Online program Overview Session: NSS-26

Abstract

Constraining a Transformer decoder to a known physical generative structure induces emergent interpretability in the encoder's self-attention without labels or supervision. We demonstrate this in gamma-ray spectroscopy, where measurements are Poisson-sparse at one-second mobile survey dwell times, environments are nonstationary, and labeled decompositions are unavailable. Observed spectra arise as nonnegative additive superpositions of source contributions, which we encode using a Nonnegative Matrix Factorization (NMF)-constrained decoder paired with a Transformer encoder. The learned representations organize into physically meaningful components, separating environmental substructures and recovering known photopeak features at correct energies. In contrast, an identical Transformer with an unconstrained decoder produces unstructured latent representations, with 48.6% of weights taking negative values that are inconsistent with photon-counting physics, isolating the decoder constraint as the causal mechanism. The proposed architecture achieves state-of-the-art anomaly detection on 6,221 field-collected spectra, outperforming linear baselines and unconstrained autoencoders while exhibiting reduced run-to-run variance, demonstrating that embedding known generative structure into decoder design can induce interpretable representations across domains governed by additive physical processes.

Abstract

Reconstruction of surface radiation contamination from sparse airborne measurements has matured rapidly with deep learning, but a validation gap persists: the surface activity is rarely measurable in operational deployments, leaving no oracle for evaluating predictions. We address this by reframing airborne mapping as an inverse deconvolution problem in which a known forward operator maps surface activity to aerial fluence via altitude blur and detector response. The framing yields a ground-truth-free validation protocol: each candidate surface reconstruction is forward-simulated through the operator and compared to held-out aerial measurements. We characterize the protocol on a synthetic benchmark with a 64-by-64 ground-truth surface of Gaussian hotspots, a fixed Gaussian forward kernel, 150 sparse aerial samples per scene, and 200 held-out validation points. Four reconstructions are compared: inverse distance weighting (IDW), ordinary kriging, IDW followed by Wiener deconvolution, and a small encoder-decoder convolutional network. Across 40 scenes per method, the round-trip aerial-space root-mean-square error correlated with the oracle surface-space root-mean-square error, with a pooled Pearson correlation of 0.871 across methods and within-method correlations from 0.71 to 0.82. The protocol identified the encoder-decoder network as the most accurate reconstruction in agreement with the oracle ranking, without access to ground truth. These findings suggest that round-trip forward-simulation residuals can serve as a practical proxy for surface reconstruction quality, supporting evaluation and benchmarking of radiation mapping methods when direct surface ground truth is limited or unavailable.

Abstract

Compton backscatter imaging (CBST) is a promising non-contact technique for nondestructive testing of complex structures, but its performance is limited by the trade-off between detection efficiency and spatial resolution. In this work, we propose a backscattered spectral network imaging framework based on a combined long–short crystal (CLSC) detector and a spectral fusion neural network. The CLSC detector integrates alternating long and short scintillation crystals to improve photon detection efficiency while retaining geometric selectivity, and segmented metal collimation is introduced to provide depth-dependent scattering information. The acquired multi-crystal spectra are processed by a dual-branch fusion network that jointly exploits spectral and spatial features for defect reconstruction. Geant4 simulations with varying defect sizes and depths under realistic noise conditions demonstrate that the proposed method outperforms conventional count-based and collimator-free approaches under low-count conditions, enabling more reliable detection of deep and small defects.

Abstract

High spatial resolution and sensitivity in Positron Emission Tomography (PET) are essential for early diagnosis and treatment monitoring. However, in modern pixelated detectors, a large fraction of coincidence events involves Inter-Crystal Scatter (ICS), where a gamma ray deposits its energy across multiple scintillators through a Compton interaction followed by photoelectric absorption.These events are difficult to exploit because the sequence of interactions cannot be directly determined, degrading the reconstruction of the Line of Response and, consequently, the spatial resolution. For this reason, ICS events are often discarded, leading to a significant loss in sensitivity. The Multiple Emission Tomography (MET) project aims to recover these events and use them to enable a hybrid imaging approach combining PET and Compton camera capabilities. In conventional approaches, ICS reconstruction relies on analytical methods based on chi-squared minimization and manually tuned selection cuts. Although effective, these methods are computationally intensive and strongly dependent on user-defined criteria, limiting both efficiency and scalability. In this work, we introduce a Machine Learning (ML) pipeline integrated into the MET prototype to automatically identify ICS events and determine their interaction sequence. By replacing the analytical reconstruction with XGBoost classifiers and a Convolutional Neural Network, the proposed approach provides a more flexible and efficient solution, improving the identification of the first interaction point and enabling a more effective use of multi-crystal events.

