ASTRAI 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 citation (plain text)

Andrea Claudio Grasso, Vincenzo Del Zoppo, Alessio Mezzina, Marco Cataldo, Giuseppe Puglisi, Fabio Spampinato, Stefano Pio Cosentino, Maria Letizia Pumo, Luca Naso,
        A dual AI framework for the automatic characterisation of low-interacting H-rich SNe within the ASTRAI (Advanced Supernova Transient Research with Artificial Intelligence) project,
        Astronomy and Computing,
        Volume 56,
        2026,
        101116,
        ISSN 2213-1337,
        https://doi.org/10.1016/j.ascom.2026.101116.
        (https://www.sciencedirect.com/science/article/pii/S2213133726000582)
        Abstract: Rapidly increasing data volumes from high cadence surveys, and the still larger streams anticipated from the Legacy Survey of Space and Time (LSST), pose a major challenge for the timely and accurate analysis of transient phenomena such as core-collapse Supernovae (CC SNe). Traditional approaches based on semi-analytical or hydrodynamical models, while physically grounded, hinder the large-scale characterisation of these transient events. We present a novel machine learning approach, focusing on low-interacting hydrogen-rich (H-rich) SNe. This approach encompasses two stages: a generative model produces synthetic light curves (LCs) to augment sparse or incomplete datasets, and deep learning models characterise LCs by regressing fundamental physical parameters. The characterisation stage uses an ensemble of neural networks, each specialised to regress a single parameter, improving interpretability, modularity and robustness. To account for observational systematics, we employ data augmentation strategies that mimic realistic noise and cadence conditions. Benchmarking against semi-analytical and hydrodynamical simulations demonstrates that the proposed generative method recovers LCs with error comparable to or smaller than observational noise, while achieving inference speeds orders of magnitude faster than traditional modelling pipelines (with a throughput beyond 105 LCs per second on a single GPU hardware). This could enable large-scale characterisation of H-rich SNe within modern and upcoming survey data streams, preparing the scientific community to handle the massive LCs flow expected from LSST.
        Keywords: Supernovae; Transient astrophysics; LSST; Light curves; Deep learning; GenAI; Synthetic 
        dataURL: https://astrai.koexai.com/resources/  Licence: CC BY 4.0.

BibTeX paper

@article{GRASSO2026101116,
        title = {A dual AI framework for the automatic characterisation of low-interacting H-rich SNe within the ASTRAI (Advanced Supernova Transient Research with Artificial Intelligence) project},
        journal = {Astronomy and Computing},
        volume = {56},
        pages = {101116},
        year = {2026},
        issn = {2213-1337},
        doi = {https://doi.org/10.1016/j.ascom.2026.101116},
        url = {https://www.sciencedirect.com/science/article/pii/S2213133726000582},
        author = {Andrea Claudio Grasso and Vincenzo {Del Zoppo} and Alessio Mezzina and Marco Cataldo and Giuseppe Puglisi and Fabio Spampinato and Stefano Pio Cosentino and Maria Letizia Pumo and Luca Naso},
        keywords = {Supernovae, Transient astrophysics, LSST, Light curves, Deep learning, GenAI, Synthetic data},
        abstract = {Rapidly increasing data volumes from high cadence surveys, and the still larger streams anticipated from the Legacy Survey of Space and Time (LSST), pose a major challenge for the timely and accurate analysis of transient phenomena such as core-collapse Supernovae (CC SNe). Traditional approaches based on semi-analytical or hydrodynamical models, while physically grounded, hinder the large-scale characterisation of these transient events. We present a novel machine learning approach, focusing on low-interacting hydrogen-rich (H-rich) SNe. This approach encompasses two stages: a generative model produces synthetic light curves (LCs) to augment sparse or incomplete datasets, and deep learning models characterise LCs by regressing fundamental physical parameters. The characterisation stage uses an ensemble of neural networks, each specialised to regress a single parameter, improving interpretability, modularity and robustness. To account for observational systematics, we employ data augmentation strategies that mimic realistic noise and cadence conditions. Benchmarking against semi-analytical and hydrodynamical simulations demonstrates that the proposed generative method recovers LCs with error comparable to or smaller than observational noise, while achieving inference speeds orders of magnitude faster than traditional modelling pipelines (with a throughput beyond 105 LCs per second on a single GPU hardware). This could enable large-scale characterisation of H-rich SNe within modern and upcoming survey data streams, preparing the scientific community to handle the massive LCs flow expected from LSST.}

Tip: if a dataset provides its own citation.txt or DOI, please use that instead of the template above.

Software — Licence & how to cite

Licence (intended): gpl v3 license (to be confirmed in the repository). A copy of the licence will be included as LICENSE in the repo. About GPL v3.

Recommended software citation (plain text)

ASTRAI Project (2025). ASTRAI Core (v0.1) — Generative models and characterisation tools. Source code. URL: https://astrai.koexai.com/resources/  Licence: GPL V3.

Software BibTeX (template)

@software{astrai_core_v0_1_2025,
      author  = {Koexai Srl},
      title   = {ASTRAI Source Code},
      year    = {2025},
      version = {0.1},
      url     = {https://astrai.koexai.com/resources/},
      license = {GPL V3},
      note    = {Replace with repository URL and tag when public}
    }