Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Cosmology with Topological Deep Learning

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Higher-order message passing over tetrahedra, clusters, and hyperedges extracts cosmological information that edge-only graphs miss, cutting error by up to 22% for matter density and 60% for clustering amplitude on Quijote.

desk verdict First real application of TNNs to cosmology with a clean internal comparison, but the E(3)-invariance claim is false as stated because the edge features use absolute coordinates. read the letter →

arxiv 2505.23904 v1 pith:KIOLGWV2 submitted 2025-05-29 astro-ph.CO

Jun-Young Lee , Francisco Villaescusa-Navarro This is my paper
classification astro-ph.CO PACS 98.80.-k98.65.-r
keywords cosmologicalparameterinferencelarge-scalestructuretopologicalneuralnetworkscombinatorialcomplexeshigher-ordermessagepassinggraphhalocatalogsN-bodysimulations
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 tries to establish that cosmological parameter inference from halo and galaxy catalogs improves when a neural network passes messages through the higher-order structure of the cosmic web — tetrahedra, clusters of tetrahedra, and hyperedges — instead of only along pairwise edges as graph neural networks do. The authors build such topological neural networks on two simulation suites and report that on Quijote the higher-order architectures reduce mean squared error by up to 22% for the matter density $\Omega_{\rm m}$ and up to 60% for the clustering amplitude $\sigma_8$ relative to their own GNNs, with the largest gains coming from the structures graph networks cannot see. On CAMELS, a smaller hydrodynamic suite, the topological models do not beat the GNN benchmark, which the authors read as evidence that graphs already saturate the information in that dataset. If the Quijote results hold, adding topological cells is a low-cost way to sharpen constraints on the universe's composition and structure growth from the same catalogs that surveys already produce.

What carries the argument

The load-bearing object is the combinatorial complex built on a point cloud of halos or galaxies: a hierarchy of cells — rank-0 vertices (halos), rank-1 edges (pairs within a linking radius $r_{\rm link}$), rank-2 cells (Delaunay tetrahedra), rank-3 cells (HDBSCAN clusters of tetrahedra), and rank-4 cells (minimum-spanning-tree edges between clusters) — together with neighborhood functions (incidence, adjacency, coadjacency) that decide which cells exchange messages. Message passing is carried out by convolutional push-forward operations, which map a cochain on one rank to a cochain on another rank through a sparse neighborhood matrix, fused by convolutional merge nodes that aggregate across all neighborhood types before a residual update; the authors also insert distance-based invariant message matrices (Euclidean and Hausdorff distances between cell centroids) into the convolution so that every message is a scalar invariant. The hierarchy is meant as a proxy for the cosmic web's scales: tetrahedra encode three- and four-point correlations on small scales, while clusters and hyperedges carry large-scale structure, and long-range messages can be routed through high-rank cells without traversing many edges.

What would settle it

Two cheap experiments settle the claims. Keep the FullTNN architecture but replace the Delaunay tetrahedra with randomly selected quadruples of halos; if the 22% to 60% mean squared error gains over the GNN persist, the improvement is not caused by the tetrahedral structure itself. And translate every halo position in a fixed test catalog by a random offset (with periodic wrapping), rebuild the complex, and rerun the trained network; a model invariant under translations must give identical predictions for $\Omega_{\rm m}$ and $\sigma_8$, and the angle features defined in Section 3.4 change under exactly such a translation.

Watch

Extended reading notes

Core claim

In the paper's own terms, the discovery is that lifting message passing from graphs to combinatorial complexes — data structures with ranked cells (vertices, edges, Delaunay tetrahedra, HDBSCAN clusters of tetrahedra, and minimum-spanning-tree hyperedges) — gives neural networks access to cosmological information in the halo distribution that pairwise interactions miss. On the Quijote N-body suite, where the task is joint inference of $\Omega_{\rm m}$ and $\sigma_8$ from the 5,000 most massive halos per simulation, ClusterTNNs lower the mean squared error by 22% ($\Omega_{\rm m}$) and 34% ($\sigma_8$) compared with the authors' GNNs, and the FullTNN selected by the integrated hyperparameter search lowers $\sigma_8$ MSE by up to 60%; the $\sigma_8$ gains also exceed those of point-cloud and several graph-based benchmarks from prior work, despite training on fewer halos. The authors attribute the improvement to the E(3)-invariant (translation-, rotation-, and reflection-invariant), topology-aware architecture: all cell features are geometric scalars, and messages between cells of different ranks are built from Euclidean or Hausdorff distances, so the network is designed to respect the symmetries of the problem. On the CAMELS suite the same architectures fail to beat the GNN benchmark for $\Omega_{\rm m}$, a reversal the authors explain by the smaller box, low galaxy counts, and baryonic feedback, concluding that the value of higher-order topology is dataset-dependent and most pronounced where large-scale structure is well resolved.

