Seeing through Matter: Optimizing Cosmic-Ray Muon Tomography for Faster and Reliable Cargo Screening

Cargo screening is essential for detecting illegal or dangerous nuclear materials in shipping containers.

Currently, X-ray systems (fig. 1) are widely used, but they are costly and struggle to detect densely shielded materials.

A safer, recent alternative is Muon Tomography (MT) (fig. 2), which uses naturally occurring cosmic-ray muons to image dense material.

Although MT is being used in real-world locations such as the Bahamas and the US–Mexico

border, it still faces key limitations. Detailed scans can take up to minutes, while anomaly detection can take up to seconds. Very small objects may also go undetected, and expert review is required to identify threats.

This project was motivated by the need to make MT faster, more reliable, and more practical for

real-world cargo screening. I investigated three main questions:

What is the minimum scan time required to detect an anomaly?

What is the smallest detectable volume of uranium?

Can machine learning assist in identifying dangerous materials?

To validate my Geant4 simulations and measure real-world muon flux, I also built a Geiger–Müller Muon Detector.

This project could help border services and transportation industries by reducing inspection

time, improving detection reliability, supporting faster decision-making and ultimately enhancing

global security.

Geant4 Simulations

I used Geant4, a C++ toolkit developed by CERN[5], to simulate Muon Tomography (MT), since it is difficult to create a physical MT system. I simulated a downscaled model of a realistic Cargo Container (fig. 3) since a real-life size setup would take a long time to simulate. My setup included a Cargo Container with the following objects: Uranium, Iron, Lead, TNT, Cocaine, Bananas and Plutonium, and detector layers in pairs in directions around the container. In each simulation, muons were simulated with varying creation (zenith) angles, based on the distribution, a formula used to approximate the zenith angles of muons when they are created from cosmic rays in the upper atmosphere.

Volume Creation

The data from the Geant4 simulations was saved into ROOT files, after which these files were used to reconstruct into a 3D volume using the Point-Of-Closest-Approach (POCA) algorithm. POCA (fig. 4) identifies where a muon bends to map 3D density.

Minimizing Time

I then tried to reduce the amount of data given as input to the reconstruction step to as little as possible, while still being able to see enough information in the final 3D volume. Then I calculated how much time the data would take to acquire in real-life. CNR and SNR were used to evaluate image quality.

Detection Limits

I then gradually decreased the volume of Uranium in the simulations until it was not visible in the volume anymore.

Machine Learning

Finally, I also created a Machine Learning model (Boosted Decision Tree) to assist in the detection of illegal objects in the volume, specifically Uranium and Plutonium.

Muon Detector

I also created a Geiger-Müller-Muon-Detector (fig. 5) and studied the average muon creation rate (flux) at different times of the day, for consecutive days.

The entire project process is shown in fig. 6.

Reducing Scan Time

The minimum scan time required to detect the presence of an anomaly was approximately

seconds (fig. 7). At this level, image quality was low, but sufficient to determine whether further

inspection was needed. This suggests MT could be used as a rapid pre-screening tool.

For higher-quality imaging, SNR and CNR analysis showed that scan times of to minutes

produced significantly clearer reconstructions, suitable for identifying material type and

structure (fig. 8).

Smallest Detectable Uranium Volume

The smallest detectable uranium volume was approximately cm³ (fig. 9). Below this threshold,

uranium could not be reliably distinguished from background noise. This defines an important

operational limit for MT systems and helps set realistic expectations for detection capabilities.

Machine Learning Model

The machine learning model (fig. 10) successfully classified materials with an accuracy of % for

uranium and % for plutonium. The lower performance for plutonium is likely due to its

similar density to uranium, making classification more difficult.

This demonstrates that AI can assist in reducing the need for manual interpretation and improve

consistency in identifying high-risk materials.

Muon Detector

The muon detector measured average muon flux was muon/min[6] and this slightly varied over time, with peak levels occurring in the morning and at dusk (fig. 11). Flux during these periods was approximately % higher than average. These results align with expected cosmic-ray behavior.

To validate the muon shower implemented in the Geant4 simulations, a Distance–Gap study was conducted. Experimentally, increasing the sensor separation from cm to cm resulted in a % reduction in detected flux. The simulation produced a comparable reduction of %. The small difference of % indicates strong agreement, confirming that the model reliably captures both the geometric acceptance and the angular distribution of cosmic-ray muons.

This project demonstrates that Muon Tomography can be optimized to operate more efficiently

while maintaining reliable detection capabilities. A key finding is that rapid scans (~ seconds)

can be used for anomaly detection, allowing for faster decision-making in cargo screening

workflows.

The identification of a cm³ detection threshold for Uranium highlights both the strengths and

limitations of MT. While effective for detecting significant threats, very small quantities may

remain undetected, emphasizing the need for complementary methods.

The integration of machine learning shows strong potential to reduce reliance on expert analysis

by automating material classification and improving consistency.

Additionally, experimental measurements of muon flux support the realism of the simulation,

strengthening confidence in the results.

Overall, this work contributes to making MT more practical for real-world use by improving

speed, reliability, and automation. These advancements could enhance border security, reduce

inspection delays, and support safer, more efficient global trade systems.

My next steps are:

This project used a downscaled simulation with ideal detector conditions, and the POCA algorithm has known spatial limitations. Future work will include simulating full-scale cargo systems and testing more advanced reconstruction methods (MLEM).

The machine learning model will be improved using more diverse and realistic datasets. I also plan to build a scintillator-based detector (e.g., CosmicWatch[10]), which is required for tracking the muon pathway to compare with my current system.

Further studies will investigate environmental effects on muon flux and explore additional physics processes, such as muonic atom formation, to enhance detection capabilities.

Thank you to my mentors, my teachers and my parents for guiding me through my project.

Physics Matters tutorial on YouTube - https://www.youtube.com/watch?reload=9&v=Lxb4WZyKeCE&list=PLLybgCU6QCGWgzNYOV0SKen9vqg4KXeVL

Cosmic Shower project on GitHub - https://github.com/maoderos/cosmic_rays

Muon Imaging IAEA - https://www-pub.iaea.org/MTCD/Publications/PDF/TE-2012web.pdf

Barnes, S., Georgadze, A., Giammanco, A., Kiisk, M., Kudryavtsev, V. A., Lagrange, M., & Pinto, O. L. (2023). Cosmic-ray tomography for border security. Instruments, 7(1), 13.

Geant4 - http://geant4.web.cern.ch/

Bae, J., & Chatzidakis, S. (2022). A new semi-empirical model for cosmic ray muon flux estimation. Progress of Theoretical and Experimental Physics, 2022(4), 043F01.

Allison, J., Amako, K., Apostolakis, J., Arce, P., Asai, M., Aso, T., ... & Yoshida, H. (2016). Recent developments in Geant4. Nuclear instruments and methods in physics research section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 835, 186-225.

Allison, J., Amako, K., Apostolakis, J. E. A., Araujo, H. A. A. H., Dubois, P. A., Asai, M. A. A. M., ... & Yoshida, H. A. Y. H. (2006). Geant4 developments and applications. IEEE Transactions on nuclear science, 53(1), 270-278.

Agostinelli, S., Allison, J., Amako, K. A., Apostolakis, J., Araujo, H., Arce, P., ... & Geant4 Collaboration. (2003). Geant4—a simulation toolkit. Nuclear instruments and methods in physics research section A: Accelerators, Spectrometers, Detectors and Associated Equipment, 506(3), 250-303.

Cosmic Watch Muon Detector - http://www.cosmicwatch.lns.mit.edu/detector

AI tools for Research Assistance (Gemini, ChatGPT)