Abstract

Abstract—The search for dark matter requires increasingly advanced detectors and data processing systems. The ARGO detector, under development, uses liquid argon as the detection medium and will deploy Single Photon Avalanche Diodes (SPAD) over approximately 200 m2. These SPAD matrices, arranged in mm2 units, will generate millions of data channels, with potential drawbacks such as high power consumption, complex cabling, and costly high-speed data storage. To address these challenges, we investigate a unified real-time machine learning (ML) approach capable of simultaneously performing particle identification and event position reconstruction. The proposed Convolutional Neural Network (CNN) architecture classifies events as nuclear recoils (NR), electron recoils (ER), or alpha decays, while concurrently reconstructing the interaction position inside the detector. This multi-task approach reduces computational resources, making it well suited for resource- constrained edge environments near the detector. Future work will focus on optimizing the models and implementing them on edge hardware.

Abstract

Full-energy peak efficiency (FEPE) is a critical quantity for quantitative gamma-ray spectrometry using HPGe detectors. However, the FEPE response may vary during long-term operation because of operational history and environmental conditions. Conventional detector characterization usually requires repeated Monte Carlo simulations, in which selected geometrical parameters are manually adjusted until the simulated FEPE curve agrees with experimental measurements. This iterative procedure is computationally expensive and strongly dependent on operator expertise.

To address this limitation, this work proposes a machine-learning-based rapid characterization framework using a sparsely connected multilayer perceptron (MLP) with multiple attention heads. By taking experimentally determined FEPE values and fixed detector parameters as inputs, the network directly determines key characterization parameters, including the crystal radius, crystal length, and dead layer thickness.

Abstract

The COMET Phase-I experiment searches for the neutrinoless muon-to-electron conversion with a target sensitivity of $7 \times 10^{-15}$. A critical challenge is identifying signal hits amid a high background occupancy of approximately 40\% in the Cylindrical Drift Chamber (CDC). Traditional methods struggle under such dense environments. We present a deep-learning-based hit filtering strategy using a U-Net architecture with an EfficientNet backbone. By optimizing image formats and multi-hit handling, we achieved a purity of 98\% and a retention of 90\%, significantly surpassing design goals. This framework establishes a robust foundation for track reconstruction in high-occupancy environments.

Abstract

Charged lepton flavor violation (CLFV) is a promising probe of physics beyond the Standard Model. The COMET Phase-I experiment searches for coherent muon-to-electron conversion (μ⁻–e⁻ conversion), which produces a mono-energetic electron with an energy close to 105 MeV. However, Decay-In-Orbit (DIO) electrons form a major background near the signal region, requiring precise track reconstruction and energy estimation.

In this study, a U-Net–based deep-learning method was applied to Cylindrical Drift Chamber (CDC) hit information represented as two-dimensional images. Hit wires with deep-learning scores above a threshold were treated as signal hits and passed to GenFit for track reconstruction. Quality cuts based on the p-value, the number of degrees of freedom (NDF), and reduced χ² were also applied to improve the reconstruction quality.

The ROC curve evaluation demonstrated effective discrimination between signal-electron hits and background hits. Furthermore, applying the selected hits to GenFit improved the momentum residual distributions compared with the conventional reconstruction method. In particular, the combination of hit selection and quality cuts suppressed events with large momentum residuals contributing to the DIO background and reduced the overlap between the DIO background and the signal region. Smearing evaluations using the obtained momentum residual distributions and the theoretical DIO spectrum indicated improved energy reconstruction performance under the condition corresponding to a conversion rate of CR = 3.0 × 10⁻¹⁵. These results suggest that deep-learning-based hit selection can improve track reconstruction and energy estimation performance in the COMET experiment.