TAP Planetary Atmospheres Initiative Computational Workshop - Luis Welbanks

TAP Planetary Atmospheres Initiative Computational Workshop

Luis Welbanks, ASU

Visit Dates:  December 7-11, 2026

Title:  Do you Believe Your Own Error Bars?  Diagnostics for Bayesian Inference in Astrophysics

Abstract / overview
Bayesian inference is now the default tool for extracting physical parameters from astronomical data, and nested sampling has made model comparison routine. Routine is not the same as reliable. A sampler will return a posterior and an evidence whether or not the model is capable of describing the data, and the resulting Bayes factor is frequently translated into a detection significance that the data do not support.
This workshop is about the step that usually gets skipped: interrogating what an inference is actually telling you. Participants will run an end-to-end retrieval on a spectrum, compute a Bayes factor, and then use pointwise diagnostics, including Bayesian leave-one-out cross-validation, to locate where a model succeeds and where it fails at the level of individual data points. We will then look at what to do when the model is missing physics or the noise is correlated, exploring concepts like Gaussian processes as a flexible way to absorb unmodeled structure without pretending it is signal.
The examples are drawn from exoplanet transmission and emission spectroscopy, but the methods and the failure modes are general. Anyone fitting a parameterized physical model to data with a likelihood should find the diagnostics directly transferable.

Learning objectives
1.  Run a complete nested-sampling inference on a spectrum and correctly interpret the resulting posterior and evidence.
2.  Compute a Bayes factor for a nested model comparison, convert it to a significance, and articulate precisely what that number does and does not establish about a detection.
3.  Apply pointwise predictive diagnostics, including leave-one-out cross-validation, to identify which data are driving a model preference and where a model is failing.
4.  Recognize when correlated noise or unmodeled physics is present, assess how these affect inferred planet properties.

Materials and software to review or install in advance
A public GitHub repository with Jupyter notebooks, synthetic and real spectra, and precomputed posterior samples will be emailed to registered participants. It will include a conda environment file.
Software: Python 3.10 or later, numpy, scipy, matplotlib, a nested sampler (multinest), arviz for the diagnostics. 

Suggested pre-reading, none of it required
Computational challenges in exoplanet atmospheric retrieval -> https://link.springer.com/article/10.1007/s41115-026-00031-9
Welbanks, McGill, Line & Madhusudhan (2023, AJ 165, 112), on leave-one-out cross-validation applied to atmospheric analysis.
Challenges in the detection of gases in exoplanet atmospheres -> https://ui.adsabs.harvard.edu/abs/2026NatAs..10..234W/abstract

REGISTER HERRE - Coming soon