Acceleration of the Convergence of a Core Thermal-Hydraulic Code Using Initialization from a Neural Network

For safety studies of nuclear reactor cores, a finite-volume thermal-hydraulic code with a porous medium approach named THYC-coeur has been developed by EDF to simulate steady-state flows of a two-phase mixture within the core of a nuclear reactor. The steady-state solution is obtained through a fictitious transient. Despite a relatively low individual computational time, many state points are considered in the safety studies, which can represent significant CPU time. To speed up the computation, one possible objective is to reduce the number of iterations (i.e. the number of time steps) required to reach convergence. In the present work, the idea is to train a neural network (NN) to predict steady-state solutions and use this prediction to initialize the transient computation. This method allows combinating the advantages of NN prediction, in terms of rapidity, with that of the THYC-coeur model, in terms of the physical validation of the solution. To evaluate the potential of this method, a simplified one-dimensional code was designed. This code simulates a two-phase water vapor flow in a heated channel. A NN was trained to predict the solution fields from imposed boundary conditions. In this paper, we present the methodology, the database selection process, the structure of the NN, and the optimization of the network’s hyperparameters. This paper examines two different physical models: the 3-equation model and the 4-equation model. Initially, a simple loss function is used to predict the fields of these two models. The observed accelerations are approximately 70% for the 3-equation model and 48% for the 4-equation model. A detailed study of the correlation between prediction accuracy and the resulting acceleration has shown that the low-frequency components in the entropy field prediction error significantly slow down the search for the steady state. The highlight of this work is that by introducing spatial frequencies in the error for the optimization of the NN, the number of iterations is reduced between 61% (4-equation model) and 83% (3-equation model) from baseline simulations.