Optimization of a cosmic muon tomography scanner for cargo border control inspection

TL;DR

This work develops and compares two overarching pathways to optimize muon scattering tomography for cargo border control: (i) GEANT4-based, physics-faithful simulations to study detector geometry and secondary-particle effects, and (ii) TomOpt, a differentiable, end-to-end optimization framework that uses gradient descent and Bayesian optimization to tune detector layouts and volume-inference methods. GEANT4 studies reveal clear design trade-offs: minimizing in-plane gaps boosts acceptance, while larger out-of-plane spacing improves angular resolution at some cost to efficiency; secondary hits contribute minimally to material discrimination and only modestly affect reconstruction. TomOpt demonstrates two volume-inference strategies, radiation-length-based $X_0$ and BCA, with the latter proving more robust to noise. Incorporating Bayesian optimization via the Ax library, BO-guided detector configurations yield notable but not overwhelming gains over intuitive baselines, with the BCA method delivering more stable improvements than the $X_0$ approach, guiding future work toward more robust ML-based volume inference for scalable border-security applications.

Abstract

The past several decades have seen significant advancement in applications using cosmic-ray muons for tomography scanning of unknown objects. One of the most promising developments is the application of this technique in border security for the inspection of cargo inside trucks and sea containers in order to search for hazardous and illicit hidden materials. This work focuses on the optimization studies for a muon tomography system similar to that being developed within the framework of the `SilentBorder' project funded by the EU Horizon 2020 scheme. Current studies are directed toward optimizing the detector module design, following two complementary approaches. The first leverages TomOpt, a Python-based end-to-end software that employs differentiable programming to optimize scattering tomography detector configurations. While TomOpt inherently supports gradient-based optimization, a Bayesian Optimization module is introduced to better handle scenarios with noisy objective functions, particularly in image reconstruction-driven optimization tasks. The second optimization strategy relies on detailed GEANT4-based simulations, which, while more computationally intensive, offer higher physical fidelity. These simulations are also employed to study the impact of incorporating secondary particle information alongside cosmic muons for improved material discrimination. This paper presents the current status and results obtained from these optimization studies.

Figures (29)

• Figure 1: A visualization of a particular configuration of the detector setup in the GEANT4 simulation. The sensitive detector panels are shown in gray, the shipping container in green and a small cargo object in red. The front view (a) shows that the coverage extends around all four long sides of the container whilst the gaps between hodoscope panels are emphasized in the side view (b).
• Figure 2: Locations of reconstructed PoCA points within the container in the $xy$-plane (looking down on the container). The scenario used a 100$\times$100$\times$50 cm cargo of lead, centered in the container.
• Figure 3: The resolution of the reconstructed scattering angle in two different example scenarios, (a) the nominal configuration and (b) where the vertical spacing between hodoscope plates has been increased.
• Figure 4: The change in efficiency (top) and angular resolution (bottom) metrics as the gap between hodoscopes in the $x$-direction is varied.
• Figure 5: The change in efficiency (top) and angular resolution (bottom) metrics as the gap between hodoscopes in the $z$-direction is varied.