This paper investigates single-event effects in enhancement-mode GaN high electron mobility transistors (HEMTs) and presents a physical mechanism for single-event leakage current (SELC) degradation. Heavy-ion irradiation performed on commercial EPC2007C power devices reveals permanent, step-like increases in drain-to-source leakage current at sub-rated operational voltages. Temperature-dependent characterization demonstrates a distinct degradation mechanism. In the off-state, the post-irradiation leakage floor transitions to a conduction path that is consistent with a three-dimensional Mott variable-range hopping (VRH) transport model. On-state measurements reveal a sharp, high-temperature divergence in threshold voltage hysteresis, indicating newly introduced or modified trapping centers above room temperature. A physical degradation model is proposed in which localized microscopic energy deposition provides a plausible framework for the dehydrogenation and depassivation of internal buffer dopants and native defects, which may be aided by generation and lateral motion of nitrogen vacancies to form a network of traps that enhance hopping transport.
Plastic scintillators are widely used in environmental monitoring and radiation detection due to their short decay time and low cost. However, the absence of well-resolved full-energy peaks makes radionuclide identification via classical peak analysis infeasible. While deep neural networks have improved identification accuracy, their computational complexity hinders deployment on resource-constrained embedded devices. To address this challenge, we propose MicroNet-1D, a lightweight multi-label classification network tailored for plastic scintillator spectra. The model employs Micro-Factorized convolution and structured sparse connectivity to replace conventional convolutions, drastically reducing parameters while preserving discriminative spectral feature extraction. A dataset of 50,000 spectra spanning eight radionuclides (22Na, 54Mn, 60Co, 131I, 137Cs, 133Ba, 152Eu, 241Am) was generated via MCNP5 simulation with Poisson noise and gain drift augmentation, and validated against measured data. Evaluated on a held-out test set with performance variability quantified via 10-fold cross-validation, MicroNet-1D achieves an overall accuracy of 99.25±0.09% and an exact match ratio (EMR) of 94.51±0.38% with only 0.33M parameters and 14.52M FLOPs, outperforming all six baselines—including both deep networks (ResNet34, VGG16, Transformer, BP) and classical classifiers (SVM, Random Forest)—while requiring substantially less computational cost than all neural-network counterparts. These results demonstrate that high-accuracy radionuclide identification does not require overparameterized architectures when task-specific inductive biases are properly incorporated. The ultra-lightweight design enables real-time inference on embedded platforms, supporting practical deployment in intelligent radiation monitoring systems.
Single-event burnout (SEB) was experimentally investigated in AlN Schottky barrier and p-n diodes. The burnout thresholds were found to decrease with increasing ion linear energy transfer (LET), consistent with trends observed in other semiconductors. To enable a materials-based comparison of burnout thresholds independent of device design parameters, the concept of critical applied electric field at SEB was introduced. Within this framework, AlN was shown to withstand close to an order of magnitude higher fields than other common semiconductors. An analytical model was developed that accurately describes the LET dependence of burnout thresholds in simple vertical devices across multiple material systems, suggesting that SEB in these devices can be understood as conventional avalanche breakdown induced by the separation of ion-generated charge.
An intense electric field and a controlled collector current tailored to the degradation level are applied between the collector and emitter under forward-bias conditions to accelerate charge detrapping and enable rapid recovery of gain degradation in irradiated bipolar COTS transistors. This approach supports hardness assurance methodologies and enables practical in situ mitigation of radiation-induced gain degradation in bipolar COTS components.
Neutron radiography provides essential and complementary sensitivity to light elements compared to X-ray imaging due to distinct physical attenuation mechanisms, but practical neutron measurements often suffer from limited spatial resolution and elevated noise. While enhancing neutron images using a co-registered, high-resolution X-ray image as guidance is a promising computational strategy, this approach implicitly assumes structural consistency between the two modalities. In industrial non-destructive testing (NDT), fundamentally different interaction physics frequently violate this assumption, causing guide-driven hallucinations that pose safety-critical risks by misleading physical inspections. This paper systematically investigates this vulnerability through a progressive two-stage diagnosis protocol—spanning controlled conflict scenarios to real-world industrial NDT datasets—and identifies a critical failure mode: Guide-Induced Structural Discrepancy (GISD). We formally define GISD as a specific, severe form of guide-driven hallucination where guide-aligned structures, virtually unobservable in the target modality, falsely emerge as salient features in the restored output. To rigorously quantify this risk and ensure the reliability of multi-modal measurement systems, we propose Conflict-Weighted Structural Leakage (CWSL), a localized diagnostic metric that explicitly isolates modality-conflict regions to measure guide-preferential structural hallucination. Through localized line profiles and robustness analyses across varying registration conditions, we demonstrate that modern guided models are deeply susceptible to GISD, and that these severe, visually salient NDT failures remain dangerously hidden under high global quality scores like PSNR and SSIM.
