Online program Overview Session: N-LT-04

Abstract

Bridging the MeV gap remains difficult due to the interplay of competing photon interaction processes, leaving a persistent sensitivity gap in gamma-ray astronomy. Compton telescopes based on 3D CZT Drift Strip Detectors (DSD) are promising candidates for bridging this gap, offering room-temperature operation, high stopping power, excellent energy and position resolution. A key open question is whether recoil electron tracking, which significantly improves angular resolution and detector sensitivity is achievable from raw detector signals. We investigate this using a machine learning approach trained on forward simulations of the 3D CZT DSD response. Principal Component Analysis of simulated electron tracks confirms that track structure is encodable in a low-dimensional representation, and neural network training to regress the recoil direction directly from detector waveforms is ongoing.

Abstract

Micro silica glass jewellery beads have emerged as a viable material for thermoluminescence dosimetry (TLD), offering a low-cost and widely accessible option for personal dose assessment. Their linear glow curve responses, governed by irradiation, thermal conditions, and sample characteristics, exhibit unique spectral elements and structure variability. While machine learning (ML) methods have been successfully applied to dose estimation and irradiation analysis, can ML be used to elevate silica glass beads beyond just decorative jewellery? In this work, a Conditional Variational Autoencoder (CVAE) is implemented to model and generate thermoluminescent glow curves from silica glass beads conditioned on experimental parameters. The model is trained on a dataset of experimentally acquired responses, augmented with Poisson noise to increase the dataset size and robustness. Conditioning variables include temperature profiles, irradiation and sample characteristics, enabling the model to learn relationships between physical conditions and the glow curve. Results show that the CVAE model can generate realistic glow curves that reflect the underlying conditioning variables. Generated predications maintain peak structure and overall shape, while also reproducing the expected linear relationship between total counts and absorbed dose. Model performance remains consistent when evaluated on previously unseen testing data with a mean figure of merit of 7.5%, although accuracy declines at low dose levels due to diminished signal intensity. The proposed approach demonstrates the potential of generative ML techniques for supporting dosimetry simulation, calibration, and data augmentation workflows for future research.

UK Ministry of Defence © Crown Owned Copyright 2026/AWE.

Abstract

Characterization of converter layer geometry is critical for determining the absolute detection efficiency of pixelated thermal neutron detectors. However, uncertainties in film deposition and source strength often lead to significant discrepancies between simulated models and actual experimental performance. This work addresses these challenges by implementing a Bayesian optimized FLUKA framework to determine the effective enriched Lithium-6 Fluoride converter thickness for an ADVACAM Timepix3 detector with a 1000 μm Silicon sensor. The FLUKA Monte Carlo model was benchmarked against experimental measurements using an Americium-Beryllium source with a nominal activity of 40 mCi and a documented tolerance of fifteen percent. The Bayesian optimizer successfully converged to an effective thickness of 5 μm after forty evaluations by minimizing the residuals between the nominal source activity and the simulation-implied activities. This methodology demonstrates a robust approach to resolving geometric uncertainties in hybrid semiconductor detectors where direct physical measurement of internal components is not feasible post-assembly. The resulting validated geometry provides a high-fidelity foundation for modeling future detection efficiency benchmarks across diverse neutron spectra. This work demonstrates a practical approach for integrating machine-learning optimization with Monte Carlo transport modeling to address geometric uncertainties in neutron detectors.

Abstract

Accurate characterization of spent nuclear fuel (SNF) is essential for verifying operator-declared information in nuclear safeguards. Conventional nondestructive assay (NDA) approaches for such characterization based on gamma-ray spectroscopy typically rely on extracting specific photopeak information (ex. 662 keV peak of ¹³⁷Cs), which compresses high-dimensional detector responses into a limited number of variables and may lead to information loss. To address this limitation, we propose a deep learning–based SNF characterization approach that directly utilizes sequential detector-response data for advanced characterization performance. The training dataset was generated through integrated ORIGEN and MCNP simulations for 5000 realistic SNF conditions, covering variations in initial enrichment, burnup, and cooling time. The proposed C2C-GRU-v2 architecture was designed to process long sequential detector-response data by combining Coordinate-to-Context embedding and GRU-based recurrent sequence modeling, enabling both event-level contextual learning and sequence-level generalization. For comparison, a conventional spectrum-based regression method using ¹³⁷Cs peak counts was also implemented under identical conditions. The proposed approach demonstrated improved characterization performance compared with the conventional method, particularly for cooling time estimation, achieving Relative Root Mean Square Errors (RRMSE) of 0.157 for burnup and 0.0456 for cooling time prediction. These results demonstrate the feasibility of deep learning–based sequential detector-response analysis for advanced SNF characterization in nuclear safeguards applications.

