search

gully / ynot

Astronomical échellogram digital twins with pixel-perfect machine learning: rehabilitating archival data and pathfinding for EPRV
https://ynot.readthedocs.io
MIT License
4 stars 0 forks source link

Reviving the project / lessons learned with added PyTorch experience #12

Closed gully closed 1 year ago

gully commented 1 year ago

Woohoo! We're reviving this project! We'll publish a paper on the overall design/performance and apply the code to a range of sources. I'll want to revisit the architecture of the code and apply some of the lessons learned from blasé. Mainly, we could/should be using more flexible functions, possibly with a Gaussian Process prior for regularization. We can also learn 2D convolution kernels and other strategies.

We should probably start by writing out the math in a formal methods section for the paper, and generally writing the background and design considerations and challenges. The API documentation has some of the equations written down.

For Keck NIRSPEC data specfically, we'll want to list the various modes: echelle orders and cross-disperser / grating settings. So far we've mostly demonstrated it for a few orders with a fixed setting.

gully commented 1 year ago

ha, I forgot I had already made a paper draft years ago-- it lives in the private zoja repo.

gully commented 1 year ago

Ok, we have a new paper draft and I've revisited the echellogram model. The main lesson learned I may want to incorporate is sparse matrices. Closing this issue since there is no more action to take.