A planet's whole life story, not just a single snapshot: Asynchronous Bayesian optimization is helping astronomers see past a longstanding retrieval problem
Exoplanet scientists usually work backward from a single moment. They take a planet's present-day mass, radius, and atmospheric spectrum, and fit a static model to explain what they see. That approach is fast, but it comes at a cost: a small planet's bulk density can be explained equally well by very different interiors; for instance, a steam atmosphere sitting over a small metallic core, or a hydrogen envelope over a dense one. No amount of additional data quality resolves that kind of ambiguity, because the model was never built to distinguish between the histories that could have produced it.
A new paper by Harrison Nicholls (Institute of Astronomy, Cambridge), Tim Lichtenberg (Kapteyn Astronomical Institute, University of Groningen), Ben Riegler (Helmholtz AI, TU Munich), Robb Calder (Cambridge), and Vincent Fortuin (Helmholtz AI PI, UT Nuremberg) takes a different approach. Instead of fitting the planet as it is today, the method fits its entire life, from the molten magma ocean it started as, all the way through to the world we observe now. Every candidate scenario becomes a fully coupled interior-atmosphere simulation, run within the PROTEUS framework, rather than a cheap static approximation. The team tested the method on three synthetic prototypes representative of the wider exoplanet population: a young sub-Neptune, an older super-Earth, and a warm terrestrial planet.
Simulating a planet's full evolutionary history thousands of times over would normally be far too expensive for a retrieval. The paper's contribution is what makes it affordable: asynchronous Bayesian optimization, an area Ben Riegler works on as an ELLIS PhD student at the Helmholtz AI Institute in Munich, supervised by Vincent Fortuin. Following their methodology published at AISTATS 2026 (arXiv), rather than waiting for an entire batch of simulations to finish before choosing the next set of parameters to try, each worker picks its next guess and starts immediately. That change means a complete retrieval, 100 simulations, can run in a few hours on fewer than ten CPU cores.
We asked Vincent and Ben more on the topic.
Asynchronous Bayesian optimization is normally applied to fairly abstract benchmark problems. What was different, or harder, about adapting it to a real multi-physics simulator like PROTEUS?
Vincent Fortuin: I always find it humbling to work with applied domain scientists because it makes you realize that many machine learning benchmarks are based on overly simplified assumptions compared to the real world. For instance, in this project, some parameters can strongly influence planetary evolution even with small changes, while others have almost no influence at all. Incorporating this into our Gaussian process prior took a lot of interaction with the physicists, who have this prior knowledge.
Ben Riegler: Making BO work on standardized benchmarks is significantly easier, because the objective function is usually well understood and not expensive to evaluate. So one can try out many tweaks of the algorithm, which always require evaluating the objective, until it works. In this work, the objective is truly black-box and expensive to evaluate, so the kind of tweaking BO researchers do for benchmarks is computationally infeasible. To overcome this, we prototyped our method with a cheaper and lower fidelity version of PROETUS, and relied on conservative choices for certain algorithm parameters.
This is described as a proof of concept on synthetic planets. What would need to happen for evolutionary retrievals like this to be run on real observational data, and is that something your collaboration is already working toward?
Vincent Fortuin: We are definitely planning to extend this to real planets; this first paper is just meant to convince the community that we can retrieve the true parameters for simulated planets, so the method works. On real planets, nobody knows the ground-truth, so the astrophysics community will have to trust our retrievals based on the fact that our method works in principle.
Ben Riegler: Even in the case of real data, we can judge the quality of the retrieval by how well the simulated observables match the real data. If the match is (near) perfect, then the only ambiguity about the unknown parameters stems from the PROTEUS simulation itself. This is not something BO could resolve, the only thing it does is match observables in a sample efficient manner.
At the end, results vary by planet type. For a warm, rocky planet, the method recovers the original volatile inventory to within about 20 percent, along with a rough picture of the mantle and core. Sub-Neptunes and super-Earths remain harder, since core size and volatile content still blur together, much like with existing methods.
The work is a proof of concept on synthetic planets, submitted to The Astrophysical Journal, and arrives ahead of a wave of new observational data from JWST, PLATO, Roman, and the ELT that will need exactly this kind of careful interpretation. The preprint is available on arXiv.