GPU-Accelerated Lagrangian Modeling for Radiological Atmospheric Dispersion and Risk Assessment in Severe Nuclear Accident Scenarios = 원자력 중대사고 시나리오에서 방사성 물질의 대기 확산 및 위험 평가를 위한 GPU 기반의 라그랑지안 모델링

국문 초록 (Abstract)

본 연구는 원자력 중대사고 시나리오에서 방사성 물질의 대기 확산 및 위험 평가를 위한 GPU 기반의 라그랑지안 모델링 기법을 제안하였다. 최근 컴퓨팅 능력의 급속한 발전으로 대기 확산 모델의 성능이 비약적으로 향상되고 있으나, 원자력 분야에서는 이러한 고성능 모델을 응용한 현실적이고 정밀한 해석이 여전히 부족한 실정이다. 특히 기존 연구자들은 주로 가우시안 플룸 모델을 이용하여 왔으나, 이 모델은 복잡한 기상 조건 및 대기 확산 현상을 충분히 묘사하지 못하는 근본적인 한계를 가지고 있어 정확한 평가에 어려움을 겪었다. 본 연구에서는 이러한 한계를 극복하기 위해 GPU 병렬화 기술을 적용한 라그랑지안 기반의 대기확산 모델을 개발하고 이를 원자력 사고 시나리오의 현실적이고 정밀한 분석에 효과적으로 응용하였다.

개발된 GPU 병렬화 라그랑지안 대기확산 모델은 독립적이고 계산량이 많은 입자 추적 및 퍼프 시뮬레이션 연산을 효과적으로 병렬화함으로써 기존 방법 대비 현저히 향상된 계산 효율성과 높은 해상도의 실시간 모의를 제공하였다. 본 연구의 주요 성과는 다음과 같다.

첫째, 기존의 연구는 방사선량 평가 시 현실적인 비상 대응 시뮬레이션을 포함하지 못해 세부적인 방사선 피폭 경로에 대한 평가가 미흡하였다. 이에 본 연구에서는 LDM과 행위자 기반 대피 모델 (MATSim)의 연계를 통해 다양한 방사선 피폭 경로를 상세하고 정확하게 평가할 수 있도록 하였으며, 방사성 결말해석 분야에서 현존하는 가장 고해상도의 해석 모델을 구축하였다.

둘째, 기존의 선원항 역추정 방법은 주로 가우시안 플룸이나 퍼프 모델과 같은 저해상도의 대기확산 모델을 이용하여 제한된 정확성을 제공하였다. 본 연구에서는 가장 고해상도의 라그랑지안 모델(LDM)을 전방 모델로 활용하여 앙상블 칼만 역추정(EKI)을 적용함으로써, 보다 정밀한 방사성 물질 방출량 추정을 수행하였다.

셋째, 기존의 CPU 기반 레벨 3 확률론적 안전성 평가(PSA)는 복잡한 계산 문제로 인해 다수 호기 시나리오를 현실적으로 처리하는 데 한계가 있었다. 본 연구에서는 GPU 병렬화를 적용하여 다수의 원자로에서 동시 발생 가능한 다양한 사고 시나리오를 포괄적으로 평가할 수 있는 효율적인 고속 모델링 환경을 구축하였다.

본 연구에서 제시한 GPU 기반의 고성능 라그랑지안 확산 모델링 기법은 복잡한 원자력 중대사고 시나리오 및 해석들을 효율적으로 처리할 수 있는 강력한 계산 도구로서, 원자력 안전 및 대기 확산 평가 분야에서 다양한 응용 가능성을 제시한다.

다국어 초록 (Multilingual Abstract)

This study proposes a GPU-based Lagrangian modeling for atmospheric dispersion analysis and radiological risk assessment under severe nuclear accident scenarios. Despite rapid advancements in computational capabilities enabling significantly improved atmospheric dispersion modeling, the nuclear safety field still lacks realistic and high-resolution analyses leveraging these advanced models. Traditional studies have predominantly employed Gaussian plume models, which inherently struggle to accurately represent complex atmospheric phenomena, thus limiting precise consequence evaluations. To overcome these limitations, this research developed and applied GPU-parallelized Lagrangian modeling for realistic, high-fidelity analysis of nuclear accident scenarios.

The GPU-accelerated Lagrangian modeling effectively parallelizes computationally intensive and independent particle tracking and puff simulations, achieving substantially enhanced computational efficiency and enabling high-resolution, real-time simulations. The key achievements of this research are as follows:
First, previous studies lacked realistic emergency response simulations integrated with detailed radiological exposure assessments, resulting in insufficient evaluations of radiation dose pathways. To address this, the current research coupled Lagrangian Dispersion Model (LDM) with an agent-based evacuation model (MATSim), allowing detailed and precise radiation dose assessments, and establishing the highest-resolution modeling framework currently available for radiological consequence analysis.

