2025 workshop program
Uncovering the Giants in the Sky - Semantic Segmentation of Galaxy Clusters in Astrophysical Surveys
The work
Abstract
This project aims to enhance our understanding of the complex multiwavelength processes shaping galaxy clusters. Traditional methods for source detection and parameter inference mainly focus on single-wavelength analyses and require significant effort in data processing and modeling. Machine learning offers a unified approach, combining multiple wavelengths and eventually bypassing conventional pipelines.
We trained a UNet-based neural network for galaxy cluster detection via semantic segmentation using the eROSITA DR1 dataset and cluster catalogs from various wavelengths. We explored how different wavelengths - such as X-ray, Sunyaev-Zel’dovich (SZ), radio, and optical - influence the network’s performance (e.g. compared to classical algorithms) and studied correlations across the different wavelengths.
The correlations learned by the network provide a set of features that describe the interplay between different wavelengths. These features can be used to enhance parameter inference of galaxy clusters, enabling the direct extraction of physical properties from raw data across multiple wavelengths.