuenzi       IUP  

UL  Ulrich Loehnert    MUB                    

JM  Jungang Miao       IUP   

SO  Satoshi Ochiai     CRL                     

UR  Uwe Raffalski      IRF                 
PR  Peter Rayer        TMO                  

BS  Birger Schimpf     DLR           
FS  Franz Schreier     DLR                
RS  Richard Siddans    RAL                   

CT  Chikako Takahashi  FUJ           
AT  Ariane Thiele      MUB                  

JU  Joachim Urban      OBX           

CV  Carmen Verdes      IUP  

IW  Ingo Wohltmann     IUP        



IUP:
Institute of Environmental Physics
University of Bremen / FB 1                
p.o. box: 330440; 28334 Bremen; Germany
fax: +421-218-4555
http://www.iup.physik.uni-bremen.de/

RSS:
Dep. Radio and Space Science
Chalmers University of Technology
412 96 Göteborg
Sweden

CRL:
Global Environment Division
Communications Research Laboratory
4-2-1 Nukuikitamachi, Koganei, Tokyo 184-8795, Japan
Fax: +81-42-327-6110
http://www.crl.go.jp

IAP:
Atmospheric Physics Group
Institute of Applied Physics, University of Bern
Sidlerstrasse 5
3012 Bern
Switzerland
Fax: +41-31-631 3765

IMK:
Institut f. Meteorologie und Klimaforschung
Forschungszentrum Karlsruhe GmbH
Postfach 3640
D-76021 Karlsruhe Germany
Fax: ++49 7247 824742

MUB:
Meteorologisches Institut Uni Bonn
Auf dem Huegel 20                 
D-53121 Bonn/Germany              
Fax    ++49 +228 735188    
ftp.meteo.uni-bonn.de
http://www.meteo.uni-bonn.de

IRF:
Swedish Institute of Space Physics
Box 812
981 28 Kiruna  Sweden
Fax:       +46 980 790 50
http://www.irf.se

TMO:
The Met. Office                       
London Road           
Bracknell, RG12 2SZ                             
United Kingdom
Fax:     +44 (0)1344 854026   
http://www.met-office.gov.uk 

DLR:
DLR - Remote Sensing Technology Institute 
Oberpfaffenhofen
D-82234 Wessling
GERMANY
Fax  +49 8153-28-1446
http://www.op.dlr.de/ne-oe/ir/                                        

RAL:
Rutherford Appleton Laboratory, 
Chilton OX11 0QX, U.K.
Fax: (44) (0)1235 445848

FUJ:
Fujitu FIP Corporation 
System Dept.   
Environmental System Business Division 
Fax  : +81-3-5531-1615 

OBX:
Observatoire de Bordeaux         
2 Rue de l'Observatoire          
B.P. 89                          
33270 Floirac                    
France                           
Fax : (++33).(0)5.57.77.61.55 



============================================================================



Monday 19.6.2000 PM (Chairman: SB)
----------------------------------

o Introduction and aims of the workshop (Chairman: SB, 10 min)

o Discussion of workshop program (Chairman: SB, 10 min)

o Workshop notes and proceedings (Chairman: SB, 5 min)

o Overview of forward models not presented last year (Chairman: PE, 2.5 h)
  Presentations:
  --------------
   1)  MIRART (FS, 25 min)
       MIRART = Modules for IR Atmospheric Radiance and Transmission
       Applications: - THOMAS (airborne FIR heterodyne spect. (strat. OH)
                     - BRUKER high resolution FTS
                     - PIRAMMYD (limb sounding with Fabry-Perot-Interferometer,
                                FTS, heterodyne)
                     - AERO-X (FTS meas. of jet engine exhausts)
                     - FOCUS (space borne (ISS) high temp. events)
       model: line-by-line (lbl) 
       catalogs: HITRAN, JPL, HITEMP
       line shape functions: Voigt (Humlicek algorithm), Lorentz, VVH*Doppler
       continuum: CKD, Liebe, FIRS (empirical)
       FOV: pencil, Box, Gauss (for limb, nadir, zenith)
       instrument response: sinc, box, Gauss  
       language: F77 with F90 extensions
       math libs: SLATEC Common Math. Lib. (www.netlib.org), BLAS
       data format: netCDF interface
       flow control: "Namelist"
       grid: fine freq. grid at line centers and wider grid in wings
       num. quadrature (integration): trapez, overlapping parabolas, spline
       intercomparison: in PIRAMHYD study with IROE, RAL, SRON
       additional WEB interface is under study
       future: - AMiL2DA MIPAS, HGF
               - aerosols, scattering
               - Voigt optimization, refraction, non-LTE
               - code improvement to enlarge spectral range
               - conversion to F90 and/or steering with Python

