Celia

Trained encoder of Celia: Catalog Embeddings Learned from cosmologIcal simulAtions (Zhang, Ho, Kumar, Di Matteo, Croft, Póczos 2026). Code: github.com/xwzhang98/celia.

Celia maps a catalog of subhalos and their galaxy properties to one embedding vector. It was pretrained without parameter labels on the IllustrisTNG, SIMBA and Astrid suites of CAMELS and their N-body twins.

celia_step18000.pt (39 MB): the student encoder after 18,000 updates, its normalisation statistics and the model configuration. This is the model of the paper.

sha256: 27bf4b2456c49f71c3219a0016a0ad2e2508a6192b31b1de3d4d343145094be6

Usage

pip install git+https://github.com/xwzhang98/celia
wget https://huggingface.co/xiaowen-zhang/celia/resolve/main/celia_step18000.pt

from celia.catalog import read_catalog
from celia.embed import embed, load_encoder

model = load_encoder("celia_step18000.pt", device="cuda")
catalog = read_catalog("CAMELS/FOF_Subfind/IllustrisTNG/L25n256/LH/LH_0/groups_090.hdf5",
                       suite="IllustrisTNG", sim="LH_0", nbody=False)
vector = embed(model, catalog, catalog_number=0)      # numpy float32, shape (32896,)

Citation

@article{zhang2026celia,
  title   = {Celia: Catalog Embeddings Learned from cosmologIcal simulAtions},
  author  = {Zhang, Xiaowen and Ho, Matthew and Kumar, Anshul and Di Matteo, Tiziana and Croft, Rupert A. C.
             and P{\'o}czos, Barnab{\'a}s},
  journal = {arXiv e-prints},
  eprint  = {2610.40218},
  archivePrefix = {arXiv},
  year    = {2026}
}