Transfer learning empowers material Z classification with muon tomography

From Muon Imaging to Transfer Learning — A Moment of Inspiration

Cosmic-ray muons, owing to their exceptional penetrating power, have long been regarded as a powerful tool for non-invasive imaging and material identification. However, conventional muon tomography methods rely heavily on complex physical reconstruction algorithms or extensive amounts of labeled data, which significantly limits their deployment in real-world applications—particularly when target materials are concealed or shielded.

To address this challenge, Professor Liangwen Chen and his collaborators have, for the first time, introduced transfer learning into muon tomography. By treating bare materials as the source domain and shielded materials as the target domain, the team enables knowledge acquired from richly labeled data to be transferred to scenarios where labels are scarce or entirely unavailable.

From Bare Materials to Shielded Materials: A New Learning Paradigm

Using detailed Geant4 Monte Carlo simulations, the researchers generated large-scale datasets of muon scattering angles for nine representative materials spanning low-, medium-, and high-Z categories. They then designed two lightweight neural network frameworks:

• a fine-tuning-based transfer learning model, suitable for cases with limited labeled target data;
• a domain-adversarial neural network (DANN), designed for fully unlabeled target domains.

Both approaches successfully learned the intrinsic physical correlations between scattering angle distributions and material atomic number classes, even after the materials were encased in common shielding substances such as aluminum or polyethylene.

Professor Chen explains: “Our goal is to move beyond idealized conditions and address the data limitations that arise in real inspection scenarios. Transfer learning allows us to preserve the fundamental physical characteristics of muon scattering while efficiently adapting to unknown environments under shielding.”

High Prediction Accuracy Enabled by Physics-Guided Sampling

In addition to the learning framework, the team also proposed a novel sampling method based on the inverse cumulative distribution function (CDF), designed to preserve the physical characteristics of muon scattering angle distributions from a physics-informed perspective. Compared with conventional random sampling, this method improves feature quality and model performance while enhancing interpretability from a physical standpoint.

When applied to shielded materials, the combined strategy of inverse CDF sampling and transfer learning improved classification accuracy by approximately 10% compared with direct prediction methods without transfer learning. Specifically, the fine-tuning-based transfer learning model achieved an overall accuracy of 98% in the more challenging task of identifying aluminum-shielded materials. In the same aluminum-shielded scenario, the more broadly applicable DANN model achieved:

• over 96% overall Z-class classification accuracy,
• nearly 99% accuracy for high-Z materials, which are of primary concern in nuclear security and safeguards.

Advancing Intelligent Muon Imaging Technologies

These findings highlight the strong potential of transfer learning in addressing long-standing challenges in muon tomography, particularly those related to limited data availability and high reconstruction costs. The proposed approach shows promise for applications in scenarios such as cargo inspection, nuclear safeguards verification, and arms control.

Professor Chen notes: “This work demonstrates that advanced machine learning can complement rather than replace physical principles. By integrating transfer learning with muon physics, we can achieve reliable material identification under realistic constraints.” Looking ahead, the research team plans to extend this framework to more complex geometries and experimental conditions, including detector effects and mixed-material scenarios. The ultimate goal is to develop intelligent, low-cost muon tomography systems capable of stable operation in real-world environments.

Professor Chen concludes: “By integrating simulation, physics, and data-driven learning, this research opens new pathways for applying artificial intelligence to nuclear science and security technologies.”

The complete study is via by DOI: https://doi.org/10.1007/s41365-026-01901-w