Prediction of UHPC mechanical properties using optimized hybrid machine learning model with robust sensitivity and uncertainty analysis

Abstract This study presents a comprehensive evaluation of three hybrid machine learning models XGB-LGB, RF-XGB, and ET-LGB for predicting the mechanical performance of Ultra-High-Performance Concrete (UHPC), including compressive strength (CS), flexural strength (FS), and tensile strength (TS). Each dataset was standardized and split into training (80%) and testing (20%) subsets. Hyperparameter optimization was conducted using a random search algorithm to improve prediction accuracy. The performance of each model was assessed using five key metrics: mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R 2 ). This work integrates hybrid ensemble models with SHAP-based explainable AI and uncertainty quantification to achieve high accuracy, model interpretability, and robustness assessment, which are rarely combined in UHPC prediction studies. Among the models, ET-LGB reliably achieved the highest accuracy across all target outputs, with R 2 values reaching 0.99 in both training and testing phases. XGB-LGB also demonstrated strong performance, particularly for CS and FS, achieving R 2 values of 0.99 and 0.98, respectively. In contrast, RF-XGB showed relatively lower accuracy, especially for TS, with R 2 values around 0.93. To improve model interpretability, SHAP-based sensitivity analysis including feature importance plots, beeswarm plots, and heatmaps were employed to analyze the contribution of input features to model predictions. Additionally, an uncertainty analysis was performed to measure the robustness of predictions. Overall, the ET-LGB model proved to be the most reliable and accurate, followed closely by XGB-LGB, demonstrating strong potential for practical UHPC property prediction.