Abstract

Gamma-ray spectroscopy is a fundamental technique for radioisotope identification and quantification across safety-critical applications, where deep learning models have recently emerged as a data-driven alternative to traditional library-dependent methods. However, most existing deep learning approaches operate under a closed-set assumption that all isotopes during deployment are present in the training set; field deployments frequently violate this assumption, and the introduction of unknown isotopes leads to silent degradation. Recent OOD-detection studies have been reported, but they treat OOD spectra as anomalies to be flagged rather than quantified, and systematic evaluation under increasing unseen contaminant complexity remains lacking. This study proposes a denoising autoencoder (DAE) framework that reconstructs the clean target spectrum (¹³⁷Cs, ¹³³Ba) from contaminated measurements while simultaneously estimating target fractions through a dual-head architecture. Using Ce:GPS scintillator pulse data under a leave-one-out unknown noise (LOO-UN) protocol, the DAE achieved a mean absolute error (MAE) of 0.024 at 20% out-of-distribution (OOD) contamination in the two-unseen configuration, outperforming NNLS-known (0.038) and MLP (0.031), with near-constant accuracy as the unseen contaminant count increased from one to two (0.025 → 0.024).

Abstract

High-purity germanium (HPGe) detector systems are widely used for gamma spectroscopy; however, increasing throughput demands, expanding measurement configurations, and a declining base of expert users have made traditional efficiency calibration and transfer workflows increasingly impractical. Efficiency transfer between counting geometries has historically required manual configuration, custom made sources, and/or execution of Monte Carlo simulations by trained computational physicists, limiting scalability in routine laboratory environments.

We present an approach that embeds Monte Carlo-based efficiency transfer directly within a commercial spectral analysis software stack, enabling automated operation without additional user intervention. At runtime, sample geometries—already specified by laboratory personnel using standard container templates such as Marinelli beakers—are used to drive a ray-tracing-based gamma transport simulation coupled to detector interaction and pulse-processing models. The approach operates over an energy range of 20 keV to 2.5 MeV and produces both full spectra and efficiency curves.

Validation is performed through comparison of simulated spectra and efficiency curves against measured data and results from the Geant4 Monte Carlo toolkit. While not intended to match the full generality of Geant4, the proposed approach demonstrates comparable practical accuracy for routine laboratory configurations while significantly reducing user complexity. Additionally, seamless integration into existing workflows is demonstrated, enabling non-expert users to apply physics-based efficiency transfer without specialized training.

This work provides a scalable solution for automated, high-throughput gamma spectroscopy, bridging the gap between rigorous Monte Carlo modeling and practical laboratory usability.

Abstract

The Monte Carlo-based Particle Heavy Ion Transport Code System (PHITS) was used to model deuterium-deuterium (D-D) and deuterium-tritium (D-T) neutron generators for preliminary radiation safety assessment of the ongoing establishment of the Neutron Irradiation Facility (NIF) of the Philippine Nuclear Research Institute (PNRI). Neutron and secondary photon dose distributions were obtained for fully shielded, partially shielded, and unshielded configurations. In both D-D and D-T cases, the counting room dose rates were acceptable with respect to occupational dose limits for a 500-hour annual operating time. However, additional shielding should be explored for the D-T setup to further reduce dose rates in the work area. The dose maps presented the significance of the configuration of moderating and shielding materials in minimizing radiological doses during operation. The findings of this work will serve as a reference for the safe and reliable installation of the country's first neutron generator laboratory.

