The urgent demand for sustainable concrete has intensified as greenhouse gas emissions increasingly threaten global environmental stability. High-performance geopolymer concrete (HPGC) presents a promising sustainable alternative to traditional cement-based materials. The mix design of HPGC considering compressive strength, cost, and carbon emissions is crucial while hard to implement using traditional non-destructive methods. To resolve the issue, this study proposes an integrated framework combining automated machine learning (AutoML) modeling and multi-objective optimization (MOO) to balance compressive strength, cost, and carbon emissions in HPGC mix-design. Our research mainly includes three innovations: (1) We integrate ML-based predictive modeling with MOO framework for HPGC; (2) we establish an AutoML-Shapley Additive Explanations (SHAP) framework that harmonizes predictive accuracy with interpretability; and (3) we introduce the Pareto non-dominated sorting into MOO for HPGC. First, we employ an advanced AutoML algorithm to automatically develop a robust predictive model for HPGC compressive strength. Based on a database containing 295 mixes, the AutoML model demonstrated comparable accuracy compared to conventional ensemble learning methods, achieving a validation dataset determination coefficient (R2) of 0.9280, root mean squared error (RMSE) of 5.2954 MPa, mean absolute error (MAE) of 4.2307 MPa, mean absolute percentage error (MAPE) of 0.0724, and a20 index of 0.9677. Subsequently, the SHAP method is applied to identify critical factors influencing HPGC performance and enhance the interpretability of the AutoML model. Finally, a Pareto non-dominated sorting algorithm is integrated into MOO to generate solutions that minimize unit cost and carbon emissions while maintaining compressive strength. The optimization framework reduces CO₂ emissions by 23–60% and unit costs by 16–36%, confirming the method's efficacy in balancing multiple objectives. This research advances eco-efficient concrete design methodologies and supports the broader adoption of green building technologies. It should be highlighted that the trained ML models are based on limited data, so the application of the models should be restricted to a certain range. The collected database will be expanded, which can resolve the ML limitations such as model generalizability.
Wu et al. (Fri,) studied this question.