A novel framework for data-driven modeling, uncertainty quantification, and deep learning of nuclear reactor simulations

Nuclear Multiphysics, Uncertainty Quantification, Deep Learning, Data-driven Modeling, Bayesian Statistics

Radaideh, Majdi Ibrahim Ahmad

关键词: Nuclear Multiphysics, Uncertainty Quantification, Deep Learning, Data-driven Modeling, Bayesian Statistics;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/106162/RADAIDEH-DISSERTATION-2019.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

This work presents a novel and modern method for reactor modeling, simulation, and uncertainty characterization through an integrated framework developed under the terminology of combining four fundamental principles in scientific modeling and computing: Physics, Models, Data, and UQ (Uncertainty Quantification). The framework houses various physical phenomena that occur inside nuclear reactors, such as neutronics, reactor kinetics, fuel depletion, thermal-hydraulics, and fuel performance as well as outside the reactor such as spent fuel and criticality safety. The framework utilizes various computer models in the nuclear area, which are already validated and known to provide accurate results. The framework is supported and validated by a wide range of experimental data from different single and multiphysics experiments, such as delayed neutron data, void fraction measurements, isotopic composition, nuclear data, and others. Many computational models to simulate the actual physical phenomena are developed under this framework, which vary in their complexity from a simple 2D pin-cell to a complex 3D lattice model with multiphysics coupling. Additionally, the framework is built based upon a wide range of mathematical and statistical methods featuring different areas such as sensitivity analysis, variance decomposition, dimensionality analysis and reduction, reduced order modeling, machine learning, data science, deep learning, Monte Carlo and deterministic uncertainty propagation, Bayesian statistics, correlation analysis, and data assimilation. All efforts in this thesis are expected to yield a better understanding of nuclear reactor simulations, which in turn can lead to improved performance, safety, and reduced costs for nuclear industry. Within this thesis, many frameworks, platforms, and models are developed to support the master framework. An integrated UQ approach is developed through the Bayesian framework, which handles various forms of uncertainty in scientific modeling such as parametric, experimental, predictive, interpolation, and model-form uncertainty. The methodology is useful to account for various uncertainty sources in nuclear computer models. This integrated UQ methodology can also be used for model selection of different physical models, through evaluating them against real data. The methodology is applied in this thesis to nuclear thermal-hydraulics and two-phase flow codes to quantify their predictive and model-form uncertainties. Data science methods are a core part of the framework. Machine learning methods are integrated to alleviate the computational burden of the complex simulations to construct cheap-to-evaluate reduced order or surrogate models. Modern deep learning methods form a major part of this thesis to analyze complex datasets resulting from the advanced simulations generated using the master framework. These machine and deep learning models are tested using real-world and benchmarked nuclear simulations with different underlying physics, from fundamental nuclear data to nuclear fuel performance. Data-driven models are constructed using simulation and experimental data to perform uncertainty propagation, surrogate modeling, model validation, and variance decomposition.Development of a new precursor-group kinetics framework is done to propagate the uncertainty into reactor kinetic parameters due to the fundamental nuclear and delayed neutron data. Coupling of single physics processes (e.g. neutronics, thermal-hydraulics, fuel performance) to form more realistic multiphysics simulations is also accomplished through FUSE platform. FUSE is verified through two test cases of two-way coupled neutronics-thermal-hydraulics and neutronics-fuel performance simulations. Spent fuel analysis and criticality safety frameworks are built as a validation object to assess the accuracy of the framework modeling approaches. The spent fuel composition discharged from the reactor core is assessed in the spent fuel cask to determine the overall system safety. A comprehensive application of the spent fuel framework on BWR spent fuel is carried out in this thesis. All the physics, data, methods, and frameworks are integrated into the master framework developed in this thesis.The major achievements of the framework developed to the nuclear area include: a set of kinetic parameters' values and uncertainties for light water reactor systems, advanced depletion models for accurate burnup credit of BWR, integrated assessment and advanced UQ of nuclear computer models, a platform for nuclear multiphysics simulations, and building deep learning models for high dimensional UQ purposes.Most of the methods and the frameworks developed here are extendable to other problems outside the nuclear area. The reader is strongly recommended to read the first chapter of this thesis as it will provide directions to efficiently access the whole document. The first chapter presents an executive summary of the work done over the whole thesis. This thesis is published in several peer-reviewed articles in premier conferences and journals specialized in nuclear engineering, system safety, uncertainty quantification, and energy resources. A summary of the framework developed in this thesis is published in Radaideh and Kozlowski (2019b).