Low Activity Tritium Detection in CCDs Using Deep Learning Techniques

TL;DR

The paper addresses in situ detection of ultra-low-energy tritium beta radiation using thick, fully depleted CCDs by integrating measured data with Geant4 simulations. It evaluates classical and deep-learning classifiers (CNN, PFN, and autoencoder) for distinguishing tritium-induced tracks from backgrounds, finding that CNN and PFN offer the strongest discrimination (AUC ≈ 0.98) and achieve minimum detectable activities around 1.8–1.96 mBq/µl in 24 h, outperforming traditional methods. The work demonstrates potential for portable, high-sensitivity tritium monitoring and informs detector design and training-data requirements for robust deployment, with implications for GRAIL-like in situ sensing. It also highlights that while deep models excel with large labeled datasets, background-agnostic approaches (like the autoencoder) remain valuable when background characteristics are uncertain or variable.

Abstract

This study explores the use of charge-coupled devices (CCDs) for detecting low-energy beta particles from tritium decay - a critical signal for nuclear safety, nuclear nonproliferation, and environmental monitoring. We employ a dual approach utilizing both measured CCD data and detailed Geant4 simulations. Our analysis compares classical techniques with advanced deep learning methods, including convolutional neural networks (CNNs), autoencoders trained exclusively on tritium data, and preliminary studies on boosted decision trees (BDTs). The CNN, trained on mixed signal/background datasets, demonstrates superior classification performance, while the autoencoder shows the potential of unsupervised, background-agnostic strategies when background characteristics are poorly defined. These results highlight the excellent sensitivity achievable thanks to the background rejection made possible by information-rich CCD data, paving the way for improved portable tritium monitoring.

Figures (16)

• Figure 1: Particle tracks measured by a CCD without (left) and with (right) a tritium source placed 3$"$ from the detector.
• Figure 2: Zoomed in 10$\times$10 pixel views of particle clusters from background (left) and tritium (right).
• Figure 3: Electron ranges in silicon from NIST NIST, simulations by Cabello et. al. cabello2010, and our Geant4 simulation. The red points show the full path length traveled, while the brown points show the maximum depth reached in the sensor for isotropic emission. The energy spectrum of tritium is overlaid for reference.
• Figure 4: Simulated electron trajectories in a CCD at different incident energies (1--50 keV). The gray region represents the detector volume, with a 10 µm pixel width shown for scale. The red band illustrates a 10 nm thick dead layer of insensitive detector.
• Figure 5: The fraction of tritium beta particles that deposit energy in the sensitive CCD volume for dead layer thicknesses of 0, 10, 100, and 250 nm. The inset plot show the percentage of incoming tritium particles that are lost due to the dead layer.