Advanced Material Identification in Muon Tomography

Abstract

Muon tomography employs cosmic ray muons to image the internal structure and material composition of target volumes, such as maritime cargo containers. Despite its potential, the high dimensionality of reconstructed muon tomography data poses challenges for material discrimination and the discernment of structural layouts. In this study, we propose a novel methodology for material identification through the insightful visualization of data embeddings. Our approach synergizes the integrated workflow provided by the Blender-to-Geant4 (B2G4) framework for the rapid simulation of a variety of scenes and the standard muon tomography reconstruction methods. Then, we utilize an unsupervised exploratory data analysis technique along with a novel structured methodology for muon data sampling aimed at discerning materials afforded by the resulting embedding spaces. The experimental results resolve the difficulties in applying machine learning for muon tomography data and highlight new venues for principled material classification.