Load-bearing premise

The load-bearing premise is that the network is invariant under translations, rotations, and reflections because every cell feature is a geometric scalar, but the angle features $\gamma_1$, $\gamma_2$, and $\eta_i$ are computed from halo positions measured from the box origin, so translating the whole catalog changes them and the claimed translation invariance does not hold as built.

Editorial extensions

If this is right

  • On Quijote-scale data, the hierarchy matters more than capacity: adding tetrahedra or clusters to the message-passing scheme buys 22% to 34% MSE reductions over the authors' GNNs without adding simulations or observables.
  • The gains are concentrated in $\sigma_8$, the amplitude of matter clustering, which is precisely the parameter where the compared point-cloud networks performed worst; this points to higher-order structure as a targeted source of clustering information.
  • The CAMELS reversal delimits the method's domain: in small, feedback-dominated boxes with sparse galaxy counts, higher-order topology adds no information for $\Omega_{\rm m}$, so TNNs are a tool for well-resolved, large-volume surveys rather than a universal upgrade.
  • Because all four architectures share the same training procedure, loss, and data splits, the error ladder from GNN to TetraTNN to ClusterTNN to FullTNN is the paper's evidence that each added cell rank contributes measurable information on Quijote.

Reading between the lines

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

  • The hierarchy construction is ad hoc — linking radius, Delaunay triangulation, HDBSCAN clustering, minimum spanning tree — so the reported gains are tied to these choices; a sweep over alternative higher-order cells (k-nearest-neighbor cliques, persistence diagrams, filaments rather than clusters) would map how much of the improvement is genuine topology and how much is engineering detail.
  • The CAMELS reversal supports a testable prediction I would want to run: TNN gains should grow monotonically with survey volume and resolution, so intermediate-scale surveys with resolved filaments but few massive clusters are where the method should pay off most; a controlled scan over box volumes spanning from CAMELS to Quijote scales would confirm or refute this.
  • The Delaunay tetrahedra are the geometric analogue of three- and four-point correlation functions, so the paper implicitly suggests that a learned hierarchical summary can capture the non-Gaussian information that the bispectrum and trispectrum chase analytically — a connection worth testing by comparing TNN errors against explicitly 2PCF-plus-3PCF-informed summaries on identical catalogs.
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

2 major / 5 minor

Summary. The paper introduces topological neural networks (TNNs) operating on combinatorial complexes built from halo/galaxy catalogs, with cells of ranks 0 through 4 (vertices, edges, tetrahedra, clusters, hyperedges). It compares TNN variants (TetraTNN, ClusterTNN, FullTNN) against GNNs on the Quijote and CAMELS simulation suites, reporting up to 60% lower MSE for sigma_8 on Quijote and comparable performance on CAMELS. The authors claim the networks are E(3)-invariant by construction, and attribute the performance gains to the combination of higher-order message passing and symmetry awareness.

Significance. The empirical comparison is carefully set up: identical train/validation/test splits across all models, bootstrap uncertainties on test metrics, and hyperparameter selection on a separate validation set. If the symmetry issue were resolved, the finding that higher-order message-passing yields significant gains on Quijote would be a valuable first demonstration of topological deep learning in cosmology, and the paper provides code to reproduce the analysis. However, the central symmetry claim is currently false, so the advertised mechanism is not established.

major comments (2)
  1. [3.4] The claim that the models are E(3)-invariant is not supported by the construction. The rank-1 edge features include gamma_1 = (x_i · x_j)/(||x_i|| ||x_j||) and gamma_2 = (d_ij · x_i)/(||x_i|| ||d_ij||), which depend on absolute positions x_i relative to the origin of the simulation box. Under a translation x -> x + t, d_ij is unchanged but gamma_1 and gamma_2 change unless t = 0. The same issue affects the rank-3 cluster features eta_i = (e_i · c)/(||e_i|| ||c||), since the cluster centroid c is not translation invariant, and the rank-4 hyperedge features Gamma_1, Gamma_2. Because these scalars are used in the message-passing updates in Eqs. (11) and (12), the network output cannot be invariant under translations. The model is at most invariant under rotations and reflections about the box origin. This should be corrected by recentering the coordinates (e.g., subtracting the catalog mean) before computing the features, or by replacing these features with genuinely translation-invariant quantities. The abstract and Section 3.4 should then be revised accordingly.
  2. [Abstract and Section 5] The reported performance gains are attributed in the abstract and conclusions to 'topology and symmetry-aware' networks. Since the E(3)-invariance claim is false, the mechanism for any generalization benefit is not established. The empirical MSE comparison on Quijote may still be valid, but the authors should either enforce translation invariance and verify it empirically, or explicitly frame the model as exploiting the fixed box convention of the simulations. A concrete test would be to apply random translations (with periodic wrapping) to the test catalogs and measure the change in predictions; as written the features guarantee this test would fail.
minor comments (5)
  1. [3.2] Equation (2) defines the k-down incidence neighborhood with the condition x ⊊ y, but for down incidence one should have y ⊊ x; as printed, the condition is identical to that of the up incidence neighborhood in Eq. (1).
  2. [3.4] The notation d_ij is overloaded: it is first defined as the vector x_i - x_j, then immediately redefined as the scalar ||x_i - x_j||. This makes the formulas for gamma_1 and gamma_2 confusing, especially because gamma_2 uses d_ij as a vector in the numerator.
  3. [Table 1] The row for '2PCF' is misaligned in the table: it contains only two numerical entries ('2.03±0.02 4.66±0.06') while the column structure requires eight. Please reformat so that the four metrics for Omega_m and sigma_8 are clearly assigned.
  4. [3.4] The phrase 'permutational invariant' should be 'permutation invariant' for consistency with standard terminology.
  5. [3.4] The rank-0 vertex features are described as random uniform draws used as a regularizer. Please clarify whether these are fixed per catalog before training (so the model is deterministic at test time) or resampled at each forward pass; the current wording is ambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the empirical TNN improvements are held-out test results anchored to external benchmarks; the E(3)-invariance flaw is a correctness issue, not circularity.

