Multi-objective optimization and predictive analytics of strength and embodied impacts of CDW-based geopolymers using various machine learning approaches

Geopolymers generally exhibit heterogeneous composition, variability and complex interactions, posing significant challenges in predicting their properties. This study focuses on the multi-objective optimization (MOO) and predictive analytics of geopolymer mortars (GPMs) incorporating construction and demolition wastes (CDWs) as recycled precursors and aggregates. The objectives are to predict and optimize the compressive strength (CS) while minimizing the embodied energy (E-energy) and CO₂ (E-CO₂) emissions. To achieve these goals, various interpretable ensemble machine learning (ML) models, including linear regressors, advanced tree-based methods, K-nearest neighbors and neural network-based multi-layer perceptron regressors, are employed. The standalone extreme gradient boosting model achieved the best performance, attaining the highest predictive accuracy (R² = 0.939, 0.937 and 0.843 for CS, E-energy and E-CO₂, respectively) with minimal error indices. Furthermore, a MOO framework utilizing advanced sorting genetic algorithms was employed to generate a Pareto front of optimal solutions for maximized CS and minimized E-energy and E-CO₂. SHapley Additive exPlanations and sensitivity analyses identified SiO₂/Al₂O₃, curing temperature and recycled concrete aggregate as the three most influential parameters contributing to the output prediction of ML models. Experimental validation confirmed a high-level predictive accuracy for the proposed multi-objective optimization framework and ML models, demonstrating the possibility of reaching strengths varying from 30 to 58 MPa, E-CO₂ from 90 to 150 kg CO₂/m3 and E-energy from 300 to 500 MJ/m3 for GPMs prepared with CDW-based binders and aggregates.