Proxy Feature-Based Interpretable Machine Learning to Predict Multi-Properties of Ultra-High-Performance Concrete

Abstract

Raw mix-design variables used to predict ultra-high-performance concrete (UHPC) properties cannot fully represent internal proportions and structural compatibility. This study develops mechanism-informed proxy-core features based on particle packing, water film thickness, and rheology, and evaluates four feature systems using literature-derived datasets for compressive strength (924 samples), flexural strength (406 samples), and slump flow (192 samples). Eight regression models were compared using five-fold cross-validation, Bayesian optimization, SHAP, and ablation analysis. Proxy-W achieved an R2 of 0.918 for compressive strength, only 0.005 higher than baseline-W on the original split. For flexural strength, baseline-WB achieved the highest R2 (0.906), whereas proxy-W yielded the lowest MAE (2.140 MPa). The largest single-split difference occurred for slump flow (R2: 0.740 to 0.838 for W-based systems), but 30 repeated splits yielded mean R2 values of 0.667 ± 0.169 and 0.738 ± 0.101 for baseline-W and proxy-W, respectively, while proxy-WB showed no average improvement over baseline-WB. Repeated and source-group validation further indicated that the predictive effects of proxy-core features were property-, representation-, partition-, and source-dependent. Overall, proxy-core features are best interpreted as physically informed relational representations rather than universally accuracy-enhancing features.