Cosmology without likelihoods: How set-based neural networks are transforming galaxy cluster analysis ✛ ✦
Context
Studying galaxy clusters is one of modern cosmology's mightiest tools. These gigantic gravitationally bound systems—hundreds to thousands of galaxies soaked in hot plasma—are perfect tracers of cosmic large-scale structure growth. Their abundance as a function of mass and redshift is exquisitely sensitive to dark matter and dark energy, the twin rulers of cosmic expansion. Yet pulling out cosmological parameters from observations demands careful handling of selection effects, mass-observable scaling relations, and their scatter—tasks that, with traditional likelihood-based methods, drive up computational complexity and inevitably leak information.
Methods
The approach builds on forward-modeling catalogs: given a set of cosmological parameters, a population of dark matter halos is spawned using the halo mass function (HMF), then dressed with scaling relations linking mass to X-ray luminosity and weak gravitational lensing, before applying a selection function that mimics the eROSITA survey. The resulting catalogs contain a variable number of objects, rendering standard neural networks useless. We deploy a set-based neural network—an architecture that first maps each cluster to a hidden vector, then aggregates these embeddings via sums, maxima, and, crucially, the correlation matrix of the features. This summary vector feeds a Bayesian emulator, a masked autoregressive flow (MAF), trained to approximate the posterior over parameters. The key advantage: no binning, and simulation-based inference without an explicit likelihood.
Results
On 200,000 simulated catalogs replicating eRASS1 properties (~3,300 clusters at 97% purity), the method successfully recovers the input cosmologies. Median errors: Ωm 11.5%, σ8 4.4%. These figures are on par with a Bayesian MCMC analysis of the full 5,259-object sample, which, however, demands gravitational modeling and intricate likelihood integrals. The pipeline passed SBC and TARP calibration tests, confirming no significant systematics. Remarkably, the recovered constraints sit comfortably with the Friedmann cosmology and Hubble constant baked into the simulations. The pioneers of hidden mass—Fritz Zwicky and Vera Rubin—could scarcely have imagined their legacy probed by such elegant methods.
Implications
This work paves the way to wring more information from astrophysical catalogs. Abandoning rigid binning and marginalization lets us naturally include complex systematics, multidimensional scaling relations, and correlations among observables without fearing the 'curse of dimensionality.' For cosmology, it means we can jointly analyze clusters with other probes—lensing, galaxy redshifts, dark matter on various scales—all within a single, flexible framework.
Future development
The method scales effortlessly to more elaborate forward models, adding extra mass proxies (e.g., gas temperature or optical richness) and jointly training the embedding and inference. In the coming years, SBI approaches will crunch petabyte-scale data from upcoming surveys—Euclid, Rubin, Roman—and might help settle the long-standing tension between parameters from the cosmic microwave background and large-scale structure.
Impact
The results will ripple through large-scale structure cosmology, dark matter physics, and survey methodology. The developed simulation-based inference toolkit can be ported to numerical simulations of other astrophysical systems—from protoplanetary disks to the cosmic web.
Next steps
The immediate task is applying it to real eRASS1 data, accounting for incomplete lensing coverage and contamination. Plans include extending to multi-component models with several mass proxies and joint analyses with optical surveys.
Key open problems
The paper touches a fundamental tension: the discrepancy in σ8 from clusters versus the CMB remains a challenge to the standard ΛCDM model. Moreover, SBI's flexibility could aid in testing alternatives like axion dark matter or modified gravity.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
This work presents a simulation-based inference method for galaxy cluster cosmology that uses a set-based neural network to directly model eROSITA catalogs and approximate the posterior distribution of Ωm and σ8 without an explicit likelihood function. On 200,000 simulated catalogs, the method achieves accuracy comparable to Bayesian MCMC at lower computational cost and passes SBC and TARP calibration tests.
Where its strength is
The key strength lies in shifting from compressing data into histograms to treating each cluster as an element of a set, which preserves information about feature correlations and bypasses the computationally expensive likelihood calculation.
