Scholarlink

We train a Graph Neural Network (GNN) on dark matter halo merger trees from TNG100, a cosmological magnetohydrodynamical simulation of galaxy formation from the IllustrisTNG project. Our trained model is designed to infer galaxy baryonic properties: stellar and gas mass. It successfully reproduces galaxy population statistics consistent with the input training data, enabling rapid application to larger dark matter only simulations. We use an attention-based GNN architecture to capture the important information from the full, unpruned merger trees. This also enables us to evaluate feature importance via the attention mechanism, revealing details about the astrophysical galaxy-halo connection.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Inferring Galaxy Baryonic Properties from IllustrisTNG Dark Matter Merger Trees with Graph Neural Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Nick Andreadis,
  • Dylan Nelson

摘要

We train a Graph Neural Network (GNN) on dark matter halo merger trees from TNG100, a cosmological magnetohydrodynamical simulation of galaxy formation from the IllustrisTNG project. Our trained model is designed to infer galaxy baryonic properties: stellar and gas mass. It successfully reproduces galaxy population statistics consistent with the input training data, enabling rapid application to larger dark matter only simulations. We use an attention-based GNN architecture to capture the important information from the full, unpruned merger trees. This also enables us to evaluate feature importance via the attention mechanism, revealing details about the astrophysical galaxy-halo connection.