Abstract: Advanced reactors are expected to fulfill a key role in next-generation nuclear power plants due to their increased safety performance and reliability. The System Analysis Module (SAM) is currently developing a reduced-order three-dimensional (3-D) module to accurately model complex thermal-fluid phenomena in advanced reactor systems. This module adopts a coarse-mesh setup to be consistent with the one-dimensional (1-D) system modeling framework, which ensures computational efficiency. A major difficulty for the coarse-mesh 3-D module in SAM is to accurately capture turbulence.
This seminar will discuss some recent efforts to develop a coarse-mesh turbulence model with a physics-informed machine learning approach. This approach leverages deep learning technologies and utilizes fine-mesh computational fluid dynamics data for the turbulence model development. The performance of different architectures of deep neural networks, including densely connected convolutional networks and long-short-term-memory networks, are evaluated. The optimized neural network model will be implemented into SAM as a data-driven turbulence model. Such an “open-box” approach puts the machine learning model within the solver so that the major physics constraints of the system can be preserved.
As a separate product, a machine learning-based Bayesian approach is also developed to quantify the uncertainty of the developed data-driven model. A case study on two-phase bubbly flows demonstrated the applicability of the proposed approach.