Integrating Experimental Durability Assessment and Explainable Artificial Intelligence for Sustainable UHPC in Smart Construction Applications

Digitalisation and explainable artificial intelligence (XAI) will reshape sustainable materials engineering by enabling data-driven design and performance prediction for smart infrastructure. In this study, experimental durability assessment and machine learning techniques were combined to investigate silica fume (SF) and ground granulated blast-furnace slag (GGBFS)-based UHPC subjected to normal curing (NC), steam curing (SC), and wrapped curing (WC). Mechanical, durability, and mineralogical characterisation tests were performed, including compressive strength, the Rapid Chloride Penetration Test (RCPT), water absorption, sorptivity, an accelerated corrosion test, and X-ray diffraction (XRD). Accelerated steam curing (SC) produced a measured early-age compressive strength of 85.97 MPa at 6 h and the lowest RCPT value of 81.09 Coulombs among the investigated curing conditions. The measured water absorption and accelerated corrosion results also varied with curing conditions, with SC showing comparatively favourable performance. These findings provide experimental evidence of curing-dependent mechanical and durability behaviour, although the pore structure was not directly quantified. XRD analysis provided qualitative evidence of differences in the crystalline and amorphous features of the UHPC matrix under the investigated curing conditions. An exploratory comparison of compressive strength and chloride-ion penetrability was performed; however, the measurements were obtained at different curing ages and from only three experimental curing conditions, so no age-matched or quantitative strength–RCPT relationship was established. An XGBoost model (R2 = 0.9591) was developed using the adopted dataset and 80/20 test partition to support digital materials engineering. SHAP-based feature attribution identified SF, curing age, superplasticizer dosage, and fibre reinforcement as important model features for compressive-strength prediction. The combined experimental–XAI framework enables a qualitative consistency assessment between observed UHPC behaviour and data-driven feature attribution. Because the experimental specimens were independent of the ML database, the experimental programme serves as a qualitative consistency check of SHAP-based feature attribution rather than a rigorous validation of the model predictions.