Interpreting cosmological simulations: Spherinator and HiPSter

Responsible Partner: HITS

The data produced by Exascale cosmological simulations will not only be large (exceeding several PBs), but also deep: hundreds of billions of resolution elements represented by hundreds of features corresponding to physical properties of galaxies and large-scale structures in a representative region of the universe evolving across hundreds of snapshots. While many ML tools have been developed to help scientists analyze large observational datasets, solutions are lacking for interpreting cosmological simulation outputs.

We have created the Spherinator and HiPSter tools (Polsterer et al 2024) to maximize the discovery potential of simulations by enabling scientists to explore and analyze the raw outputs produced by any cosmological code. Our tools use Representation Learning to automatically learn a compact projection of the simulated objects in a low-dimensional space that naturally describes their intrinsic characteristics. The data is seamlessly projected onto this latent space for interactive inspection, visual interpretation, sample selection, and local analysis.

Simulations alone do not lead to groundbreaking discoveries. Our tools are designed to bridge the gap between the complex model predictions produced by simulations, and observational data from large galaxy surveys. By seamlessly integrating the observed and simulated data into a single low-dimensional representation, these ML tools enable powerful simulation based inference and model selection techniques that maximally exploit the information contained in the raw data.

An interactive web demo using more than 60k synthetic images of simulated galaxies from the IllustrisTNG project can be found here.

Repo: github.com/HITS-AIN/Spherinator | github.com/HITS-AIN/HiPSter

Documentation: spherinator.readthedocs.io/en/latest/index.html

Multimedia material: video of demo: drive.google.com/file/d/1al4hJgt_sPSFu4q10bSrZaCcw_OVSrfs/view?usp=sharing

Publications: Polsterer K.L., Doser B., Fehlner A., Trujillo-Gomez S., 2024, arXiv, arXiv:2406.03810. doi:10.48550/arXiv.2406.03810

Generalizing ICL Predictions Across Simulations: A Velocity Dispersion-Based Deep Learning Approach

Responsible Partner: BSC

One of our machine learning projects focuses on a long-standing challenge in astrophysics: detecting intracluster light (ICL), a faint but important tracer of the dynamical history of galaxy clusters. Instead of relying on direct light-based measurements, which can be tricky to extract from simulations, we're exploring an alternative approach—inferring ICL from velocity dispersion maps.

Our pipeline uses deep learning models to uncover patterns between the kinematic structure of galaxy clusters and the distribution of ICL. We train various neural network architectures—including UNet, AttentionUNet, and Mamba—on mock velocity dispersion images generated from a diverse set of cosmological simulations, including DIANOGA, Illustris, Magneticum, MillenniumTNG, and Flamingo. The dataset is augmented with different projections and spatial shifts to help the models generalize across environments and simulation types.

At the heart of this workflow is a scalable, simulation-agnostic pipeline that extracts complex spatial correlations between velocity dispersion and ICL morphology. Our results show that the model can successfully reconstruct ICL maps from velocity dispersion data, with strong generalization between datasets like DIANOGA and Illustris.

Repo and documentation available soon.