Dynamic Frost Heave Susceptibility of Loess Under Climate Change: A Physics-Constrained Machine Learning Framework Integrating SFCC Prior Knowledge and CMIP6 Projections

Frost heave in seasonally frozen loess regions is fundamentally governed by pore water migration towards the freezing front driven by temperature gradients, forming ice lenses that damage engineered infrastructure. Because both freezing intensity and moisture availability evolve with climate, frost heave susceptibility is itself dynamic, yet existing assessments remain static and ignore future climate trajectories. This paper presents a physics-constrained machine learning framework that couples soil freezing characteristic curve (SFCC) prior knowledge with multi-source open data and CMIP6 climate projections to achieve dynamic frost heave susceptibility mapping for the Loess Plateau. Monotonicity constraints derived from the coupled phase-transition and cryosuction mechanisms described by the SFCC and the segregation potential theory are enforced during gradient-boosted tree training, ensuring that predictions respect the established relationships among freezing intensity, fine-grained content and ice segregation potential. An ordinal decomposition strategy is adopted to guarantee that the monotonicity constraint on each binary sub-model translates into monotonicity of the predicted ordinal susceptibility level. The best performer, physics-constrained XGBoost, reaches an overall accuracy of 88.7% and an AUC of 0.942 on a four-class susceptibility scheme. Independent validation against 156 field records and Sentinel-1 InSAR observations confirms that the model captures genuine frost heave patterns. Under SSP5-8.5, the area classified as high or very-high susceptibility contracts by approximately 38% by the 2080s owing to warming, while under SSP1-2.6 the reduction is only 12%, and transitional zones of moderate risk expand in both scenarios. These findings provide a temporally explicit and physically grounded basis for climate-adaptive infrastructure planning in cold loess regions.