Multi-dimensional CFD-Mask R-CNN and CFD-watershed segmentation approach for multiphase non-catalytic gas-solid reactions: A case study for hydrogen reduction of porous iron oxide pellets

Abstract

The direct reduction of iron oxide (DRI) using hydrogen is a promising method for clean steelmaking. Previous studies used one-dimensional models to explore this non-catalytic gas–solid reaction, but they lacked accuracy in assessing the impact of pellet shape. To tackle this issue, a multi-dimensional computational fluid dynamic (CFD) method has been employed, integrating a novel porous three-interface unreacted shrinking core model (USCM) to explore the effects of pellet shape through multiphase reactions of Fe2O3 → Fe3O4 → FeO → Fe by pure hydrogen. The model incorporates Knudsen, molecular, and effective diffusion mechanisms, alongside a continuous porosity function. Validation against previous experiments within a temperature range of 800–1000 °C and varying porosity is performed by solving four partial differential equations (PDEs) of mass and momentum balance in a porous media using the finite element method (FEM). Subsequently, a dataset comprising 754 images of pellets from the industrial DRI unit is collected and applied for the mask region-based convolutional neural network (Mask R-CNN) algorithm coupled with CFD, to draw geometries and simulate DRI to study indentations and protrusions of pellets. Furthermore, watershed segmentation is applied to investigate the shapes and real sizes of pellets from their images. The Mask R-CNN achieves intersection over union (IoU), precision, and recall percentages of 86.1 %, 87.3 %, and 87.2 % on average, respectively. Pellet shape investigations reveal notable discrepancies in wustite phase profiles and porosity. A nuanced impact is discerned regarding the DRI in 2D shape deviation, while 3D model errors in solid conversion time range from 3 % to 12.8 %.