Pith. sign in

REVIEW 4 major objections 5 minor 64 references

Variational autoencoder for generating realistic $N$-body simulations for dark matter halos

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A variational autoencoder trained on one cosmological simulation can generate new dark-matter density maps whose matter power spectrum matches the reference simulation over the resolved scales, offering a fast, lightweight alternative to…

desk verdict A plausible VAE proof of concept undercut by circular validation and a wrong resolution limit; deserves revision, not rejection. read the letter →

arxiv 2507.18054 v2 pith:HJBEVDRV submitted 2025-07-24 astro-ph.CO

classification astro-ph.CO PACS 95.35.+d98.80.-k
keywords variationalautoencodercosmologicalsimulationsdarkmatterdensityfieldspowerspectrumgenerativemodelingdataaugmentationN-body
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a convolutional variational autoencoder trained on two-dimensional projected slices of dark matter density from a Lambda-CDM N-body simulation learns a compact latent representation of large-scale structure. Decoding random latent vectors yields new synthetic density fields whose matter power spectrum agrees with the reference simulation and the Lambda-CDM expectation over the resolved range, while offering a more than order-of-magnitude speedup compared to the simulation cost. The authors also show that the small-scale smoothing inherent to the VAE can be repaired by a scale-dependent filter and by extrapolating the spectrum with a physically motivated fitting function. If right, this establishes the VAE as a practical generative emulator for projected dark matter fields, supporting fast simulation generation and data augmentation.

What carries the argument

The load-bearing object is the variational autoencoder itself: an encoder that maps each 256x256 density slice to a mean and log-variance in a low-dimensional latent space, a KL-regularized Gaussian prior that encourages the latent space to follow a standard normal distribution, and a decoder that maps sampled latent vectors back into density fields. The reparameterization trick makes the sampling differentiable, so the ELBO loss can be optimized with Adam; hyperparameters are selected by NSGA-II, a multi-objective evolutionary search, and training uses random rotations and Gaussian noise masks applied to slices taken along the x-y, x-z, and y-z projections. The argument is carried by the matter power spectrum $P(k) = \langle |\tilde{\delta}(k)|^2 \rangle$ computed from the generated fields: it is the summary statistic that demonstrates agreement with the reference simulation and with Lambda-CDM, and it motivates the two post-processing stages, a scale-dependent filter and a parametric or Gaussian-process extrapolation, that repair the small-scale suppression.

What would settle it

Train the same VAE on one Enzo run with a given initial-condition seed, then compare generated power spectra and images against a second independent N-body simulation of the same cosmology and box size. If the generated spectra deviate beyond the spread among true simulation realizations, or if a two-sample test cannot distinguish generated fields from held-out training fields while an independence test fails, the claim that the VAE generates realistic new dark-matter maps would be refuted.

Watch

Extended reading notes

Core claim

The central discovery is that a VAE with a convolutional encoder, a Gaussian latent prior, and a transposed-convolution decoder, trained on 256x256 projected density slices from a 128 Mpc/h Enzo simulation, compresses those fields into a low-dimensional latent space while retaining enough information to reconstruct them and to generate novel samples that are statistically consistent with the training data in Fourier space. The power spectra of both reconstructed and unconditionally generated images follow the reference simulation at intermediate scales and the expected Lambda-CDM behavior, while the smallest scales show an artificial suppression due to the model's smoothing. The paper shows this suppression can be corrected by post-hoc filtering and by a parametric extrapolation $P(k) = A k^{-n} e^{-\alpha k} (1+\beta k)^{-\gamma}$ that recovers the $P(k) \propto k^{-3}$ asymptotics expected for cold dark matter. The authors also report that the whole VAE pipeline uses roughly three real-time hours on a single NVIDIA T4 GPU, against about 720 core-hours for the Enzo run, a reduction of more than an order of magnitude.

Load-bearing premise

The load-bearing premise is that validating generated images against the same Enzo simulation used for training is a sufficient test of realism; if the VAE has merely memorized the finite set of training slices with their rotations and noise masks, the power-spectrum agreement would persist while the generated samples would not be independent new realizations.

