Pillar embedding visualization for muon-scattering tomography

Abstract

Muon-scattering tomography (MST) utilizes the deflection of cosmic-ray muons to non-invasively reconstruct the three-dimensional internal structure and material composition of concealed objects, such as those in maritime cargo. Yet, the high dimensionality of reconstructed MST volumes and sparsity of muon hits hinder reliable material discrimination and structural interpretation. We present an unsupervised workflow that visualizes learned data embeddings for material identification. The pipeline couples the Blender-to-Geant4 simulation framework, enabling the rapid prototyping of complex 3D scenes with a standard and widely adopted MST reconstruction algorithm, the Point of Closest Approach (PoCA), to reconstruct the scenes. A structured muon-data sampling grid, termed pillars, feeds an exploratory embedding technique that reveals discriminative material patterns in the reconstructed outputs. Experimental results demonstrate that the proposed approach mitigates key machine-learning challenges in MST; at the same time, they reveal the intrinsic limitations of PoCA estimates for mainstream material classification with machine-learning approaches, and we introduce corrections that enhance visualization and enable data-driven analysis in practical MST deployments.