full rationale

The central empirical claims are self-contained: TNNs are trained on simulation labels and evaluated on held-out test sets, with hyperparameters selected on a validation set (Sections 3.6 and 3.8, Tables 1 and 2). The reported MSE improvements over GNNs are measured out-of-sample results, not quantities that reduce by construction to the architecture definition or to any fitted parameter. The edge and cell features follow standard geometric invariants from prior work including Villanueva-Domingo & Villaescusa-Navarro (2022), but this self-citation only imports feature choices; the higher-order message-passing mechanism is implemented explicitly and the results are additionally compared against independent external benchmarks (Balla et al. 2024; Chatterjee & Villaescusa-Navarro 2024). No load-bearing claim reduces to a self-citation chain, and no uniqueness theorem is invoked. The E(3)-invariance defect identified in Section 3.4, where gamma_1 and gamma_2 depend on absolute halo coordinates and are therefore not translation-invariant, is a genuine correctness flaw in the stated symmetry property, but it is not a circular step: the empirical MSE comparisons do not presuppose the invariance, and the advertised design goal failing does not make the prediction equivalent to its inputs. No fitted input is renamed as a prediction, and no known result is repackaged as novel. The paper is therefore free of significant circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

All physical parameters (Omega_m, sigma_8, Omega_b, h, n_s) are inputs from simulations used as labels, not free parameters fitted by the paper. The free parameters are instead the engineering choices of the combinatorial complex and the training setup: linking radius, tetrahedron cut, HDBSCAN parameters, random vertex features and Optuna hyperparameters. These are tuned on the validation set or fixed by hand, and the central MSE gains depend on them. The axioms are the modeling assumptions that the constructed hierarchy captures cosmologically relevant information, that the chosen scalars are E(3)-invariant, and that a single fixed data split is sufficient for comparison. No new physical entities are introduced.

free parameters (4)
  • r_link (linking radius) = Quijote: 0.015 fixed; CAMELS: searched over {0.010, 0.015, 0.020}
    Controls which halo pairs become rank-1 edges, affecting all higher-order cells. Tuned by Optuna on validation, not given by first principles.
  • N_cut (max tetrahedra) = Quijote: {3,000, 4,000, 5,000}
    Keeps only the smallest-volume Delaunay tetrahedra; chosen by validation performance. A free selection that may bias results toward small scales.
  • HDBSCAN clustering parameters = unspecified (likely library defaults)
    Determines rank-3 tetrahedra clusters; parameters are not listed in the paper, so the cluster construction is not fully specified.
  • random rank-0 vertex features = u ~ Uniform(0,1)
    Introduced as a regularizer (Section 3.4); adds stochastic noise to vertex cells rather than physical information.
assumptions (3)
  • domain assumption The cosmic web's information relevant for Omega_m and sigma_8 is captured by Delaunay tetrahedra, HDBSCAN clusters and MST hyperedges built from halo catalogs.
    The paper provides no theoretical guarantee or ablation showing this specific hierarchy is the right inductive bias; it is motivated by intuition about filaments and clusters (Section 3.3).
  • domain assumption The scalar cell features used in messages are E(3)-invariant.
    Section 3.4 asserts this, but the rank-1 features gamma_1 and gamma_2 depend on absolute coordinates and are not translation invariant.
  • domain assumption A fixed 80/10/10 split with bootstrapped test metrics is sufficient to compare architectures.
    The paper states that bootstrapping over multiple splits is computationally prohibitive (Section 4), so no split-to-split variance is reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cosmology with Topological Deep Learning." pith.science (2026). https://pith.science/paper/KIOLGWV2

