BeAR: The Bern Atmospheric Retrieval code - A GPU-accelerated framework for low-resolution, high-resolution, and joint retrievals of exoplanet atmospheres

arXiv:2611.08489 · astro-ph.EP, astro-ph.IM · Submitted 2026-11-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics Radio. Generated commentary on the latest astrophysics papers.

Vera: We're looking at the paper: "BeAR: The Bern Atmospheric Retrieval code - A GPU-accelerated framework for low-resolution, high-resolution, and joint retrievals of exoplanet atmospheres".

Jocelyn: Here's the gist of the paper. Joint retrievals of ground-based high-resolution and space-based low-resolution exoplanet spectra are becoming more common This paper presents BeAR, the Bern Atmospheric Retrieval code and its new high-resolution crosscorrelation framework for emission and transmission spectra because it makes tests requiring many retrievals practical by performing joint retrievals by adding the log-likelihoods of highand low-resolution data computed from the same model atmosphere The framework implements established cross-correlation log-likelihoods, with the line-depth scaling either fitted analytically or sampled as a free parameter, their per-pixel generalization, fast model filtering, and model re-injection adapted to this filtering BeAR performs joint retrievals by adding the log-likelihoods of highand low-resolution data computed from the same model atmosphere.

Vera: Framework Implementation BeAR is an open-source Bayesian retrieval code for the atmospheres of exoplanets and brown dwarfs It was first developed under the name Helios-r2 (Kitzmann et al.

Jocelyn: 2020) as part of the Helios family of codes, which also includes the atmospheric model Helios (Malik et al.

Vera: 2017) and the opacity calculator Helios-k (Grimm & Heng 2015; Grimm et al.

Title and authors: Vera: So we're looking at this paper now, "BeAR: The Bern Atmospheric Retrieval code - A GPU-accelerated framework for low-resolution, high-resolution, and joint retrievals of exoplanet atmospheres." It’s about making tests involving many different types of data practical by doing joint retrievals.

Jocelyn: Right. This paper tackles the problem where you have these high-resolution ground observations and these lower resolution space observations, and testing them separately is too slow because you need a low-resolution counterpart to break degeneracies.

Subrahmanyan: It’s interesting because it suggests that combining the likelihoods from both data types computed using the same atmosphere model can constrain parameters like the absolute abundance scale or even the temperature gradient better than either dataset alone.

Vera: Exactly. The paper describes BeAR as an open-source Bayesian retrieval code, and it’s built with a lot of GPU acceleration, which is key to making these long tests feasible on modern hardware.

Jocelyn: I noticed they mention the core of the code is written in C++ and NVIDIA’s CUDA language, with all the heavy lifting on the GPU for things like opacity interpolation and likelihood evaluation. That sounds like it could really speed up those model evaluations.

Subrahmanyan: From a theoretical side, that computational speed implies we can explore parameter spaces much more thoroughly when we are testing atmospheric models, which is crucial for understanding things like composition.

Vera: The paper details how they implement established cross-correlation log-likelihoods, and they mention fitting the line-depth scaling either analytically or as a free parameter, along with fast model filtering adapted to that filtering.

Jocelyn: That sounds like a real efficiency gain because repeating those slow detrending steps for every model would take up all the time. It’s about reducing the computational cost of setting up each test run significantly.

Subrahmanyan: And they also mention how they adapt model re-injection to this filtering, which shows a careful thought went into ensuring the speed doesn't compromise the accuracy of what we're inferring about the planet.

Vera: When you look at their validation results, they used separate retrievals of Gemini South/IGRINS and JWST/NIRSpec day spectra for WASP-77A b, and BeAR reproduced published abundances, C/O ratios, and metallicities to within one point five sigma <ref:2611.08489#pg2,C/O ratios, and metallicities>.

Jocelyn: So they confirmed a sub-solar metallicity while also finding a near-solar C/O ratio using this joint approach. That's a pretty strong result for confirming the atmospheric structure of that specific hot Jupiter.

