fvillaescusa/CosmoBench

CosmoBench: A Multiscale, Multiview, Multitask Cosmology Benchmark for Geometric Deep Learning

Paper: https://arxiv.org/abs/2507.03707 See Table 1 in Paper for details of each dataset Github: https://github.com/nhuang37/cosmology_benchmark

Data

Point cloud datasets

• Quijote: large-scale point clouds simulated with box size 1000 cMpc/h
  • top5000_halos*: each h5 file contains a train/validation/test split of point clouds, where each point cloud contains the top-5000 halos (sorted by mass) within the simulation
  • ALL_halos_*: each h5 file contains a train/validation/test split of point clouds, where each point cloud contains all halos within the simulation
• CAMELS-SAM/galaxies: medium-scale point clouds simulated with box size 100 cMpc
  • top5000_galaxies*: each h5 file contains a train/validation/test split of point clouds, where each point cloud contains the top-5000 galaxies (sorted by mass) within the simulation
  • ALL_galaxies_*: each h5 file contains a train/validation/test split of point clouds, where each point cloud contains all galaxies within the simulation
• CAMELS: small-scale point clouds simulated with box size 25 cMpc/h
  • ALL_halos_*: each h5 file contains a train/validation/test split of point clouds, where each point cloud contains all galaxies within the simulation

Merger tree datasets

• CAMELS-SAM/trees: merger trees constructed via the merging history of CAMELS-SAM halos
  • CS_tree*: each pt file contains a list of Pytorch Geometric (PyG) data from the train/validation/test split, where each data describes a merger tree intended for tree-level regression tasks
  • infilling_trees_25k_200*: each pt file contains a list of Pytorch Geometric (PyG) data from the train/validation/test split, where each data describes a merger tree intended for node-level classification tasks.