Dortmund 2021 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

T: Fachverband Teilchenphysik

T 43: Neutrino physics without accelerators II

T 43.8: Talk

Tuesday, March 16, 2021, 17:45–18:00, Tr

Likelihood-free inference for low-energy reconstruction in IceCube DeepCore — •Jan Weldert¹, Philipp Eller², and Sebastian Böser¹ for the IceCube collaboration — ¹JGU, Mainz, Germany — ²TUM, Munich, Germany

DeepCore, the low energy extension of the IceCube neutrino observatory at the geographic South Pole, detects neutrinos at a rate on the order of mHz resulting in unprecedentedly large event samples. Reconstructing the latest generation of these samples (∼300.000 νs) is currently computationally expensive (∼40s per event). In addition, the employed max. LLH method includes simplifications in the photon propagation in ice which limit the reconstruction accuracy but are hard to overcome in the current approach.

Machine learning techniques can solve problems at a tremendous speed. But they also come with disadvantages, e.g. most will just give you a point estimation without any information about the uncertainty. Likelihood-free inference is a possibility to combine the speed of neural networks with a likelihood-based approach, which is very well understood. The main idea is to let a network learn a function proportional to the likelihood, which can then be used for a max. LLH reconstruction. While this is slower than a pure deep learning approach, it offers the possibility to perform likelihood scans for error estimation or test coverage.

In my talk I will present the application of likelihood-free inference to the reconstruction of low-energy events in IceCube-DeepCore. We achieve speed-ups up to a factor of 100 at comparable resolutions.