@misc{pith2026250523904,
  author       = {Pith},
  title        = {Pith review of: Cosmology with Topological Deep Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KIOLGWV2}},
  note         = {Machine review of arXiv:2505.23904}
}
abstract

The standard cosmological model with cold dark matter posits a hierarchical formation of structures. We introduce topological neural networks (TNNs), implemented as message-passing neural networks on higher-order structures, to effectively capture the topological information inherent in these hierarchies that traditional graph neural networks (GNNs) fail to account for. Our approach not only considers the vertices and edges that comprise a graph but also extends to higher-order cells such as tetrahedra, clusters, and hyperedges. This enables message-passing between these heterogeneous structures within a combinatorial complex. Furthermore, our TNNs are designed to conserve the $E(3)$-invariance, which refers to the symmetry arising from invariance against translations, reflections, and rotations. When applied to the Quijote suite, our TNNs achieve a significant reduction in the mean squared error. Compared to our GNNs, which lack higher-order message-passing, ClusterTNNs show improvements of up to 22% in $\Omega_{\rm m}$ and 34% in $\sigma_8$ jointly, while the best FullTNN achieves an improvement of up to 60% in $\sigma_8$. In the context of the CAMELS suite, our models yield results comparable to the current GNN benchmark, albeit with a slight decrease in performance. We emphasize that our topology and symmetry-aware neural networks provide enhanced expressive power in modeling the large-scale structures of our universe.

Figures

Figures reproduced from arXiv: 2505.23904 by the authors.

Figure 1
Figure 1. Visualization of a combinatorial complex constructed from a halo catalog of a Quijote simulation. Given a halo catalog, the halos represent rank 0 cells, and the collection of halos are referred to as point clouds (top-left panel). Two halos are linked by an edge (rank 1) when their separation is less than rlink. Creating edges within the point cloud forms a graph (top-right panel). A Delaunay triangulation is perfo… view at source ↗
Figure 2
Figure 2. This diagram illustrates the identification of neighborhoods within combinatorial complexes and the process of message-passing. The panels highlight a singular red vertex, x, alongside its neighboring cells (orange), and non-neighboring cells (gray). For simplicity, we depict cells of rank 0 (vertices), rank 1 (edges), and rank 3 (tetrahedra). The dotted lines represent disconnected, or combinatorially discarded edg… view at source ↗
Figure 3
Figure 3. This diagram shows different cells of a combinatorial complex and their associated E(3) features. The combinatorial complex exhibits vertices (rank 0) in gray, edges (rank 1) in blue, tetrahedra (rank 2) in purple, clusters of tetrahedra (rank 3) in yellow, and hyperedges (rank 4) in black solid line. Vertices and edges are colored based on the highest rank of the cells out of vertices, edges, and tetrahedra they ar… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Depicted are tensor diagrams illustrating the structures of the neural networks utilized in this research: the GNN layer (top left), the TetraTNN layer (top center), the ClusterTNN layer (top right), and the FullTNN layer (bottom). These layers may be stacked verticall…
Figure 5
Figure 5. Figure 5: The evaluation results for our most effective model on the Quijote suite, ClusterTNN, derived from its Isolated Run are presented. Accuracy metrics with bootstrapped standard deviations on the test set for Ωm (left, blue) and σ8 (right, green) are depicted. ClusterTNNs…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Learning Cosmology from Nearest Neighbour Statistics

    astro-ph.CO 2025-11 conditional novelty 6.0 of 10

    Nearest-neighbour distance maps, combined with kNN-CDFs in a hybrid neural network, constrain Ωm and σ8 from Quijote halos with R2=0.80 and 0.93, matching or beating point-cloud methods at a fraction of the compute.

Reference graph

Works this paper leans on

67 extracted references · 13 canonical work pages · cited by 1 Pith paper

  1. [1]

    Abbott, T. M. C., Aguena, M., Alarcon, A., et al. 2022, PhRvD, 105, 023520, doi: 10.1103/PhysRevD.105.023520

    doi:10.1103/physrevd.105.023520 2022

  2. [2]

    2019, in The 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2623–2631

    Akiba, T., Sano, S., Yanase, T., Ohta, T., & Koyama, M. 2019, in The 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 2623–2631

    work page 2019

  3. [3]

    2017, MNRAS, 470, 2617, doi: 10.1093/mnras/stx721

    Alam, S., Ata, M., Bailey, S., et al. 2017, MNRAS, 470, 2617, doi: 10.1093/mnras/stx721

    doi:10.1093/mnras/stx721 2017

  4. [4]

    2020, Phys

    Allys, E., Marchand, T., Cardoso, J.-F., et al. 2020, Phys. Rev. D, 102, 103506, doi: 10.1103/PhysRevD.102.103506

    doi:10.1103/physrevd.102.103506 2020

  5. [5]