Editorial extensions

If this is right

  • If the central claim holds, a cosmological VAE can act as a fast emulator: thousands of synthetic projected density fields can be produced in minutes on a single GPU instead of by rerunning an N-body code.
  • The learned latent space gives a compact, differentiable representation of large-scale structure, enabling interpolation and exploration between density-field states and mapping similar inputs to nearby latent regions.
  • Generated fields can serve as training data for downstream machine-learning models in cosmology, easing data-augmentation bottlenecks when full simulations are too expensive.
  • With the post-processing filter and extrapolation, generated spectra can be compared to theoretical Lambda-CDM predictions up to the nonlinear regime around $k \sim 1$ to $10\,h/\mathrm{Mpc}$, though the directly resolved range is intermediate scales.
  • The reported cost reduction of more than an order of magnitude makes this a reproducible baseline for generative modeling of projected large-scale structure.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Independence test: the paper validates generated samples against the same Enzo run that produced the training slices; a stronger test would train on one simulation of a given cosmology and compare the generated fields against a different simulation code or a different initial-condition seed.
  • Power-spectrum agreement is necessary but not sufficient: the smoothed VAE output could pass the two-point statistic while failing higher-order diagnostics such as the bispectrum, halo mass function, or filament connectivity, which a reader should check before relying on the fields for covariance estimation.
  • The same architecture is likely extendable to conditional generation, conditioning the latent space on cosmological parameters such as $\Omega_m$ and $\sigma_8$, turning the emulator from a single-cosmology generator into a parameterized surrogate for simulation-based inference.
  • Because the directly resolved range is limited, most of the claimed small-scale agreement comes from the extrapolation rather than from the network output; training on higher-resolution maps of the same simulation would be the clean test of whether smoothing is the true bottleneck.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The manuscript presents a convolutional variational autoencoder (VAE) trained on two-dimensional projected dark-matter density slices from a single Enzo ΛCDM simulation with 64^3 particles in a 128 Mpc/h box. The authors claim the VAE learns a compact latent representation that enables accurate reconstruction and fast generation of new synthetic realizations, and they validate the generated fields by comparing matter power spectra with the reference simulation and with ΛCDM predictions. The paper also reports an NSGA-II hyperparameter search, applies post-hoc filtering and Gaussian-process/parametric extrapolations of the power spectrum, and provides publicly available code.

Significance. If the central claims were substantiated, a lightweight generative emulator of projected dark-matter maps would be valuable for data augmentation and rapid production of synthetic cosmological fields, and the public code is a reproducible asset. The reported computational cost difference (720 CPU-core hours for Enzo vs. 3 GPU hours for the VAE pipeline) is suggestive, though not a controlled benchmark. However, the significance as demonstrated is currently limited: the validation is performed against the same simulation used for training, the quantitative comparisons rely on post-hoc filtering and fitted extrapolations rather than raw measured spectra, no residuals or error bars are provided, and the wavenumber convention is ambiguous. These issues directly affect the load-bearing claim of generating independent, realistic new realizations.

