MSc Thesis Defense: Atakan Saraçyakupoğlu, GENERATING DUST SCATTERING HALO IMAGES WITH CONDITIONAL DENOISING DIFFUSION PROBABILISTIC MODELS USING PHYSICS-INFORMED AUXILIARY LOSS

GENERATING DUST SCATTERING HALO IMAGES WITH CONDITIONAL DENOISING DIFFUSION PROBABILISTIC MODELS USING PHYSICS-INFORMED AUXILIARY LOSS

Atakan Saraçyakupoğlu
Computer Science & Engineering, MSc. Thesis 2025

Thesis Jury

Assoc. Prof. Kamer Kaya (Thesis Supervisor)

Prof. Emrah Kalemci (Thesis Co-Supervisor)

Prof. Erchan Aptoula

Asst. Prof. Onur Varol

Asst. Prof. Arkadaş İnan Özakın

Date & Time: December 18th, 2025 – 10:30 AM

Place: FENS L048

Keywords: Dust scattering halo, X-ray astronomy, diffusion models, cDDPM, RefleX

Abstract

Dust grains scatter X-ray photons at small angles, producing diffuse halos around bright X-ray sources. Despite their importance, modeling these halos remains challenging. Analytical models such as NewDust lack geometric flexibility and neglect multiple scattering, whereas Monte Carlo ray-tracing codes like RefleX capture these effects accurately but are computationally expensive. To address these limitations, we develop a physics-informed Conditional Denoising Diffusion Probabilistic Model (cDDPM) trained on RefleX-generated simulations to approximate dust-scattering halo images conditioned on hydrogen column density (n_H) and cloud distance ratio (x). To assess the impact of data representation and physics-based loss, we evaluate several model variants and find that variance stabilization (via the Anscombe transform) of the training data substantially improves accuracy, while auxiliary physical loss accelerates convergence. The resulting Anscombe-based models achieve radial reduced-chi^2 ≲ 6 on the testing dataset, while requiring several orders of magnitude less computational time.