       Questions:
        - PE: num. quad. rule for temperature?
          FS: explanation of the integration way.
        - SB: HITEMP?
          FS: HITEMP is HITRAN for higher temperatures: H2O, CO2, CO
        - PR: HITRAN recommended for IR?
          FS: some wrong O2 isotope data in HITRAN96 in FIR region
              but otherwise fine.

   2)  Karlsruhe (GK, 5 min)
       - hardware:
         base: ground based
         F range: 268-280 GHz
         spec. resolution: 1.2 MHz
         Spectrometer: AOS
       - FM:
	 author: M. Kuntz
	 language: F77
	 cat: HITRAN96, Voigt, VVW line shapes
         cont: Liebe
         model: lbl
         refraction: included
         layers: Curtis-Godson means
         sideband filter: sinusoidal       
         not incl: Dopper broadening, pressure shift, scattering, 
                   antenna pattern
         inversion: semi-analytical calc. for homog. thin layers
                    Thikonov-Phillips regularization or optimal estimation
	 
	 Questions:
         - JU: what is Curtis-Godson means?
         - PR: elevation angles possible?
           GK: adjustable
         - PE: retrieval with TP regularization how it works?
           GK: each constituent has its own regularization parameter
         - UL: any climatology needed for retrieval?
           GK: we use standard profiles
	 - JU: any progress in limb sounding?
!          GK: no, KOPRA for MIPAS can perform limb sounding

   3)  MPM89/92 and RTTOV-6 (Peter Rayer, R. Saunders, 15 min)
       model: (a) lbl (absorption archive)
                  water vapor (44 lines), oxygen (30 lines) 
                  and continuum term (O2, N2, H2O)
                  based on Liebe model MPM89/92
              (b) RTTOV-6, fast model for transmittance coeff.
	          satellites of interest: TRMM , METEOSAT, GOES, NOAA 
		  because of the weather prediction purpose, no 
                  attempt is made to retrieve vmr of trace gases 
                  except H2O and ozone
       FOV: nadir
       thermodynamics: LT equilibrium assumed
       the model is maintained in the next years for NWP
       the code is available since Feb 2000

       Questions:
       - CJ: linear terms in the absorption archive accurate enough?
         PR: the archive includes also some predictors to minimize errors.
       - SB: where is the uncertainty larger lbl or fast model?
             in the lbl model except for some channels.
       - JU: AMSU-B instrument?
	 PR: there are 60 GHz oxygen band channels etc.
             1013 to 0.1 hPa is the altitude region.
             the measurement is the input of the transmittance models
             and this output is going into the NWP together with other
             information. There is no attempt to do separate retrievals
             with this data and model.
       - SB: Cirrus clouds?
         PR: no information for this because others do this job.

   4)  MWMOD (UL, AT, 50min)
       MicroWave MODel from Simmer (1994)
       spec region: 1 to 1000 GHz
       dimensions: one dim. model and azimuthally isotropic
       scattering: aspherical droplets are implemented from Czekala (1999)
       products: TB, scattering, clouds, etc.
       MWMOD is separated in three modules: 
           (a) "ENVGEN": for the setting of the atmospheric and 
               boundary environment like e.g.:
	       p,T, humidity profiles
               clouds, ice, rain  content
               wind speed, ocean temperature
           (b) "IAPGEN": specifications of FM parameters e.g.:
               freq. and FOV, phase function with Lorentz-Mie theory
               ocean/ice surface specification
           (c) "RADTRA": radiative transfer calculations 
               with the rad. trans. equation incl. all the
               selected topics of the forward model setting
       the analysis of clouds and liquid water from profiles is with 
       this model possible.

       MWMOD with non-spherical particles especially for particles 
       above 1mm size necessary. Cross polarization is remarkably
       in this region. T-matrix method is used instead of Lorentz-Mie
       theory for spherical particles. See Master Thesis of AT, 
       Univ. of Bonn, for more and detailed information of modeling and 
       retrieval of water liquid paths from rain and clouds.
       Remarkably is the difference in polarization if one looks from the 
       space to the ground or from the ground to space.