    2022, arXiv e-prints, arXiv:2211.12346, doi: 10.48550/arXiv.2211.12346

    Anagnostidis, S., Thomsen, A., Kacprzak, T., et al. 2022, arXiv e-prints, arXiv:2211.12346, doi: 10.48550/arXiv.2211.12346

    arXiv doi:10.48550/arxiv.2211.12346 2022

  6. [6]

    2024, arXiv e-prints, arXiv:2410.20516, doi: 10.48550/arXiv.2410.20516

    Smidt, T. 2024, arXiv e-prints, arXiv:2410.20516, doi: 10.48550/arXiv.2410.20516

    arXiv doi:10.48550/arxiv.2410.20516 2024

  7. [7]

    2024, arXiv e-prints, arXiv:2405.15429, doi: 10.48550/arXiv.2405.15429

    Battiloro, C., Karaismailo˘glu, E., Tec, M., et al. 2024, arXiv e-prints, arXiv:2405.15429, doi: 10.48550/arXiv.2405.15429

    arXiv doi:10.48550/arxiv.2405.15429 2024

  8. [8]

    E., Villaescusa-Navarro, F., Massara, E., et al

    Bayer, A. E., Villaescusa-Navarro, F., Massara, E., et al. 2021, ApJ, 919, 24, doi: 10.3847/1538-4357/ac0e91

    doi:10.3847/1538-4357/ac0e91 2021

Show all 67 references
  1. [9]

    2019, in NeurIPS 2019 Workshop on Machine Learning for Physical Sciences

    Beck, R., & Sadowski, P. 2019, in NeurIPS 2019 Workshop on Machine Learning for Physical Sciences. https://ml4physicalsciences.github.io/2019/files/ NeurIPS ML4PS 2019 80.pdf

    2019

  2. [10]

    S., Wechsler, R

    Behroozi, P. S., Wechsler, R. H., & Wu, H.-Y . 2013, ApJ, 762, 109, doi: 10.1088/0004-637X/762/2/109

    2013 doi

  3. [11]

    R., Faber, S

    Blumenthal, G. R., Faber, S. M., Primack, J. R., & Rees, M. J. 1984, Nature, 311, 517, doi: 10.1038/311517a0

    1984 doi

  4. [12]

    R., Kofman, L., & Pogosyan, D

    Bond, J. R., Kofman, L., & Pogosyan, D. 1996, Nature, 380, 603, doi: 10.1038/380603a0

    1996 doi

  5. [13]

    2021, arXiv e-prints, arXiv:2110.02905, doi: 10.48550/arXiv.2110.02905

    Welling, M. 2021, arXiv e-prints, arXiv:2110.02905, doi: 10.48550/arXiv.2110.02905

    2021 doi arXiv

  6. [14]

    Campello, R. J. G. B., Moulavi, D., & Sander, J. 2013, in Advances in Knowledge Discovery and Data Mining, ed. J. Pei, V . S

    2013

  7. [15]

    2024, arXiv e-prints, arXiv:2405.13119, doi: 10.48550/arXiv.2405.13119

    Chatterjee, A., & Villaescusa-Navarro, F. 2024, arXiv e-prints, arXiv:2405.13119, doi: 10.48550/arXiv.2405.13119

    2024 doi arXiv

  8. [16]

    2020, in Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS ’20 (Red Hook, NY , USA: Curran Associates Inc.)

    Chen, Z., Chen, L., Villar, S., & Bruna, J. 2020, in Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS ’20 (Red Hook, NY , USA: Curran Associates Inc.)

    2020

  9. [17]

    2020, arXiv e-prints, arXiv:2004.05718, doi: 10.48550/arXiv.2004.05718

    Corso, G., Cavalleri, L., Beaini, D., Li`o, P., & Veliˇckovi´c, P. 2020, arXiv e-prints, arXiv:2004.05718, doi: 10.48550/arXiv.2004.05718

    2020 doi arXiv

  10. [18]

    R., Philcox, O

    Coulton, W. R., Philcox, O. H. E., & Villaescusa-Navarro, F. 2023, arXiv e-prints, arXiv:2306.11782, doi: 10.48550/arXiv.2306.11782

    2023 doi arXiv

  11. [19]

    R., Villaescusa-Navarro, F., Jamieson, D., et al

    Coulton, W. R., Villaescusa-Navarro, F., Jamieson, D., et al. 2022, arXiv e-prints, arXiv:2206.01619. https://arxiv.org/abs/2206.01619

    2022 arXiv

  12. [20]

    2021, in Proceedings of Machine Learning Research, V ol

    Cranmer, M., Melchior, P., & Nord, B. 2021, in Proceedings of Machine Learning Research, V ol. 148, NeurIPS 2020 Workshop on Pre-registration in Machine Learning, ed. L. Bertinetto, J. F

    2021

  13. [21]