major comments (4)
  1. [Section 4.1, Figures 3-5] The quantitative validation is performed exclusively against the same Enzo simulation that provided the training slices. The comparisons labeled 'original simulation data' and 'training images' are drawn from the training distribution, so P(k) agreement is expected for any decoder that maps the regularized latent space back to the training manifold; it does not demonstrate that unconditionally sampled latent vectors produce statistically independent new realizations. Please add a held-out test (e.g., a different simulation volume, a train/test split of slices, or an independent N-body run such as CAMELS) and report residuals with error bars rather than visual overlap. In addition, provide a quantitative diversity measure (e.g., nearest-neighbor distances between generated and training images in pixel or latent space) to support the assertion in Section 5 that the model 'generalizes well rather than memorizing the training data.'
  2. [Sections 4.2-4.3, Eq. (15)] The paper does not specify whether k is in radians per unit length (standard cosmological convention k=2π/λ) or cycles per unit length. Under the standard convention, the fundamental mode of a 128 Mpc/h box is k ≈ 2π/128 ≈ 0.049 h/Mpc, so the stated fit interval 10^{-3} < k < 0.025 h/Mpc in Section 4.3 lies entirely below the first nonzero Fourier mode of the maps; the resulting fit is unconstrained by any data and the extrapolation to k=10 h/Mpc is not 'well-sampled' as claimed. If a cycles-per-Mpc convention is intended, it must be stated explicitly and all quoted values (e.g., the k∼0.2 h/Mpc nonlinear transition and the k≲0.3 h/Mpc resolution bound in Section 4.2) must be rescaled accordingly. As written, the two conventions are conflated and the quantitative claims in Figure 5 are not interpretable.
  3. [Section 4.1, Figure 4; Section 4.3, Figure 5] The central validation relies on post-processed quantities rather than the raw generated fields. Section 4.1 applies a 'scale-dependent filter' and reports improved agreement, but the filter is not specified and the raw (unfiltered) residuals are not shown with error bars; a corrected output is not a property of the generative model. Similarly, Figure 5's 'excellent agreement' over k ∼ 10^{-1}-10^1 h/Mpc is an extrapolation of the fitted function Eq. (15) or of the Gaussian-process regression, not a measurement of the power spectrum of generated maps. Please report the raw, measured P(k) of generated fields over the full accessible range with error bars, and clearly separate measured points from fitted or extrapolated curves in all figures.
  4. [Section 4.2] The claim that P(k) is limited to k ≲ 0.3 h/Mpc for 256×256 maps of a 128 Mpc/h box is inconsistent with the pixel Nyquist frequency: with pixel size 0.5 Mpc/h, the Nyquist wavenumber is about 6.3 h/Mpc (radians per unit length). If the intended statement is that only scales below some fraction of the map size are reliably resolved, this needs to be justified quantitatively (e.g., by convergence tests or a comparison with the simulation's native resolution); as written, the stated resolution bound appears arbitrary and is in tension with the subsequent extrapolation up to k=10 h/Mpc.
minor comments (5)
  1. [Eq. (13)] The loss function is written as E[||x - x̂^2|| + D_KL(...)]; presumably the intended expression is E[||x - x̂||^2] + D_KL(...). Please correct the notation.
  2. [Abstract and Section 1] The sentence 'Variational autoencoders, in particular, present promising capabilities to generate realistic cosmological simulations, accurately capturing essential features such as dark matter halos' appears nearly verbatim in both the abstract and the introduction; please remove the duplication.
  3. [References] Reference [44] cites a diffusion-model tutorial by Calvin Luo for the ELBO in Eq. (11); the canonical citation for the VAE ELBO is Kingma and Welling (2019), which is already reference [41]. Please update the citation to avoid introducing an irrelevant reference.
  4. [Figure 3 caption] The caption says the top row shows 'a VAE-generated sample conditioned on a real input,' but the text describes the model as unconditional and the bottom row as an unconditional generation. Please clarify whether the model is conditional or unconditional, and define what 'conditioned on a real input' means in this context.
  5. [Section 4.1, runtime comparison] The runtime comparison between 720 core-hours on a 16-core CPU and 3 hours on a single NVIDIA T4 GPU is not a controlled benchmark because the hardware, implementation, and optimization levels differ. The caveat in the text is appreciated, but the 'reduction of over an order of magnitude' phrasing should be softened to reflect that the comparison is only indicative.

Circularity Check

2 steps flagged · score 6.0 of 10

The quantitative 'agreement with the reference simulation' partly reduces to a fit: the high-k power-spectrum comparison is produced by a scale-dependent filter and by Eq. (15) fitted to a low-k interval and extrapolated with an ansatz chosen to reproduce the k^-3 target, so the central validation is not independent of the VAE output.

  1. fitted input called prediction [Section 4.3, Eq. (15) and Figure 5 caption]
    "we adopted a parametric approach, fitting a smooth, empirical function to the portion of the spectrum that lies within a well-sampled and reliable intermediate range, specifically 10−3 < k < 0.025 h/Mpc. ... The calibrated model was then extrapolated up to k = 10 h/Mpc, allowing comparison with theoretical spectra of the form P (k) ∝ k−3."

    The Figure 5 caption reports 'excellent agreement' at k ∼ 10−1 to 10^1 h/Mpc, but those wavenumbers were not measured from the VAE images; they are the output of Eq. (15) fitted below k = 0.025 h/Mpc and then extrapolated. The extrapolating function was itself selected because of its ability to produce the expected small-scale slope (the paper later states 'we employ a model with a −3 slope to adequately represent the expected small-scale behavior'). Thus the claimed consistency with the ΛCDM k^-3 model is inherited from the fitting ansatz rather than independently demonstrated by the generated density fields. The 'prediction' being compared is therefore a fitted extrapolation, not an emergent property of the VAE.

  2. other [Section 4.1, paragraph on statistical filtering after Figure 3]
    "To address these deviations, we implemented a statistical filtering approach to selectively suppress noise-dominated modes at high k-values. After applying this scale-dependent filter, the agreement between the generated and target power spectra improved significantly."