Subrahmanyan: For the bigger picture, seeing confirmation on both those parameters simultaneously from different observational regimes really builds confidence in the model’s ability to resolve these chemical and thermal structures under real-world constraints.

Vera: The paper suggests several improvements for future work, like comparing temperature profiles at fixed pressures instead of through their free parameters, and reporting the line contrast of joint fits as a consistency check.

Title and authors: Jocelyn: That makes sense; you need a way to check if the different regimes are talking about the same physical thing when you combine them. Comparing evidence only between runs using the same sampler is another point they bring up for checking reliability.

Subrahmanyan: And calibrating the posterior widths by injecting data into real residuals is a solid way to get a more honest measure of how uncertain we really are about those inferred parameters.

Vera: The paper notes that BeAR can perform joint retrievals by simply adding the log-likelihoods of high and low-resolution data computed from the same model atmosphere. This is their main mechanism for making these complex tests practical.

Jocelyn: So, to put it plainly, they’re taking the likelihoods from both datasets and summing them up to get a better constraint on the atmosphere than either dataset could give independently. It's about what you can constrain when you combine things.

Subrahmanyan: This combination is particularly powerful because as they point out, fitting high- and low-resolution data with one atmosphere can break degeneracies that neither regime resolves alone, which directly impacts our understanding of planetary chemistry and structure.

Vera: Their computational performance has also improved a lot; they say retrieval calculations that used to take about two hours in earlier work can now be finished in roughly two minutes when using modern GPUs. With sixty-four GPUs, they generate more than three million forward models per minute.

Jocelyn: That speed is what makes the whole thing practical, Vera. It moves these complex tests from being a theoretical exercise to something you can actually run and iterate on quickly.

Subrahmanyan: And that kind of speed allows us to test the model against a much wider range of atmospheric scenarios, which is necessary when we're trying to pin down things like the mean molecular weight or scale height if we don't know the background gas composition.

Vera: So, to wrap up this discussion on "BeAR: The Bern Atmospheric Retrieval code - A GPU-accelerated framework for low-resolution, high-resolution, and joint retrievals of exoplanet atmospheres," the implication is that we now have a much faster and more robust way to test atmospheric models by combining ground and space data.

Jocelyn: It gives us a better tool to check if our atmospheric physics assumptions hold up when we look at different observational setups, which is something I always want to do when I’m looking at time-domain signals.

Subrahmanyan: For cosmology and structure formation, this approach shows how combining different scales of data can yield stronger constraints on fundamental parameters than relying on just one type of measurement.

Vera: We'll keep an eye on how this framework evolves, but for now, BeAR seems like a solid step forward in making these joint retrieval tests much more accessible.

The paper's summary: Vera: So, we just talked about how BeAR makes testing atmospheric models practical by combining high and low resolution data using joint likelihoods from the same model atmosphere and GPU acceleration to speed things up a lot.

Jocelyn: Right, so it’s like taking two different pictures of a planet—one super sharp, one blurry but wide—and instead of looking at them separately, you stack their best guesses together when they use the same underlying physical model.

Subrahmanyan: And that stacking isn't just math; it’s about using the constraints from both scales to lock down parameters like the temperature gradient and the absolute abundance scale simultaneously, which neither dataset could do on its own.

Vera: Exactly, and this framework allows them to test analysis choices that used to take hours or even days by generating millions of forward models per minute on a GPU cluster.

Jocelyn: It’s about efficiency for testing; they can run way more tests in the same amount of time, which means we can explore a much wider space of atmospheric possibilities.

Subrahmanyan: And that speed lets them probe those parameter spaces more thoroughly, which is crucial for understanding the chemical and thermal structure if we don't have those constraints otherwise.

Vera: The paper shows they reproduced published abundances and C/O ratios for WASP-77A b within one point five sigma when doing these joint retrievals with both Gemini South data and JWST data <ref:2611.08489#pg2>.

Jocelyn: That’s a pretty strong validation, showing that the joint method actually works to confirm things like sub-solar metallicity while also finding a near-solar C/O ratio.

Subrahmanyan: That dual confirmation of chemical ratios from different regimes really lends weight to the physical structure they're inferring.

