Development of Data-Driven Eddy Viscosity Closure to Support MSR Flow and Heat Transfer Modeling Based on SAM-ML

Abstract

Safety analysis of advanced non-LWRs often involves predicting sophisticated thermal-fluid (T-F) phenomena that occur in various reactor components with complex geometric setups. An example is the liquid-fueled molten salt reactor (MSR). The System Analysis Module (SAM), a modern nuclear system code, has developed a multi-dimensional (multi-D) flow model to address the modeling of complicated T-F phenomena in advanced reactors. SAM’s multi-D flow model employs a coarse mesh setup to ensure computational efficiency and consistency with its 1-D components. Consequently, this requires constitutive relations in the model for unresolved fine-scale physics, such as turbulence.

In this paper, we introduce a novel neural network architecture that serves as a data-driven eddy viscosity closure to complement SAM's multi-dimensional module, supporting the modeling and simulation of MSR. The newly proposed neural network draws on flow features from neighboring flow regions to predict local eddy viscosity. This novel architecture satisfies the physical constraints of the fluid system (i.e., Galilean invariance) and is mesh-agnostic, thus significantly enhancing its flexibility and generality by accommodating irregular computation domains. We have integrated the developed data-driven closure with SAM through the SAM-ML capability, facilitating the modeling of flow and heat transfer in the MSR's primary fuel circuit. The integrated SAM-ML model demonstrated good agreement with high-resolution CFD results. The network utilizes non-dimensional local flow features as inputs, the developed capability showed good performance for various Reynolds number flow conditions within the fully turbulent flow regime.