Machine Learning Seminar: Georg Schwefer

Likelihood-free inference and gamma-ray astronomy

Estimating (physical) model parameters θ from experimental data x is one of the core tasks in any field of physics. Statistical estimation procedures frequently make use of the likelihood function p(x|θ). Often, the likelihood can not be evaluated explicitly, but can be efficiently sampled from. In these cases, methods from the field of likelihood-free inference are an attractive option.
I will present the FreePACT gamma-ray reconstruction algorithm for IACTs based on the likelihood-free method of neural ratio estimation and and showcase its performance compared to classical reconstruction methods and the ImPACT method that uses an analytical approximation of the likelihood on simulations of the planned CTAO southern array. Then, I will talk about potential extensions of the method to include image timings and to improve gamma-hadron separation as well as its application to data from water Cherenkov arrays.
Finally, I will mention an example for the enhanced science potential that comes with improved gamma-ray angular resolution, namely the study of compact X-ray Pulsar Wind Nebulae with CTAO.