I got 1529 MCMC samples in 2h22m, projecting about 7.7 hours for 5000 samples on my 12 thread laptop. I'd expect TACC's 40+ threads to be about 3 times as fast for 40 emcee walkers, so about 2.5 hours per spectral order.

With 40 orders per IGRINS spectrum, 4 epochs, and 2 components per epoch, that's 320 spectral orders.

All together that's 800 hours on TACC--- a single node for 33 days, or 33 nodes for 1 day.

Typically we'd need to re-run certain failed runs, or with new boundaries or recipes. In my experience this fudge factor is at least 2. So 66 node-days.

TESS paper figure and prose:

• How did we analyze and process the TESS data
• How does our method compare to Apai et al. team's, in terms of systematic trends and SNR
• Figure overplotting Apai lightcurve with ours
• Figure of periodograms
• optional GP analysis
• TESS fluxes at the time of IGRINS acquisition

We want to label the brightness temperature per echelle order to get a coarse sense of characteristic depth probed.

We have the cloudy models spanning visible to near-IR, at low-ish resolution (R~10,000 or less).

If you look closely at the spectrum in #9 you can see places where the model fit is so good that it implies the underlying template would be adequate for a deconvolution approach: infer the convolution kernel that gives rise to the line profile perturbations. I am inclined to try this out. If we do this on a few trustworthy regions and get about the same kernel then we can have some assurance that the signal is genuine. My scientific expectation is that the kernel should vary smoothly with wavelength, so comparing adjacent orders may be fine, but H- and K-band could very well have different characteristic vsini values depending how their equators and poles are illuminated.

Then associate each component with its IGRINS spectra filenames.

Bonus: plot the TESS FFI image on the same scale for a head-to-head comparison.

Most practitioners, accustomed to low resolution SpeX spectra, get disoriented by the shear grasp of IGRINS spectra. We propose to orient readers by presenting flux-calibrated low resolution spectral models with selected zoom-in cutouts. We can label the broad features in this low resolution view, then possibly again at finer resolution.

Place the times of the observations, probably in BTJD, the component A or B, and the nod A or B, and possibly the TESS flux at the time of the acquisition.

We want to label the conspicuous spectral features. My first guess for how to do this is with radis, with a characteristic temperature and pressure. That might be overkill, but it is certain to work and can be run uniformly for a range of species. Besides, I'm fond of the radis community and their great algorithmic advances, including GPU and recent exomol support.

Still, this feels like a lot of work, maybe a few days just to get everything working. It'll be worth it, but it's not free. So any suggestions for free alternatives are welcomed!

We now have a working interactive spectral analysis tool intuition.py built on bokeh. This tool will allow us to get "by-eye" best fit parameters and build an intuition for the spectra, either in-place-of or in-advance-of a full-on spectral inference procedure. Here are the science questions I brainstormed with a coarse classification for how easy/hard the task is.

🟢 = Easy
🟡 = Takes some work
🔴 = Research Project

1. 🟢 Do the raw spectra show variations conspicuous to the human eye? If so what are the wavelength locations of the largest such perturbations?
2. 🟢 What is the overall performance of the custom cloudy Sonora models?
3. 🟡 What is the best fit temperature and surface gravity? To what extent do these values vary with spectral order (i.e. wavelength/species)?
4. 🟡 What is the overall character of the spectral variations? Is there evidence for line profile perturbations?
5. 🟡 With what filling factor of clouds is each component most consistent?
6. 🟡 To what extent are the pairs of parameters degenerate, for example resolution and surface gravity, or spectral resolution and filling factor, or resolution and filling factor? Are there discrete spectral features or groups of features that break these degeneracies?
7. 🔴 Can the heritage of spectral perturbations be convincingly assigned to astrophysical variation and not, say, telluric correction artifacts or instrumental/wavelength-dependent slit-loss?
8. 🔴 Is there conspicuous evidence for two distinct spectral components (cloud and cloud-free) coexisting in the single composite spectrum?
9. 🔴 How do the observed spectral changes relate to the instantaneous TESS flux (available April 9)?
10. 🔴 Does the overall pattern of best fit values comport with theoretical expectations? Overall what do we learn from the spectra?

This project has been dormant for a while. I propose we open source it, namely make it public instead of its current private. This way folks can benefit from the analysis even if it is unpublished. I emailed Yifan Zhou to see if he was interested in having students work on it, and he said maybe/probably, and agreed that open sourcing it would be helpful either way.

The default IGRINS plp leaves divides the target spectrum by an unprocessed A0V template. This step can imbue large humps in the spectrum for reasons I pointed out in this 2017 plp Issue comment.

version controlling jupyter notebooks can be annoying. Let's finally setup the nbstripout command to strip all output before committing.

We now have custom cloudy models working in our interactive spectral analysis dashboard. Overall the fits look really good and the intrinsic and extrinsic physical properties(Teff, logg, fsed, vsini, RV) are mostly consistent from spectral-order-to-spectral-order. Caroline and I examined the fits--- it's tricky to see the visit-to-visit variation by eye, but we flagged some locations that may be more variable than others. The next step is to quantify these changes and and quantify the degree to which spurious artifacts could mimic them. Below we list the specific takeaways from the meeting.

Tasks for the next week:

• 1. Label conspicuous spectral features
• 2. Overlay low-res spectrum with high-res cutouts
• 3. Label brightness temperature per echelle order (you can coarsely compute it with a scaling relation)
• 4. Refine the telluric correction (the default plp one leaves broad residuals)