关键词: 计算流体力学, 填充床, 乙烯氧化, 神经网络, 图卷积架构

Abstract:

As a crucial intermediate in the petrochemical industry, the collaborative optimization of catalyst morphology and operating parameters during the ethylene oxide production process represents the core challenge in enhancing reactor performance. In response to the high cost bottleneck of traditional experimental and simulation methods in analyzing the structure-activity relationship of catalysts, this study integrated particle-resolved computational fluid dynamics (PRCFD) and graph convolutional neural networks (GCN) to construct an intelligent prediction framework for reactor multi-physics fields. A high-fidelity CFD model was established based on the COMSOL platform to investigate the flow-reaction coupling process of cylindrical, single-hole, and five-hole structure catalysts in a randomly packed system. A comprehensive research scenario covering three typical particle morphologies, random packing configurations, and four inlet gas rates was constructed. Compared with real ethylene conversion data, the effectiveness of the simulation parameter settings in COMSOL was verified. The simulations revealed that the effects of catalyst particle shape and inlet gas rate on the ethylene conversion rate and the bed pressure drop exhibited a strong non-linear relationship. Based on the valid simulation data, a graph convolutional neural network was employed to learn the mapping relationships between the geometric shapes of catalyst particles and pressure and concentration. The trained model could rapidly predict the pressure and concentration distributions under different catalysts and inlet gas rates, with a correlation coefficient R² greater than 0.9. This study provided a new paradigm which combines physical interpretability and computational efficiency for the intelligent design of chemical reactors.

Key words: computational fluid dynamics, packed bed, ethylene oxidation, neural networks, graph convolution architecture