Replies: 1 comment 1 reply
-
Hi Paco,
You can follow these tutorials on adding Gaussian Processes to the model: docs example, workshop tutorial. |
Beta Was this translation helpful? Give feedback.
-
Hi Paco,
You can follow these tutorials on adding Gaussian Processes to the model: docs example, workshop tutorial. |
Beta Was this translation helpful? Give feedback.
-
Hello,
So I've been working with the different estimators to see which gets me a better fit and I've been having some issues with lc geom. It says that primary is overflowing at periastron. I tried removing requivsumfrac from adopt_parameters in an attempt to resolve the issue but it did not work. Running a run_checks() tells me this:
Additionally, it does not want to fit ecc and per0 (I removed them from adopt_parameters for the image above). When I run_checks() with ecc and per0 in adopt_parameters, I get this:

I tried flipping ecc and per0 with esinw and ecosw respectively and it still won't work:

I'm currently running this using two-gaussian as the analytical_model. I tried doing it with polyfit, and had different issues. With polyfit, teffratio was a nan and had to be removed. I also had the following problems with run_checks():
It's possible that lc geometry is struggling with this dataset as the system observed has tidally excited pulsations. The other estimators gave more promising values:

I just wanted to make sure that I didn't do any mistakes using lc geometry before abandoning it and moving on with ebai. Any help would be greatly appreciated!
Thanks for your time,
Paco
Beta Was this translation helpful? Give feedback.
All reactions