A Structure-based Machine Learning Approach for Spatiotemporal Fluctuations in Nuclear Thermal Fluids

Abstract

In nuclear energy systems, flow-assisted damage can arise at locations where complex turbulent structure interaction drives enhanced mechanical loads and mixing of different temperature flows drives thermal-induced mechanical failures. Thermal striping is a such phenomenon characterized by the turbulent mixing of non-isothermal streams, which can induce thermal stress fluctuations and fatigue damage in the critical components. The concern over such damage mechanism cannot be fully addressed via plant instrumentation due to the high frequencies involved, as well as the complex interaction between the source of the flow oscillation and the affected locations. High-fidelity models and simulations can play a significant role in predicting the performance of critical components under thermal fatigue. While scale-resolving models are capable of capturing complex unsteady flow features, such models are often computationally overburdening for industrial applications, making them inapplicable to online monitoring.

This thesis proposes an industrial-scale prognosis machine learning (ML) tool for thermal striping applications. A two-level ML framework based on turbulence coherent structures is developed. In the first level, well-organized coherent structures are extracted by performing proper orthogonal decomposition on local parameters, and then a tree-based machine learning model is used to down-select the reference structures for the field reconstruction. In the second level, a parameterized convolution neural network is trained to predict the bias introduced by reference structures approximation. The two-level design leverages vortex identification and local bias correction techniques, which largely increase the data efficiency for training.

The demonstration of the methodology shows that the proposed framework can successfully capture the fluctuation frequencies and amplitudes of the spatiotemporal fields. Through three test cases, the proposed framework has shown its capability to work with a notably limited training sample size (between 20 to 30 in each instance) in a highly variational setting. A speed-up factor of an order of 10⁶ ∼ 10⁷ was achieved. Based on the vortex identification method, the methodology is expected to be applicable to general phenomena driven by large coherent structures. The framework is also shown to have the ability to identify fatigue limiting regions for the spectrum of operating conditions.