Vera: The paper also lays out some specific improvements for future work, like comparing temperature profiles at fixed pressures instead of relying on free parameters, and checking if the joint fits are consistent by looking at line contrast.

Jocelyn: It’s smart to check that consistency; you need to make sure the constraints from the high resolution data align with what you get from the lower resolution data when you put them together.

Subrahmanyan: And calibrating those posterior widths by injecting data into real residuals is a good way to get a more honest feel for how much uncertainty we really have about those inferred values.

Vera: So, BeAR is essentially this open-source, GPU-accelerated tool that lets astronomers run many more tests much faster by intelligently summing the likelihoods from high and low resolution data computed from one single atmosphere model.

Jocelyn: It’s about making these complex comparisons practical for a lot more people using modern hardware.

Subrahmanyan: This is moving us toward a new way of testing atmospheric physics where we don't have to wait weeks for one massive computation, but can iterate on things much quicker across multiple observational constraints.

Vera: It shows that the combination of high and low resolution data, when handled this way, gives us tighter constraints on planetary properties than either measurement could provide alone.

Jocelyn: It’s a big step for how we interpret exoplanet spectra because it lets us see the interplay between different observational techniques more clearly.

Subrahmanyan: This kind of framework is what we need to move beyond just finding individual planets and start building robust models of planetary systems in context.

The paper's improvements: Tom: So, we're shifting gears to talk about how BeAR suggests ways to make this retrieval code even better in the future by changing how they handle things like line broadening and temperature profiles.

Vera: Right, so they suggest moving away from just fitting temperature profiles using free parameters and instead comparing them at fixed pressures.

Jocelyn: That makes sense because it gives you a more consistent physical comparison across different atmospheric layers rather than just fitting a curve to the data points.

Subrahmanyan: If you fix the pressure levels, you’re testing the underlying physics of how energy moves through that gas, which is what we really want to know about planetary interiors.

Vera: And they also want us to report the line contrast when we do these joint fits; that should be a check to see if the two datasets are actually giving you consistent results.

Jocelyn: So, if the high resolution and low resolution data are talking about different things, you’ll see it in how much those spectral lines stand out when you combine them.

Subrahmanyan: It's a good way to test if the physics they are assuming holds up across these different observational scales.

Vera: They also recommend comparing evidence only between runs that use the same sampler, which helps you figure out if your results are stable or just random noise from the sampling process itself.

Jocelyn: And calibrating those posterior widths using injections into real residuals is a way to get a more accurate handle on how much uncertainty is actually in the inferred abundance ratios.

Subrahmanyan: That’s important because knowing how uncertain we are about chemical abundances directly impacts how much weight we give those measurements when building bigger models of planet formation.

Vera: So, these suggestions focus on tightening up the statistical rigor and making sure the physical interpretations across different data regimes actually match up.

Jocelyn: It’s about moving from just getting a number to really understanding the reliability and consistency of that number across all the inputs.

Subrahmanyan: This level of refinement is what's needed if we want to use these atmospheric retrievals to truly constrain planetary evolution models for things like brown dwarfs later on.

Conclusion: Vera: So we’re wrapping up our look at BeAR: The Bern Atmospheric Retrieval code - A GPU-accelerated framework for low-resolution, high-resolution, and joint retrievals of exoplanet atmospheres.

Jocelyn: Basically, this paper shows how combining high resolution data with lower resolution data using the same atmosphere model allows them to test many more scenarios much faster than before.

Subrahmanyan: It’s about making these complex tests practical for a lot more people so we can explore the parameter spaces for planet chemistry and structure more thoroughly.

Vera: The main implication is that we can get tighter constraints on things like C/O ratios and metallicities by using both types of observations together.

Jocelyn: And they validated this with results on WASP-77A b, confirming a sub-solar metallicity and a near-solar C/O ratio through that joint approach <ref:2611.08489#pg2>.

Subrahmanyan: If we can reliably get those kinds of constraints from combining these different scales of data, it becomes a key tool for understanding how different planets evolve in their environments.

