3.12 עמודים/קפה זה יחס טוב?

I got no tips on the aesthetics 😭 but I did have an absolutely awesome team support me by checking content and helping me practice and that made all the difference

This paper is now accepted in ApJ and the latest version is on arXiv as of last night, so go check it out! A huge thanks to the coauthors on here: @luiswel.bsky.social @astropeter.bsky.social @nixonmatthew.bsky.social

arxiv.org/abs/2503.21702

(13/13)

So does all this mean everything we've ever done is wrong? Absolutely not! GPs just present a new, statistically robust, more flexible modeling methodology for exoplanetary spectra that can provide insights into not just our data, but our models themselves!

(12/n)

In fact, we also see that the GP identifies a distinct global correlation in the NIRISS data, which is not unexpected from instrumental effects. At ~75 ppm, this is not a massive oversight, but something for future NIRISS enthusiasts to be aware of.

(11/n)

In fact, if a local kernel is placed, it falls on the CO2 feature at 2.7 microns, telling us that perhaps CO2 characterization with NIRISS is less robust than previously expected. That doesn't mean this is necessarily the case, but it's a good diagnostic already!

(10/n)

Very possibly! We find that including GP kernels hint at wider H2O and CO2 abundance distributions than traditional retrievals. The latter even has a tail that contains the values from *both* previous estimates!

(9/n)

We then analyze the *real* NIRISS spectrum of WASP-96b! Previous analyses (Radica et al. 2023, Taylor et al. 2023) largely agree, but found discrepancies in their CO2 abundance. Can GPs help?

(8/n)

This can help us understand our modeling deficiencies and/or data wonkiness at a much better level than before, so we can push towards more complete models!

"Well that's great Yoav," you say, "but what about real observations?"

Glad you asked!

(7/n)

Not only that, but the parameters of the GP ("hyperparameters") actually contain information! By analyzing the GP kernels, we can get an understanding for how much correlation there is in our data and *where* our model cannot fit it.

(6/n)

Yes! The traditional retrieval (black histograms) not only can't account for the injected feature, but it actually biases all our inferences. Different GP parameterizations (blue/pink/purple) give significantly less biased, more reliable answers.

(5/n)

We kick this off by creating a synthetic spectrum of WASP-96b, where we inject some underlying correlation and a "mystery" feature at 1.65 microns, that we *know* our model can't account for. Will a GP work better?

(4/n)

We use a GP that looks for a) global correlations throughout the spectrum that may come from instrumental effects and b) localized correlations that manifest where the model can't explain the data!

(3/n)

We set out to fix this problem with Gaussian Processes (GPs). GPs help account not only for the model but also for any underlying correlation. By "GP-ifying" retrievals of transmission spectra, we can actually *relax* assumptions about model completeness and noise.

(2/n)