Abstract

Stand-off detection of special nuclear materials (SNM) in safeguards and non-proliferation applications demands imaging systems capable of tracking sources across dynamic environments. In this work, we introduce a Bayesian-Filtering (BF) algorithm for simultaneous  source localization and tracking where existing reconstruction methods, such as simple back-projection (SBP) and maximum likelihood expectation maximization (MLEM), are insufficient. We evaluate the BF algorithm using synthetic data generated with MCNPX-PoliMi for a simulated Compton imager consisting of a 4 x 7 array of LYSO scintillator crystals. Two independent sources are simulated, one sweeping in azimuth and one in elevation, each traveling from -90° to +90° at 1 m/s. The BF algorithm successfully tracked both sources across the full angular sweep, achieving mean angular errors of 6.4° and 6.03° for the azimuth and elevation cases, respectively. Additionally, preliminary results for multi-source tracking are successful with the algorithm tracking both sources up until their paths intersect which is a known challenge we are currently investigating.

Abstract

Radiomics analysis in total-body PET imaging is highly sensitive to acquisition variability, particularly count reduction, which directly affects noise characteristics and feature reliability. In this study, we investigate the impact of count reduction and multi-parametric harmonization on the robustness of multi-organ radiomics. A retrospective dataset of 550 patients from two centers was analyzed. Multiple count levels were generated from list-mode data, including standard and reduced counts (reduction factors up to 100). Multi-organ segmentation was performed automatically, and 107 IBSI-compliant radiomics features were extracted from SUV-normalized images. Feature harmonization was applied using ComBat under different configurations, including scanner, voxel size, radiopharmaceutical, and scatter correction. Reproducibility was evaluated using intraclass correlation coefficients across count levels and anatomical regions. Results show that radiomics stability is strongly dependent on both count level and feature type. Approximately 31% of features achieved good reproducibility (ICC ≥ 0.75) in at least one organ. Shape and first-order features remained relatively stable under moderate count reduction (CRF ≤ 20), while higher-order texture features exhibited substantial degradation at extreme count levels (CRF ≥ 50). Harmonization improved feature consistency, particularly for scanner- and resolution-related variations, but had limited impact under severe noise conditions. Taken together, the results indicate that preserving quantitative reliability in low-count PET requires careful consideration of both feature selection and acquisition variability. Rather than relying solely on harmonization, robust radiomics analysis should incorporate count-aware strategies, especially when extending to multi-center or ultra-low-count imaging scenarios.

Abstract

The Compton camera has emerged as a pivotal gamma-ray imaging technique, with critical applications in astrophysical observation, nuclear safety monitoring, and environmental radiation detection. In nuclear astrophysics, the detection and localization of gamma rays in the MeV range are indispensable for probing high-energy cosmic phenomena. Despite its importance, the advancement of MeV gamma-ray imaging instrumentation remains constrained by low interaction probabilities and high imaging complexity. This study presents a single-layer Compton camera prototype designed for enhanced spatial imaging, building upon the mission heritage of the BNU-120—a space-borne detector successfully deployed for monitoring gamma-ray bursts (GRBs). To transcend the original system's lack of spatial imaging capability, we propose a high-performance architecture that leverages the established BNU-120 framework while integrating a dual-ended readout to achieve a 4π detection solid angle. The prototype uses a 8×8 Ce:GAGG array coupled to Hamamatsu Multi-Pixel Photon Counters (MPPC, S14161-3050HS-08) at both ends. Each crystal measures 3×3×25 mm³ and is optically isolated by 0.2-mm-thick BaSO₄ reflective layers. System characterization demonstrated an energy resolution of approximately 5% (FWHM) at 511 keV. Depth-of-interaction (DOI) information, extracted via the energy ratio method, yielded a DOI resolution of roughly 2 mm (FWHM) along the crystal axis. Computational validation using GATE simulations of a U-shaped 511 keV planar source showed that the source geometry was clearly distinguishable through a back-projection (BP) algorithm. Subsequent point-source experiments verified the detector's capability to localize discrete sources. Future work will focus on realistic physical imaging experiments and the exploration of more advanced reconstruction algorithms to fully evaluate the module's Compton imaging performance.