Second, existing inverse source-term estimation approaches commonly relied on lower-resolution forward models such as Gaussian plume or puff models, offering limited accuracy. This research employed the high-resolution Lagrangian model as a forward model integrated with Ensemble Kalman Inversion (EKI), enabling more accurate estimation of radionuclide source terms.

Third, traditional CPU-based Level 3 probabilistic safety assessments (PSA) faced significant computational challenges when dealing with multi-unit accident scenarios. By implementing GPU parallelization, this study established a realistic and efficient modeling environment capable of comprehensively assessing multiple concurrent accident scenarios involving several reactor units.

The GPU-accelerated Lagrangian modeling developed in this research thus provides a powerful computational tool capable of efficiently handling complex nuclear accident scenarios, highlighting significant potential for broader applications in nuclear safety and atmospheric dispersion assessment.

목차 (Table of Contents)

• Abstract i
• Table of Contents iii
• List of Figures viii
• List of Tables xiii
• Chapter 1. Introduction 1
• 1.1. Background and Motivation 1
• 1.2. Objective and Scope 2
• Chapter 2. Atmospheric Dispersion Modeling Approaches 5
• 2.1. Gaussian Plume Model 5
• 2.2. Gaussian Puff Model 7
• 2.2.1. Concentration Formulation of Gaussian Puff 7
• 2.2.2. Advection of Puff Using Meteorological Fields 7
• 2.2.3. Pasquill Stability Category Determination 8
• 2.2.4. Virtual Distance for Dispersion Calculations 9
• 2.2.5. Dispersion Coefficient Calculation 10
• 2.2.5.1. Pasquill-Gifford Formulation 10
• 2.2.5.2. Briggs-McElroy-Pooler Formulation 11
• 2.2.6. Iterative Virtual Distance Calculation: Newton-Raphson Method 12
• 2.2.7. Removal Processes 13
• 2.2.7.1. Dry Deposition 13
• 2.2.7.2. Wet Scavenging 13
• 2.2.7.3. Radioactive Decay 14
• 2.2.7.4. Remaining Concentration after Removal Processes 14
• 2.3. Lagrangian Dispersion Model 15
• 2.3.1. Wiener Process 17
• 2.3.2. Turbulence modeling and the Langevin equation 19
• 2.3.3. Determination of Turbulence Parameters 21
• 2.3.3.1. Definitions and Dimensionless Variables 21
• 2.3.3.2. Neutral Atmospheric Conditions 22
• 2.3.3.3. Stable Atmospheric Conditions 22
• 2.3.3.4. Unstable Atmospheric Conditions 23
• 2.3.4. Coupling of Lagrangian Stochastic Models with CFD-Generated Turbulence Fields 25
• 2.3.5. Modeling of Removal Processes 27
• 2.3.5.1. Dry Deposition Parameterization 28
• 2.3.5.1.1. Aerodynamic Resistance 29
• 2.3.5.1.2. Quasi-laminar Sublayer Resistance 30
• 2.3.5.1.3. Surface Resistance 30
• 2.3.5.1.4. Particle Deposition 31
• 2.3.5.2. Wet Deposition Parameterization 32
• 2.3.5.2.1. In-cloud Scavenging 33
• 2.3.5.2.2. Below-cloud Scavenging 34
• 2.3.5.3. Gravitational Settling 35
• 2.3.5.4. Radioactive Decay 36
• Chapter 3. GPU-Based Computational Methods 38
• 3.1. Architectural Differences between GPU and CPU and Principles of Parallelization 39
• 3.2. Computational Patterns for GPU Optimization 40
• 3.3. GPU Implementation for Gaussian Puff Models 41
• 3.3.1. Puff-wise Parallelization 42
• 3.3.2. Concentration Accumulation on the Grid 42
• 3.4. GPU Implementation for Lagrangian Dispersion Models 45
• 3.5. GPU Performance Optimization via Double-Buffering Pipelining 48
• 3.6. Multi-GPU Parallelization using MPI 49
• Chapter 4. Validation of Atmospheric Dispersion Models 52
• 4.1. Validation of Gaussian Puff Models 52
• 4.1.1. Accuracy Verification Based on Puff Number 53
• 4.1.2. Validation of Removal Processes in Gaussian Puff Models 55
• 4.1.3. Validation Under Different Atmospheric Stability Conditions 57
• 4.2. Validation of Lagrangian Dispersion Models 59
• 4.2.1. Simulation Setup for LDM Validation 60
• 4.2.2. Concentration Comparison at Receptor Points: LDM vs. FLEXPART 61
