Muon event localisation with AI

Abstract

Low-cost muon detectors utilising cheap plastic scintillators and a limited number of individual silicon photomultipliers (SiPMs) offer a compelling approach to cheap experimental designs, provided the event localisation of a traversing particle can be accurately determined. In this theoretical work, we use Geant4 to simulate a diverse range of detector configurations, shapes and SiPM photosensors, predicting the light intensity received at a given SiPM. Testing a range of methods to localise muon events we determine that machine learning techniques outperform analytic models, and of these, a simple gradient boosted framework is the most reliably accurate localisation technique for our simulated scintillators. We find that a simple square scintillator outperforms other geometries and that AI performs, when applied to this shape, with a linear relationship between the positional accuracy of the event recovery and the average distance between photosensors around the detector perimeter.