Themes that could go in the paper, but currently are left out

[gully]

Here are some themes that could go in the paper, but currently are left out.

1. Injection/recovery tests
   We did these, but left them out for now because of time/space/priorities. The code had changed somewhat since they were run and so we'd likely have to redo them. I think they're better suited to a follow-on paper anyways.

2. Earth Science Remote Sensing
   The newfound precision semi-empirical telluric capability may be so good as to rival Earth Science remote sensing techniques. We could catalog a grid of semi-empirical high-bandwidth Earth Atmosphere spectra observed under a range of conditions, with the astronomical spectrum removed (use A0V stars so it is easy). I am not aware of such a resource, and many Earth science practitioners don't have access to 10 meter telescopes or such high-bandwidth spectrographs in this wavelength range at night time. We could/should mention these themes. Since it's not "Astrophysics" I left it out for now, but it is still "Science". In particular, remote sensing of Methane would have both climate change and policy enforcement implications. This topic is big enough to be explored in its own paper, probably not for an astronomy journal.

3. Plots of the clustering of line properties
   I made these but did not put them in the paper. As a result all the figures (other than Fig 1) are of Spectra and not derived properties. While the spectra are nice, showing that we have access to the full catalog of line properties is only stated (albeit repeatedly) in the paper. Showing it with a graphic would emphasize this important aspect of the approach. There are many choices for how to generate that figure, upto and including a corner plot. I am fine to leave these plots for a separate paper on EW-finding or something like that. The paper is long enough as it is!

[gully]

4. Benchmarking and scaling
   We could have a plot of, say, training time versus $N_\mathrm{lines}$, or training time versus bandwidth for fixed number of lines, or training time versus data size. We could have performance benchmarks, such as RMS versus number of training epochs (i.e. a Loss curve), or RMS versus P_rom (we made this already, once!). Those are nice-to-have, but may bloat the paper beyond the key concepts? That is, those are mostly computational considerations and not scientific ones.

[gully]

5. Transfer-transfer learning
   We don't say it explicitly, but it's implied that the semi-empirical output of a blasé transfer learning step can serve as the input of a new blasé transfer learning step. Not only is this possible, but it's a good idea.
   Basically, if you have real data on a stellar template that you believe to be very similar to a target-source-of-interest, you simply use the semi-empirical template derived from that real stellar template data. Then, the lines only have to move a little bit to get to their final resting place, rather than moving far from their synthetic values. This step retires the "synthetic gap" by Calendar and Bello and colaborators.