       Questions:
       - JM: way is the polarization difference for nadir and zenith 
             direction?
         AT: why this difference exists is not known jet. This is a
             subject of the Master Thesis to investigate.
         UL: there is a validation campaign this summer to validate
             the model with measurements.
       - JM: 3. and 4. Stokes parameter can be calculated with MWMOD?
         AT: no
       - PR: Liebe MPM9* model can calculate transmittance above 1
             caused by adjustments of parameters by Liebe, Hufford 
             and Rosenkranz. So be carful by using MPM9* models!
       - PE: why only positive polarization differences for spherical
             particles?
         UL: perhaps due to assumed T profiles.
       - SB: why goes the polarization difference so negative at 
             120 zenith angle degree?
         UL: Czekala assumed special conditions for this calculations.
             This is not a general feature.

o Updates of forward models since last year (Chairman: PE, 40 min)
  Presentations:
   JEM/SMILES (SO and CT, YK, 15 min)
   hardware:
   instrument must be shield from electromagnetic interferences from ISS
   FOV resolution: 3.8 km
   bands: LSB 624.32-625.52, 625.12-626.32
          DSB ?
   launch: 2006(?)
   software:
   MAES = Millimeter-wave Atmospheric Emission Simulator
          for ground based, balloon based and space born platforms
   FM1: rad. transfer in the atmosphere
        cat: HITRAN/JPL and measurements with Lorentz and Voigt line shape
        cont: Liebe MPM89
        refraction implemented
        FOV direction: limb
        contributions from antenna side-lob directions to the signal 
        are possible to calculate
   FM2: instrument part
        included topics: Doppler shift, beam efficency, calib. loads, 
        optical loss, standing waves, SSB filter, noise, gain, 
        freq. conversion, AOS, power to temperature conversion, 
        freq. calibration,
   IM:  inversion model
        optimal estimation method, OEM, for retrieval with a priori 
        information 

   Questions:
   - GK: from where are the line broadening parameters?
     YK: from HITRAN and own measurements
   - PE: non-linearities of the instruments from where?
     SO: estimation of 1 per cent non-linearity, mostly from the
         optical system
   - GK: how to treat standing waves in the inversion model?
     SO: -

   Moliere-4 (JU, 10 min)
   Moliere originally for the Odin instrument developed
   special version of Moliere for real data processing, which
   is different of the main-stream development of Moliere for
   model calculations.
   cat: HITRAN/JPL
   cont: Liebe MPM93, CKD-89
   refraction included
   weighting function calculation is done analytically
   language: F90
   flow control: ascii input file
   differences in the intercomparison with FM of iup-Bremen not yet
   solved completely.



============================================================================



Tuesday 20.6.2000 AM (Chairman: PE)
-----------------------------------

Presentation of ARTS
--------------------

o General introduction (SB)
  - increasing time to change already existing FM is not efficient
  - open source software of ARTS for wider development of the model
  - present status of the code: not ready for operational data analysis 
  - code in modern C++ with GNU tools and compiler
  - one has to be aware of the memory consumption during the calculations
  - to be independent, no existing higher language like idl or matlab
    is needed to run ARTS

o Concept of ARTS (SB)
  - distinction between workspace variables and workspace methods;
    workspace methods have workspace variables and control 
    parameters as input
  - also generic workspace methods can be created
  - flow control is performed with an ascii input file which will be 
    interpreted with the help of the ARTS parser (internal lookup tables)

o Absorption (SB)
  - not really finished this work, so only a simplified version 
    can be demonstrated
  - the catalog format is different from other models
  - SI units of the catalog entries are used 
  - flexibility in handling of the records is the aim of the catalog
  - tag groups are also implemented. With this one can distinguish 
    distinct lines of one species for different treatment (e.g. weighting 
    functions)
  - the absorption can be calculated per species with the help of the 
    tag groups
  - every tag group will have its one line list

o Rad. transfer (PE)
  - line of sight (LOS) 1D, elevation angle (angle with respect to zenith 
    direction) and platform altitude are the parameters to define limb LOS
  - LOS are divided into equally distant steps or calculations
  - T and absorption are linearly interpolated

