2025 workshop program
Neural Posterior Estimation for MYTorus Decoupled: Training on Observation-Driven Parameter Grids
The work
Abstract
We present the development of an automated inference tool tailored to extract key physical parameters from obscured AGN X-ray spectra by means of more complex physical models than ever before with machine learning. For our pilot work, we use the decoupled MYTorus model in a toroidal or clumpy geometry, with separate direct and scattered (“reflected”) continua, as well as Fe K fluorescense. Such a complex model poses a significant computational challenge for traditional inference techniques. To address this, we construct a physically informed, observation-driven training grid, based on the parameter space spanned by nearby AGN observed with NuSTAR. We use this grid to train a Neural Posterior Estimation (NPE) model within the framework of simulation-based inference (SBI). The parameters inferred are the photon index (Γ), the global and line-of-sight equivalent hydrogen column densities (N_Hs and N_Hz), and the reflection scaling factor (A_S), each with associated uncertainties. This approach demonstrates a path to likelihood-free posterior estimation using neural networks, providing a scalable alternative to traditional methods for parameter inference in complex astrophysical models.