    The raw VAE spectrum deviates at high k; the paper then applies a scale-dependent filter specifically designed to reduce those deviations and reports the improved agreement as evidence that the model captures the statistics of the training data. Because the correction is derived from the target spectra, the post-filter 'agreement' is partly manufactured from the comparison quantities rather than measured from the VAE's own output. This makes the quantitative validation of the generated fields weaker than the text suggests, and it contributes to the same pattern in which the good power-spectrum agreement is a post-processed or fitted quantity.

full rationale

The VAE training itself is a standard machine-learning procedure, and the paper does not rely on a load-bearing self-citation chain: the self-citations to Refs. [11, 15, 32, 36, 45, 49] are methodological or contextual rather than the source of the central claim. However, the paper's central quantitative validation is partially circular in the sense of pattern 2. The only reference simulation used for comparison is the same Enzo run that supplied the training images, so P(k) agreement at resolved scales is checked against the model's own training distribution; no held-out simulation or independent realization test is provided. More specifically, the high-k 'excellent agreement' in Figure 5 is not a property of raw VAE samples: Section 4.1 applies a scale-dependent filter to improve agreement, and Section 4.3 fits Eq. (15) to the interval 10^-3 < k < 0.025 h/Mpc and extrapolates to k = 10 h/Mpc, comparing against the same k^-3 asymptotic behavior that motivated the fit. The abstract's statement that the generated fields show 'good agreement with the reference simulation' therefore incorporates fitted and filtered quantities, so the strength of the claim exceeds what the independent VAE output demonstrates. There is also an unaddressed k-convention ambiguity (the fitted interval lies below the 2π/L fundamental mode of the box), which compounds the uncertainty but is more a correctness concern than a circularity concern.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the fidelity of one Enzo simulation, on the power spectrum as the sole quantitative validator, and on the assumption that prior-space sampling produces independent realizations. No new physical entities are introduced; the free parameters are the architecture choices and the spectral fit parameters used in post-processing.

free parameters (3)
  • VAE hyperparameters (batch size, kernel size, dense units) = batch=4, kernel=5, dense units=256 (Table 1)
    Chosen by NSGA-II optimization; they set model capacity and affect reconstruction fidelity and latent-space coverage.
  • Parametric power-spectrum fit parameters A, n, alpha, beta, gamma (Eq. 15) = not reported numerically
    Fitted by nonlinear least squares over 1e-3 < k < 0.025 h/Mpc and used to extrapolate to k=10 h/Mpc; these values, not the VAE, generate the small-scale agreement with theory.
  • Latent dimension = not stated
    The number of latent variables is never reported; it directly controls how much information can be retained and how smooth generated samples are.
assumptions (5)
  • domain assumption The single Enzo simulation (Omega_DM=0.25, Omega_b=0.05, Omega_Lambda=0.7, h=0.71, 64^3 particles, 128 Mpc/h box) is representative of ΛCDM dark matter fields.
    Section 3 defines the training data; all validation is against this simulation, so errors in the simulation become errors in the claim.
  • domain assumption The matter power spectrum is a sufficient statistic for the realism of dark matter density fields.
    Section 4.1 uses only P(k) and histograms; no bispectrum, halo mass function, or higher-order statistics are checked.
  • domain assumption Sampling z ~ N(0,I) yields valid new realizations after KL-based training.
    Section 2.4 states this generative procedure; posterior collapse or unmatched priors would make prior draws unrealistic, and the paper does not test latent coverage.
  • domain assumption The nonlinear dark matter power spectrum approaches P(k) ~ k^{-3} at small scales.
    Section 4.3 adopts this asymptotic from refs. [57-59] and uses it as the benchmark for the extrapolated spectra.
  • standard math Standard VAE mathematics (ELBO, reparameterization trick) is valid.
    Standard background in Section 2.4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Variational autoencoder for generating realistic $N$-body simulations for dark matter halos." pith.science (2026). https://pith.science/paper/HJBEVDRV

@misc{pith2026250718054,
  author       = {Pith},
  title        = {Pith review of: Variational autoencoder for generating realistic $N$-body simulations for dark matter halos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HJBEVDRV}},
  note         = {Machine review of arXiv:2507.18054}
}
abstract

In this paper, we present a deep-learning approach to generate synthetic cosmological images by training a convolutional variational autoencoder on two-dimensional dark matter density slices projected from $\Lambda$CDM $N$-body simulations. The model learns a compact latent representation that enables accurate reconstructions and fast generation of new synthetic realizations through a single forward pass through the decoder. We validate the generated fields using cosmology-based summary statistics, focusing on the matter power spectrum and related Fourier space diagnostics, and found good agreement with the reference simulation across the range of scales where the maps exhibit good resolution. Thanks to its low inference cost and stable training target, this variational-autoencoder approach provides a lightweight and reproducible basis for generative modeling of large-scale projected structures and can support downstream tasks such as fast simulation generation and data augmentation.

Figures

Figures reproduced from arXiv: 2507.18054 by the authors.

