nm_solver ignores eclipses and returns flat line

[Sabcapack]

Hello,

I am using PHOEBE 2 to model a detached eclipsing binary (which is one pair of a doubly eclipsing system).

I have constructed a forward model using my estimated parameters, and it matches the data reasonably well visually (the eclipses are deep and aligned).

The problem is that when I run the nm_solver to fine-tune just a single parameter like t0_supconj, the solver fails to recognise the eclipses. The fitted parameter remains identical to the initial value. The resulting model light curve becomes a flat line or a simple sinusoid, missing the deep eclipses entirely. I have tried fitting different parameters (incl, t0, mass) individually and together, but the results are always the same.

Thank you for your time and help.

Below is the code I am using. I have included the data that I am using as well.

phoebe2_code.txt
parA_detrend_4.txt

[kecnry]

eclipse_method='only_horizon' will ignore eclipses and should only be used for non-eclipsing systems. Try removing that (both when calling run_solver and run_compute) and see if that helps!

-Kyle

[Sabcapack]

Hi, thank you very much for your fast reply!

Removing eclipse_method='only_horizon' worked perfectly - the eclipses are now showing up in the model.

However, I am still facing a problem where the solver finishes, but the parameters do not actually change. The values stay exactly the same as the initial guess, even though the fit is clearly not satisfying (the eclipse width and depth are off).

Thanks again for your help.

[abokon]

Hi Sabina,

my name is András, and although I have been inactive with phoebe2 recently, I remember a similar problem from my work one or two years ago.

Without seeing any figures, I would suggest fit only the t0_supconj parameter, handling it separately from other parameter(set) could benefit for you. My other 'tip' for you would be to use the differential evolution solver. It searches for absolute minima in a predefined distribution which the user can set after estimating a reasonable small but good interval for the t0.

(1) First, you need to set a distribution. I sadly don't remember whether the uniform_around function is working here or not. (For emcee it certainly does, but there were some complications two years ago.) The safest approach can be uniform(a,b) function, where the a and b is the end of interval would you search for the t0.
See for distributions:
https://phoebe-project.org/docs/2.4/tutorials/distribution_types
(2) After this you can add a new solver, which use the differential evolution, the fit parameter is only the t0_subconj in question, with distribution you set up in step (1) as bounds@....
See:
https://phoebe-project.org/docs/2.4/tutorials/differential_evolution

Be aware that:

• you can utilize the number of initialization (initial number of search) as well which is used for search of the minima in the interval of (a,b). The keyword for this is popsize@.... The default value for this 8.
• Multiparallel process (i.e. using multiple core of your processor) is used by default here. You can check the number of process by the function of
  phoebe.multiprocessing_get_nprocs()
  and modifying it with e.g.
  phoebe.multiprocessing_set_nprocs(2)

functions. You need to set this up in every run, so don't forget it

(3) At last, you can run it with the known b.run_solver(...) function,

and you can check its solution with the usual way.
For me, that was a reliable way to fit the exact time of superior conjunction in my case, which was an essential step for the optimizations in the next step.

This worked for me, and I really hope that this approach will be successful in your case as well. If not, can you show a figure of the fit? Sometimes it can help to detect the possible cause of the problem in question.

András

[kecnry]

Thanks @abokon! The other thing that comes to mind is that it may be worth checking the message (in the solution) for the reason that nelder-mead stopped - if it stopped quickly because the tolerance was already met, then you may want to adjust xatol/fatol.