o Sensor modeling (PE)
  - FM is divided into two parts, one for the atmospheric part (FM1) 
    and one for the sensor part (FM2). FM1 is the input of FM2
  - the sensor is treated as matrix multiplications, for each sensor part
    a single matrix is build. Nonlinearities are therefore not implemented
  - data reduction is also considered (binning, eigenvector decomposition)
  - weighting functions (WF) are calculated analytically where possible and
    numerically else
  - decomposition of the WF into appropriate parts which are easy to
    handle in the calculations (constant part, vertical variations, 
    changes with changing state vector)
  - hydrostatic equilibrium is used for WF of the temperature
  - the vertical grid is expressed in pressure not in km.
  - because of geometrical reasons, the correspondence of pressure and 
    altitude is internally used in ARTS for limb sounding
  - the spherically symmetric atmosphere which is assumed at the moment
    is not severe problem if one would like to include 2D modeling
  - at the moment it is not definitely decided where all the different 
    measurement errors should enter the forward model

o practical demonstration of ARTS (SB)


Presentation of BEAM
--------------------

o BEAM, Bernese Atmospheric Model (D. Feist)
  - general purpose model which is also optimized for fast calculations
    for certain conditions with numerical optimization
  - the accuracy of the optimized calculations is very good
  - spectral line catalog BAMCAT especially developed for BEAM
  - the interpolation in the predefined frequency grid is done with
    splines



============================================================================



Wednesday 21.6.2000 AM (Chairman: JU)
-------------------------------------

small group work about the following topics:

a) Spectroscopic matters
   (SB, YK, UR, GK)
   - with new measurements one can compare data and catalog entries
     for crucial parameters
   - it should be explicitely stated if the data in a catalog is 
     calculated or measured
   - catalog format as it is stated in the proceedings of the first
     Bredbeck workshop is in principal good
   - difficulties to merge catalog data into one operational list 
     for the FM. UL, PE, DF, SB have done this in different ways 
     for their FM. Perhaps a Japanese groups will do it also for SMILES
   - another point is how new information should/will be integrated
     into existing structures
   - DF has a well worked out algorithm how to merge different 
     catalogs and to put the information into a flexible structure.
     This module will be public in near future
   - DF propose that it should be possible to overcome the old
     computer capacity restricted format which is now the state.
     It should be possible to establish a general data base for
     spectral data which will contain every information possible.
     The user should then select himself which data set with which
     reliability level and which format he will select/take. This
     can be done via Internet.
     All participants agree and encourage DF to set up such a data
     base
  

b) Extension to IR
   (PE, FS, BS, JU)
   - motivation:
     it is not a big step to enlarge the freq. range to IR
   - applications:
     ground based FTIR
     space based limb sounder (MIPAS)
     space based nadir sounder (IASI)
     other balloon, airborne based instruments
     airplane exhaust investigations
   - physics problems:
     scattering, non-LTE, spec. (CO, CO2).
     continua, dry air, H2O: FIR uncertainties.
     refraction is also a matter of investigation.
     line strengths information is a question (vibrational states).
   - computational problem:
     narrow freq. grids which leads bulk of frequency based calculations
     Therefore an intelligent frequency grid should be used.
     Line shape function calculated by decomposition of Lorentz in subfunctions
     by Clough and Kneizys in Applied Optics July 1979 is a fast method
     D. Edwards, 1991 has proposed another algorithm which is used 
     in GENLN2.
     MIPAS retrieval uses a microwindow approach for fast calculations
   - sensor:
     FOV coherent detectors : Gauss
         incoherent detectors : box function
   - FM code:
     + University of Oxford is writing a detailed FM for MIPAS. 
     + KOPRA is a general FM which is used for MIPAS from the the
       Univ. of Karlsruhe.
     + It is not really clear if FASCODE is still maintained or not.
       In the proceedings a list of references will be given to
       different models and intercomparisons.

