An Interpretable and Comparative Uncertainty-Aware Machine Learning Approach for UHPC Mix Design Optimization

<title>Abstract</title> <p>Ultra-high-performance concrete (UHPC) is an advanced cementitious composite offering superior strength, durability, and extended service life. However, its complex mix design involves interactions among multiple constituents, making conventional trial-and-error approaches time-consuming, labour-intensive, and costly. This study proposes a comprehensive and interpretable machine learning (ML) framework for predicting the compressive strength of UHPC. Seven ML models—linear regression (LR), support vector regression (SVR), backpropagation neural network (BPNN), k-nearest neighbors (KNN), decision tree (DT), random forest (RF), and gradient boosting (GB)—are systematically evaluated using a unified dataset and consistent performance metrics. Beyond conventional accuracy assessment, predictive uncertainty is quantified using a 95% confidence interval to facilitate robust model selection. Shapley Additive exPlanations (SHAP) are further employed to interpret the optimized model and quantify the influence of individual mix constituents on compressive strength. The results demonstrate that ensemble-based models generally outperform linear and single-learner models in terms of prediction accuracy and robustness. Uncertainty analysis further reveals differences in model reliability that are not captured by conventional accuracy metrics alone. To facilitate practical implementation, a Python-based graphical user interface (GUI) is developed, enabling engineers to predict compressive strength and efficiently adjust UHPC mix proportions. The proposed framework integrates prediction accuracy, uncertainty quantification, interpretability, and practical usability, providing a reliable data-driven tool for UHPC mix design and supporting the broader adoption of ML-assisted sustainable concrete design.</p>