Development of a GPU-Based Time-Dependent Monte Carlo Code System for High-Fidelity Large- Scale Nuclear Reactor Transient Analysis

Abstract

This thesis details the development and performance evaluation of a CPU and GPU-accelerated multi-physics Monte Carlo dynamics code tailored for reactor transient analysis. Understanding the time-dependent neutron behavior is crucial for various reactor core applications, such as start-up analyses, reactivity control and measurements, reactor transient with accident analysis, and research reactor dynamics experiments. Reactor transient analysis typically demands more computational resources than steady-state analysis, leading to a reliance on deterministic or quasi-static methods, which can introduce inaccuracies due to differential approximations in space, energy, or time domains. However, recent advancements in computing power and high-reliability methodologies have made the time-dependent Monte Carlo (TDMC) neutron transport method a viable and accurate alternative for reactor transient analysis. This thesis emphasizes the development of Monte Carlo (MC) transient analysis methods to serve as essential tools for analyzing various reactor types, including gen-IV reactors, research reactors, and commercial reactors. It aims to enhance Time-Dependent Monte Carlo (TDMC) algorithms, enabling accurate and practical transient analyses. The TDMC method introduces discrete time intervals into conventional MC simulations, managing neutron populations and employing methods like the analog Monte Carlo branching and combing methods for precise population control. Significant advancements include eliminating scale factors for weight normalization and introducing dynamic weight windows to improve computational efficiency. Key methodological focuses include simulating delayed neutrons and steady-state conditions. Effective simulation of delayed neutrons, crucial due to their longer generation times compared to prompt neutrons, is achieved through a forced decay algorithm with precursor simulation. The TDMC steady-state simulation method ensures consistency from initial to transient states, facilitating accurate modeling without changing calculation modes. Additional enhancements include improved precursor normalization and novel approaches for handling complex three- dimensional transient scenarios using moving geometry treatments. Validation against the C5G7-TD reactor transient benchmark, including 2D and 3D problems, demonstrates the capability of the developed MCS transient analysis. Results show close agreement with STREAM3D, particularly in scenarios involving control rod manipulations and dynamic reactor core behavior. However, challenges such as biased sample variance in TDMC tally means highlight the need for refined variance estimation techniques to ensure accurate reliability assessments in MC-based transient analyses. The developed MCS transient analysis capability is validated against 2D and 3D problems of the C5G7-TD reactor transient benchmark and compared with STREAM3D. The simulations accurately represent the axial insertion and withdrawal of control rods in 3D problems, with core dynamic reactivity and relative fission rate trends aligning well with STREAM3D results within stochastic errors. Typically, Monte Carlo calculations ensure result accuracy and reliability through sample means and variance. However, verification of the MCS transient analysis modules revealed that the sample variance of a TDMC tally mean, obtained using conventional stochastic processing, is highly biased. This bias arises from neutron correlations during the branching process or population control in TDMC calculations, leading to distorted accuracy and reliability assessments. To address these issues, a history-based batch method for TDMC simulation is developed. Neutrons and precursors are grouped into several batches for separate simulation, with results processed batch- wise to eliminate correlation between estimates. This method also naturally resolves the allocation problem of delayed neutron contributions by assigning them to the corresponding batch tally. Verification in infinite homogeneous two-group problems and C5G7-TD benchmark problems confirms that this method provides unbiased variances for tally means if the batch size is adequate. Additionally, error propagation in TDMC simulation is observed, originating from the weight normalization scheme in population control and propagating through surviving neutron weights. However, in systems where delayed neutrons are more dominant, error propagation is minimal due to shorter survival intervals for neutrons and significant delayed neutron contributions. Finally, the MCS/TH1D coupled transient analysis system is established to incorporate thermal- hydraulic feedback for more practical transient analysis problems. By coupling the MCS code with the one-dimensional TH feedback code TH1D developed by SNU, comprehensive transient analyses are possible. Validation using the VERA HFP core problem shows good agreement with the STREAM3D/TH1D system, with a maximum difference of approximately 2% in power distribution. A mockup rod ejection accident in the modified VERA core problem with HFP conditions is analyzed, demonstrating the Doppler effect of fuel temperature feedback and verifying the integrity of the coupled transient analysis system. This research focuses on advancing Monte Carlo (MC) techniques for high-precision and high- performance analysis of reactor core transient states and safety assessments. Key objectives include evaluating analysis codes to meet rigorous commercial reactor standards and assessing their performance across different computing resources, particularly CPU and GPU-based systems. The study aims to apply and validate MC Time-Dependent Monte Carlo (TDMC) methodologies for producing precise reactor core solutions and benchmarks, demonstrating their potential to achieve commercial reactor standards effectively. Key findings include the development of MC-based programming codes, successful verification through benchmark problems like C5G7-TD, and evaluations showing minimal errors in critical parameters such as maximum pin power distribution. Ongoing research addresses further optimization of computing resource utilization to maximize code performance efficiency.