Key points are not available for this paper at this time.
This paper focuses on analyzing the thermal stratification in the upper plenum of a lead-bismuth fast reactor after an emergency shutdown, which threatens the safety of internal components and makes it difficult to remove residual heat. First, high-precision full-order snapshots of this phenomenon are obtained using the computational fluid dynamics program FLUENT. Second, a graph neural network is employed to construct a graph autoencoder for reducing the nonlinear dimensionality of the full-order snapshots, and the findings are compared with the linear dimensionality reduction results obtained using proper orthogonal decomposition (POD). Finally, a multilayer perceptron is used for online state recognition and conducting a predictive analysis of the thermal stratification snapshots. The graph neural network exhibits high nonlinearity and can inherently handle the unstructured grid data used in computational fluid dynamics, thus achieving a comparable first-order modal reconstruction accuracy to that of POD with 30–50 basis functions. During online processes, the feature recognition and prediction of thermal stratification snapshots can be completed within 472 ms, with accuracies similar to those achieved via reconstruction. This research holds great significance for illustrating the nonlinear dimensionality reduction of large-scale computational fluid dynamics data, elucidating the thermal stratification mechanism, and rapidly predicting the evolution of this mechanism.
Zeng et al. (Wed,) studied this question.