21cmEMUv3: a hybrid diffusion-LSTM emulator of 21cmFAST summary observables

arXiv:2606.00219 · astro-ph.CO, astro-ph.GA, cs.LG · Submitted 2026-05-29 · 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: "21cmEMUv3: a hybrid diffusion-LSTM emulator of 21cmFAST summary observables".

Title and authors: Vera: So, to recap where we are, this paper introduces "21cmEMUv3: a hybrid diffusion-LSTM emulator of 21cmFAST summary observables <ref:2606.00219#pg1,21cmEMUv3: a hybrid diffusion-LSTM emulator of 21cmFAST summary observables>." The main idea is using a combination of a score-based diffusion model for the cylindrical power spectrum and LSTM networks for the other six summaries.

Jocelyn: And what they’re really showing us is that this setup achieves sub-percent median accuracy across all those seven different summary observables. It’s not just one number, it’s consistently good across the board.

Subrahmanyan: This consistency is a strong indicator that the model has successfully captured the underlying physics governing both atomic and molecular cooling processes in these simulated galaxies. The complexity of modeling two distinct populations—ACGs and MCGs—is being handled effectively.

Vera: They detail how they set up their galaxy model, showing how the stellar mass of an ACG or a MCG is related to the halo mass using equations involving baryon and matter energy densities and some scaling relations for stellar fractions. That shows the physical framework they’re working within.

Jocelyn: And because they trained this on nearly forty-seven thousand simulations, including some with higher signal-to-noise ratios, their results are being used to inform priors about star formation inside molecularly cooling galaxies when interpreting HERA data.

Subrahmanyan: That connection is crucial for cosmology. By using these simulation-informed priors instead of just arbitrary assumptions, we can better constrain the evolution of structure in the universe during this cosmic dawn period, which is what the 21cm signal probes <ref:2606.00219#pg1>.

Vera: They are also tackling sample variance by naturally marginalizing over it in both their one dimensional and two dimensional power spectrum calculations. That means their inference isn't just sensitive to one specific noise realization from a simulation.

Jocelyn: So, when you combine the fast emulation with the careful handling of those statistical uncertainties, you get a very robust tool for extracting cosmological information from noisy data sets.

Subrahmanyan: It’s about moving away from simply fitting parameters to data and toward using detailed simulations to build a prior that respects known physical laws about galaxy formation. That's where the real power is for structure formation studies.

The paper's summary: Vera: Now, let’s talk about what makes this version v3 better than the previous one. They point out several improvements in the methodology that lead to these better performance numbers and broader applicability.

Jocelyn: One big thing is that they improved the dimensionality handling for the twenty-one-cm power spectrum emulation <ref:2606.00219#pg1>. Instead of just doing a simple one dimensional PS, they’re emulating it with thirty two bins along the sky plane wavemode axis and sixty four bins along the line of sight wavemode axis.

Subrahmanyan: That expansion to higher resolution is significant because interferometric data often gives us those spatial details, and this emulator is designed to handle that complexity much better than simpler one dimensional PS emulation did in v1.

Vera: And they mentioned that the mean error for the log two dimensional power spectrum eight dropped to zero point four five percent, which was compared against a previous log one dimensional power spectrum error of zero point seven three percent.

Jocelyn: That jump is driven by the enhanced performance of both those diffusion and LSTM architectures working together as a hybrid system. It’s the combination that really pays off here.

Subrahmanyan: From my side, I’m looking at how they handle the input space. They are using eleven input parameters to condition their emulator, which suggests a well-defined set of astrophysical uncertainties that drive the galaxy properties we care about.

Vera: And they also mentioned that this entire process runs incredibly fast, producing SKA parameter forecasts in about one GPU hour, which is over six orders of magnitude faster than running direct simulations on CPUs. That’s a massive leap in computational feasibility for us.

Jocelyn: That speed is what makes it useful for real-time analysis or rapid exploration of the parameter space across different SKA layouts, like AA4, AA star, and P1. It lets us test many more scenarios quickly than we could before.

