ILLINOISROCSTAR LLC — Department of Energy SBIR Phase I: 33d

Description

Computational fluid dynamics methods have found increasing use in high fidelity modeling of complex systems, particularly within nuclear power systems. For example, the phenomenon of thermal stratification in a sodium-cooled fast reactor under natural circulation conditions is resolvable with computational fluid dynamics, but not traditional system-level models. The high accuracy of computational fluid dynamics incurs increased computational cost, however. Therefore, methods that can provide high quality results at a reduced cost are in high demand. The goal of this project is to develop a software module that extends existing capabilities, enabling the production of high quality reduced order models using a parameterized proper orthogonal decomposition approach. This novel technology will reduce the burden placed on the user to make certain modeling choices, such as closure. Large volumes of high fidelity simulation data will be used to train machine learning algorithms to best assist with low order model stability and accuracy. Several data-driven models will be created for use with proper orthogonal decomposition based reduced order models. Large eddy simulation example problems will be carried out to be used as test cases. Open-source implementations of key machine learning algorithms, such as artificial neural networks, will be leveraged to reduce technical risk. Models trained on sample data will be evaluated for accuracy, stability, and efficiency. Recommendations for full models and nuclear engineering workflow integration will be made to inform Phase II efforts.Proper orthogonal decomposition and similar methods have been well understood for decades. However, these models have not been widely adopted for reasons including that they are difficult and time-consuming to compute, their stability properties are unfavorable, and their calculation requires ad hoc software development. By leveraging recent developments and internal expertise, each of these concerns is addressed in the proposed work. The models produced by this effort will have applications in nuclear reactor design and safety analysis, as well as drag force computation for airfoils and automotive frames, to name just a few. Nearly any industry with large modeling and simulation components will benefit. A key application in the nuclear engineering community is the production of surrogate computational fluid dynamics models to couple with system level codes for safety analyses. The tool will be usable to rapidly compute detailed boron concentrations in the lower reactor plenum after certain accidents.