• 4.2.3. Mass Distribution and Computational Performance Comparison 63
• 4.2.4. Application to Fukushima Accident Scenario 72
• 4.3. Validation of CFD-coupled LDM simulations 74
• 4.3.1. Two-dimensional CFD-LDM Validation 74
• 4.3.2. Three-dimensional CFD-LDM Validation 77
• 4.4. Analytical Advantages of LDM over Gaussian Puff Models 79
• Chapter 5. Application 1: Radiological Consequence Analysis 83
• 5.1. Limitations of Current Radiological Consequence Analysis Codes 83
• 5.2. Evacuation Modeling using MATSim 86
• 5.3. Modeling Radiological Exposure Pathways 87
• 5.3.1. Cloudshine Dose 87
• 5.3.2. Inhalation Dose 88
• 5.3.3. Groundshine Dose 89
• 5.4. Analysis Results for Hanbit Nuclear Power Plant Scenario 89
• 5.4.1. Simulation Settings for Hanbit Nuclear Power Plant Unit 5 Scenario 90
• 5.4.2. Validation Results and Comparison with RASCAL 93
• 5.4.3. Advanced Scenario for High-Resolution Atmospheric Dispersion Modeling 94
• 5.4.3.1. Simulation Setup 94
• 5.4.3.2. Simulation Results for Advanced Meteorological Scenario 97
• 5.5. Conclusion of Radiological Consequence Analysis 101
• Chapter 6. Application 2: Source Term Estimation 103
• 6.1. Theoretical Background 104
• 6.2. Literature Review 106
• 6.3. Coupling Methodology and GPU Parallelization for LDM-EKI 108
• 6.4. Scenario Setup and Result Analysis for Source Term Estimation 111
• 6.4.1. Simulation Setup 111
• 6.4.2. Result Analysis for Source Term Estimation 113
• 6.5. Summary for Source Term Estimation using LDM-EKI 116
• Chapter 7. Application 3: Multi-Unit Level 3 Probabilistic Safety Assessment 118
• 7.1. Differences between Consequence Analysis and Level 3 PSA 120
• 7.2. Integrated Computational Framework for Multi-Unit Level 3 PSA 122
• 7.2.1. Overall Framework Description 122
• 7.2.2. Evacuee Movement Mechanism 123
• 7.2.3. Exposure Pathways of Gaussian Puff 126
• 7.2.3.1. Cloudshine Dose Rate 126
• 7.2.3.2. Groundshine Dose Rate 127
• 7.2.3.3. Inhalation Dose Rate 127
• 7.3. GPU-Parallelized Methods for Multi-Unit Level 3 PSA 128
• 7.3.1. Parallelized Puff Advection and Dispersion 129
• 7.3.2. Parallelized Evacuation Dynamics 131
• 7.3.3. Parallel Dose Assessment Methods 133
• 7.3.3.1. Three-Dimensional Parallelization Structure 134
• 7.3.3.2. Parallel Computation and Dose Summation 136
• 7.3.3.3. Exposure Pathway and Dose Calculation 137
• 7.3.4. Overall Parallel Computing Framework 138
• 7.4. Performance Evaluation and Results Analysis of Multi-Unit Level 3 PSA 140
• 7.4.1. Performance Analysis Across GPU Models 140
• 7.4.2. Comparison of CPU-based and GPU-based Code Execution Efficiency 142
• 7.4.3. Performance Comparison with RCAP under Identical Multi‑Unit Scenario 148
• 7.4.3.1. Simulation Configuration and Parameters 148
• 7.4.3.2. Comparison Results and Analysis 150
• 7.5. Wind Directional Sensitivity Analysis of Multi-Unit Dispersion Scenarios 152
• 7.6. Summary of GPU Accelerated Multi-Unit Level 3 PSA Framework 158
• Chapter 8. Conclusions 159
• 8.1. Summary of Research 159
• 8.2. Research Limitations and Recommendations for Future Research 160
• Bibliography 162
• Appendix 180
• Appendix A. Validation Results and Comparison with RASCAL 180
• A.1. Inhalation Dose Validation: Volatile Iodines 180
• A.2. Inhalation Dose Validation: Semi-volatile Nuclides and Cesium 182
• A.3. Inhalation Dose Validation: Non-volatile Metallic Particulates 184
• A.4. Inhalation Dose Validation: Insoluble Actinides and Transuranic Nuclides 186
• A.5. Inhalation Dose Validation: Ruthenium and Molybdenum Nuclides 188
• A.6. Ground Deposition Validation: Volatile Iodines 189
• A.7. Ground Deposition Validation: Cesium Isotopes 190
• A.8. Ground Deposition Validation: Ruthenium and Related Metallic Nuclides 191
• A.9. Ground Deposition Validation: Alkaline Earth and Lanthanide Nuclides 193
• A.10. Ground Deposition Validation: Tellurium Nuclides 195
• Abstract in Korean (국문 초록) 197