Abstract

Precision physics at future e+e− Higgs factories set stringent requirements on calorimetric performance, with 3%/√E for electromagnetic particles and ∼ 30%/√E for neutral hadrons identified as key benchmarks. A hybrid dual-readout approach, combining a homogeneous crystal electromagnetic section with a fiber-based hadronic section, offers a cost-effective path toward simultaneously achieving both targets. Reaching such performance requires enhancing the longitudinal and transverse granularity of the crystal section beyond what is available in conventional homogeneous calorimeters and equipping each crystal cell with two independent readout channels to simultaneously collect Cherenkov and scintillation light. These features enable advanced particle-flow dual-readout reconstruction algorithms to exploit the maximum available information from each event. This work will present extensive laboratory characterizations together with beam-test measurements carried out at CERN on single-crystal PWO, BGO, and BSO cells instrumented with optical filters and SiPMs. These studies demonstrate event-by-event separation of Cherenkov and scintillation light, with both scintillation and Cherenkov yields exceeding the thresholds required for optimal electromagnetic and hadronic energy resolution. The results have guided the technological choices for the ongoing construction of full-containment calorimeter prototypes

Abstract

ePIC is the general purpose detector for the Electron-Ion Collider (EIC)
that is being built at Brookhaven National Laboratory and will be ready for first physics 
when the Electron-Ion Collider turns on.  ePIC is based
on a new 1.7T superconducting solenoid magnet.  The detector
consists of precision tracking, particle identification and 
calorimeter (electromagnetic and hadronic) to measure the
properties of electron-Ion collisions that will occur at the
EIC. In this paper we will present the over all design of the ePIC
detector, with a focus  on the Barrel Hadronic Calorimeter (BHCal)
which is the repurposed sPHENIX Outer Hadronic Calorimeter (HCal).
We will also present the current status of the BHCal design, test stand
and preliminary results of the new electronics readout system based on the 
CaloROC chip designed for ePIC

Abstract

The ATLAS Liquid Argon (LAr) calorimeter readout electronics are being upgraded for operation at the High-Luminosity Large Hadron Collider (HL-LHC), where up to approximately 200 proton-proton interactions per bunch crossing will significantly increase detector occupancy, data throughput, and trigger rates. While the calorimeter detectors themselves remain unchanged, the full readout electronics chain is being replaced. The upgraded off-detector processing system continuously receives digitized calorimeter signals from approximately 182000 channels at the 40 MHz bunch-crossing rate through high-speed optical links. The processing architecture is implemented in Advanced Telecommunications Computing Architecture (ATCA) systems hosting large FPGAs performing real-time reconstruction of energy and timing while satisfying stringent trigger latency requirements. Prototype hardware and firmware have been developed and validated in integration setups combining front-end electronics, timing distribution, and data acquisition systems. In parallel, advanced reconstruction methods including machine-learning approaches optimized for FPGA implementation are being investigated to improve robustness against pileup and overlapping pulses.

Abstract

An upgrade of the LHCb electromagnetic calorimeter (ECAL) is planned for Long Shutdown 3
(LS3) to address significant aging effects in the central region caused by radiation damage and
high particle flux. The innermost ECAL modules will be replaced with a new design based on
Spaghetti Calorimeter (SpaCal) technology, optimized for radiation tolerance and improved
performance under high occupancy conditions. A total
of 176 SpaCal modules will be installed. These are based on scintillating plastic fibers embedded
in dense absorbers. The innermost 32 modules will use 3D-printed tungsten absorbers to
withstand the highest radiation levels, while the remaining 144 modules will employ lead
absorbers produced through an innovative casting technique. Both configurations have been
prototyped and extensively validated in test beam campaigns at
DESY-II and CERN SPS. The measured energy resolution for electron beams is consistent with
performance requirements and shows good agreement with detailed Monte Carlo simulations. The
scintillating fibers will be read out using radiation-hard hollow light guides coupled to fast photon
detectors. In total, 3072 new readout channels will be implemented with dedicated front-end
electronics. In parallel with the installation of the new SpaCal modules, the current Shashlik
calorimeter modules will be rearranged to better match the spatial distribution of particle
occupancy. Simulation studies demonstrate a significant reduction in per-cell occupancy after the
LS3 upgrade, improving detector performance in Run 4 conditions. Benchmark analyses of key
physics channels confirm that this enhancement leads to a substantial gain in reconstruction
efficiency and measurement precision. Overall, the LS3 ECAL upgrade represents a crucial step
toward coping with higher luminosity conditions and anticipates key design features of the future
PicoCal system foreseen for implementation during LS4.