Forecasting three-dimensional unsteady multi-phase flow fields in the coal-supercritical water fluidized bed reactor via graph neural networks

The study of three-dimensional unsteady multi-phase flows inside the coal-supercritical water fluidized bed (SCWFB) reactor coupled with fluid dynamics, heat transfer and chemical reactions usually relies on time-consuming numerical simulation. Recently, to provide quick spatio-temporal estimation of the SCWFB reactor, the data-driven deep learning techniques have been adopted to build a digital yet fast reactor forecasting model. However, due to the irregular 3D geometry, it usually uses unstructured meshes to discrete the domain of reactor and builds the corresponding simulation model, which poses difficulties for the convolutional neural networks based deep learning models. Hence, this article develops a 3D unstructured spatio-temporal forecasting model, named GraphReactorNet, via graph neural networks for learning the dynamics under multi-phase flow fields inside the SCWFB reactor. It utilizes an efficient strategy to tackle large-scale 3D transient flow graphs and proposes an effective unstructured spatio-temporal learning module to extract the underlying flow dynamics. This model is trained on data collected from the transient simulator at different conditions. Numerical experiments reveal that in comparison to the counterparts, the GraphReactorNet not only achieves accurate multi-step-ahead predictions of the multi-phase flow fields at unseen conditions, but also performs thousands of times faster than the numerical simulator.