RSM and Hybrid Machine Learning for Multi-Objective Optimization of Sustainable Concrete with Plastic and Sawdust Wastes

Background: This study presents a comprehensive predictive–optimization framework that integrates Response Surface Method: (RSM), advanced machine learning (ML) models, and metaheuristic multi-objective optimization to design eco-efficient concrete mixtures incorporating multiple waste materials, namely regular-gray ABS (GR), irregular-black ABS (BI) plastic granules, and sawdust (SD). Unlike previous studies that examined individual waste additives, this research evaluates their combined effects on concrete performance, enhanced with silica fume (SF) and superplasticizer (SP). Thirty-four concrete mixtures were prepared using thirteen input parameters: W/C, W/B, C, SF, FA, CA, W, SP, GR, BI, SD, PW, and TW, to predict compressive strength at 7, 14, and 28 days (CS7, CS14, CS28), workability, and density. Results: RSM and Gradient Boosting achieved the highest predictive accuracy, with Gradient Boosting models reaching R² values up to 0.997 and root mean square errors (RMSE) as low as 0.99 MPa for CS28, 4.83 mm for workability, and 11.54 kg/m³ for density. A composite score-based evaluation integrating R², RMSE, mean absolute error (MAE), average absolute deviation (AAD), and overfit gap identified the best-performing models, which were coupled with Differential Evolution (DE) for multi-objective optimization. The optimal mixture with W/C = 0.617, W/B = 0.552, SP = 0.011 m³, GR = 0.000 m³, BI = 0.086 m³, and SD = 0.054 m³ achieved CS7 = 29.95 MPa, CS14 = 36.87 MPa, CS28 = 39.69 MPa, workability = 107 mm, and density = 2,182.8 kg/m³, with an overall desirability of 0.99. Experimental validation confirmed low residuals (0.31–0.67 MPa for strength, 8.83 mm for workability, and 10.34 kg/m³ for density), demonstrating the reliability and reproducibility of the hybrid RSM–ML–metaheuristic framework for converting multiple waste streams into high-performance eco-friendly concrete.