Figure 1
Figure 1. Comparison between a dark matter density field from the training dataset (left) and a sample generated [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. A 5×5 grid of dark matter density fields generated by the VAE using random samples from the latent space. The diversity and consistency of the structures demonstrate that the model learns a meaningful latent representation, enabling the generation of novel, statistically coherent realizations that reflect the large-scale features present in the training data. it is crucial to evaluate how well the generated images p… view at source ↗
Figure 3
Figure 3. Raw power spectra for generated dark matter density fields. The top row shows a reconstruction: the left panel displays a VAE-generated sample conditioned on a real input, and the right panel shows its corresponding power spectrum (red dashed line), compared with the original sample (blue solid line). The bottom row shows an unconditional generation from a random latent vector (left), along with its power spectrum (… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Post-processed matter power spectrum comparison between the original simulation data and a sample generated from the VAE’s latent space. A reference ΛCDM model, characterized by a P(k) ∝ k −3 behavior at small scales, is overlaid for comparison. In the intermediate ran…
Figure 5
Figure 5. Figure 5: The extrapolated power spectrum, derived from both the training and VAE-generated images, is compared with the theoretical prediction from the ΛCDM model. Left: Gaussian process regression. Right: Parametric extrapolation with damping function. At intermediate scales (…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 55 canonical work pages

  1. [1]

    The cosmological simulation code gadget-2.Monthly notices of the royal astronomical society, 364(4):1105–1134, 2005

    V olker Springel. The cosmological simulation code gadget-2.Monthly notices of the royal astronomical society, 364(4):1105–1134, 2005

  2. [2]

    Eisenstein

    Svetlin Tassev, Matias Zaldarriaga, and Daniel J. Eisenstein. Solving large scale structure in ten easy steps with cola. Journal of Cosmology and Astroparticle Physics, 2013(06):036, jun 2013

  3. [3]

    Simulating cosmic structure formation with the gadget-4 code

    V olker Springel, Rüdiger Pakmor, Oliver Zier, and Martin Reinecke. Simulating cosmic structure formation with the gadget-4 code. Monthly Notices of the Royal Astronomical Society, 506(2):2871–2949, 2021. 12 A PREPRINT - AUGUST 23, 2025

  4. [4]

    Enzo: An adaptive mesh refinement code for astrophysics

    Greg L Bryan, Michael L Norman, Brian W O’Shea, Tom Abel, John H Wise, Matthew J Turk, Daniel R Reynolds, David C Collins, Peng Wang, Samuel W Skillman, et al. Enzo: An adaptive mesh refinement code for astrophysics. The Astrophysical Journal Supplement Series, 211(2):19, 2014

  5. [5]

    Deep learning

    Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning. MIT press, 2016

  6. [6]

    Generative adversarial nets

    Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014

  7. [7]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020

  8. [8]

    On the design fundamentals of diffusion models: A survey

    Ziyi Chang, George Alex Koulieris, and Hubert PH Shum. On the design fundamentals of diffusion models: A survey. arXiv preprint arXiv:2306.04542, 2023

Show all 64 references
  1. [9]

    M.A. Kramer. Autoassociative neural networks. Computers & Chemical Engineering, 16(4):313–328, 1992. Neutral network applications in chemical engineering

  2. [10]

    Autoencoders, minimum description length and helmholtz free energy

    Geoffrey E Hinton and Richard Zemel. Autoencoders, minimum description length and helmholtz free energy. In J. Cowan, G. Tesauro, and J. Alspector, editors,Advances in Neural Information Processing Systems, volume 6. Morgan-Kaufmann, 1993

  3. [11]

    Analysis of dark matter halo structure formation in n-body simulations with machine learning

    Jazhiel Chacón, Isidro Gómez-Vargas, Ricardo Menchaca Méndez, and J Alberto Vázquez. Analysis of dark matter halo structure formation in n-body simulations with machine learning. Physical Review D, 107(12):123515, 2023

  4. [12]

    A deep-learning model for the density profiles of subhaloes in illustristng

    Luisa Lucie-Smith, Giulia Despali, and V olker Springel. A deep-learning model for the density profiles of subhaloes in illustristng. Monthly Notices of the Royal Astronomical Society, 532(1):164–176, 2024

  5. [13]

    Predicting dark matter halo formation in n-body simulations with deep regression networks

    Mauro Bernardini, Lucio Mayer, Darren Reed, and Robert Feldmann. Predicting dark matter halo formation in n-body simulations with deep regression networks. Monthly Notices of the Royal Astronomical Society , 496(4):5116–5125, 2020

  6. [14]