   
c) 2D forward model
   (RS, AE, CV, CJ, SO, CT)
   - RS is presenting the 2D results of the ESTEC MASTER extension study.
     Detailed information about this study can be given by RS and 
     SB. The final report will soon be available (2 months)


d) Scattering
   (UL, JM, AK, AT)
   - when?
     when the wave length is of the order of the scatterer size the
     scattering is of importance.
     clouds: 100 micrometer
     rain  : 500 micrometer
     frequency 3    GHz correspondence to 100 mm in wave length
     frequency 1000 GHz correspondence to 0.3 mm in wave length
   - how?
     + for spheres Mie theory is applicable:
       Mie parameter is chi = 2 * pi * r / lambda 
       if chi is smaller than 1, Rayleigh scattering can be used
       if chi is larger than 1, Mie scattering can be used
     + for non-spherical particles:
       axial symmetry: quasi analytical treatment (T-matrix method)
       arbitrary form: volume integration method (DDA)
   - radiative transfer:
     extension of the simple Beer-Lambert equation for scattering
     terms (extinction and source term are different from
     non-scattering)
   - solving methods:
     successive order of scattering, Monte Carlo, doubling adding (Evans)
   - implementation:
     integration into ARTS is difficult because of the different 
     way of processing the radiative transfer.


e) Application to meteorology/assimilation
   (PR, IW, DF, TK)
   - frequency band of interest for the ,met. operation
     20 to 200 GHz approximately for T, humidity, total column of H2O
   - nadir sounding is the normal way
   - only LTE is considered
   - Operation requirements:
     + fast calculations is the dominant requirement
     + if this can be done lbl calculation is an open question. At 
       the moment an approximation and pre-calculations are taken into
       the FM of transmittance calculations
     + an adjoint FM module will calculate the Jacobian matrix numerically
       If an analytic way of calculation is comparable in speed
       is not clear. 
     + The NWP standard is to divide the FM into 
       two parts: one which calculates fast the rad. transfer
       and one part for fast Jacobian matrix calculation 
       (best estimate of the atmospheric state vector similar to OEM)
   - ground emissivity is still an open question
   - additional trace gases which are not included into the FM at 
     the present can give a systematic effect which can be treated
     by including these gases into the FM 
   - scattering is another matter which should be concerned if 
     the speed is not going down of the FM
   - Future applications:
     + models must cope with the NWP speed and the increasing amount of 
       data from satellites
     + including trace gases for met. products (e.g. ozone)
     + models have to be comparable with the accuracy of the future
       sensors and the improved NWP models.
     + NWP people has to be convinced to take satellite data into
       account for their calculations. It is not a priori given that 
       the NWP people will use this essential input in future.
       This is completely different from the atmospheric science
       community



============================================================================



Wednesday, 21.06.2000 PM  
------------------------

  Retrieval issues: (Chairman: AE, 2.5 h)
  Existing strategies and approaches

  Presentations:
  o Tikhonov retrieval code (BS, 25 min)
    - regularization with quadratic constraints
    - additional qualitative information comes from the constraints
      of the smoothness of the solution
    - numerical stability with generalized singular value
      decomposition (GSVD)
    - the determination of the regularization parameter lambda comes
      from the L-curve criterion
    - current retrieval code is a stand alone package in F90
    - in future Python will be used for the scripting language
    - ground based FT measurement are processed with this retrieval
      package
    - future: additional constraints: profiles larger than zero
              automation of simultaneous retrieval of several profiles
              comparability of different retrieval methods
              application to data of FOCUS, AERO-X, MIPAS, SCIAMMACY, GOME
    - questions:
      SB: what is the operator for positiveness constraints
      BS: current search in the math. literature for solutions,
          but the implementation will be not straightforward
      PE: log retrieval for positiveness constraints
      BS: with this approach the problem will be then non-linear
          which makes it difficult
      GK: simultaneous retrieval: how to determine lambda then
      BS: relative values from a priori because n-dimensional
          L-curve solutions are difficult to find.
          Mathematicians will help to solve this


  o Neural nets (nn), (CJ, 20 min)
    - introduction into neuronal nets and its structures
    - the setup of a neuronal net for a retrieval problem is basically
      to be found by trial and error. There is no fixed structure
    - input: spectra, output: profiles
    - advantage of neuronal nets: 
      is the handling of non-linear problems
      computational cost is lower than with OEM
    - model calculations performed for Odin
    - the retrieval results are very similar for both methods for O3, ClO
    - also averaging kernels and errors can be calculated with nn
      in the case of nn, the AK oscillate stronger around zero than
      OEM. The cause of this is not understood now
    - Discussion:
      AE: how many training sets?
      CJ: 500 training sets are used
      AE: different nn for different latitude and seasons ranges?
      CJ: yes
      UL: additional information input of the nn?
      CJ: not yet but in the future like platform high etc
      DF: how is the nn implemented?
      CJ: with matlab today but different in future
      RS: how is the AK calculated?
      CJ: contribution function with nn and afterwards build up with 
          Jacobian to AK. 
      JM: what about noise in the nn, will the nn recognize noise?
      CJ: it is in principle no general problem
      BS: what will be if an actual spectra is out of the training
          set variability?
      CJ: there will be of course problems in this case. Therefore
          the training set including the FM should be very carefully 
          selected