    2024, PhRvD, 109, 123531, doi: 10.1103/PhysRevD.109.123531

    Cuesta-Lazaro, C., & Mishra-Sharma, S. 2024, PhRvD, 109, 123531, doi: 10.1103/PhysRevD.109.123531

    2024 doi

  14. [22]

    S., & White, S

    Davis, M., Efstathiou, G., Frenk, C. S., & White, S. D. M. 1985, ApJ, 292, 371, doi: 10.1086/163168 de Santi, N. S. M., Shao, H., Villaescusa-Navarro, F., et al. 2023a, ApJ, 952, 69, doi: 10.3847/1538-4357/acd1e2 —. 2023b, ApJ, 952, 69, doi: 10.3847/1538-4357/acd1e2 DESI Colla...

    1985 doi arXiv

  15. [23]

    2022, arXiv e-prints, arXiv:2204.07646, doi: 10.48550/arXiv.2204.07646

    Eickenberg, M., Allys, E., Moradinezhad Dizgah, A., et al. 2022, arXiv e-prints, arXiv:2204.07646, doi: 10.48550/arXiv.2204.07646

    2022 doi arXiv

  16. [24]

    2023, arXiv e-prints, arXiv:2305.07100, doi: 10.48550/arXiv.2305.07100 Garcia Satorras, V ., Hoogeboom, E., & Welling, M

    Eijkelboom, F., Hesselink, R., & Bekkers, E. 2023, arXiv e-prints, arXiv:2305.07100, doi: 10.48550/arXiv.2305.07100 Garcia Satorras, V ., Hoogeboom, E., & Welling, M. 2021, arXiv e-prints, arXiv:2102.09844, doi: 10.48550/arXiv.2102.09844

    2023 doi arXiv

  17. [25]

    K., Jegelka, S., & Jaakkola, T

    Garg, V . K., Jegelka, S., & Jaakkola, T. 2020, arXiv e-prints, arXiv:2002.06157, doi: 10.48550/arXiv.2002.06157

    2020 doi arXiv

  18. [26]

    2021, JCAP, 2021, 029, doi: 10.1088/1475-7516/2021/04/029

    Hahn, C., & Villaescusa-Navarro, F. 2021, JCAP, 2021, 029, doi: 10.1088/1475-7516/2021/04/029

    2021 doi

  19. [27]

    2023, Topological Deep Learning: Going Beyond Graph Data

    Hajij, M., Zamzmi, G., Papamarkou, T., et al. 2023, Topological Deep Learning: Going Beyond Graph Data. https://arxiv.org/abs/2206.00606

    2023 arXiv

  20. [28]

    Harrison, E. R. 1970, Phys. Rev. D, 1, 2726, doi: 10.1103/PhysRevD.1.2726

    1970 doi

  21. [29]

    J., Chartier, N., et al

    Ho, M., Bartlett, D. J., Chartier, N., et al. 2024, The Open Journal of Astrophysics, 7, 54, doi: 10.33232/001c.120559

    2024 doi

  22. [30]

    Y ., Sabiu, C

    Hwang, S. Y ., Sabiu, C. G., Park, I., & Hong, S. E. 2023, JCAP, 2023, 075, doi: 10.1088/1475-7516/2023/11/075

    2023 doi

  23. [31]

    M., Philcox, O

    Ivanov, M. M., Philcox, O. H. E., Cabass, G., et al. 2023, PhRvD, 107, 083515, doi: 10.1103/PhysRevD.107.083515

    2023 doi

  24. [32]

    M., Simonovi´c, M., & Zaldarriaga, M

    Ivanov, M. M., Simonovi´c, M., & Zaldarriaga, M. 2020, Journal of Cosmology and Astroparticle Physics, 2020, 042, doi: 10.1088/1475-7516/2020/05/042

    2020 doi

  25. [33]

    2022, MNRAS, 516, 2406, doi: 10.1093/mnras/stac2083

    Jagvaral, Y ., Lanusse, F., Singh, S., et al. 2022, MNRAS, 516, 2406, doi: 10.1093/mnras/stac2083

    2022 doi

  26. [34]

    Jeffrey, N., & Wandelt, B. D. 2020, in 34th Conference on Neural Information Processing Systems. https://arxiv.org/abs/2011.05991

    2020 arXiv

  27. [35]

    2024, ApJ, 975, 38, doi: 10.3847/1538-4357/ad73d4 17

    Lee, J.-Y ., Kim, J.-h., Jung, M., et al. 2024, ApJ, 975, 38, doi: 10.3847/1538-4357/ad73d4 17

    2024 doi

  28. [36]

    H., Hahn, C., et al

    Lemos, P., Parker, L. H., Hahn, C., et al. 2023, in Machine Learning for Astrophysics, 18, doi: 10.48550/arXiv.2310.15256

    2023 doi arXiv

  29. [37]

