DeepCosmoNet Resource Hub

Licence & how to cite

Licence (default): Creative Commons Attribution 4.0 International (CC BY 4.0). You must provide appropriate credit and indicate if changes were made. Read the licence.

Recommended paper Halos citation (plain text)

Fabio Spampinato, Vincenzo Del Zoppo, Giuseppe Puglisi, Alessio Mezzina, Marco Cataldo, Jean Marc Christille, Matteo Calabrese, Luca Naso, Carmelita Carbone,
        HALOS: Hierarchical Aggregation Learning for Overdensity Search,
        Astronomy and Computing,
        Volume 56,
        2026,
        101114,
        ISSN 2213-1337,
        https://doi.org/10.1016/j.ascom.2026.101114.
        (https://www.sciencedirect.com/science/article/pii/S2213133726000569)
        Abstract: The identification of gravitationally bound substructures (subhalos) within cosmological simulations is a cornerstone for understanding galaxy formation and evolution. Traditional algorithms, while accurate, are often computationally intensive, posing a significant bottleneck for the analysis of next-generation cosmological simulations and limiting the feasibility of on-the-fly processing. In this work, we introduce HALOS (Hierarchical Aggregation Learning for Overdensity Search), a novel deep learning pipeline for subhalo identification in 3D point clouds. Our method employs a multi-stage approach that decouples particle classification from instance segmentation. First, we engineer a set of physically motivated features for each particle. Second, a multi-layer perceptron simultaneously performs two tasks: (i) a semantic segmentation to classify particles as either bound to a subhalo or part of the unbound background, and (ii) a regression to predict the 3D coordinates of the parent subhalo centroid for each bound particle. Finally, the HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) algorithm performs a density-based clustering exclusively on the pre-filtered set of bound particles, significantly reducing computational complexity. We train and validate our model using catalogues generated from cosmological N-body simulations by the SUBFIND algorithm. HALOS achieves a semantic classification accuracy of 95%, an Adjusted Rand Index for instance segmentation >90%, and an overall Completeness of 90%, demonstrating a close alignment with SUBFIND, while reducing computational time by a factor of ∼ 16.
        Keywords: Subhalos; Cosmology; Deep learning; Clustering; Particle segmentation; SUBFIND; Halo-finders; 3D point clouds
        dataURL: https://deepcosmonet.koexai.com/resources/  Licence: CC BY 4.0.

BibTeX Paper Halos(template)

@article{SPAMPINATO2026101114,
          title = {HALOS: Hierarchical Aggregation Learning for Overdensity Search},
          journal = {Astronomy and Computing},
          volume = {56},
          pages = {101114},
          year = {2026},
          issn = {2213-1337},
          doi = {https://doi.org/10.1016/j.ascom.2026.101114},
          url = {https://www.sciencedirect.com/science/article/pii/S2213133726000569},
          author = {Fabio Spampinato and Vincenzo {Del Zoppo} and Giuseppe Puglisi and Alessio Mezzina and Marco Cataldo and Jean Marc Christille and Matteo Calabrese and Luca Naso and Carmelita Carbone},
          keywords = {Subhalos, Cosmology, Deep learning, Clustering, Particle segmentation, SUBFIND, Halo-finders, 3D point clouds},
          abstract = {The identification of gravitationally bound substructures (subhalos) within cosmological simulations is a cornerstone for understanding galaxy formation and evolution. Traditional algorithms, while accurate, are often computationally intensive, posing a significant bottleneck for the analysis of next-generation cosmological simulations and limiting the feasibility of on-the-fly processing. In this work, we introduce HALOS (Hierarchical Aggregation Learning for Overdensity Search), a novel deep learning pipeline for subhalo identification in 3D point clouds. Our method employs a multi-stage approach that decouples particle classification from instance segmentation. First, we engineer a set of physically motivated features for each particle. Second, a multi-layer perceptron simultaneously performs two tasks: (i) a semantic segmentation to classify particles as either bound to a subhalo or part of the unbound background, and (ii) a regression to predict the 3D coordinates of the parent subhalo centroid for each bound particle. Finally, the HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise) algorithm performs a density-based clustering exclusively on the pre-filtered set of bound particles, significantly reducing computational complexity. We train and validate our model using catalogues generated from cosmological N-body simulations by the SUBFIND algorithm. HALOS achieves a semantic classification accuracy of 95%, an Adjusted Rand Index for instance segmentation >90%, and an overall Completeness of 90%, demonstrating a close alignment with SUBFIND, while reducing computational time by a factor of ∼ 16.}
          }

DeepCosmoNet Project (2025). DeepCosmoNet Core (v0.1) — Deep Learning for cosmic web Analysis.
Source code. URL: https://deepcosmonet.koexai.com/resources/  Licence: GPL V3.

@software{deepcosmonet_core_v0_1_2025,
      author  = {Koexai Srl},
      title   = {DeepCosmoNet Source Code},
      year    = {2025},
      version = {0.1},
      url     = {https://deepcosmonet.koexai.com/resources/},
      license = {GPL V3},
      note    = {Replace with repository URL and tag when public}
    }