freedom

Likelihood-free reconstruction method for Cherenkov neutrino detectors

Example likelihood contours for a toy neutrino event as a function of vertex position. Compared are the freedom NN approximations (solid lines) compared to the true likelihood (dashed lines).

While in the project retro we built a reconstruction for IceCube DeepCore based on an event model together with pre-tabulated photon propagation values to build an approximate, fast likelihood.

In the project graphnet in contrast, we are building a framework to use GNNs as regressors (and classifiers) for event reconstruction.

In this project, we explore a middle ground between the two. We use a technique of likelihood-free inference to build a NN surrogate likelihood, that can then be used in conjunction with optimizers or samplers for the inference of reconstruction quantities, such as an event’s energy.

Further Information

Collaborators: Jan Weldert (JGU), Garrett Wendel (PSU), Dr. Aaron Fienberg (PSU), Prof. Dr. Sebastian Böser (JGU), Prof. Dr. Doug Cowen (PSU)

Preprint available here: https://arxiv.org/abs/2208.10166