    2016, arXiv e-prints, arXiv:1608.03983, doi: 10.48550/arXiv.1608.03983

    Loshchilov, I., & Hutter, F. 2016, arXiv e-prints, arXiv:1608.03983, doi: 10.48550/arXiv.1608.03983

    2016 doi arXiv

  30. [38]

    2022, arXiv e-prints, arXiv:2202.07123, doi: 10.48550/arXiv.2202.07123

    Ma, X., Qin, C., You, H., Ran, H., & Fu, Y . 2022, arXiv e-prints, arXiv:2202.07123, doi: 10.48550/arXiv.2202.07123

    2022 doi arXiv

  31. [39]

    L., Charnock, T., Lemos, P., et al

    Makinen, T. L., Charnock, T., Lemos, P., et al. 2022, The Open Journal of Astrophysics, 5, 18, doi: 10.21105/astro.2207.05202

    2022 arXiv

  32. [40]

    2023, ApJ, 951, 70, doi: 10.3847/1538-4357/acd44d

    Massara, E., Villaescusa-Navarro, F., Hahn, C., et al. 2023, ApJ, 951, 70, doi: 10.3847/1538-4357/acd44d

    2023 doi

  33. [41]

    2019, Computational Astrophysics and Cosmology, 6, 2, doi: 10.1186/s40668-019-0028-x

    Nelson, D., Springel, V ., Pillepich, A., et al. 2019, Computational Astrophysics and Cosmology, 6, 2, doi: 10.1186/s40668-019-0028-x

    2019 doi

  34. [42]

    2023, ApJ, 959, 136, doi: 10.3847/1538-4357/ad022a

    Ni, Y ., Genel, S., Angl´es-Alc´azar, D., et al. 2023, ApJ, 959, 136, doi: 10.3847/1538-4357/ad022a

    2023 doi

  35. [43]

    J., Yuan, S., & Garrison, L

    Ntampaka, M., Eisenstein, D. J., Yuan, S., & Garrison, L. H. 2020, ApJ, 889, 151, doi: 10.3847/1538-4357/ab5f5e

    2020 doi

  36. [44]

    2023, Architectures of Topological Deep Learning: A Survey on Topological Neural Networks

    Papillon, M., Sanborn, S., Hajij, M., & Miolane, N. 2023, Architectures of Topological Deep Learning: A Survey on Topological Neural Networks. https://arxiv.org/abs/2304.10031

    2023 arXiv

  37. [45]

    Philcox, O. H. E., Massara, E., & Spergel, D. N. 2020, PhRvD, 102, 043516, doi: 10.1103/PhysRevD.102.043516

    2020 doi

  38. [46]

    2018, MNRAS, 473, 4077, doi: 10.1093/mnras/stx2656 Planck Collaboration, Aghanim, N., Akrami, Y ., et al

    Pillepich, A., Springel, V ., Nelson, D., et al. 2018, MNRAS, 473, 4077, doi: 10.1093/mnras/stx2656 Planck Collaboration, Aghanim, N., Akrami, Y ., et al. 2020, A&A, 641, A6, doi: 10.1051/0004-6361/201833910 PYT-Team. 2024, arXiv preprint arXiv:2402.02441

    2018 arXiv

  39. [47]

    G., Filippenko, A

    Riess, A. G., Filippenko, A. V ., Challis, P., et al. 1998, AJ, 116, 1009, doi: 10.1086/300499 Sahl´en, M. 2019, PhRvD, 99, 063525, doi: 10.1103/PhysRevD.99.063525

    1998 doi

  40. [48]

    2006, PhRvD, 74, 023522, doi: 10.1103/PhysRevD.74.023522

    Sefusatti, E., Crocce, M., Pueblas, S., & Scoccimarro, R. 2006, PhRvD, 74, 023522, doi: 10.1103/PhysRevD.74.023522

    2006 doi

  41. [49]

    2023, ApJ, 944, 27, doi: 10.3847/1538-4357/acac7a

    Shao, H., Villaescusa-Navarro, F., Villanueva-Domingo, P., et al. 2023, ApJ, 944, 27, doi: 10.3847/1538-4357/acac7a

    2023 doi

  42. [50]

    N., Bean, R., Dor´e, O., et al

    Spergel, D. N., Bean, R., Dor´e, O., et al. 2007, ApJS, 170, 377, doi: 10.1086/513700

    2007 doi

  43. [51]

    2005, MNRAS, 364, 1105, doi: 10.1111/j.1365-2966.2005.09655.x —

    Springel, V . 2005, MNRAS, 364, 1105, doi: 10.1111/j.1365-2966.2005.09655.x —. 2010, MNRAS, 401, 791, doi: 10.1111/j.1365-2966.2009.15715.x

    2005

  44. [52]

