Luis Welbanks

What are your thoughts?

We need to understand the limits of our models and our data for reliable interpretations of exoplanetary spectra.

First paper by my first PhD student! Happy advisor moment. Congrats @yoavrotman.bsky.social

Check out his paper on the Arxiv and his thread below #exoplanets 🔭

I think the full title was an unfortunate Ceres of events btw, but I am no Lemoony Snicket expert

And sanctioned by the journal :P

Call them out ;)

Great show. 10/10

I agree with you that saying K2-18 b “can’t” have an ocean or “isn’t” an ocean world is a stretch - we can’t totally rule it out with the present data, but it does appear that Neptune-like or gas dwarf models are consistent with what we know about the planet and require much less fine-tuning

"Relies on the obfuscation of how, exactly, they are defining [...] in order to garner press coverage. "

&
"An exceedingly generous observer might chalk up this divergence to the perennial conflict between scientists and their PR machines"

Great title: [Fill in the blank] Can’t Have It Both Ways

There’s Probably No Life on K2-18b After All, Three Studies Conclude In April, astronomers said they had detected a possible signature of life on the exoplanet K2-18b. Now, three independent analyses discount the evidence.

Here's my follow-up story on K2-18b, the distant planet where scientists claimed to see a possible sign of life last month. In three preprints, other researcher argue that the signal is noise. nyti.ms/4jaqQRv

What do you mean by original hypothesis? Just to make sure I understand correctly, are you suggesting that the discourse (online/in papers?) should say that the original hypothesis (which?) is not ruled out but neither is Wogan's or X?
Not being facetious, legitimately asking.

In the low SNR, it is possible for the inferred properties to be shaped as much by preconceived notions for what a planet ought to be like, as by data (if we are not careful).

Our take-home on this specific point could be: "Reliance on
Bayesian evidence alone, coupled with exploration
of only a narrow part of the model space, has led
to contradictory interpretations."

On that we say "Conversely, when all candidate models
adequately fit a spectrum, a preference for one model
over another does not rule out the worse-performing
model."

2) At the same time - you could ask: can the data rule out X hypothesis? The answer may be no, and that's ok too!

e.g., K2-18b MIRI - current data cannot rule out the scenario of a planet under radiative-convective-photochemical equilibrium (section Self-Consistent Models)

We say in our paper "When all considered candidate models are poor representations of reality, the best-performing model is simply the least inadequate and may not necessarily lead to reliable interpretations of the data."

Not 100% sure I understand your question(s) but let me try.

1) Constraining a parameter does not equal 'detecting' that parameter. I can add a none-sense parameter and get a tight constrain. This is the whole point of the cheese v. sponge example.

There is no "Bayesian police" to say what to compare or not. Any paper can compare any two models (but please contextualize!).
However, if we are going to argue about 'standard practices', the "consensus" (somewhat arbitrary) is to compare relative to the full hypothesis 2/2
Thanks for engaging!

Sounds good! Two distinct yet complementary points.
In general Welbanks & Nixon+ is not arguing against model comparisons but it is making an appeal to contextualize them. What two models did you compare? From that perspective Chubb+20 is doing that in the abstract: X sigma relative to blah. 1/2

🤩🤩🤩🤩

Say hi to Ian and Jasmina! If you haven't done so, visit the Sheikh Zayed grand mosque. I was blown away by the beautiful interiors.

But how certain are we that we can ignore the uncertainties?

The scientist's paradox

@viciykevin.bsky.social one could argue whether the "full hypothesis space" is valid or not - we discuss that in our paper. However, the comparison is performed against this full hypothesis space. This is how the comparison got its connection to 'detection'.

What @distantworlds.space said. Section 4 of Gasman says "In Tables 6–10 we specify the Bayes
factor, B01, for each retrieval set-up, comparing the retrieval
with the specified molecule (C2H2, C2H4, CH4) included versus not included" - you want to compare against your full hypothesis space.

Following the conventional practices of the field (nested models) and his exact results there is no detection of DMS and no detection of DMDS. #exoplanets 🔭 #K218b

Check out "Interpretation According to Conventional
Practices" in our paper and let us know what you think arxiv.org/abs/2504.21788

What are your impressions of the reaction within Europe? Is the community planning to react in that way?

Ah! Great! How do I get a copy of the newsletter?

@philplait.bsky.social - is there a way to help you / crash course / @hankgreen.bsky.social? His video on the topic was great but failed to catch some of the nuance. This is declaring a detection of cheese on the Moon from only considering a cheese model and a sponge model.

This is a major change in interpretation within just one month, with no new data, no new retrieval framework, and no newly available cross-section data. That such a dramatic revision was needed reinforces that the original inference (detection) was overstated.

Offsets between instruments - vertical shifts in transit depth. Observations at different epochs may require offsets and indeed, they seem inevitable, even for big planets like WASP-39b