Subrahmanyan: The fact that they can do this joint eleven parameter inference with the mock one dimensional power spectrum and recover the EoR history with a confidence interval of delta xHI of zero point zero nine eight, zero point zero eight five, and zero point zero seven seven for P1, AA star, and AA four. That’s a concrete result we can actually use to check against other cosmological models.

The paper's improvements: Vera: So wrapping up this discussion on "21cmEMUv3: a hybrid diffusion-LSTM emulator of 21cmFAST summary observables," we see a system that is highly accurate, fast, and capable of handling complex input parameters <ref:2606.00219#pg1,21cmEMUv3: a hybrid diffusion-LSTM emulator of 21cmFAST summary observables>. It provides simulation-informed interpretations for current HERA upper limits.

Jocelyn: And it gives us the ability to rapidly test various observational scenarios, especially when looking at SKA forecasts. It’s a tool that bridges the gap between heavy simulation work and actual observational data analysis in a very practical way.

Subrahmanyan: For me, this is about making our theoretical models more predictive by using these emulator priors derived directly from the simulations they run. It moves us closer to understanding structure formation physics with higher fidelity than we could get otherwise.

Vera: So, it grounds those molecular galaxy priors in theoretical predictions from state-of-the-art hydrodynamical simulations, which is a very strong step forward for how we interpret 21cm data <ref:2606.00219#pg1>.

Jocelyn: It’s an emulator that produces SKA parameter forecasts in about one GPU hour, showing that the scientific return of SKA-Low remains strong even with some deployment schedule deferral.

Subrahmanyan: We have to remember that this work uses these simulation-informed priors to weaken the constraints on X-ray luminosity per unit star formation when reinterpreting those HERA upper limits. That’s a specific physical implication we need to keep in mind for future studies.

Conclusion: Vera: So this paper presents 21cmEMUv3, which is this hybrid diffusion LSTM emulator trained on 21cmFAST simulations to model both atomic and molecularly cooling galaxies with sub-percent accuracy across seven summary observables <ref:2606.00219#pg1>.

Jocelyn: It’s pretty impressive because they’re using a score-based diffusion model for the cylindrical power spectrum and then LSTM networks for the other six summaries, and they hit sub-percent accuracy on all of them.

Subrahmanyan: That level of consistency across those different statistics suggests the emulator is really capturing how these different cooling processes interact in these simulated galaxies.

Vera: And this isn't just about running a simulation; it’s about using that simulation data to inform priors for star formation inside molecularly cooling galaxies when we look at HERA upper limits.

Jocelyn: That’s the practical application, right? Using the emulator to reinterpret those 21cm PS upper limits from HERA, and then using those priors to constrain things about star formation that we couldn't pin down before <ref:2606.00219#pg1>.

Subrahmanyan: It means we're grounding our assumptions about molecular cooling in detailed hydrodynamical simulations from 21cmFAST, which is a solid way to build those priors <ref:2606.00219#pg1>.

Vera: And the speed is what really gets me—this emulator produces SKA parameter forecasts in about one GPU hour, which is over six orders of magnitude faster than running direct simulations on CPUs.

Jocelyn: That speed makes it super useful for testing different SKA layouts quickly, like AA4 or P1, and seeing how the sensitivity changes fast.

Subrahmanyan: It’s a major computational win for anyone trying to explore the parameter space of cosmological structure formation with these kinds of simulations.

Vera: They also handled sample variance by naturally marginalizing over it in both their one dimensional and two dimensional power spectrum calculations, which keeps the errors smaller than just looking at individual simulation noise realizations.

Jocelyn: That’s smart because it means the inferences are based on the mean signal rather than being overly sensitive to a specific noise realization from a particular simulation.

Subrahmanyan: Overall, this paper presents 21cmEMUv3, which jointly predicts seven summary observables using that diffusion and LSTM setup, providing a powerful tool for simulation-informed interpretation of 21cm upper limits <ref:2606.00219#pg1>.

Vera: It’s been really interesting to see how they’ve put these two very different modeling techniques together to get such high accuracy on both the power spectrum and the time-evolving summaries.

