Recycled aggregate concrete based on an improved SVM method a study on the prediction of compressive strength

The mix design of recycled aggregate concrete (RAC) has long relied on extensive trial-and-error testing, which is costly and time-consuming; moreover, existing empirical models struggle to accurately capture the complex nonlinear relationship between material components and compressive strength. To address these challenges, this study aims to develop a method for predicting RAC compressive strength that combines high accuracy with interpretability, thereby providing data-driven support for the engineering application of recycled materials. To this end, this study proposes a hybrid kernel support vector regression (SVR) model, whose core innovation lies in the weighted combination of linear and radial basis function (RBF) kernels. This enables the model to simultaneously capture both the global linear trends (dominated by the linear kernel) and the local nonlinear fluctuations (dominated by the RBF kernel) between RAC mix design parameters and compressive strength, thereby overcoming the limitations of a single kernel function in characterizing complex material systems. Based on 178 sets of experimental data, the study constructed a comprehensive database. Nine input variables were systematically selected across three levels—cementitious system, aggregate content, and aggregate characteristics—and the model hyperparameters (C, σ, λ) were systematically optimized using grid search combined with five-fold cross-validation. The results show that the optimal hybrid kernel SVR model achieved a coefficient of determination (R 2 p) of 0.99506 and a root mean square error (RMSEP) of 0.52589 MPa on the test set, with prediction accuracy significantly superior to that of the linear kernel SVR model (R 2 p = 0.88116, RMSEP = 2.9519 MPa). Ten rounds of five-fold cross-validation showed that the mean R 2 of the hybrid kernel SVR was 0.993, and its 95% confidence interval [0.989, 0.997] did not overlap with that of the linear kernel model [0.786, 0.852] ( p < 0.01), confirming the statistical significance and robustness of the model’s accuracy. This study marks the first systematic introduction of the hybrid kernel strategy into the field of RAC strength prediction, filling a research gap regarding this method in highly variable recycled concrete material systems; through a permutation importance analysis, cement content, water content, and recycled coarse aggregate content were identified as the dominant factors influencing strength, thereby enhancing the model’s physical interpretability; The established model achieves high-accuracy predictions under limited sample conditions, providing a viable tool for reducing the number of mix design trials and lowering material R&D costs. This method can be further extended to the prediction of properties in other construction solid waste materials, holding positive significance for promoting the resource utilization of solid waste and low-carbon construction.