Ensemble Super Learner Model for predicting strain at maximum stress and energy absorption capacity of UHPFRC

Accurately predicting the mechanical performance of ultra-high-performance fibre-reinforced concrete (UHPFRC) remains challenging due to the complex, nonlinear interactions among its constituents. This study presents a Super Learner model using various ensemble techniques to simultaneously predict strain at maximum stress ( ɛ p c ) and energy absorption capacity ( g ) based on 34 input variables across 980 UHPFRC samples. Exploratory data analysis and leverage-based outlier detection were employed to refine five optimised datasets and improve predictive accuracy in regions of data misalignment. The top-performing base learners, namely Gradient Boosting Regressor (GBR), XGBoost and CatBoost, were integrated into a Super Learner framework (SL-GXC), which achieved a high coefficient of determination (R 2 = 0.973), surpassing existing modelling approaches. Model interpretability was strengthened through SHapley Additive exPlanations (SHAP) and Partial Dependence Plot (PDP) analyses. Results identify the fibre reinforcement index as the most influential variable, positively affecting both outputs. While higher matrix compressive strength generally enhances mechanical performance, values above 130 MPa reduce ductility due to increased brittleness. Additionally, longer fibres and larger equivalent diameters improve strain capacity and energy absorption by promoting better mechanical interlocking. Among fibre types, twisted steel fibres were most effective for energy dissipation, whereas smooth high-strength steel microfibres excelled in maximising strain at peak stress. This work introduces a robust and interpretable predictive framework that offers valuable insights for optimising UHPFRC mix design in high-performance structural applications. • Super Learner model using ensembles to predict strain at max stress & energy absorption of UHPFRC. • GBR, XGBoost, and CatBoost combined in Super Learner (SL-GXC) for superior results. • Further insights were gained using PDP and SHAP analyses. • Results of original and filtered datasets showed superior performance over prior studies.