Microstructural Evolution and Durability of Fly Ash–Bentonite Cementitious Composites Under Freeze–Thaw and Wet–Dry Cycling in Sulfate Environments: An Interpretable Machine Learning Approach

Sulfate-bearing cyclic environments induce coupled degradation of hydration gels and pore structure, but the differences between freeze–thaw and wet–dry deterioration remain unclear. Nine fly ash–bentonite cementitious composites with fly ash-to-cement ratios of 1.2, 1.8, and 2.4 and bentonite contents of 0%, 3%, and 6% were exposed to water, 10% Na2SO4, and 10% MgSO4 for up to 150 cycles. Macroscopic durability was evaluated using the squared ultrasonic pulse-velocity index (plotted as apparent RDEM), compressive-strength retention, and peak-stress retention; hydration products and pore evolution were characterized by X-ray diffraction, scanning electron microscopy, and low-field nuclear magnetic resonance. Freeze–thaw cycling was mainly associated with physical cracking and a redistribution toward longer-T2 water-accessible relaxation components, whereas wet–dry cycling intensified salt crystallization and sulfate-product accumulation. Sodium sulfate exposure was associated with stronger ettringite and gypsum diffraction peaks, whereas magnesium sulfate exposure showed a stronger gypsum peak and a weaker portlandite peak. Among the tested complete formulations, moderate fly ash contents and low bentonite dosages (with corresponding water-reducer adjustments) provided the most balanced resistance, whereas FA/C = 2.4 with 6% bentonite generally deteriorated more rapidly. Based on pooled out-of-fold predictions, support vector regression performed best for the apparent RDEM indicator (R2 = 0.929), while random forest gave the highest accuracy for compressive-strength retention (R2 = 0.937) and peak-stress retention (R2 = 0.994); the first two targets had 378 condition records each and peak stress had 216 author-confirmed numerical values. In leave-one-mixture-out sensitivity analysis, the corresponding pooled R2 values were 0.897, 0.942, and 0.992, respectively; this analysis assesses interpolation to an unobserved combination within the tested 3 × 3 composition grid, not extrapolation beyond its factor levels. Recomputed grouped SHAP attribution identified cycle count as the leading input for apparent RDEM (44.94% of relative mean absolute attribution), while the exposure-regime input dominated strength retention (53.36%) and peak-stress retention (56.65%); cycle count × regime was the strongest pairwise interaction in both mechanical targets. At limit states of 60% RDEM and 75% compressive-strength retention, the predicted durability thresholds ranged from 17 to more than 150 cycles, with RDEM governing most cases. The proposed framework pairs crystalline-phase and pore-structure observations with interpretable durability prediction within the tested cyclic sulfate environments. These are exploratory laboratory-cycle thresholds. A 300-resample trajectory bootstrap quantified model-refit variation; neither the thresholds nor their intervals are calibrated to field service time.