Multiscale effects of freeze thaw cycling on cement stabilized aeolian sand from mechanical behavior to pore and bond evolution

Freeze–thaw cycling can alter the mechanical response, pore structure, and cemented contacts of stabilized aeolian sand, but these changes may not evolve consistently across observation scales. This exploratory study examined aeolian sand stabilized with 6% cement after 0, 1, 3, 6, 10, and 15 freeze–thaw cycles. Static triaxial tests were conducted under confining pressures of 25, 50, and 100 kPa, while nuclear magnetic resonance (NMR) and SEM–EDS were used to characterize pore-structure and local cementation changes. One triaxial specimen was tested for each freeze–thaw and confining-pressure condition; therefore, the mechanical results are presented as individual observations rather than statistically averaged material responses. Strength and stiffness varied non-monotonically during the early and intermediate cycles, followed by an overall reduction at later cycles. The NMR and SEM–EDS observations likewise showed that pore-scale and bond-scale changes did not always evolve in parallel with the mechanical response. Taken together, the results are consistent with early structural disturbance, a heterogeneous intermediate condition, and later localized deterioration of cemented contacts, although this sequence is interpreted as exploratory rather than as a fixed set of material thresholds. An exploratory curve-grouped machine-learning analysis was then used to test whether freeze–thaw history and cycle-level NMR descriptors provided predictive information beyond axial strain and confining pressure. Relative to the mechanical-loading model, the macro–micro feature set reduced the median leave-one-curve-out RMSE from 159.55 to 113.82 kPa and increased the median R 2 from 0.831 to 0.904. However, the improvement varied among individual curves and was not retained when FT15 was completely withheld, indicating that the available state-level pore descriptors provide condition-dependent rather than uniformly transferable information for point-wise stress prediction.