    F. et al. Villaescusa-Navarro. The camels project: Cosmology and astrophysics with machine-learning simulations. ApJ, 915:71, 2021

  7. [15]

    Deep learning and genetic algorithms for cosmological bayesian inference speed-up

    Isidro Gómez-Vargas and J Alberto Vázquez. Deep learning and genetic algorithms for cosmological bayesian inference speed-up. Physical Review D, 110(8):083518, 2024

  8. [16]

    A deep learning model to emulate simulations of cosmic reionization

    Jonathan Chardin, Grégoire Uhlrich, Dominique Aubert, Nicolas Deparis, Nicolas Gillet, Pierre Ocvirk, and Joseph Lewis. A deep learning model to emulate simulations of cosmic reionization. Monthly Notices of the Royal Astronomical Society, 490(1):1055–1065, 2019

  9. [17]

    Cosmoflow: Using deep learning to learn the universe at scale

    Amrita Mathuriya, Deborah Bard, Peter Mendygral, Lawrence Meadows, James Arnemann, Lei Shao, Siyu He, Tuomas Kärnä, Diana Moise, Simon J Pennycook, et al. Cosmoflow: Using deep learning to learn the universe at scale. In SC18: International Conference for High Performance Comp...

  10. [18]

    Linna: Likelihood inference neural network accelerator

    Chun-Hao To, Eduardo Rozo, Elisabeth Krause, Hao-Yi Wu, Risa H Wechsler, and Andrés N Salcedo. Linna: Likelihood inference neural network accelerator. Journal of Cosmology and Astroparticle Physics, 2023(01):016, 2023

  11. [19]

    Cosmological n-body simulations: a challenge for scalable generative models

    Nathanaël Perraudin, Ankit Srivastava, Aurelien Lucchi, Tomasz Kacprzak, Thomas Hofmann, and Alexandre Réfrégier. Cosmological n-body simulations: a challenge for scalable generative models. Computational Astrophysics and Cosmology, 6:1–17, 2019

  12. [20]

    Encoding large-scale cosmological structure with generative adversarial networks

    Marion Ullmo, Aurélien Decelle, and Nabila Aghanim. Encoding large-scale cosmological structure with generative adversarial networks. A&A, 651:A46, July 2021

  13. [21]

    Yin Li, Yueying Ni, Rupert A. C. Croft, Tiziana Di Matteo, Simeon Bird, and Yu Feng. Ai-assisted superresolution cosmological simulations. Proceedings of the National Academy of Sciences, 118(19):e2022038118, 2021

  14. [22]

    Superresolution emulation of large cosmological fields with a 3d conditional diffusion model

    Adam Rouhiainen, Moritz Münchmeyer, Gary Shiu, Michael Gira, and Kangwook Lee. Superresolution emulation of large cosmological fields with a 3d conditional diffusion model. Physical Review D, 109(12):123536, 2024

  15. [23]

    Can denoising diffusion probabilistic models generate realistic astrophysical fields? arXiv preprint arXiv:2211.12444, 2022

    Nayantara Mudur and Douglas P Finkbeiner. Can denoising diffusion probabilistic models generate realistic astrophysical fields? arXiv preprint arXiv:2211.12444, 2022

  16. [24]

    Cosmological field emulation and parameter inference with diffusion models

    Nayantara Mudur, Carolina Cuesta-Lazaro, and Douglas P Finkbeiner. Cosmological field emulation and parameter inference with diffusion models. arXiv preprint arXiv:2312.07534, 2023. 13 A PREPRINT - AUGUST 23, 2025

  17. [25]

    Stochastic super-resolution of cosmological simulations with denoising diffusion models

    Andreas Schanz, Florian List, and Oliver Hahn. Stochastic super-resolution of cosmological simulations with denoising diffusion models. arXiv preprint arXiv:2310.06929, 2023

  18. [26]

    D. G. et al. York. The sloan digital sky survey: Technical summary. AJ, 120:1579, 2000

  19. [27]

    T. M. C. et al. [DES Collaboration] Abbott. Dark energy survey year 1 results: Cosmological constraints from galaxy clustering and weak lensing. Phys. Rev. D, 98:043526, 2018

  20. [28]

    R. et al. Laureijs. Euclid definition study report. arXiv preprint, 2011

  21. [29]

    Lsst science book, version 2.0

    LSST Science Collaboration. Lsst science book, version 2.0. arXiv preprint, 2009

  22. [30]

    M. et al. V ogelsberger. Introducing the illustris project: Simulating the coevolution of dark and visible matter in the universe. Nature, 509:177, 2014

  23. [31]

