Influence of feature-to-feature interactions on chloride migration in type-I cement concrete: A robust modeling approach using extra random forest

Understanding the chloride transport mechanism is crucial for enhancing the durability of structures exposed to chloride-rich environments. This study assessed various machine learning (ML) algorithms for predicting the chloride migration coefficient in type I Portland cement-based concrete. The fine-tuned extra random forest model demonstrated satisfactory predictive capability, with the majority of tested-predicted data points falling within the +/- 30 % tolerance margin. This calibrated model was employed to explore feature-to-feature interactions affecting concrete resistance to chloride ingress. Examination of partial dependence plots revealed that an increased sensitivity with decreasing water-binder ratio (WB), reaching saturation beyond 0.40 WB, and the increase in curing age substantially reduced chloride diffusion (DN). Further, an augmented dosage of silica fume (SF) decreased DN, with a more pronounced effect of slag (SL) at higher WB, and the impact of superplasticizer (SP) decreased DN at high WB ratios. Concrete incorporating SF showcased optimal resistance to chloride permeability with the highest SF content (20-30 kg/m(3)). Similarly, high SL content (>150 kg/m(3)) enhanced resistance. The interactive influence of SF and SP on DN was marginal at low SF dosages (<20 kg/m(3)) but became complex at higher levels, with potential minimum values observed in the range of 0.5-1.0 kg/m(3) SP. For SL-containing concrete, achieving the lowest DN was possible with PC and SL dosages around 200-400 and exceeding 100 kg/m(3), respectively. For concrete incorporating Type I PC, an increase in SP content generally had a negligible impact on DN, with a slight tendency to raise values observed at higher cement amounts (exceeding 380 kg/m(3)). The optimal combination for achieving the minimum DN involved compositions of PC, fine, and coarse aggregate at 200-300, 100-150, and 250-500 kg/m(3), respectively.