AI-Based Detection of Dust-Scattering Halos in eROSITA Survey

Astrophysics has always been data-driven, and in today’s era of big data and artificial intelligence, automated tools are becoming essential for discovery. At the same time, many astrophysical phenomena are subtle and spread across vast datasets, making them well-suited for machine learning approaches. Dust-scattering halos (DSHs), faint X-ray rings or diffuse structures produced when interstellar dust grains scatter photons from bright X-ray sources, are a prime example. These halos provide valuable information about both the distance to the source and the distribution of dust along the line of sight.
 
Our group has previously studied the DSH around the black hole binary 4U 1630–47, combining Chandra and Swift data to place new constraints on its distance (Kalemci et al. 2018).  we extended this work by using Chandra and APEX data to refine 3D molecular cloud representations by machine learning approaches and ray tracing simulations, resulting in improved distance estimates to the same source (Kalemci et al. 2025).
Building on this foundation, the project will focus on generating synthetic dust scattering halo datasets, incorporating instrumental effects with SIXTE, and training deep learning–based image detection models to recognize DSHs and distinguish them from the point spread function of X-ray telescopes. These models will then be applied to wide-field survey data such as eROSITA, where the challenge lies in distinguishing faint halo structures from background and noise. The project aims to deliver an initial detection pipeline and benchmark results, demonstrating the feasibility of automated halo searches in wide-field X-ray datasets. This proof-of-concept will open the door to scalable applications and give students hands-on exposure to modern techniques in astrophysics and deep learning.