    V . et al. Springel. Simulations of the formation, evolution and clustering of galaxies and quasars.Nature, 435:629, 2005

  24. [32]

    Neural network reconstructions for the hubble parameter, growth rate and distance modulus

    Isidro Gómez-Vargas, Ricardo Medel-Esquivel, Ricardo García-Salcedo, and J Alberto Vázquez. Neural network reconstructions for the hubble parameter, growth rate and distance modulus. The European Physical Journal C, 83(4):304, 2023

  25. [33]

    Cosmo vae: Variational autoencoder for cmb image inpainting

    Kai Yi, Yi Guo, Yanan Fan, Jan Hamann, and Yu Guang Wang. Cosmo vae: Variational autoencoder for cmb image inpainting. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1–7. IEEE, 2020

  26. [34]

    Accelerating lensed quasar discovery and modeling with physics-informed variational autoencoders

    Irham T Andika, Stefan Schuldt, Sherry H Suyu, Satadru Bag, Raoul Cañameras, Alejandra Melo, Claudio Grillo, and James HH Chan. Accelerating lensed quasar discovery and modeling with physics-informed variational autoencoders. Astronomy & Astrophysics, 694:A227, 2025

  27. [35]

    Bayesian parameter estimation using conditional variational autoencoders for gravitational-wave astronomy.Nature Physics, 18(1):112–117, 2022

    Hunter Gabbard, Chris Messenger, Ik Siong Heng, Francesco Tonolini, and Roderick Murray-Smith. Bayesian parameter estimation using conditional variational autoencoders for gravitational-wave astronomy.Nature Physics, 18(1):112–117, 2022

  28. [36]

    Alberto Vazquez, and Ruslan Gabbasov

    Jazhiel Chacon, J. Alberto Vazquez, and Ruslan Gabbasov. Dark matter with n-body numerical simulations. Rev. Mex. Fis. E, 17(2):241–254, 2020

  29. [37]

    A hierarchical O(N log N) force-calculation algorithm

    Josh Barnes and Piet Hut. A hierarchical O(N log N) force-calculation algorithm. Nature, 324(6096):446–449, December 1986

  30. [38]

    Alexander A. Vlasov. The vibrational properties of an electron gas. Uspekhi Fizicheskih Nauk, 1967

  31. [39]

    M. L. Norman and G. L. Bryan. Cosmological Adaptive Mesh Refinement, page 19–28. Springer Netherlands, 1999

  32. [40]

    Gewers, Gustavo R

    Felipe L. Gewers, Gustavo R. Ferreira, Henrique F. De Arruda, Filipi N. Silva, Cesar H. Comin, Diego R. Amancio, and Luciano Da F. Costa. Principal component analysis: A natural approach to data exploration. ACM Comput. Surv., 54(4), May 2021

  33. [41]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. An Introduction to Variational Autoencoders. Foundations and Trends® in Machine Learning, 12(4):307–392, 2019

  34. [42]

    O’Reilly Media, Inc

    David Foster. Generative deep learning. " O’Reilly Media, Inc.", 2022

  35. [43]

    On information and sufficiency

    Solomon Kullback and Richard A Leibler. On information and sufficiency. The annals of mathematical statistics, 22(1):79–86, 1951

  36. [44]

    Understanding Diffusion Models: A Unified Perspective, August 2022

    Calvin Luo. Understanding Diffusion Models: A Unified Perspective, August 2022

  37. [45]

    Neural networks optimized by genetic algorithms in cosmology

    Isidro Gómez-Vargas, Joshua Briones Andrade, and J Alberto Vázquez. Neural networks optimized by genetic algorithms in cosmology. Physical Review D, 107(4):043509, 2023

  38. [46]

    Optuna: A next-generation hyperparameter optimization framework

    Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 2623–2631, 2019

  39. [47]

    Reconstruction of dark energy and expansion dynamics using gaussian processes

    Marina Seikel, Chris Clarkson, and Mathew Smith. Reconstruction of dark energy and expansion dynamics using gaussian processes. Journal of Cosmology and Astroparticle Physics, 2012(06):036, 2012

  40. [48]

    Gaussian processes reconstruction of dark energy from observational data

    Ming-Jian Zhang and Hong Li. Gaussian processes reconstruction of dark energy from observational data. The European Physical Journal C, 78:1–12, 2018

  41. [49]

    Escamilla, Purba Mukherjee, and J

    José de Jesús Velázquez, Luis A. Escamilla, Purba Mukherjee, and J. Alberto Vázquez. Non-Parametric Reconstruction of Cosmological Observables Using Gaussian Processes Regression. Universe, 10(12):464, 2024. 14 A PREPRINT - AUGUST 23, 2025