Jocelyn: Yeah, it shows that combining a diffusion model for spatial structure with an LSTM for temporal evolution is a good strategy when you have data like this.

Subrahmanyan: It’s really showing how these hybrid approaches can handle complex physical modeling while still being fast enough to be useful in actual analysis pipelines.

D. Breitman, A. Mesinger, S. G. Murray, I. Nikolić, R. Trotta

Research Center for the Early Universe, Graduate School of Science, The University of Tokyo · Department of Physics, Graduate School of Science, The University of Tokyo · Scuola Normale Superiore (SNS) · Department of Physics and Astronomy “Ettore Majorana”, University of Catania · Physics Department, Stellenbosch University · Cosmic Dawn Center (DAWN) · Niels Bohr Institute, University of Copenhagen · SISSA, INFN Sezione di Trieste & INFN Sezione di Trieste & Centro Nazionale di Ricerca in High Performance Computing, Big Data e Quantum Computing · Physics Department, Blackett Lab, Imperial College London

astro-ph.CO, astro-ph.GA, cs.LG

Submitted: 2026-05-29

Updated: 2026-05-29

Comments: 12 pages, 6 figures

Journal ref: A&A 715, A324 (2026)

DOI: 10.1051/0004-6361/202661186

Code: https://github.com/21cmfast/21cmFAST

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

Importance score: 100/100

The gist: 21cmEMUv3 is a hybrid diffusion-LSTM emulator trained on 21cmFASTv3 simulations that model both atomically and molecularly cooling galaxies…

Key concepts

21cmFASTv3.0.4
This is the specific hydrodynamical simulation code used to generate the training data for the emulator. It models how galaxies evolve and cool, including two types of galaxy populations: ACGs and MCGs, which are crucial for understanding 21cm signals.
Score-based Diffusion Model
This generative AI component is used to emulate the cylindrical 21cm power spectrum. Instead of just predicting a single value, it learns how to 'score' the data distribution, allowing it to create new, realistic 2D PS samples by conditioning on input parameters.
LSTM Networks
These are recurrent neural networks designed to capture temporal structures in time-series data, such as the global 21cm signal. They are used to emulate the six summary observables that describe how the signal changes over time or space within a galaxy.

Terminology

Summary

21cmEMUv3 is a hybrid diffusion-LSTM emulator trained on 21cmFASTv3 simulations that model both atomically and molecularly cooling galaxies, achieving sub-percent median accuracy across seven summary observables

The cylindrical 21cm power spectrum is emulated via score-based diffusion, while the remaining six summaries are emulated via long-short term memory (LSTM) networks, all achieving sub-percent median accuracy This emulator is used to reinterpret current 21cm PS upper limits from HERA, for the first time using state-of-the-art hydrodynamical simulations to inform priors on star formation inside molecularly cooling galaxies

Model and Training Data

The simulations use the code 21cmFASTv3.0.4 with a galaxy model that includes two distinct galaxy populations: ACGs and MCGs The free parameters include seven ACG astrophysical parameters, three MCG astrophysical parameters, and the cosmological parameter sigma8 The database consists of approximately 47k simulations total, with 44k having standard sampling and an additional 3k with higher SHMRs The priors for the X-ray parameters are based on Park et al. p. 2 (2019), while the prior for the three MCG parameters is based on Qin et al. p. 3 p. 3 (2021b)

Emulator Architecture

The emulator consists of two independent parts: a generative score-based diffusion model for the cylindrical 21-cm PS and an LSTM plus multilayer perceptron (MLP) architecture that emulates the remaining six summary observables The score-based diffusion network is trained to learn the score function in order to solve the reverse SDE, and conditioning it on the eleven input parameters allows for generating new 2D PS samples The LSTM networks are designed to capture temporal structure in timeseries data, such as the global 21-cm signal

Performance and Applications

The median (68% CL) error is 0.45% (0.81%) for the log 2D PS8 and 0. p. 6 p. 673% (2.57%) for the log 1D PS This improvement over previous versions is likely attributed to the enhanced performance of the LSTM and diffusion architectures The emulator is applied to re-interpret the HERA23 measured upper limits on the 21-cm PS, where including an additional population of MCG-hosted HMXBs weakens the implied constraints on the X-ray luminosity per unit star formation

