Speaker
Description
Inferring theory parameters starting from observed events is a difficult task in high energy physics. This becomes particularly troublesome when dealing with events whose observables are not precisely measured and we want to understand the inference capability of multiple experimental setups. As a representative scenario, we will consider the production of ALPs and their subsequent decay into photons at beam-dump experiments. This type of signal presents issues due to the difficulty in reconstructing the photon properties, especially their directionality. In addition to this, the design is not finalized, which means that we would like to understand how modifying the detector properties changes our inference power. We construct a set of conditional invertible neural networks for the different experimental setups under scrutiny: by learning the posterior starting from the observed events, we are able to correctly state the model parameters and most importantly the uncertainty on them. The uncertainty on the model parameters can be seen as a measure of the inference power of the experimental setups, thus providing a way to compare them.
