benmontet
commented
4 years ago Hey,
Thanks for writing this up! I have lots of thoughts. There's still so much low-hanging fruit in the FFIs, I really want me or a student to go back and dig in deeply.
I had an undergrad last summer automate all the hand-holding I had to do previously (which was mostly in aperture selection, trying to pick something big enough to encompass the entire PSF but small enough to not include any other stars) and we have light curves from that for the full 200,000 star sample. They're a little noisier than the artisanal light curves one might create doing aperture selection by hand, but they do decently well and exist for everything. I want to do something with those! Which might be using some population-level information or additional housekeeping data that we didn't have previously to do better detrending, as you mentioned in your final point.
To your specific questions:
Is it possible to calibrate the C5, C16, C18 FFIs in the same way as the Kepler prime mission?
I think this is really going to be a challenge. The reason that this is a tractable problem in the first place is because every year the star comes back and lands on the same pixel. We then have a free parameter which is the offset between the four orientations, which in practice matters at the 1-2 percent level. Just having the one visit in the C16 orientation means there will be an offset that should be hard to correct for in that visit. C5 and 18 might be doable. FFIs at the start of campaigns rather than the end will have thermal issues that will be challenging, and it is better if the pointing is consistent between the two, rather than just before and just after a thruster fire, for example. But maybe! From my perspective it's more work for less effort than mining the Kepler field further.
Applications to non-solar-like stars, and non-stars in general
I want to do non-solar-like stars very much, and also look at the effects of e.g. metallicity on photometric variability (see Karoff et al. 2017 for example). Krista Smith and a student are looking at AGN in the FFIs with a different tool. We compared ~10 targets, five which I selected and five which they selected. I was happy with the performance of f3 for bright stars from that comparison; they might be doing better on very faint objects, which is perhaps not surprising given that's where their science is and ours isn't. But I definitely want to go back at hit up the rest of the stars. I'm imagining something like a random forest classifier that gives the relative importance of various stellar parameters on observed variability.
Joint modeling the long- and short- timescales in Kepler prime?
You've hit on another of my dreams. I would like to be able to do this. The main issue is that our errors are largely dominated by systematics in our uncertainties of the detectors, rather than in photon noise (which we know at the 10ppm level, since Kepler works!) So writing down a likelihood function is challenging. I think you'd also have to go back to the pixel-level data and try to build your own light curve, since the smaller apertures for PDC will have smaller contamination so your astrophysical signals (eg from an EB) will look different in the FFIs than in the long cadence data. I've not quite been convinced by any of the science cases that a probabilistic light curve at the 30-minute level that treats all that and incorporates the FFIs leads to an interesting science result that we don't have now---but I'm open to being convinced that one exists and it's worth pursuing!
Is it worth investing in modernization?
I think it depends on the ultimate userbase. It might be 2 people plus their students, in which case a lot of the ease of use things become less important. But I think gradient based methods/GPU computing could be a useful thing to get the best photometry out of the FFIs---if we can model all the stars in the detector simultaneously, that might give us a big gain! Similarly housekeeping data, especially on pointing, might give us the opportunity to have a big gain.
There's another problem I'm interested in. Luan Ghezzi in Brazil has a Master's student working on this a bit, but he's also applying for phd programs and might leave. There are 200,000 stars observed by Kepler but 4.5 MM in the FOV. Many of the other 4.3 million we will be sensitive to them as long-term variables, eclipsing binaries, or in some cases hot Jupiters. Something in a 3 day orbit should randomly catch a couple points in transit, which could be followed up with TESS data in other special cases (or eventually with PLATO). I think there are ways to efficiently find the stars with outliers and then categorize them. It'll get complicated with SPSDs, but I think they're separable from astrophysics at the pixel level.
Also, I'm sure you're trying to limit your travel right now but if you wanted at some point to come down for extended time and hack on some of this we have the resources to make that happen!
gully
I recently had another look at f3 and wondered about conceivable extensions and applications.
Is it possible to calibrate the C5, C16, C18 FFIs in the same way as the Kepler prime mission?
As it stands, f3 is not equipped for this task, since it is A) hardcoded to Kepler prime, and B) many of the "ubercalibration"-style likelihood assumptions require a large enough sample of FFIs to break significant degeneracies.
Applications to non-solar-like stars, and non-stars in general
Kepler Prime observed Active Galactic Nuclei (AGN, Smith et al. 2018), and it'd be meaningful to constrain the long-term behavior of this AGN sample.
Joint modeling the long- and short- timescales in Kepler prime?
Imagine a likelihood function consisting of a flexible Gaussian Process with many tuning parameters that fits all Q1-17 long cadence data and simultaneously fits the F3-derived photometry. You could then warp the mean model of the long cadence similar in spirit to forward modeling spectral continuum normalization (e.g. Czekala et al 2015, Figure 3).
Is it worth investing in modernization?
Could porting some parts of the code to lightkurve alleviate the brittle dependency of kplr? Could recent innovations in gradient-based methods help with the large number of parameters for joint modeling projects described above? Could some tutorials for more objects and application be helpful to expand the usefulness of f3? Could recently released and reformatted thermal telemetry data help as regressors for long term systematics? These modernization steps could help, but are they worth the investment?