Online program Overview Session: NSS-11

Abstract

The maintenance and dismantling of high-energy accelerators at CERN generate activated waste with significant variability in irradiation history, cooling time, and material composition. If activity levels are negligible, waste can be cleared from regulatory control in Switzerland and reused as conventional material. We present a method for radiological classification which is conservative, without being as penalizing as more traditional approaches based on a worst-case scenario. More specifically, we classify accelerator components as conventional or radioactive based on a radiation measurement (e.g. with a bismuth germanate detector), and on the application of a predictive model. The models here presented were trained over ~60,000 synthetic waste items and their radiological properties. Each synthetic item represents a specific activation scenario, generated through probabilistic sampling of operational parameters. The associated radiological properties were computed with ActiWiz, a CERN-developed activation code based on FLUKA Monte Carlo simulations and evaluated neutron data libraries. Focusing on the classification of metallic accelerator components, we compare the performance of logistic regression based on a bismuth germanate detector response, with that of a neural network model incorporating both the detector response and the item mass. Results show that the neural network improves classification performance and robustness relative to logistic regression, reducing misclassification of borderline cases while maintaining compliance with Swiss regulatory criteria. The proposed framework establishes a robust and scalable clearance workflow, adaptable to the evolving requirements of large-scale research facilities.

Abstract

Radiological characterization is a key step in nuclear decommissioning, particularly during final operations where contaminated concrete structures must be assessed prior to decontamination. Conventional approaches combine surface measurements with radiochemical analyses of samples at different depths, and use geostatistical methods to estimate 3D contamination distributions. Despite their efficacy, these methods remain intrusive, time-consuming, and costly. In this work, we investigate a complementary non-destructive methodology based on in situ gamma spectrometry conducted at the surface using a HPGe detector. The primary objective of this methodology is to reduce the need for destructive sampling and improving the understanding of contamination phenomena in concrete. This study focuses on the analysis of ¹³⁷Cs, a predominant gamma-emitting radionuclide frequently encountered in nuclear facilities. The proposed approach relies on inverse reconstruction of depth contamination profiles, where the volume of concrete is discretized into layers. The feasibility of the approach was confirmed through simulations. The detector response is modeled using Monte Carlo simulations with MCNP and is combined with the ML-EM algorithm to solve the inverse problem. However, investigations also revealed limitations, especially ambiguities and non-unique solutions in complex configurations. We propose to integrate prior knowledge on the contamination profile (Fickian diffusion) and demonstrate its benefits in addressing the aforementioned limitations. In controlled laboratory conditions, the experimental validation was executed with a calibrated ¹³⁷Cs point source shielded with a concrete barrier. With this configuration, the ability to estimate source position by using relevant energy ranges were demonstrated. These results highlight both the potential and inherent limitations of ML-EM, thus motivating hybrid approaches combining physics-based modeling and learning methods.

Abstract

Charge-comparison pulse-shape discrimination in organic scintillators fails at high irradiation rates due to the effects of pulse pileup. While pileup rejection reduces particle
misclassification, this severely degrades detector efficiency. We present a new machine-learning pipeline to extract particle-type information from piled-up trans-stilbene waveforms using a novel joint autoencoder–classifier design. Events are classified as single or pileup, then routed to separate multi-task convolutional autoencoders, which extract the important features of the input waveform and perform n/γ classification. Trained on 519,910 labeled Time-of-Flight waveforms and synthetically generated pileup data, the system achieves 97.66% accuracy for single/pileup detection, 99.42% n/γ classification accuracy on single pulses, and 90.15% classification accuracy of both constituents in pileup waveforms. This approach is shown to substantially outperform charge-comparison-based analytical pulse-shape discrimination by improving pileup detection and preserving neutron and gamma-ray counts.

Abstract

Neutron source localization and identification are critical for nuclear safeguards and security, yet current scatter-based imaging pipelines typically decouple spatial reconstruction from spectral analysis, treating both as separate post-processing steps. This work introduces a unified machine-learning framework for joint spectral-directional neutron source inference using energy-resolved simple back-projection (SBP) images. The proposed framework combines unsupervised Gaussian mixture model localization with a lightweight k-nearest neighbors classifier. Despite its simplicity, the approach achieves strong performance on controlled laboratory data involving Cf-252 and PuBe sources in single- and mixed-source configurations, with a mean angular localization error <5º and a source classification accuracy >98%. These results demonstrate that low-complexity statistical learning applied to energy-resolved SBP data enables fully automated, end-to-end inference of both source direction and source identity without manual region selection or separate spectral analysis, offering a practical pathway toward more scalable and deployable neutron imaging systems for real-world safeguards and security operations.

Abstract

In this study, a real-time gamma-ray contribution direction imaging system using a Fractal Radiation Imaging Element (FRIE) detector and a Multi-Layer Perceptron (MLP) was developed and evaluated. The system consists of an FRIE detector composed of sixteen GFAG scintillators, a 360-degree camera, and a PC for real-time analysis and visualization. Radiation contribution directions were estimated from count-rate data of the sixteen detectors using an MLP trained with Cs-137 point-source irradiation data including statistical fluctuations. Evaluation results showed that the average angular estimation error was approximately five degrees, and the estimation accuracy improved as the count rate increased. Experiments using measured data demonstrated that the estimated source direction gradually converged through temporal accumulation of measurement data, even under low count-rate conditions. Compared with the conventional similarity-based method, the proposed method reduced the direction estimation error while enabling real-time gamma-ray contribution direction estimation. These results demonstrate the effectiveness of the proposed system for real-time localization of gamma-ray sources in nuclear decommissioning environments.

Abstract

Radiation detector networks provide a potential means of monitoring the movement of radiological material within urban environments. Such systems require automated methods capable of identifying non-background isotopes within collected spectra, particularly for count-starved measurements acquired in complex backgrounds. In 2018, the London-based SIGMA trial collected approximately 1.5 billion one-second gamma spectra using a combination of stationary and mobile scintillation detectors. This work presents a comparison between three viable isotope identification tools, to ascertain which is best placed to be used in environments with count-starved gamma spectra, such as those typically seen in large datasets like SIGMA. Representative spectra were created by combining laboratory collected spectra and Geant4 simulated spectra with real-world SIGMA background. Tranches of spectra were tested in each of the chosen methods: InterSpec, wavelets, and a convolutional neural network. Although all three methods showed promise when identifying a range of ¹³⁷Cs spectra, the CNN outperformed the others when presented with spectra with the smallest photopeaks, making it the preferred approach for isotope detection in urban landscapes.
UK Ministry of Defence © Crown Owned Copyright 2026/AWE