원자력 안전조치를 위한 우라늄 검증용 딥러닝 기반 감마분광분석 기술 개발 = Development of Gamma-ray Spectrometry Techniques with Deep Learning for Uranium Verification in Nuclear Safeguards

다국어 초록 (Multilingual Abstract)

Nondestructive assay (NDA) techniques are essential for the timely detection of uranium diversion in international nuclear safeguards. These methods enable rapid and accurate on-site verification of nuclear materials. However, existing NDA systems face two primary technical challenges. First, portable detectors often lack the resolution required for precise enrichment verification. Second, unknown counting geometry in samples, such as waste drums, limits the accuracy of mass quantification. This study proposes an integrated verification framework to address these limitations. It combines a hardware-optimized detection system with simulation and deep learning-based analysis methods.
First, the Korea Inspection Multi-Channel Analyzer (K-IMCA) was developed to improve the reliability of field enrichment verification. This system utilizes a dual-mode configuration comprising NaI and HPGe detectors. Additionally, a deep learning-based spectral transformation technique was introduced to address the resolution limits of field-portable NaI detectors. A 1D U-Net–like fully convolutional network (1D FCN) model was constructed to convert low-resolution spectra into HPGe-like high-resolution spectra. This approach enables quantitative enrichment analysis of uranium in thick-walled containers, such as UF6 cylinders, which is difficult to analyze by using conventional methods with low-resolution systems. Field verification results for UO2 pellets and fuel rods satisfied the IAEA acceptance criteria, whereas UF6 cylinder assays—particularly NaI measurements and the high-enrichment case—showed noncompliance of the enrichment-meter approach when linear calibration curves were applied under severe attenuation conditions. Second, an evaluation methodology combining Monte Carlo simulation (MCNP) and machine learning was established for uranium mass quantification in radioactive waste drums. A large-scale efficiency calibration dataset covering various matrices and densities was constructed. A verification procedure was developed using regression analysis and iterative optimization to estimate the physical characteristics of samples. Furthermore, an end-to-end Random Forest model was implemented to predict uranium mass directly from spectral data. This approach aims to minimize error propagation associated with physical modeling steps. The data-driven model achieved improved mass prediction accuracy for unknown waste samples without requiring complex geometry correction factors. This research indicates that software-based approaches can compensate for physical hardware limitations. The results suggest that reliable uranium verification is feasible even in environments with missing geometric information or spectral interference. The proposed framework provides an effective verification tool for advanced nuclear fuel cycle facilities and safeguards inspections. It contributes to the advancement of technical capabilities for nuclear non-proliferation verification.

목차 (Table of Contents)

• TABLE OF CONTENTS ⅰ
• LIST OF FIGURES ⅳ
• LIST OF TABLES ⅵ
• ABSTRACT
• Ⅰ. INTRODUCTION 1
• 1.1 Background of Nuclear Safeguards and NDA 1
• 1.2 Technical Limitations of Conventional NDA Systems 1
• 1.3 Research Objectives 3
• Ⅱ. Theoretical Background 5
• 2.1 Uranium Enrichment Verification Techniques 5
• 2.1.1 Peak Ratio Analysis 5
• 2.1.2 Enrichment Meter Principle (Infinite Thickness Method) 10
• 2.2 Uranium Mass Quantification Techniques 16
• 2.2.1 In Situ Object Counting System (ISOCS) 16
• 2.2.2 Advanced ISOCS Iteration System (A-ISOCS) 18
• Ⅲ. DEVELOPMENT OF ADVANCED URANIUM ENRICHMENT VERIFICATION SYSTEM AND DEEP LEARNING METHODOLOGY 20
• 3.1 Development of the Korea Inspection Multi-Channel Analyzer (K-IMCA) 20
• 3.1.1 Dual-Mode Hardware Configuration and Design 21
• 3.1.2 Advanced Spectral Analysis Algorithms and Integrated Software Architecture 27
• 3.2 Performance Evaluation of the K-IMCA System 41
• 3.2.1 Experimental Setup and Verification Standards 41
• 3.2.2 Verification Results for UO2 Pellets and Fuel Rods 46
• 3.2.3 Verification Results for UF6 Cylinders 46
• 3.2.4 Performance Comparison with Conventional IMCA 48
• 3.3 Machine Learning-based Spectral Transformation for Enrichment Verification 50
• 3.3.1 Simulation Data Generation and Training Strategy 50
• 3.3.2 Comparative Evaluation of Machine Learning Models 52
• 3.3.3 Optimization of Neural Network and Validation Results 58
• Ⅳ. DEVELOPMENT OF URANIUM MASS QUANTIFICATION SYSTEM AND MACHINE LEARNING METHODOLOGIES 65
• 4.1 Construction of a Simulation-based Efficiency Database via Sensitivity Analysis 65
• 4.1.1 Physicochemical Characterization of Uranium Waste Drums 66
• 4.1.2 Sensitivity Analysis of Key Parameters on Detection Efficiency 73
• 4.1.3 Construction of Realistic Efficiency Database using Statistical Inverse Transform Sampling 81
• 4.2 Development of Regression-based Iterative Verification System 84
• 4.2.1 Statistical Analysis of Simulation Database 86
• 4.2.2 Correlation between Detection Efficiency and Physical Parameters 90
• 4.2.3 Development of Linear Regression Model 98
• 4.2.4 Development of Non-Linear Regression Models 134
• 4.2.5 Iterative Optimization Algorithm for Independent Physical Parameter Estimation 179
• 4.2.6 Performance Evaluation of the Regression-based Iterative Verification Method 183
• 4.3 Advanced Machine Learning Models for Uranium Mass Quantification 187
• 4.3.1 Universal Models Incorporating Effective Atomic Number 187
• 4.3.2 ECF-based Correction Model for Empirical Data 193
• 4.3.3 End-to-End Random Forest Model for Direct Mass Prediction 197
• Ⅴ. CONCLUSION 203
• REFERENCES 206
• ABSTRACT (In Korean) 213