Residual-corrected stacking ensemble learning for concrete strength prediction and optimization of low-carbon mix design

To improve sustainability of concrete materials, this study proposes an advanced method for optimizing concrete mix design, combining ensemble learning with NSGA-II algorithm. A weight reduction strategy is employed to manage outliers, enhancing the robustness of the model. Support vector regression, extreme gradient boosting, and gradient boosting regression are used as base learners, integrated through a stacking ensemble approach with ridge regression as the meta-model. A multi-layer perceptron (MLP) neural network is introduced as a residual correction model to improve prediction accuracy for concrete compressive strength. Monte Carlo simulations are used to evaluate the uncertainty and robustness of the model under input perturbations. The NSGA-II algorithm is applied for multi-objective optimization, balancing concrete strength, cost, and carbon emissions. The results demonstrate that the MLP residual model effectively captures nonlinear features, improving prediction performance. Monte Carlo analysis reveals that the C20 (17.5–22.5 MPa) strength grade is the most reliable prediction interval, with an average confidence interval width of 11.23 MPa. The NSGA-II optimization identifies an optimal solution for C20 concrete, with cost, carbon emissions, and strength values of 298.73 CNY, 87.70 kg, and 22.42 MPa, respectively. This study provides a sustainable, low-carbon alternative for concrete mix design, offering valuable insights into sustainable construction practices.