Hello Simon Pfreundschuh,
Thank you for your reply.
I have employed both of the suggestions and still get the irregular errors.
From the same controlfile modified in the previous email, I have included and
changed to the following lines:
OEM(method="lm",
max_iter=15,
display_progress=1,
lm_ga_settings=[100,5,2,1000,1,99])
and
AgendaSet( inversion_iterate_agenda ){
xClip(ijq = 0, limit_low = 150.0, limit_high = 350)
Ignore(inversion_iteration_counter)
# Map x to ARTS' variables
x2artsStandard
# To be safe, rerun checks dealing with the atmosphere
atmfields_checkedCalc
atmgeom_checkedCalc
# Calculate yf and Jacobian matching x.
yCalc( y=yf )
# Add baseline term
VectorAddVector( yf, yf, y_baseline )
# This method takes cares of some "fixes" that are needed to get the Jacobian
# right for iterative solutions. No need to call this WSM for linear
inversions.
jacobianAdjustAndTransform
}
I included both of the suggestions, and from there increased the maximum number
of iterations to 15 and set the temperature limits to be between 150 K and 350
K. All other lines remain the same, including the input absorption line and
atmospheric fields data.
Now I irregularly get the errors during the OEM computation as before, of
t_field having NaN values or of bad partition function warning (which happens
for any t_field value outside between 150 K and 300 K). While the bad partition
function warning could be suppressed in the code, it seems that the same
problem as before persists.
Could you please look into this issue a little more? I am not sure how to
tackle the problem.
Thank you for your service.
Regards,
Byungsuk Lee
ARA Consulting & Technology
30 Songdomirae-ro D-1510, Yeonsu-gu, Incheon 21990, South Korea
-----Original Message-----
From: "Simon Pfreundschuh"<[email protected]>
To: "이병석"<[email protected]>;
"[email protected]"<[email protected]>;
Cc:
Sent: 2018-09-14 (금) 15:13:55
Subject: Re: [arts-users] ARTS OEM
Hi Byungsuk Lee,
It is a bit surprising that these errors occur irregularly, since there is no
randomness
in the test setup. However, if you are running retrievals with external
observation data
it may well happen that the iteration ends up in a state that produces invalid
values.
For your temperature retrieval I would recommend one or both of the following
solutions:
1. Use the Levenberg-Marquardt (LM) iteration scheme: Using the LM scheme with
a high start
value for the gamma parameter can help avoid reaching invalid values when
the forward
model is non-linear:
OEM(method="lm",
max_iter=5,
display_progress=1,
lm_ga_settings=[100,5,2,1000,1,99])
2. Clip the x-vector: Using the xClip WSM inside your inversion iterate agenda
you can force the
x vector to contain only temperatures within a certain range. If you use
this approach you
should check that none of the retrieved temperatures in your final state are
clipped otherwise your
results may not be valid.
AgendaSet( inversion_iterate_agenda ){
xClip(ijq = 0, limit_low = 100.0)
Ignore(inversion_iteration_counter)
# Map x to ARTS' variables
x2artsStandard
# To be safe, rerun checks dealing with the atmosphere
atmfields_checkedCalc
atmgeom_checkedCalc
# Calculate yf and Jacobian matching x.
yCalc( y=yf )
# Add baseline term
VectorAddVector( yf, yf, y_baseline )
# This method takes cares of some "fixes" that are needed to get the
Jacobian
# right for iterative solutions. No need to call this WSM for linear
inversions.
jacobianAdjustAndTransform
}
Regards,
Simon
From: [email protected]
<[email protected]> on behalf of 이병석
<[email protected]>
Sent: Friday, September 14, 2018 2:31:25 AM
To: [email protected]
Subject: [arts-users] ARTS OEM
Dear ARTS users,
I am writing to address a question with the use of OEM in the ARTS development
version.
From the "TestOEM.arts" in "ARTS/controlfiles/artscomponents/oem", this time I
changed the codes as follows.
1) From (line 115):
retrievalAddAbsSpecies(
species = "O3",
unit = "vmr",
g1 = p_ret_grid,
g2 = lat_grid,
g3 = lon_grid
)
To:
retrievalAddTemperature (
g1 = p_ret_grid,
g2 = lat_grid,
g3 = lon_grid,
hse = "off" )
2) From (line 125):
VectorSetConstant(vars, nelem, 1e-12)
DiagonalMatrix(sparse_block, vars)
covmat_sxAddBlock(block = sparse_block)
To:
VectorSetConstant(vars, nelem, 0.1)
DiagonalMatrix(sparse_block, vars)
covmat_sxAddBlock(block = sparse_block)
3) From (line 37):
NumericSet( f_start, 110.436e9 )
NumericSet( f_end, 111.236e9 )
IndexSet( nf, 801 )
IndexSet( np, 81 )
To:
NumericSet( f_start, 110.436e9 )
NumericSet( f_end, 111.236e9 )
IndexSet( nf, 18 )
IndexSet( np, 18 )
4) From (line 42):
VectorNLinSpace( f_grid, nf, f_start, f_end )
VectorNLogSpace( p_grid, 361, 500e2, 0.1 )
VectorNLogSpace( p_ret_grid, np, 500e2, 0.1 )
To:
VectorNLinSpace( f_grid, nf, f_start, f_end )
VectorNLogSpace( p_grid, 18, 500e2, 0.1 )
VectorNLogSpace( p_ret_grid, np, 500e2, 0.1 )
Overall, I changed the ozone VMR retrieval to temperature retrieval and reduced
the number of frequencies and pressure layers.
When I run the controlfile multiple times, I irregularly get errors. The error
messages are
Run-time error in oem computation: Forward Model Evaluation Error
Run-time error in agenda: inversion_iterate_agenda
Run-time error in method: atmfields_checkedCalc
The variable *t_field* contains one or several NaNs. This is not allowed!
or
All temperatures in *t_field* must be > 0.
or seldom the bad partition function warning.
Here are my questions:
1) Is this an anticipated error, simply coming from failing to find a
converging solution? Is it expected that the OEM sometimes throws an error as
above?
2) Are there mechanisms (e.g. workspace methods) within the ARTS that can
prevent the t_field or vmr_field from becoming either negative or NaN during
the OEM iterations?
Thank you for your assistance.
Regards,
Byungsuk Lee
ARA Consulting & Technology
30 Songdomirae-ro D-1510, Yeonsu-gu, Incheon 21990, South Korea
Email: [email protected]
_______________________________________________
arts_users.mi mailing list
[email protected]
https://mailman.rrz.uni-hamburg.de/mailman/listinfo/arts_users.mi