  o Interfaces between forward model and retrieval algorithm
    Presentations:
    Urd (PE, 10 min)
    - interface to Sculd and retrieval program
    - four different variable levels
    - correlation functions are modeled Gaussian, exponential or 
      tenth function 
    - Cholesky decomposition method used
    - Discussion:
      PE: be aware of non-linearities in the calculation of the
          retrieval
      FS: Python (www.python.org)
          - full object oriented free for use scripting language
          - the program will be interpreted by the interpreter
          - intensive calculations should be done in FORTRAN, C, C++
            because of its slow speed
      PE: set up of the retrieval with matlab
      RS: set up of the retrieval with idl


  o Towards the ultimate forward model: (Chairman: SB, 1 h)
    Possible discussion topics:
    Demands:
     - ARTS should be on a level of a reference model
       + how far can be the approximation ?
     - on which level should the scattering implemented?
       + application:
         AMSU-B, MASTER, up-looking mm instruments for H2O column 
       + basic scattering model for the first step of implementation:
         spherical particles, 3D
       + for what purpose should the scattering included?
         cirrus cloud detection/correction, etc.
         This determines the complexity of the scattering model
         useful levels of scattering model:
         . 4 Stokes 3D (full complexity)
         . sources for phase functions from Uni Bonn 
         . MASTER: 2D important,   polarization unimportant
         . AMSU-B: 2D unimportant, polarization important
     - instrumental characteristics implementation
       + no derivatives with respect of instrumental parameters
       + at the moment matlab is used to setup the matrix H (PE)
       + H must be inside ARTS in future
       + it has to be considered that e.g. instrumental modules 
         are divided into several components (e.g. several AOS in 
         parallel)
       + frequency shift should be possible to implement into the 
         instrument model
     - practical demands: 
       + (PR) by calculating the absorption coefficients:
         mixing table approach (see GENLN2) for different cases 
         simultaneously
       + (DF) will the line selection be done?
         . cooperation between Uni Berne and Uni Bremen is suggested
           for the line data base (DF, SB)
         . line selection is to some extend included in ARTS in future
           (SB)	 
         . ARTS should be intelligent to select only the really
	   necessary lines for the user defined accuracy (JU) 
         . in BEAM the line selection is automatically done by the
           forward model itself, but it is not recommended (DF)
         . (SB) suggest that it should be possible in ARTS to write 
           out the lines which are selected by a single FM run to use 
           this line catalog in other FM runs as well
         . (BS) ARTS should provide some intermediate results for the
           user
       + (DF) which continuum models will be included?
         (SB) MPMXX, Rosenkranz, CKD
     - atmospheric setup
     - surface emissivity

    2D setup
     - interpolation (RS RAL uses bi-linear approach for MASTER 
       extension study)

    Speed / accuracy:
     - including scattering should not slow down the FM in 
       a severe manner 
     - parallelization of the calculation on a small scale should be
       easily possible (e.g. split up in respect of frequencies)
     - use parameterized calculations if possible (e.g. lookup tables)