    2024, arXiv e-prints, arXiv:2402.14239, doi: 10.48550/arXiv.2402.14239

    Tanimura, H., Bonnefous, A., Liu, J., & Ganguly, S. 2024, arXiv e-prints, arXiv:2402.14239, doi: 10.48550/arXiv.2402.14239

    2024 doi arXiv

  45. [53]

    L., Sheldon, E

    Tinker, J. L., Sheldon, E. S., Wechsler, R. H., et al. 2012, ApJ, 745, 16, doi: 10.1088/0004-637X/745/1/16

    2012 doi

  46. [54]

    S., Guzzo, L., Granett, B

    Tosone, F., Cagliari, M. S., Guzzo, L., Granett, B. R., & Crespi, A. 2023, A&A, 672, A150, doi: 10.1051/0004-6361/202245369

    2023 doi

  47. [55]

    2020, MNRAS, 495, 4006, doi: 10.1093/mnras/staa1155

    Uhlemann, C., Friedrich, O., Villaescusa-Navarro, F., Banerjee, A., & Codis, S. 2020, MNRAS, 495, 4006, doi: 10.1093/mnras/staa1155

    2020 doi

  48. [56]

    2022, PhRvD, 105, 103534, doi: 10.1103/PhysRevD.105.103534

    Valogiannis, G., & Dvorkin, C. 2022, PhRvD, 105, 103534, doi: 10.1103/PhysRevD.105.103534

    2022 doi

  49. [57]

    2021, ApJ, 915, 71, doi: 10.3847/1538-4357/abf7ba

    Villaescusa-Navarro, F., Angl´es-Alc´azar, D., Genel, S., et al. 2021, ApJ, 915, 71, doi: 10.3847/1538-4357/abf7ba

    2021 doi

  50. [58]

    2023, ApJS, 265, 54, doi: 10.3847/1538-4365/acbf47

    Villaescusa-Navarro, F., Genel, S., Angl´es-Alc´azar, D., et al. 2023, ApJS, 265, 54, doi: 10.3847/1538-4365/acbf47

    2023 doi

  51. [59]

    2022, PabloVD/CosmoGraphNet:v1.0, v1.0, Zenodo, doi: 10.5281/zenodo.6485804

    Villanueva-Domingo, P. 2022, PabloVD/CosmoGraphNet:v1.0, v1.0, Zenodo, doi: 10.5281/zenodo.6485804

    2022 doi

  52. [60]

    2022, ApJ, 937, 115, doi: 10.3847/1538-4357/ac8930

    Villanueva-Domingo, P., & Villaescusa-Navarro, F. 2022, ApJ, 937, 115, doi: 10.3847/1538-4357/ac8930

    2022 doi

  53. [61]

    2022, ApJ, 935, 30, doi: 10.3847/1538-4357/ac7aa3

    Villanueva-Domingo, P., Villaescusa-Navarro, F., Angl´es-Alc´azar, D., et al. 2022, ApJ, 935, 30, doi: 10.3847/1538-4357/ac7aa3

    2022 doi

  54. [62]

    2023, PhRvD, 107, 103003, doi: 10.1103/PhysRevD.107.103003

    Villanueva-Domingo, P., Villaescusa-Navarro, F., Genel, S., et al. 2023, PhRvD, 107, 103003, doi: 10.1103/PhysRevD.107.103003

    2023 doi

  55. [63]

    W., Storey-Fisher, K., Yao, W., & Blum-Smith, B

    Villar, S., Hogg, D. W., Storey-Fisher, K., Yao, W., & Blum-Smith, B. 2021, arXiv e-prints, arXiv:2106.06610, doi: 10.48550/arXiv.2106.06610

    2021 doi arXiv

  56. [64]

    2020, The Astrophysical Journal Supplement Series, 248, 32, doi: 10.3847/1538-4365/ab908c

    Weinberger, R., Springel, V ., & Pakmor, R. 2020, The Astrophysical Journal Supplement Series, 248, 32, doi: 10.3847/1538-4365/ab908c

    2020 doi

  57. [65]

    2017, MNRAS, 465, 3291, doi: 10.1093/mnras/stw2944

    Weinberger, R., Springel, V ., Hernquist, L., et al. 2017, MNRAS, 465, 3291, doi: 10.1093/mnras/stw2944

    2017 doi

  58. [66]

    F., & Kragh Jespersen, C

    Wu, J. F., & Kragh Jespersen, C. 2023, arXiv e-prints, arXiv:2306.12327, doi: 10.48550/arXiv.2306.12327

    2023 doi arXiv

  59. [67]

    Zeldovich, Y . B. 1972, MNRAS, 160, 1P, doi: 10.1093/mnras/160.1.1P 18 Table 3.Full List of Hyperparameters Hyperparameter Description Values Data-Related data mode(Quijote) Maximum number of tetrahedra (N tetra){3,000,4,000,5,000} data mode(CAMELS) Linking radius criterion fo...

    2022 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.