Microstructure analysis of corundum spinel material based on AI models

Artificial intelligence analysis of micro-CT images provides a new paradigm for uncovering complex nonlinear relationships between microstructure and properties of refractory materials. A hybrid deep learning model named ResVi is proposed to classify heat-treatment conditions from microstructural images of CAC-bonded corundum-spinel castables. ResVi integrates the local feature extraction capability of ResNet50 with the global dependency modelling strength of Vision Transformer. A microstructure image dataset covering a temperature gradient from 110 °C to 1600 °C was constructed from micro-CT scans of laboratory-provided samples with a slice spacing of 2 μm, followed by systematic preprocessing and feature quantification. Comparative experimental results show that ResVi achieves a top accuracy of 99.65% and an F1-score of 99.62% on the test set, with stable performance across different batch settings and an accuracy fluctuation below 1.4%. Grad-CAM visualisation and physics-consistent analysis confirm that the model predictions are consistent with high-temperature sintering thermodynamics and pore structure evolution, offering clear physical interpretability. This method provides a data-driven solution for microstructure analysis and process optimisation of refractory materials.