Vera: They also highlighted some important future steps, like comparing temperature profiles at fixed pressures instead of just using free parameters, and checking line contrast as a way to confirm the joint fits are working correctly.

Jocelyn: It’s about making sure the physical interpretations we get from combining data actually align with what we expect from atmospheric physics.

Subrahmanyan: If those refinement steps work, it opens up a path for using these tools to build more robust models of planetary systems over time.

Vera: So, BeAR gives us a much more efficient and powerful way to test our atmospheric assumptions using both ground and space data sets simultaneously.

Jocelyn: It’s a solid piece of software that makes complex retrieval tests accessible to a wider community in the field.

Subrahmanyan: This work on BeAR shows how combining different observational inputs can yield stronger constraints on fundamental parameters than relying on just one type of measurement.

E. Wang, D. Kitzmann, H. J. Hoeijmakers, J. P. Wardenier, H. Lecoq-Molinos

Space Research and Planetary Sciences, Physics Institute, University of Bern · Center for Space and Habitability, University of Bern · Lund Observatory, Division of Astrophysics, Department of Physics, Lund University

astro-ph.EP, astro-ph.IM

Submitted: 2026-11-05

Updated: 2026-11-05

Comments: Submitted to A&A

Code: https://github.com/NewStrangeWorlds/BeAR

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 83/100

The gist: Joint retrievals of ground-based high-resolution and space-based low-resolution exoplanet spectra are becoming more common This paper presents BeAR…

Key concepts

Joint Retrievals
This technique combines data from two different observation resolutions—low and high—using the same underlying atmospheric model. By summing their respective log-likelihoods, the retrieval can constrain parameters that neither dataset could determine alone, such as absolute abundance scales and temperature gradients.
GPU Acceleration
BeAR is written in C++ using NVIDIA's CUDA language to run computationally intensive steps on a GPU. This allows for massive parallel processing, enabling the generation of millions of forward models per minute and significantly speeding up retrieval calculations that previously took hours.
Cross-Correlation Framework
This framework implements established log-likelihood methods for comparing different spectral resolutions. It includes features like line-depth scaling (fitted or sampled), per-pixel generalization, and fast model filtering, which are adapted to efficiently handle the combination of high and low-resolution data.
Bayesian Inference
BeAR uses nested sampling as its Bayesian inference method. This technique provides not only the posterior distributions for atmospheric parameters but also the Bayesian evidence, which is crucial for comparing different models and determining which ones are statistically preferred.

Terminology

Summary

Joint retrievals of ground-based high-resolution and space-based low-resolution exoplanet spectra are becoming more common This paper presents BeAR, the Bern Atmospheric Retrieval code and its new high-resolution crosscorrelation framework for emission and transmission spectra because it makes tests requiring many retrievals practical by performing joint retrievals by adding the log-likelihoods of highand low-resolution data computed from the same model atmosphere The framework implements established cross-correlation log-likelihoods, with the line-depth scaling either fitted analytically or sampled as a free parameter, their per-pixel generalization, fast model filtering, and model re-injection adapted to this filtering BeAR performs joint retrievals by adding the log-likelihoods of highand low-resolution data computed from the same model atmosphere. p. 0

Framework Implementation

BeAR is an open-source Bayesian retrieval code for the atmospheres of exoplanets and brown dwarfs It was first developed under the name Helios-r2 (Kitzmann et al. 2020) as part of the Helios family of codes, which also includes the atmospheric model Helios (Malik et al. p. 1 2017) and the opacity calculator Helios-k (Grimm & Heng 2015; Grimm et al. p. 1 2021). The core of BeAR is written in C++ and NVIDIA’s CUDA language, with all computationally heavy steps running on the GPU, namely opacity interpolation, optical depths, radiative transfer, convolution with the instrument profile, binning to the observed wavelengths, and likelihood evaluation. p. 1 Bayesian inference uses nested sampling (Skilling 2006), which provides posterior distributions together with the Bayesian evidence for model comparison. p. 1

