Multi-output machine learning framework for predicting mechanical properties of rice husk ash blended concrete.

The use of rice husk ash (RHA) as a supplementary cementitious material provides a sustainable route to reducing cement consumption and CO2 emissions in concrete. However, accurate prediction of its mechanical properties is essential for effective design, while most existing studies focus solely on compressive strength (CS) and lack integrated interpretability. A data-driven framework is developed to simultaneously predict CS, tensile (TS), and flexural strength (FS) using a curated dataset of 105 RHA-based concrete mixtures compiled from 17 literature sources. Five regression models, including linear, ensemble, and kernel-based approaches, were evaluated under both multi-output and single-output formulations. The results show that the multi-target framework achieves predictive performance comparable to independent single-target models, without loss of accuracy. This indicates that the input variables sufficiently capture the governing relationships among strength parameters. Ensemble models, particularly Gradient Boosting (GB), achieved the best overall performance (R2 = 0.99, RMSE = 1.74 MPa), with consistent accuracy across all targets and negligible overfitting in comparison to other models. Cross-validation (CV) confirmed model stability, with minimal variation across data splits. In contrast, HGB and SVR exhibited weaker generalization, while XGB showed signs of overfitting. SHAP-based interpretability reveals physically consistent trends, with coarse aggregate content identified as the dominant factor, water-related parameters influencing matrix density, and RHA exhibiting a limited direct contribution at 28 days due to its delayed pozzolanic reactivity. The proposed framework provides both predictive capability and mechanistic insight, supporting sustainable mix design and future model development.