Statistical and machine learning analysis of filler type and dosage in relation to 28-day UHPC compressive strength

Abstract Ultra-high-performance concrete commonly relies on high binder contents, which has encouraged the use of mineral and waste-derived fillers. This study examined database-level associations and model-attributed contributions of broad filler category, absolute dosage, and filler fraction to 28-day compressive strength using 2073 mixture records from 165 publications. Mixtures with and without fillers differed by only − 1.69 MPa on average, and filler category explained approximately 3% of the unadjusted strength variation. OLS with HC3 heteroscedasticity-consistent standard errors showed no conventionally significant pooled filler-dosage coefficient, while publication-clustered inference showed that several category-specific associations were sensitive to source dependence. Record-level XGBoost performance was high on the fixed internal holdout (R 2 = 0.843), but publication-grouped cross-validation reduced mean R 2 to 0.380, indicating substantially weaker transfer to unseen source studies. Nested publication-grouped tuning materially improved Random Forest but changed Ridge, CatBoost, and XGBoost only slightly; publication-grouped performance nevertheless remained well below record-level performance. Within the fitted XGBoost models and adopted feature representations, the broad filler-category descriptor received larger model attribution than filler fraction or absolute dosage, while all filler-related variables remained less influential than the leading mixture, curing, specimen, and testing variables. The results therefore provide source-conditioned database evidence for hypothesis generation rather than universal material rankings, causal effects, or engineering-ready prediction.