    frequency interpolation:
     - (DF) reduction in CPU time can be achieved by having a fine 
       grid at line centers and a wide grid in regions where no spectral
       feature is
     - (FS) the frequency grid is for every altitude level separately
       chosen in FASCODE, GENLN2 (valid for FASCODE, for GENLN2 I don´t know
       it but I can look it up if necessary). The Voigt line shape
       function is necessarily approximated. In FASCODE, the Lorentz
       line shape function is decomposed but the sum of the parts
       gives at the end the original Lorentz function.
       Suggestion:
       In FASCODE the frequency grid is set up appropriately for each layer
       and the transmission/radiance given on a coarse grid (lower altitudes)
       is interpolated before combined with transmissions/radiances
       given on a fine grid (higher altitudes).
       GENLN2 uses a fine mesh (grid) for line center contributions and a
       wide mesh for line wings.
       FASCODE uses a decomposition of the Lorentz function in fast, medium and
       slowly varying subfunctions (with the sum still being a Lorentz)
       and accumulates these different contributions line by line. 
       The contributions are combined AFTER all lines have been calculated
       using appropriate interpolation.
     - ARTS should not be restricted in the choice of the line shape 
       function. By investigating influences of the line shape an 
       approximation will be a severe drawback
     - the absorption output will be on a common frequency grid for all 
       altitudes in ARTS at least at the beginning (PE)

    Discretization:
     - how to interpolate in general?
       + log interpolation seems superior than linear (DF)
       + there should be a single interpolation scheme for one variable
         over the whole interval of interest (PE) 
       + atmospheric variables like altitude, temperature, VMR,
         absorption linear sufficient in log(pressure) (SB)
       + frequency more sophisticated interpolation schemes make 
         sense (RS)
       + BS suggest that the interpolation schemes should be
         hidden in an object oriented (OO) code so that the user 
         can choose
       + (BS) ARTS should provide the information of the interpolation
         schemes used
       
    Distribution stuff:
     - distribution can be taken from the Internet
     - distribution includes all the code and all the documentation
     - a user and a developer version will be available.
       The developer version will have a version control

    Control mechanisms:
     - (DF) at the moment BEAM will work from matlab, idl
       + input/output is given in the matlab environment to 
         BEAM
     - for retrieval it is needed to perform loops of the FM,
         therefore the control file mechanism must support this
     - PE has a similar approach for Skuld as DF for BEAM
     - with respect to a completely separated retrieval program, 
       input/output should be as ascii files.
     - The FM should provide as much information directly to the
       retrieval algorithm as possible
     - communication between FM retrieval algorithm is a difficult 
       task. BS uses python to build the interface of the FM and 
       the core of the retrieval program
     - it is a matter of taste how strong the FM should be linked /
       should depend on the retrieval program. It is suggested to 
       have the FM as independent as possible from specific retrieval
       programs (FS, BS, JU).
 


============================================================================ 

 

Minutes Thursday 22.06.2000 AM
------------------------------

o (FS) Web interface at DLR for FASCODE3 (and others)
  - the input control file is very hard to understand and manipulate
    for users not really familiar with FASCODE
  - FASCODE uses many I/O files 
  - PFUI = Python FASCODE3 User Interface
    the interface is written in Python
  - introduction page is standard for all applications and 
    the following pages are dynamically build up for the 
    specific input
  - WIMP = Web Interface to MODTRAN using Phyton
  - it is planned to have a similar interface for MIRART
  - the Python code for PFUI is several thousands of lines long
  - Discussion:
    PE: is it open for all?
    FS: in principle yes, but a password is to  be given
    AE: can different calculations be performed simultaneously?
    FS: yes, for every user a unique directory is temporarily created


o Towards the ultimate forward model, cont. (Chairman: PE, 2.5 h)
  - partition functions
    + for ARTS SB has asked Bob Gamache for his program "tips".
      Documentation will be available soon.
      It will be perhaps used as a black box in ARTS.
  

o Best algorithms for the different parts:
  - Presentation: van Vleck & Huber (PR, 15 min)
    + this is the line shape on which Clough et al base the CKD model
    + just for thermal equilibrium (LTE) developed
    + negative frequency expressions can be understood as processes
      in negative time direction (going backwards in time to the past)
    + the abs. coeff. satisfies the demand of being an even function of
      frequency
    + It also fulfills the generalized Nyquist theorem
    + the fluctuation-dissipation theorem is also fulfilled
    + in the microwave region the VVH will reduce to the VVW line
      shape function
    + in the IR the VVH will be in a form which is also used 
      by the HITRAN catalog for stating the line intensity
    + PR is writing a book about this topic
    + Discussion
      PR: the oxygen coupling derived of Rosenkranz is also based on VVH
      SB: how to convolute VVH and Doppler line broadening function?
      PR: the GENLN2 is also a source for looking how to do the
          convolution of the line shape function with the Doppler 
          line broadening function
      IW: where to put the frequency shift?
      PR: frequency shift will only be formulated in the line shape 
          function because the shift is caused by the pressure 
          broadening so that this is the only proper place to 
          introduce it in the formulas
      PR: MPM93 has played at the line strength and line coupling 
          parameters to fit the data better. But this leads to
          unphysical results if one is looking to the absorption or
          transmittance of oxygen. So one has to be careful by using
          MPM93 for this particular calculations
      PR: eventually the Lorentzian line shape function gives too
          much absorption in the far wing of H2O

