Luis Welbanks

Research

Reading the atmospheres of other worlds

With JWST, the study of exoplanet atmospheres has gone from being limited by data to being limited by models. JWST now records a single planet's atmosphere from 0.5 to 12 microns, and the conclusion drawn from such a spectrum can depend on what the model was allowed to include.

My group works on both sides of this problem. We analyse spectra from JWST, Hubble and ground-based telescopes to measure composition, temperature, clouds and internal heat, and we develop statistical tests that show which parts of a spectrum drive each conclusion.

01 · JWST spectra

Reading atmospheres with JWST

I coordinated the modelling for the JWST Transiting Exoplanet Community Early Release Science Program, in which more than 50 researchers interpreted the first JWST spectra of the giant planet WASP-39 b. That program identified carbon dioxide in an exoplanet atmosphere, and its combined spectrum now spans 0.5 to 12 microns. Since then I have led or helped lead the interpretation of JWST spectra of giant planets such as WASP-80 b and HD 189733 b and of the sub-Neptune GJ 3470 b.

In the warm Neptune WASP-107 b, Hubble, NIRCam and MIRI data together showed water, methane, carbon monoxide, carbon dioxide, sulfur dioxide and ammonia. Methane is far scarcer than expected at about 750 K, which requires an interior hotter than 345 K, and tidal heating from the planet's slightly elliptical orbit can supply that heat and explain why the planet is so inflated.

Key papers

More JWST papers in an ADS library

Transmission spectrum of WASP-107 b with the best-fitting model, where coloured shading marks the contribution of water, methane, carbon monoxide, carbon dioxide, sulfur dioxide and ammonia.
The transmission spectrum of WASP-107 b and the best-fitting model, with shading for the contribution of each gas. Welbanks et al. 2024, Nature, Fig. 3.
Cover of Nature from 27 June 2024, titled Popcorn Planet, about tidal heating puffing up an exoplanet's atmosphere.
Cover of Nature from 23 February 2023, titled Brave New Worlds, about JWST observations of exoplanets.
The WASP-107 b results made the cover of Nature on 27 June 2024, and the first WASP-39 b papers from the Early Release Science Program appeared in the issue of 23 February 2023.

02 · Statistics

Testing whether a detection is real

A molecule is usually reported as detected when a model that includes it fits the spectrum better than the same model without it, and that comparison depends on which other models were tried. In 2023 I introduced Bayesian leave-one-out cross-validation to exoplanet spectroscopy. It scores a model by how well it predicts each data point when that point is left out of the fit, which shows which wavelengths carry a claimed detection.

For the sub-Neptune K2-18 b, the reported detections of dimethyl sulfide and related gases, proposed as possible signs of life, disappear once the set of candidate molecules is widened. We tested 90 hydrocarbons one at a time, and for one reduction of the data, model comparison favored 59 of them at 2σ or more, six of them more strongly than dimethyl disulfide. We recommend treating model comparisons as tests of relative adequacy, supported by theory and other statistical checks, rather than as detections in themselves. My PhD student Yoav Rotman extends this work with Gaussian processes that absorb features the models miss.

In a review led by Sara Seager, we proposed three tests for any claimed sign of life: a robust detection, attribution to the right gas, and an interpretation that excludes non-biological sources. The figure below shows where each step of an analysis can open up more than one answer.

Key papers

Work that uses leave-one-out cross-validation, in an ADS library

The Bayes factor written as the ratio of the evidence for a model with a molecule to the evidence for a model without it, above the words Ceci n'est pas une détection.
After Magritte's The Treachery of Images. A Bayes factor compares two models and is not, on its own, the detection of a gas.
JWST MIRI transmission spectrum of K2-18 b with several model fits that use different combinations of gases, all passing through the data.
Several models with different gases fit the JWST MIRI spectrum of K2-18 b, and statistical tests cannot tell them apart with these data. Welbanks et al. 2025, Nature Astronomy.
Transmission spectrum of HAT-P-41 b from 0.2 to 5 microns with each data point coloured from blue to red by the change in predictive performance.
Each data point of HAT-P-41 b coloured by how much better a model with H− predicts it than a model without (red) or the reverse (blue). Welbanks et al. 2023, AJ.
Diagram of the steps from a transit to an interpretation, in two rows. In the expected case, a quiet star, a clean detector image, a precise spectrum and narrow abundance estimates lead to one definitive interpretation. In practice, starspots and flares, detector defects, larger error bars and different results from three scientists leave a water world, a magma surface or a featureless planet as possible answers.
From a transit to an interpretation, as expected (top row, inspired by WASP-107 b) and in practice (bottom row, motivated by TOI-270 d). Starspots and flares, instrument noise and different analysis choices can leave several kinds of planet consistent with the same data. Seager, Welbanks et al. 2025, PNAS, Fig. 4. Credit: Seager, Welbanks, Tilke.

