A transfer learning-driven paradigm for understanding cryogenic freezing mechanisms in low water/binder cement-based composites

Abstract

Data scarcity and dispersion constrain the accurate assessment of cryogenic concrete performance and impede the investigation of its microstructural evolution. This study first developed a physics-informed transfer learning framework to predict low water/binder cement-based composites (LWCC) behavior across a wide temperature range and reveal coupled macro–micro freezing mechanisms. The results demonstrated that transfer learning, by integrating hydration and freezing features, effectively overcame data scarcity and enabled accurate prediction of LWCC low-temperature performance (R² > 0.90). The model exhibited self-adaptive capability, extending its applicable temperature range to be extended from the training range of 0 to −80°C down to −196°C. Based on the data-driven model, a freezing model was proposed, identifying strength gain, damage–densification balance, secondary strengthening, and interfacial debonding stages, with transitions influenced by w/b ratio and solid skeleton structure. This work achieves AI-driven precise prediction of cryogenic concrete, offering insights into material design and freezing mechanisms under extreme conditions.