Data-driven investigations and advanced modeling with extra random forest on the chloride migration in type I cement-based materials

Abstract

In this comprehensive investigation, a data-driven extra random forest algorithm was employed to model concrete's non-steady-state chloride migration coefficient (DN). The model uses an extensive dataset of 540 sets and 12 features. The calibrated model demonstrates exceptional predictive capability, emphasizing the influential factors in the studied complex system. The fine-tuned XRF model highlights the importance of water-binder ratio (WB) and curing age, with mean SHAP values of 3.69 and 2.58, respectively. Further, the investigation reveals valuable insights into the impact of individual factors on the migration coefficient, offering valuable quantitative conclusions. Key findings include an 18% increase in DN with cement content from 200 to 400 kg/m³, a 30% decrease in DN with slag content up to 100 kg/m³, and marginal influences of fly ash and lime filler. Additionally, the study explains a complex relationship between water content and DN and underscores the WB ratio's increased sensitivity beyond 0.45. Chemical admixtures (superplasticizer and Air-entraining-agent) to binder ratios positively influence reducing DN, and curing age significantly decreases DN, highlighting enhanced impermeability through hydration age development.