gully
commented
10 months ago To spell it out: this strategy could be the best way to preserve the institutional knowledge, as a living, breathing evaluable model, that gets added to over time. The constantly growing stream of data gets added as test cases that should satisfy each module.
I just had a pleasant hallway chat with Greg Zeimann, touching on aspects of tons of different data analysis themes for spectroscopy, including spectrogram forward modeling. Greg shared his experience and enumerated various instrumental effects that make forward modeling spectrograms so difficult. Lots of effects begin to matter that the few tens of counts level, and writing those down will be daunting.
This reaffirmed my feeling that a true digital twin should be destined for future spectrographs, where the in-lab detector testing and metrology can reveal and quantify all of these effects, and those products can be saved in a re-usable format. Specifically, the digital twin would start at that early stage, with a PyTorch (or equivalent) function made to be simulate all of those effects before any starlight even hits the detector.