Simulation-based inference with neural posterior estimation applied to X-ray spectral fitting II -- High-resolution spectroscopy with the X-ray Integral Field Unit

Abstract

X-ray spectral fitting in high-energy astrophysics can be reliably accelerated using Machine Learning. In particular, Simulation-based Inference (SBI) produces accurate posterior distributions in the Gaussian and Poisson regime for low-resolution spectra, much faster than other exact approaches such as Monte Carlo Markov Chains or Nested Sampling. We now aim to highlight the capabilities of SBI for high-resolution spectra, as what will be provided by the newAthena X-ray Integral Field Unit (X-IFU). The large number of channels encourages us to use compressed representations of the spectra, taking advantage of the likelihood-free inference aspect of SBI. Two compression schemes are explored, using either simple summary statistics, such as the counts in arbitrary bins or ratios between these bins. We benchmark the efficiency of these approaches using simulated X-IFU spectra with various spectral models, including smooth comptonised spectra, relativistic reflexion models and plasma emission models. We find that using simple and meaningful summary statistics is much more efficient than working directly with the full spectrum, and can derive posterior distributions comparable to those from exact computation using nested sampling. Multi-round inference converges quickly to the good solution. Amortized single round inference requires more simulations, hence longer training time, but can be used to infer model parameters from many observations afterwards. Information from the emission lines must be accounted for using dedicated summary statistics. SBI for X-ray spectral fitting is a robust technique that delivers well calibrated posteriors. This approach shows great promises for high-resolution spectra, offering its potential for the scientific exploitation of the X-IFU. We now plan to apply it to the current era of high-resolution telescopes, and further challenge this approach with real data.