flow3_input.txt

FLOW采用有限单元法fortran90编写的求解不可压缩流体的稳态流速和压力场的程序

#  Weights:    (1, 0, 0, 0.01, 0)
#  Reynolds:   1.0
#  Parameters: 3 inflow, 0 bump
#  Type:       3, sensitivity
#
With this run, I added a small weight to the inflow control cost.
This produced a much better solution.  After ten steps, we had

  0.36956  0.50138  0.36954


#  INPUT.26
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0.01)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 3 bump
#  Type:       3, sensitivity
#
#  I want to see what controlling the bump cost does, with a
#  sensitivity run.  (I found a slight mistake in BCOST that
#  would affect such a run, but apparently I've never made
#  such a run before...)
#

#  INPUT.27
#
#  Mesh:       21 by 7
#  Weights:    (1, 1, 1, 0, 0)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 1 bump
#  Type:       3, sensitivity
#
#  I want to do a run with just two parameters, so I can plot
#  the optimization history.
#
*********************************************************************

#  INPUT.101
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 3 bump
#  Type:       3, optimization using sensitivities
#  Ipred:      0, use 0 for start.
#

This was the first of a set of three comparison runs, in which
I was trying out the new parameter IPRED.  I am trying to
use the Euler prediction to get a good Newton starting point.

However, this set of three runs (101, 102, 103) was very
disappointing, with almost no difference.

#  INPUT.102
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 3 bump
#  Type:       3, optimization using sensitivities.
#  Ipred:      1, use previous solution for start.
#
#  INPUT.103
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 3 bump
#  Type:       3, optimization using sensitivities.
#  Ipred:      2, use Euler prediction for start.
#


#  INPUT.104
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 0 bump
#  Type:       3, optimization using sensitivities
#  Ipred:      0, use 0 for start.
#
This set of runs was similar to (101, 102, 103) except that
I cut down to a single, flow parameter, because I was hoping
that the code was OK, and the previous problem was too hard.
Or else that the difficulties occurred in the shape calculations.

Indeed, that seems to be the case.  The 106 run took only one
Newton step to converge, instead of four.  So now I have to
concentrate on what's going wrong with the shape calculations.

I will do that by going to 1 flow, 1 bump parameter.

#  INPUT.105
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 0 bump
#  Type:       3, optimization using sensitivities
#  Ipred:      1, use previous solution for start.
#
#  INPUT.106
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 0 bump
#  Type:       3, optimization using sensitivities
#  Ipred:      2, use Euler prediction for start.
#
#  INPUT.107
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 0 inflow, 1 bump
#  Type:       3, optimization using sensitivities
#  Ipred:      0, use 0 for start.
#
These runs were made to try to focus on the behavior of the bump.
If there was a problem with the bump calculations, it would be
easier to see it if there were no flow parameters, and only one
bump parameter.  

In all three runs (107, 108, 109), there was very little variation.
There were 6 optimization steps, and 27 or 28 Newton steps.

The computation on the second optimization step may be indicative
of whether things are going well or not:

  IPRED=0            IPRED=1            IPRED=2

  9.03               0.38               0.95e-1
  0.59               0.33e-2            0.19e-2
  0.85e-4            0.16e-8            0.10e-7
  0.31e-10           0.33e-12           0.62e-12
 
#  INPUT.108
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 0 inflow, 1 bump
#  Type:       3, optimization using sensitivities
#  Ipred:      1, use previous point for start.
#
#  INPUT.109
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 0 inflow, 1 bump
#  Type:       3, optimization using sensitivities
#  Ipred:      2, use Euler prediction for start.
#
#  INPUT.110
#
#  Mesh:       41 by 13
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 0 inflow, 1 bump
#  Type:       3, optimization using sensitivities
#  Ipred:      0, use 0 for start.
#
These runs were made to see what effect a finer mesh would have on 
the results.

Each of the runs (110, 111, 112) took 6 optimization steps, with a 
total of 28, 26 and 27 Newton steps respectively.

Again, we look at step number 2 of the optimization to watch the
behavior of the Newton iteration:

  IPRED=0     IPRED=1     IPRED=2

  9           0.5         0.12
  0.6         0.3e-2      0.2e-2
  0.1e-4      0.1e-8      0.1e-7
  0.1e-11     0.9e-12     0.1e-11

#  INPUT.111
#
#  Mesh:       41 by 13
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 0 inflow, 1 bump
#  Type:       3, optimization using sensitivities
#  Ipred:      1, use previous point for start.
#
#  INPUT.112
#
#  Mesh:       41 by 13
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 0 inflow, 1 bump
#  Type:       3, optimization using sensitivities
#  Ipred:      2, use Euler prediction for start.
#
#  INPUT.113
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 1 bump
#  Type:       3, optimization using sensitivities
#  Ipred:      0, use 0 for start.
#
This family of runs was made to examine the behavior when 1 inflow
and 1 bump parameter were used.  

