Quantifying Uncertainty in Particle Physics using Probabilistic Graphical Models

Measurements in particle physics have inherent uncertainties, which may be caused by intrinsic stochasticity or introduced through the experimental setup. These uncertainties are critical to differentiating scientific theories and are of particular interest when solving inverse problems, where causes are determined from observations – such as localization of an interaction within a detector. Neural networks and some machine learning algorithms can be used to calculate uncertainties using parameter estimation, but are limited in their ability to quantify uncertainties on an event by event basis. Alternatively, probabilistic graphical models use a graph-based representation and make use of the independences between variables to more compactly represent complex probability distributions. We present a method which uses a probabilistic graphical model to infer posterior probabilities and then demonstrate its capabilities using the example of localization of interactions within a detector in a dark matter direct detection experiment.