Opened 13 years ago
Closed 13 years ago
#235 closed defect (fixed)
Larger incs in bangle retrievals when using -comp option in IT-1DVAR-06
| Reported by: | Ian Culverwell | Owned by: | Ian Culverwell |
|---|---|---|---|
| Priority: | normal | Milestone: | 5.1 |
| Component: | ropp_1dvar | Version: | 4.1 |
| Keywords: | compressivility retrieval | Cc: |
Description
ROPP5 beta testing revealed that the incs to {T, q, p} resulting from a bangle retrieval in IT-1DVAR-04 are 5-10 times larger (esp for q) when including non-ideal gas effects via the -comp option to ropp_bangle_1dvar:
Why should this be, given that this option has such small effects elsewhere?
Attachments (6)
Change history (8)
by , 13 years ago
| Attachment: | IT-1DVAR-04_comp-vs-ideal.png added |
|---|
comment:1 by , 13 years ago
| Summary: | Larger incs in bangle retrievals when using -comp options in IT-1DVAR-06 → Larger incs in bangle retrievals when using -comp option in IT-1DVAR-06 |
|---|
by , 13 years ago
| Attachment: | it_pp_06_plot_bangle_minropp_profall.gif added |
|---|
by , 13 years ago
| Attachment: | it_pp_06_plot_bangle_minropp_prof1.gif added |
|---|
by , 13 years ago
| Attachment: | it_pp_06_plot_bangle_levmarq_prof1.gif added |
|---|
by , 13 years ago
| Attachment: | it_pp_06_plot_refrac_minropp_prof1.gif added |
|---|
by , 13 years ago
| Attachment: | it_pp_06_plot_refrac_levmarq_prof1.gif added |
|---|
comment:2 by , 13 years ago
| Milestone: | 6.0 → 5.1 |
|---|---|
| Resolution: | → fixed |
| Status: | new → closed |
Note:
See TracTickets
for help on using tickets.
Here's a neater pic of the bangle 1dvar retrievals (cntl=default, test=-comp, bkgr=background):
Note that it's only really an issue for (the first) one of the 5 profiles:
Examination of the output showed that the minimisation algorithms stopped when they met different convergence criteria in the two cases:
Cntl converges after 7 iters with: "INFO (from ropp_1dvar_cost): Convergence assumed to be achieved as the state vector did not change by more than 0.1000 relative to the assumed background errors for the last 2 iterations."
Test converges after 10 iters with: "INFO (from ropp_1dvar_cost): Convergence assumed to be achieved as the cost function is increasing."
Profiles 2, 3 and 4 converged for the same reasons in both cases. Profile 5 converged in the cntl (resp test) because dx (resp dJ) were as small as required, but the retrievals were similar in this case anyway.
Playing with the convergence criteria didn't help to reconcile profile 1 retrievals much. But trying the LevMarq minimiser (with the default convergence criteria) did:
I take this as further evidence of LevMarq's improved robustness and integrity over the default MINROPP solver. This is being addressed in Ticket #234.
The changes in the T, q, and p retrievals are now much smaller, and similar to what we would expect from the compressibility changes (see ticket #214).
For refractivity, the default and -comp retrievals are much closer even with MINROPP:
They're brought even closer by LevMarq:
(The different sensitivities of bangle and refrac retrievals has been found before. Perhaps it's a reflection of the diagonal obs correlations in the former case: there's less "vertical coherence" in the retrievals. More likely (I think) is that the bangle and refrac sigmas are not consistent - I don't think they've been engineered to be so.)
Closing the ticket, as the finger seems to have been firmly pointed at MINROPP, which is being examined elsewhere.