No suprises occurred.  The behavior was very similar to 1 inflow 
and 3 bump parameters.

#  INPUT.114
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 1 bump
#  Type:       3, optimization using sensitivities
#  Ipred:      1, use previous point for start.
#
#  INPUT.115
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 1 bump
#  Type:       3, optimization using sensitivities
#  Ipred:      2, use Euler prediction for start.
#
*********************************************************************

This series of runs was a test of the new IJAC option, to see
if I could get good results while holding the Newton matrix
fixed.  Using IJAC=10, for instance, seemed to slow down the
rate of convergence, but speed up the calculation, since I
avoided reforming and factoring the jacobian.  IJAC=10 ran
in about 1/4 the time of IJAC=1.

#  INPUT.201
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 3 bump
#  Type:       3, optimization using sensitivities.
#  ipred:      2, use Euler prediction for start.
#  ijac:       1, always update Newton matrix.
#
#  INPUT.202
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 3 bump
#  Type:       3, optimization using sensitivities.
#  ipred:      2, use Euler prediction for start.
#  ijac:       4, Update Newton matrix every IJAC steps.
#
#  INPUT.203
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 3 bump
#  Type:       3, optimization using sensitivities.
#  ipred:      2, use Euler prediction for start.
#  ijac:       10, Update Newton matrix every IJAC steps.
#
*********************************************************************

#  INPUT.301
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 1 bump
#  Type:       3, optimization using sensitivities.
#  ipred:      2, use Euler prediction for start.
#  ijac:       1, update Newton matrix every step.
#
#  This is a "Baseline" input file, with Reynolds number 1.
#  We will use this for comparisons against a Reynolds number
#  of 100.
#
#  INPUT.302
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   100.0
#  Parameters: 1 inflow, 1 bump
#  Type:       3, optimization using sensitivities.
#  ipred:      2, use Euler prediction for start.
#  ijac:       1, update Newton matrix every step.
#
#  We try a straight Newton method, which we expect will fail
#  at this high Reynolds number.  Indeed, the method failed
#  immediately, right at the target point!
#
#  INPUT.303
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   100.0
#  Parameters: 1 inflow, 1 bump
#  Type:       3, optimization using sensitivities.
#  ipred:      2, use Euler prediction for start.
#  ijac:       1, update Newton matrix every step.
#
#  This is a high Reynolds number problem. 
#
#  A "straight" Newton method will not generally converge for
#  such a high Reynolds number, so we use simple iteration
#  with a tolerance of 0.001, followed by Newton iteration
#  with a tolerance of 0.0000000001.
#
#  This method converged nicely, and the optimization was
#  successful.
#
#  78 simple iterations and 53 Newton iterations were required.
#
#  INPUT.304
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   100.0
#  Parameters: 1 inflow, 1 bump
#  Type:       3, optimization using sensitivities.
#  ipred:      2, use Euler prediction for start.
#  ijac:       1, update Newton matrix every step.
#
#  With this example, I turned off Newton iteration, using only
#  simple iteration.  I also reduced the simple iteration tolerance
#  to what the Newton iteration tolerance was.
#
#  The results were as expected: simple iteration by itself
#  converged, but much more slowly than Newton.
#
#  500 simple iteration steps were required, in total.
 
 
#  INPUT.305
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   500.0
#  Parameters: 1 inflow, 1 bump
#  Type:       3, optimization using sensitivities.
#  ipred:      2, use Euler prediction for start.
#  ijac:       1, update Newton matrix every step.
#
#  This is a high Reynolds number problem. 
#
#  A "straight" Newton method will not generally converge for
#  such a high Reynolds number, so we use simple iteration
#  with a tolerance of 0.001, followed by Newton iteration
#  with a tolerance of 0.0000000001.
#
 
 
#  INPUT.306
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   200.0
#  Parameters: 1 inflow, 1 bump
#  Type:       3, optimization using sensitivities.
#  ipred:      2, use Euler prediction for start.
#  ijac:       1, update Newton matrix every step.
#

#  INPUT.307
#
#  Mesh:       31 by 10
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   200.0
#  Parameters: 1 inflow, 1 bump
#  Type:       3, optimization using sensitivities.
#  ipred:      2, use Euler prediction for start.
#  ijac:       1, update Newton matrix every step.
#
#  The story so far: I've been trying to run higher Reynolds number
#  problems.  I can get to 100, but not 200.  But that was while
#  running with a fairly coarse mesh (21 by 7).  Now I will rerun
#  this case, but using the next finer grade mesh, and see what
#  happens.  If this still fails, I'll try 41 by 13.
#
 
