Deep learning-enabled characterization of freeze-thaw induced microstructural deterioration in cement-based materials using metal-impregnated BSE imaging

The performance evaluation and regulation of cement-based materials under freeze-thaw cycles are crucial for the long-term stability of engineering structures in cold regions. However, existing research still has deficiencies in the precise identification and location of microscopic damage. Meanwhile, the mechanisms by which nano-modification improves the freeze resistance of cement-based materials remains unclear. Here, we proposed an innovative metal intrusion method and backscattered electron characterization combined with deep learning-based analysis to extract the microstructure damage feature of cement-based materials under two factors, freeze-thaw cycles and graphene oxide modification, and to evaluate the performance of freeze-thaw damage of cement-based materials. The results show that the recognition accuracy of the trained model for microstructure images reaches 83.4%, with a particularly high accuracy of 92.6% in identifying graphene oxide -modified features. The class activation mapping algorithm reveals the core characteristics of freeze-thaw cycle damage recognition in cement-based materials, mainly concentrated in the pore connectivity zone, the edge of micro-cracks, and the porosity area of hydration products. With the increase in the number of freeze-thaw cycles, the regions of interest identified by the model exhibit a corresponding expansion, with the proportion of the affected area rising from 11.2% in the control group to 37.0% after 30 freeze-thaw cycles. This clearly illustrating the deterioration process characterized by pore propagation and crack interconnection. This research achievement not only promotes the development of freeze-thaw damage identification methods, but also provides new technical means and theoretical basis for freeze-thaw durability design and life assessment.