  - Presentation: MAES WF improvement (SO, 15 min)
    + analytical WF of limb sounding with finite antenna pattern:
      WF is first calculated for pencil beam antenna and after that
      integration of the WFs with the antenna pattern 
    + the limb path is divided into five different parts along the LOS
      and the WF into three parts which add up together to
      the final one
    + estimation of number of steps required to calculate WF for a 
      typical case: 1.6 10^6
    + this way of calculation has the possibility to reduce the 
      calculational burden in some simple cases
    + Problems:
      complicated calculations, use of approximations, hardly to extend
      to 2D cases
    + Discussion:
      PE: the thinking for ARTS goes in some details in the same 
          direction (e.g. "LOS-WF")

  - continuum absorption
    + MPMXX are empirically derived so that the frequency range for its
      use is limited below 1000 GHz
    + dry air continuum is a point for itself. It is not really clear
      from which species and which physics it is coming from in some
      aspects. On which level it is really accurate is an open question
    + (SB) Bremen is using Rosenkranz formulation in FM
    + (PE) is it possible to treat all different continuum models
           in the same way in the FM?
      (SB) not really it is only possible by parameterizing some
           models (e.g. CKD) and loosing the original form and with it 
           some features
    + (SB) it is perhaps the best to write for every model a separate
           workspace method.
      (PE) then one runs into problems with the different line
           catalogs which are assumed in different models
    + (JU) which model is the recommended?
      (SB) difficult to say, but Rosenkranz seems to be acceptable
    + (PE) there was an error in the description in MPM92
           at a freq. of around 500 GHz with an specific isotopic 
           O2 line. Fore details ask P. Baron, Observatoire de Bordeaux 
    + (PR) there are several typos in the papers of Liebe, so be 
           careful
    + (PR) in MPM93 the definition of the line strength changed, so is
           it compatible with the way ARTS will treat the continuum?
      (SB) it will be clearly documented how these things are treated
           in ARTS

  - surface emissivity
      PE: is there a simple model/emissivity values which can be
          used?
      KK: the emissivity of the ocean surfaces is between .5 
          (plain surface) and 1 (foam)
      PR: S. English has also specific knowledge about this topic
      KK: The participants of the specific COST working group also
          could give some information 

  - refractive index (ri)
      SB: frequency dependence of the ri should be known for 
          wide band calculations
      PE: influence of the solar light is notable in IR region
           for some geometries. CJ has done some work on this

  - sparse matrices
      PE: TNT is used for ARTS at the moment but it is not so promising
      RS: lapack is used at RAL
      AK: C++ package DIFPACK used
      FS: NETLIB, C++, has also such features www.netlib.org


o Open questions since last workshop (Chairman: SB, 20 min)
    Are they solved?
  - spectroscopy:
    + ARTS line catalog content and format, solved (Proc. 1. Bredbeck WS)
    + partition functions, solved (TIPS of Bob Gamache)
  - radiative transfer:
    + interpolation solved but needs more discussions (linear)
  - WF:
    + transform from LOS grid to retrieval grid, solved (PE, ARTS user
      guide)
  - sensor treatment:
    + H-matrix approach for linear effects, solved but further 
      discussions needed (PE, ARTS development status)


o Summary (Chairman: SB, 20 min)
  - (PE) discussion who is contributing what to the proceedings
    ***********************************************************
    + spectroscopy: YK
    + FM: RS (2d), DF (BEAM), SO (MEAN),  FS (IR), PR (met app)
    + scattering: JM
    + retrieval interface: RS, PE
    + retrieval: CJ
    + line shape functions: PR
    + optimization: DF
    + Web interface: FS
    + ARTS: SB, PE
    + layout will be distributed by PE, postscript or PDF format

  - Thank you for coming!

============================================================================
minutes written by Thomas Kuhn, iup, 29.06.2000
============================================================================