作为一种大型洁净能源装备，煤炭超临界水制氢反应器内部是耦合传热传质和化学反应的多相环境。制氢反应器的实验和数值模拟研究产生了大量的多维、瞬时流场数据。本文首先通过空间插值处理所获得的非结构化多相流场仿真数据，并利用深度学习技术构建了数据驱动的三维多相流场时空预测模型3DReactorNet，对三维制氢反应器内复杂的多相流场进行学习，从而实现对未知工况下反应器内三维多相流场时空演变的快速准确预测。进一步地，本文通过MC Dropout 策略度量了3DReactorNet模型预测结果的置信度。测试结果表明，3DReactorNet 模型的预测结果与CFD 计算结果高度一致，但预测速度远优于CFD仿真，有利于高效的反应器设计和优化。

The coal-supercritical water fluidized bed (SCWFB) reactor is a large-scale clean energy equipment that operates in a multi-phase environment with coupled heat and mass transfer and chemical reactions. Studies on the SCWFB reactor have generated a significant amount of multidimensional, transient flow fields data through experimental and numerical simulations. This paper presents a data-driven 3D multi-phase flow fields spatio-temporal prediction model, 3DReactorNet, which uses deep learning technology to learn the complex multiphase flow fields inside the SCWFB reactor. The obtained unstructured multi-phase flow fields simulation data is processed by spatial interpolation to achieve fast and accurate prediction of the spatio-temporal evolution of the 3D multiphase flow fields inside the reactor under unknown operating conditions. In this paper, the confidence level of the prediction results of the 3DReactorNet was measured using the MC Dropout method. The test results indicate that the prediction results of the 3DReactorNet are highly consistent with the CFD simulation results. Additionally, the prediction speed is much faster than CFD simulation, which is beneficial for efficient reactor design and optimization.