Dr. Han Bao (advised by Dr. Dinh)
Department of Nuclear Engineering
North Carolina State University

Abstract

The scaling issues have become the stumbling blocks to build the credibility of system modeling and simulation that supports risk-informed safety analysis on reactor systems. Current computer codes have limited capabilities to simulate real plant conditions, especially for extrapolative conditions, since the empirical correlations applied are mostly determined by curve fitting and strongly depending on geometry and boundary conditions. Although some advanced CFD-like coarse-mesh codes are widely used in system-level safety analysis due to their balance on computational efficiency and simulation accuracy, the Verification & Validation (V&V) of these codes still suffers from the lack of prototypic validation data. Considering mesh size is one of the model parameters for these coarse-mesh codes with simplified boundary-layer treatment, the mesh-induced error and model error are tightly connected which makes it difficult to analyze the mesh effect or the code/model scalability separately. Han proposes a data-driven approach to establish a technical basis to overcome these difficulties by exploring local patterns with the usage of machine learning. The underlying local patterns in multi-scale data are assumed to be represented by a set of physical features, which integrate the information from the physical system of interest, empirical correlations and the effect of mesh size. After performing limited high-fidelity numerical simulations and sufficient fast-running coarse-mesh simulations to generate an “error database”, advanced techniques in data analytics and machine learning are applied to explore the relationship between the local physical features and local simulation errors in the database to develop a data-driven model that bridges the global scale gap. Similarity of training data and testing data is measured and visualized using several metrics, and three extrapolative case studies based on mixed convection show that the prediction by well-trained data-driven model has higher accuracy as the similarity of training data and testing data increases.

Biography

Han Bao just graduated as a Ph.D. from the Department of Nuclear Engineering at NC State in October 2018. His research work is developing a data-driven framework using machine learning to optimize system thermal-hydraulic simulation. Han will continue his career as a Research Associate in the System Integration Department at Idaho National Laboratory.