  42. [50]

    R. E. Smith, J. A. Peacock, A. Jenkins, S. D. M. White, C. S. Frenk, F. R. Pearce, P. A. Thomas, G. Efstathiou, and H. M. P. Couchman. Stable clustering, the halo model and non-linear cosmological power spectra. MNRAS, 341(4):1311–1332, June 2003

  43. [51]

    Revising the halofit model for the nonlinear matter power spectrum

    Ryuichi Takahashi, Masanori Sato, Takahiro Nishimichi, Atsushi Taruya, and Masamune Oguri. Revising the halofit model for the nonlinear matter power spectrum. The Astrophysical Journal, 761(2):152, 2012

  44. [52]

    Machine learning unveils the linear matter power spectrum of modified gravity

    J Bayron Orjuela-Quintana, Savvas Nesseris, and Domenico Sapone. Machine learning unveils the linear matter power spectrum of modified gravity. Physical Review D, 109(6):063511, 2024

  45. [53]

    Hunting down systematics in baryon acoustic oscillations after cosmic high noon

    Francisco Prada, Claudia G Scóccola, Chia-Hsun Chuang, Gustavo Yepes, Anatoly A Klypin, Francisco-Shu Kitaura, Stefan Gottlöber, and Cheng Zhao. Hunting down systematics in baryon acoustic oscillations after cosmic high noon. Monthly Notices of the Royal Astronomical Society, ...

  46. [54]

    Measurement of the power spectrum turnover scale from the cross-correlation between cmb lensing and quaia

    David Alonso, Oleksandr Hetmantsev, Giulio Fabbian, Anze Slosar, and Kate Storey-Fisher. Measurement of the power spectrum turnover scale from the cross-correlation between cmb lensing and quaia. arXiv preprint arXiv:2410.24134, 2024

  47. [55]

    Ya. B. Zel’dovich. Gravitational instability: An approximate theory for large density perturbations. A&A, 5:84–89, March 1970

  48. [56]

    Mock galaxy catalogues using the quick particle mesh method

    Martin White, Jeremy L Tinker, and Cameron K McBride. Mock galaxy catalogues using the quick particle mesh method. Monthly Notices of the Royal Astronomical Society, 437(3):2594–2606, 2014

  49. [57]

    On the asymptotic behaviour of cosmic density- fluctuation power spectra of cold dark matter

    Sara Konrad, Yonadav Barry Ginat, and Matthias Bartelmann. On the asymptotic behaviour of cosmic density- fluctuation power spectra of cold dark matter. Monthly Notices of the Royal Astronomical Society, 515(4):5823– 5835, 2022

  50. [58]

    Widrow, Pascal J

    Lawrence M. Widrow, Pascal J. Elahi, Robert J. Thacker, Mark Richardson, and Evan Scannapieco. Power spectrum for the small-scale Universe. MNRAS, 397(3):1275–1285, August 2009

  51. [59]

    Gravitational turbulence: the small-scale limit of the cold-dark-matter power spectrum

    Yonadav Barry Ginat, Michael L Nastac, Robert J Ewart, Sara Konrad, Matthias Bartelmann, and Alexander A Schekochihin. Gravitational turbulence: the small-scale limit of the cold-dark-matter power spectrum. arXiv preprint arXiv:2501.01524, 2025

  52. [60]

    Raissi, P

    M. Raissi, P. Perdikaris, and G.E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, 2019

  53. [61]

    Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang

    George Em Karniadakis, Ioannis G. Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics- informed machine learning. Nature Reviews Physics, 3(6):422–440, June 2021

  54. [62]

    C. Modi, F. Lanusse, and U. Seljak. Flowpm: Distributed tensorflow implementation of the fastpm cosmological n-body solver. Astronomy and Computing, 37:100505, 2021

  55. [63]

    Physics-constrained deep learning for high-dimensional surrogate modeling and uncertainty quantification without labeled data

    Yinhao Zhu, Nicholas Zabaras, Phaedon-Stelios Koutsourelakis, and Paris Perdikaris. Physics-constrained deep learning for high-dimensional surrogate modeling and uncertainty quantification without labeled data. Journal of Computational Physics, 394:56–81, 2019

  56. [64]

    Latentpinns: Generative physics-informed neural networks via a latent representation learning

    Mohammad H Taufik and Tariq Alkhalifah. Latentpinns: Generative physics-informed neural networks via a latent representation learning. Artificial Intelligence in Geosciences, page 100115, 2025. 15

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.