SKA Forecasts

The SKA forecasts suggest that the 21-cm PS is detected at 14sigma, 13sigma, and 12sigma for AA4, AA∗, and P1 layouts at the most sensitive wavemode and redshift The impact of the 50-station deferral is expected to be minimal on CD/EoR sensitivity, suggesting that the scientific return of SKA-Low remains strong even under the deferred deployment schedule

Conclusion

The paper presents 21cmEMUv3, which jointly predicts seven summary observables using a score-based diffusion model for the cylindrical 21-cm PS and LSTM networks for the five time-evolving summaries The results provide a simulation-informed interpretation of current 21-cm upper limits that grounds MCG priors in theoretical predictions from state-of-the-art hydrodynamical simulations The emulator is showcased by producing SKA parameter forecasts in ~1 GPU-hour, over 6 orders of magnitude faster than with direct simulation on CPUs

Pruning redundant reasoning tokens

The cylindrical 21cm PS is emulated via score-based diffusion, while the remaining six summaries are emulated via long-short term memory (LSTM) networks, all achieving sub-percent median accuracy The cylindrical 21cm PS is emulated at a better accuracy than the 1D PS from 21cmEMUv1, despite being a significantly more complex summary statistic and being trained on a database nearly two orders of magnitude smaller The mean (68% CL) error is 0.45% (0.81%) for the log 2D PS8 and 0.73% (2.57%) for the log 1D PS The sub-percent median accuracy achieved across all summaries and redshifts is consistent with the results reported in the main text

How the training signal is built

The simulations use the code 21cmFASTv3.0.4 with a galaxy model that includes two distinct galaxy populations: ACGs and MCGs The free parameters include seven ACG astrophysical parameters, three MCG astrophysical parameters, and the cosmological parameter sigma8 The priors for the X-ray parameters are based on Park et al. p. 2 (2019), while the prior for the three MCG parameters is based on Qin et al.

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. Enhanced Inference Speed and Robustness: The 21cmEMUv3 emulator can perform forecasts for the detection of the cosmic 21cm PS with the Square Kilometre Array for different array configurations in about 1 GPU-hour, over 6 orders of magnitude faster than with direct simulation on CPUs. This allows AI systems to rapidly assess constraint capabilities across various SKA layouts.

  2. Joint Multi-Observable Inference: The system can perform joint 11-parameter inference with the SKA mock 1D PS and recover the EoR history with a 95% CI of delta xHI = 0.098, 0.085, and 0.077 for P1, AA∗, and AA4. This capability allows AI to interpret complex observational constraints by combining complementary data like the cylindrical PS with other summaries.

  3. Simulation-Informed Prior Integration: The system can use a simulation informed prior for star formation in MCGs motivated by Renaissance hydrodynamical simulations, which results in a posterior that is consistent with low-metallicity extrapolations of empirical HMXB trends and disfavours the local HMXB relation at ~93% CI. This allows AI to ground its scientific conclusions in detailed theoretical predictions from hydrodynamical simulations rather than arbitrary priors.

  4. Improved Dimensionality Handling: The system can emulate the cylindrical (2D) 21-cm PS with 32 bins along the sky-plane wavemode axis and 64 bins along the line-of-sight wavemode axis using a scorebased diffusion emulator, which is noted to be a significant expansion to previous versions both in terms of the physical model as well as the network architecture. This enables higher resolution and accuracy for interferometric data compared to 1D PS emulation.

  5. Sample Variance Mitigation: The system can naturally marginalise over sample variance in both the 1D and 2D PS, since each training sample has a different Gaussian IC realisation, leading to an emulator error that is smaller than the sample variance error of individual realizations (by 7% and 5% for the 2D and 1D PS respectively). This ensures that inferences are based on the mean signal rather than being overly sensitive to specific simulation noise realizations.

Sources

Related papers