#  INPUT.308
#
#  Mesh:       41 by 13
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   200.0
#  Parameters: 1 inflow, 1 bump
#  Type:       3, optimization using sensitivities.
#  ipred:      2, use Euler prediction for start.
#  ijac:       1, update Newton matrix every step.
#
#  #306 and #307 failed to converge with 21x7 and 31x10 meshes.
#  This is our last try.  And lo, it converges!
#
********************************************************************

#  INPUT.401
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Reynolds:   1.0
#  Parameters: 1 inflow, 1 bump
#  Type:       3, optimization using sensitivities.
#  ipred:      2, use Euler prediction for start.
#  ijac:       1, update Newton matrix every step.
#
#  This is the first run using a target bump with
#  different X values than the space of solutions.
#
********************************************************************

#  INPUT.501
#
#  The standard test problem.
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Parameters: 1 inflow, 3 bump, 1 flow strength
#

#  INPUT.502
#
#  This data file sets up a 3D march.
#  
#  The 2D march, for WATEB=0, corresponds to INPUT.20, which
#  produced the banana shaped cost function contour plot.
#
#  The expectation is that as WATEB increases to 1, the banana
#  shape will disappear.
#

I was able to compute 800 points in about 10 minutes on the ALPHA.
I still haven't been able to plot them, since I can't get stupid
DICER to work yet!FLOW3.INP  20 June 1994
 
********************************************************************


List of input files:
-------------------

#  INPUT.501
#
#  The standard test problem.
#
#  Mesh:       21 by 7
#  Weights:    (1, 0, 0, 0, 0)
#  Parameters: 1 inflow, 3 bump, 1 flow strength
#

#  INPUT.502
#
#  This data file sets up a 3D march.
#  
#  The 2D march, for WATEB=0, corresponds to INPUT.20, which
#  produced the banana shaped cost function contour plot.
#
#  The expectation is that as WATEB increases to 1, the banana
#  shape will disappear.
#

I was able to compute 800 points in about 10 minutes on the ALPHA.


#  INPUT.503  14 June 1994
#
#  This problem is related to the standard input problem, except
#  that the discrepancy in the VERTICAL flow has been included
#  in the cost, along with the horizontal flow.  I expect to
#  find a local minimum, roughly where I found it for the
#  standard problem.
#

#  INPUT.504  15 June 1994
#
#  This data file sets up a 3D march.  
#
#  This problem is identical to INPUT.502, except that WATEV=1,
#  so that vertical velocity discrepancy is added to the cost.
#
#  We expect that the banana shape that showed up in one slice
#  of the 3D plot for 502 will disappear in this problem.
#  

#  INPUT.505  16 June 1994
#
#  This problem is related to the standard input problem, except
#  that only the pressure discrepancy is included in the cost.
#

#  INPUT.506  16 June 1994
#
#  This data file sets up a 3D march.  
#
#  This problem is similar to #502, except that only pressure 
#  discrepancies are used to calculate the cost.
#

#  INPUT.507  16 June 1994
#
#  Set the cost using U, V and P.  
#  Use only 1 inflow and 1 bump parameter, and start with REYNLD=1.
#  

#  INPUT.508  17 June 1994
#
#  This data file sets up a 3D march.  
#
#  We only use one flow and one bump parameter.
#
#  Our cost function involves U+V+P, and we are looking at
#  varying WATEB from 0 to 1.
#

This input file was used to generate a movie for MAX#4.

#  INPUT.509  17 June 1994
#
#  This data file sets up a 2D march.  
#
#  We only use one flow and one bump parameter.
#
#  Our cost function involves U+V, and we are looking at
#  varying WATEB from 0 to 1.
#
#  I'm just trying to see whether to modify run 508 to drop
#  the pressure cost, in order to get nice round functional
#  contours.
#

#  INPUT.510  18 June 1994
#
#  This data file sets up a 2D march.  
#
#  We only use one flow and one bump parameter.
#
#  Our cost function involves U+V+P, and we are looking at
#  varying WATEB from 0 to 10.
#
#  I'm just trying to see whether to modify run 508 to increase
#  the final bump cost, in order to get nice round functional
#  contours.
#

#  INPUT.511  20 June 1994
#
#  This file is similar to the standard test problem,
#  except that the optimization is carried out with
#  WATEB=1.
#

I want to see WATEB making it easier to solve the problem quickly.
Then I will try to modify the program so that it is easy to solve
in the "vertical" direction, that is, to let WATEB gradually go
to zero.