Cool Embeddings: Predicting Galaxy Cluster Cooling Times in IllustrisTNG and TNG-Cluster with AstroCLIP
Abstract
Foundation models trained on optical astronomical data have not previously been applied to X-ray cluster thermodynamics. We show that embeddings from AstroCLIP, a contrastive vision-language model trained on optical galaxy images and spectra, retain physically meaningful structure when applied to mock Chandra X-ray images of simulated galaxy clusters - a significant shift in observational modality and physical scale. We target the central cooling time (tcool,0) of the intracluster medium, a key diagnostic of cluster thermodynamics traditionally requiring X-ray spectroscopy. We present a machine learning framework that predicts tcool,0 directly from X-ray images, applied to 711 projections from TNG300 and 993 from TNG-Cluster from the IllustrisTNG simulations. AstroCLIP embeddings form a structured latent space in which clusters with similar thermodynamic properties are naturally grouped, with no equivalent separation by halo mass. Building on this, we introduce a dot-product k-nearest neighbour attention model that retrieves similar clusters from the training set and adaptively reweights their cooling times. This model achieves root-mean-square errors of 0.24 (TNG300) and 0.23 (TNG-Cluster) in log(tcool,0), and demonstrates strong out-of-distribution generalisation, outperforming a fine-tuned convolutional neural network trained on raw images (RMSE 0.50 versus 0.73). These results demonstrate that pretrained foundation model embeddings transfer across astrophysical domains and encode thermodynamic information sufficient for inference of cluster cooling properties - offering a scalable path toward cool-core characterisation in all-sky surveys such as eROSITA.
Related
- Kinematics of the Weak Cool-Core Cluster A3571 Observed with XRISM: Low Cooling Rate Balanced by Low Heating Rate?
- Something green beneath the surface: The dynamical nature of fossil structures in IllustrisTNG-100
- The connection between surface brightness and satellite systems for central galaxies through IllustrisTNG (Corrigendum)
- Turbulent Transport of Galactic Magnetic Fields into Cosmic Voids: Insights from IllustrisTNG
- The Dynamical Mass Profiles of Galaxy Groups - Application to IllustrisTNG-300
BibTeX
@misc{karam2026cool,
title = {{Cool Embeddings: Predicting Galaxy Cluster Cooling Times in IllustrisTNG and TNG-Cluster with AstroCLIP}},
author = {Rawan Karam and Carter Lee Rhea and Julie Hlavacek-Larrondo and Laurence Perreault-Levasseur and Marine Prunier and Adam Alexandre and Hyunseop Choi},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.32582},
eprint = {2609.32582},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.32582},
}
RIS
TY - UNPB
TI - Cool Embeddings: Predicting Galaxy Cluster Cooling Times in IllustrisTNG and TNG-Cluster with AstroCLIP
AU - Rawan Karam
AU - Carter Lee Rhea
AU - Julie Hlavacek-Larrondo
AU - Laurence Perreault-Levasseur
AU - Marine Prunier
AU - Adam Alexandre
AU - Hyunseop Choi
PY - 2026
DA - 2026/09/26
PB - arXiv
DO - 10.48550/arxiv.2609.32582
UR - https://arxiv.org/abs/2609.32582
AB - Foundation models trained on optical astronomical data have not previously been applied to X-ray cluster thermodynamics. We show that embeddings from AstroCLIP, a contrastive vision-language model trained on optical galaxy images and spectra, retain physically meaningful structure when applied to mock Chandra X-ray images of simulated galaxy clusters - a significant shift in observational modality and physical scale. We target the central cooling time (t_cool,0) of the intracluster medium, a key diagnostic of cluster thermodynamics traditionally requiring X-ray spectroscopy. We present a machine learning framework that predicts t_cool,0 directly from X-ray images, applied to 711 projections from TNG300 and 993 from TNG-Cluster from the IllustrisTNG simulations. AstroCLIP embeddings form a structured latent space in which clusters with similar thermodynamic properties are naturally grouped, with no equivalent separation by halo mass. Building on this, we introduce a dot-product k-nearest neighbour attention model that retrieves similar clusters from the training set and adaptively reweights their cooling times. This model achieves root-mean-square errors of 0.24 (TNG300) and 0.23 (TNG-Cluster) in log(t_cool,0), and demonstrates strong out-of-distribution generalisation, outperforming a fine-tuned convolutional neural network trained on raw images (RMSE 0.50 versus 0.73). These results demonstrate that pretrained foundation model embeddings transfer across astrophysical domains and encode thermodynamic information sufficient for inference of cluster cooling properties - offering a scalable path toward cool-core characterisation in all-sky surveys such as eROSITA.
ER -