03 · Methods

Retrieval methods

A retrieval fits a flexible atmosphere model to a spectrum and returns every composition and temperature structure consistent with the data. During my PhD I wrote Aurora, a retrieval code for hydrogen-rich and hydrogen-poor atmospheres that handles patchy clouds and hazes and several sampling algorithms.

I have also studied where retrievals go wrong. One paper maps the degeneracies between parameters in transmission spectra of hot Jupiters, and another tests the biases attributed to one-dimensional models of planets whose morning and evening sides differ, which we traced to specific model assumptions. Matthew Nixon led a comparison of ways to combine the results of several models instead of choosing one.

Key papers

Papers on gas giants, in an ADS library

Diagram of a planet's terminator divided into clear, cloudy, hazy, and cloudy and hazy sectors.
Aurora treats the terminator as a mix of clear, cloudy, hazy, and cloudy and hazy sectors. Welbanks & Madhusudhan 2021, ApJ.
Mission-style patch reading Aurora 2021, Welbanks and Madhusudhan.

04 · Interiors

Interiors and sub-Neptunes

The composition of an atmosphere depends on what lies beneath it. In WASP-107 b the chemistry required both a hot interior and a core larger than earlier estimates. Sub-Neptunes, planets between Earth and Neptune in size, are the most common planets found so far and have no counterpart in the Solar System. My group models how a magma ocean reacts with the gas above it and how clouds change the temperature at the boundary between atmosphere and mantle, and we test these models against JWST spectra of planets such as TOI-270 d.

Key papers

Papers on small and temperate planets, in an ADS library

Six panels of pressure against abundance for water, methane, ammonia, carbon monoxide, carbon dioxide and sulfur dioxide in WASP-107 b, comparing retrieved values with equilibrium and disequilibrium predictions.
Abundances of the six gases in WASP-107 b against pressure, compared with chemical equilibrium (solid lines) and with disequilibrium chemistry for a cool 200 K interior (dotted lines). Welbanks et al. 2024, Nature, Fig. 4.

05 · Populations

Planets as a population

Many planets observed in the same way can test ideas about formation that a single planet cannot. Using homogeneous retrievals of 19 planets, from mini-Neptunes to hot Jupiters, we found that water abundances rise toward lower planet masses but stay below the trend that carbon follows in the Solar System's giant planets. A later study that added host-star abundances showed the sample was still too small to tell such a trend from no trend, and estimated how many planets a decisive test needs.

JWST spectra now allow statistical studies across many planets, and we contribute to them. For the proposed Nautilus Space Observatory, we make the case for a survey of sub-Neptune atmospheres large enough to map the population.

Key papers

Artist's impressions of ten hot Jupiters in two rows, ranging from cloudy to clear.
Artist's impressions of ten hot Jupiters observed with Hubble. Credit: NASA, ESA, and D. Sing (University of Exeter).
19

planets in our 2019 study of water, sodium and potassium abundances

45

planets with JWST-quality water abundances needed for a decisive mass–metallicity test (Sun et al. 2024)

06 · Ground-based

High-resolution spectroscopy from the ground

Spectrographs on large ground-based telescopes resolve thousands of individual molecular lines. As the planet orbits, its lines shift in wavelength while those of the star and of Earth's atmosphere stay put, and cross-correlating the spectrum with a model picks out the planet's signal. I lead observing programs with the IGRINS spectrographs on the Gemini telescopes, and with the Roasting Marshmallows program and other teams we measure abundances and temperatures in hot Jupiters.

Key papers

High-resolution papers, in an ADS library

07 · Community and teaching

Missions, community projects and teaching

NASA's Pandora SmallSat, launched in January 2026, watches host stars in visible light while it measures near-infrared transmission spectra, so that starspots can be separated from planetary signals. Yoav Rotman led the study of what Pandora will measure for planets from hot Jupiters to temperate sub-Neptunes.

I also organise EXOMINTS, the EXOplanet Model Interrogation and New Techniques Sprint, which brings graduate students, postdocs and faculty together to work on methods for characterizing exoplanets. The 2023 meeting at ASU, funded by the Heising-Simons Foundation and the Other Worlds Laboratory at UC Santa Cruz, has already produced published papers.

In 2025 I taught these methods as one of the two lecturers at Exoplanets by the Lake III, a summer school on JWST spectroscopy of exoplanet atmospheres at Lake Ammer in Germany. My four lectures went from modeling atmospheres to K2-18 b as a case study, with hands-on notebooks on least-squares fitting, Bayesian model comparison and nested sampling. That year I also lectured at the Vatican Observatory Summer School on the first three years of JWST.

Key papers

Papers from EXOMINTS, in an ADS library

EXOMINTS logo