Optimized hybrid XGBoost-CatBoost model for enhanced prediction of concrete strength and reliability analysis using Monte Carlo simulations

Abstract

Building on our previous work demonstrating the exceptional potential of the Extreme Gradient Boosting model (XGBoost) for predicting the uniaxial compressive strength of concrete, this study introduces several significant advancements. First, we develop a novel optimized hybrid model that synergistically combines XGBoost, CatBoost (one of the most advanced tree-boosting models), and the Optuna algorithm to achieve unprecedented prediction accuracy. Second, we apply this hybrid model in Monte Carlo simulations to conduct a pioneering reliability analysis of concrete strength, capturing the effects of input uncertainty. Third, we propose an innovative technique for estimating tree leaf values, which fundamentally improves prediction accuracy. Our optimized hybrid model delivers outstanding performance, as evidenced by a five-fold cross-validation showing a coefficient of determination (R²) of 0.953, a root mean squared error (RMSE) of 3.603 MPa, and a mean absolute error (MAE) of 2.261 MPa—metrics that surpass the best results reported in the existing literature. Additionally, our Monte Carlo simulations reveal a substantial error range of 10–20 MPa for a ±5 % variation in input features, underscoring the critical impact of input uncertainty on prediction reliability. Furthermore, our new leaf value estimation technique significantly outperforms traditional averaging methods, offering a transformative improvement in model accuracy. These findings are crucial for broadening the scope of machine learning applications in civil engineering and other engineering disciplines.