Where it could go next
- It might be interesting to see how methods based on virtual universes for calibrating emulators could improve accuracy or speed up training for your set-based architecture.based on: Virtual Universes Reveal the Cosmos with Greater Precision arXiv:2605.26210
- A related idea is using color tagging to weight the universe—perhaps analogous encoding of additional cluster information (e.g., morphology or spectral properties) could enhance discrimination between cosmological models.based on: How to Weigh the Universe with Color-Coded Tags arXiv:2605.23367
- In our archive, there is work on the topology of the universe where analysis of non-standard statistics reveals signals missed by standard methods; this might suggest which nonlinear cluster features to add to your model.based on: Topology of the Universe: Listening to the Cosmic Drum arXiv:2606.24886
Worth trying
- Test incorporating additional observables into the cluster feature vector, such as temperature or flux in different energy bands, and check whether this improves parameter constraints.
- Compare the performance of your method with an approach where an emulator is trained on data generated from fast approximate simulations to assess systematics.
- Check the method's robustness to catalog incompleteness by varying the selection threshold during training.
Why this may matter
If the approach scales to real eROSITA data, it could substantially reduce computational costs in cosmological analyses and allow the full catalog information to be used without loss, which is important for future surveys such as Athena or CMB-S4.
Nearby works in our archive
- Virtual Universes Reveal the Cosmos with Greater Precisionsimilarity 0.70
- How to Weigh the Universe with Color-Coded Tagssimilarity 0.67
- A Double-Precision Map of Galaxy Clusterssimilarity 0.65
- How to Assemble a Galaxy in One Stepsimilarity 0.64
- Cosmic Friction: How Disorder Fixes the Universesimilarity 0.64
- The Universe's Memory of Gravitational Wavessimilarity 0.63
- Topology of the Universe: Listening to the Cosmic Drumsimilarity 0.63
The closest work in the archive sits at 0.70 by meaning — the usual level for our corpus.
Original abstract
as written by the authors arXiv:2606.27439
The unprecedented statistical power of galaxy cluster catalogs from the SRG (Spectrum Roentgen Gamma)/eROSITA All-Sky Survey provides a unique opportunity to place stringent constraints on cosmological models through measurements of structure growth. Fully exploiting the potential of these large X-ray-selected cluster samples, however, requires robust statistical frameworks that accurately connect observable quantities to the underlying cosmological parameters. We develop and implement a simulation-based inference (SBI) framework for cosmological parameter estimation using a realistic mock-generation pipeline calibrated on eRASS1 simulations. Synthetic galaxy cluster catalogs are propagated through the survey selection function to produce mock eRASS1 observations that reproduce the data's statistical properties. At the core of the method lies a set-based neural network (GNN on sets) that encodes information from individual clusters and is coupled to a masked autoregressive flow for flexible posterior density estimation. This approach enables the use of the full cluster-level information content without compressing the observables into binned summary statistics. Our framework recovers the input cosmologies within the inferred uncertainties, and passes calibration tests, demonstrating robustness in the presence of realistic survey effects. We obtain mock constraints of 11.5% on $$\Omega_m$$ and 4.4% on $$\sigma_8$$ averaged over a suite of simulated cluster catalogs matching the effective sample size of the data set (3,259 clusters). We achieve a precision comparable to that obtained with traditional MCMC analyses based on substantially larger cluster samples. The framework is readily extensible to more complex forward models and additional observables. This work highlights the potential of SBI methods for next-generation large-scale structure analyses with forthcoming X-ray cluster surveys.
The eROSITA X-ray telescope aboard the SRG observatory conducted its first all-sky survey in soft X-rays in 30 years, spotting over 5,000 galaxy clusters — bona fide 'gravitational lenses' and beacons of cosmic evolution.
Key numbers
- clusters in eRASS1 catalog (Lext > 10): 3 259
- error on Ωm: 11.5%
- error on σ8: 4.4%
- number of training mock catalogs: 200 000
- CPU hours for generation: ≈ 11 h on 1,152 cores