Data-driven understanding of macro-micro feature evolution in cement-based composites under extreme low-temperature environments

Cementitious infrastructure systems operating in polar regions, cryogenic storage facilities, and other extreme environments require reliable long-term performance under severe thermal and environmental conditions. However, limited availability of low-temperature experimental data restricts the understanding of coupled hydro-thermal-mechanical behavior and hinders data-driven material design. To address this challenge, this study develops a knowledge-guided transfer learning framework for small-sample learning, data augmentation, and microstructure-informed system interpretation of cement-based composites under extreme low-temperature conditions. Conventional and Bayesian-optimized machine learning models are first compared using room-temperature data to construct a robust source-domain model. The optimized model is subsequently fine-tuned using limited low-temperature data to establish a target-domain predictor and enable physically consistent dataset augmentation. Based on the augmented data, the coupled evolution of mechanical properties, pore structure development, and freezing water behavior under different water-to-binder ratios is systematically analyzed. The results show that source-model optimization improves cross-domain transfer stability and predictive reliability. More importantly, the augmented dataset reveals intrinsic coupling among cryogenic exposure, pore network evolution, ice formation, and mechanical performance, providing mechanistic insights into low-temperature degradation processes. This work establishes a unified data-driven and knowledge-guided framework for the system-level design and optimization of cementitious infrastructure materials in extreme environments.