High-Resolution Retrieval

The high-resolution extension of BeAR is described in Section 3, where each high-resolution observation is read from its own observation file and a separate high-resolution spectral grid is computed. The emission model returns the planetary flux Fp(λ), which is broadened (Sect. p. 2 3.3) and then divided by a stellar spectrum. p. 2 The posterior sampler proposes θatmosphere and broadening parameters, δKp, δVsys, δϕ, and optionally α. p. 3

Joint Likelihood Evaluation

The total log-likelihood is given by ln L = Pk Pj ln Lkj + ln LLR. For joint retrievals, the low-resolution data, same atmosphere, is evaluated with its own likelihood and the log-likelihoods are summed. p. 5 The joint retrieval adds the two NIRSpec files to the observation list of the IGRINS retrieval and keeps its free parameters. p. 9

Validation Results

In separate retrievals of Gemini South/IGRINS and JWST/NIRSpec dayside spectra of the hot Jupiter WASP-77A b, BeAR reproduces published abundances, C/O ratios, and metallicities to within 1.5σ. The joint retrieval confirms a sub-solar metallicity and a near-solar C/O ratio. p. 0

Conclusion and Recommendations

BeAR is an open-source, GPU-accelerated code for low-resolution, high-resolution, and joint retrievals that makes tests requiring many retrievals practical. We recommend stating the line-broadening treatment and the spectral grid, comparing temperature profiles at fixed pressures rather than through their parameters, reporting the line contrast of joint fits as a consistency check, comparing evidences only between runs with the same sampler, and calibrating the posterior widths with injections into the real residuals. The full IGRINS retrieval takes 27 to 42 min on a single NVIDIA GeForce RTX 5080 and about 14 min on eight RTX 4090 GPUs. p. 0

Pruning redundant reasoning tokens

The paper presents BeAR, the Bern Atmospheric Retrieval code and its new framework for high-resolution crosscorrelation retrievals because it makes tests requiring many retrievals practical by performing joint retrievals by adding the log-likelihoods of highand low-resolution data computed from the same model atmosphere.

How the training signal is built

The posterior sampler proposes θatmosphere and broadening parameters, δKp, δVsys, δϕ, and optionally α. The free profiles can be described by piecewise polynomials on a finite-element discretization in log-pressure (Kitzmann et al. p. 2 2020), by cubic B-splines, or by a monotone piecewise-cubic Hermite interpolant (PCHIP; Fritsch & Carlson 1980). p. 2

Where the gains come from

The computational performance of the model has been steadily improved to a point where retrieval calculations that took about two hours in Kitzmann et al. (2020) can now be completed within roughly two minutes. p. 5 p. 5 With 64 GPUs, BeAR generates more than three million forward models per minute in both scenarios. p. 6 p. 6 The full retrieval with up to eight GPUs on UBELIX takes about 14 minutes. p. 0

The computational performance of the model has been steadily improved to a point where retrieval calculations that took about two hours in Kitzmann et al. (2020) can now be completed within roughly two minutes. With 64 GPUs, BeAR generates more than three million forward models per minute in both scenarios.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. GPU acceleration for atmospheric retrieval allows for millions of models per minute on a GPU cluster, making tests requiring many retrievals practical. This enables efficient testing of analysis choices that previously took hours to days.

  2. The framework implements established cross-correlation log-likelihoods, with the line-depth scaling either fitted analytically or sampled as a free parameter, and fast model filtering adapted to this filtering. This reduces retrieval runtimes significantly compared to repeating slow detrending steps for every model.

  3. Joint retrievals of high and low resolution data are performed by adding the log-likelihoods of high and low-resolution data computed from the same model atmosphere, allowing the joint fit to constrain parameters that neither regime resolves alone, such as the absolute abundance scale and the temperature gradient.

  4. The system can perform retrieval tests by using injection–recovery tests show that the posteriors underestimate the uncertainties of abundance ratios by about a factor of two, providing a calibrated measure of parameter uncertainty based on real data residuals.

  5. The framework can be used to compare different retrieval codes, as it allows for comparisons of individual temperature-profile parameters and demonstrates that the gain from adding low-resolution data depends on the forward model as well as on the data.

Sources

Related papers