2025 workshop program
Learning compact representations from Chandra X-ray spectra
The work
Abstract
Understanding the physical nature of astronomical X-ray sources benefits from the analysis of their spectral properties. In this work, we apply deep learning to learn compact, informative representations of Chandra X-ray spectra to support unsupervised classification and interpretation of sources’ characteristics. We develop a transformer-based autoencoder that compresses input spectra into an 8-dimensional latent space, and we evaluate the learned features through reconstruction accuracy, clustering performance, and correlation with physical quantities such as hardness ratios.
Clustering in the latent space leads to a balanced classification accuracy of ~40% across eight source classes, rising to ~70% when restricted to AGNs and X-ray binaries. Furthermore, the latent dimensions correlate with non-linear combinations of fluxes, indicating that the model captures physically meaningful patterns. These results suggest that deep latent representations can facilitate an interpretable analysis of large X-ray source catalogs without relying on manual feature engineering or labeled training data.