For the "dual carbon" (i.e., carbon peak and carbon neutrality) goals and global sustainable development strategies, the construction materials sector rapidly transitions from conventional cement-based materials toward low-carbon, environmentally friendly, and high-performance alternatives. Engineered geopolymer composites (EGC) as a new generation of green building materials exhibit superior mechanical properties, thermal stability, and environmental compatibility. However, the performance of EGC is synergistically governed by multiple factors (i.e., alkali activator concentration, liquid-to-solid ratio, aggregate characteristics, and fiber geometry). The intricate, nonlinear interactions among these parameters present significant challenges for precise mix design and performance optimization. Conventional trial-and-error approaches are prohibitively expensive and time-consuming, rendering them fundamentally incapable of effectively exploring the expansive combinatorial space of EGC mix formulations. Furthermore, EGC mix design often requires the simultaneous satisfaction of multiple, potentially conflicting performance targets, such as high compressive strength and high tensile strain capacity. Therefore, this study was to develop an intelligent design framework for EGC integrating performance prediction and multi-objective optimization. This framework directly addressed the limitations of conventional "black-box" and "experience-driven" approaches, enabling the intelligent and precise design of sustainable cementitious materials.

This study established a closed-loop intelligent design framework for EGC based on machine learning (ML), incorporating "prediction-optimization-feedback". Firstly, a comprehensive EGC database was systematically integrated, comprising 309 sets of experimental data from domestic and international literature. Eleven key mix proportion parameters (i.e., ground granulated blast furnace slag (GGBS), fly ash (FA), silica fume (SF), sand-to-binder ratio (S/B), water-to-binder ratio (W/B), alkali activator modulus (AAM), alkali equivalent (AAD), fiber content (FC), fiber length (FL), fiber length-to-diameter ratio (FLD), and fiber tensile strength (FTS)) were selected as input variables. The 28-d compressive strength, tensile strength, and tensile strain capacity were chosen as target output indicators. Five ML models (i.e., Support Vector Regression (SVR), Gradient Boosting Regression Tree (GBRT), XGBoost, Artificial Neural Network (ANN), and Invertible Neural Network (INN)) were developed to predict the key mechanical properties of EGC from its mix design parameters. Hyperparameter optimization was conducted using grid search and experience-based tuning. The SHapley Additive exPlanations (SHAP) method was employed to quantitatively analyze the influence of each mix parameter on the EGC performance. Finally, the CatBoost model was integrated with the NSGA-III (Non-dominated Sorting Genetic Algorithm III) multi-objective evolutionary algorithm to achieve targeted performance optimization, while automatically searching for optimal balanced solutions under various engineering scenarios.

Among the five ML models evaluated, the INN model exhibits the maximum predictive accuracy, achieving R² values of 0.99, 0.98, and 0.99 for compressive strength, tensile strength, and tensile strain, respectively. The results of SHAP-based global sensitivity analysis reveal that increasing GGBS content significantly enhances compressive strength via promoting secondary hydration reactions. The fiber tensile strength (FTS) is the most influential parameter for tensile strength, and the fiber length (FL) critically determines the tensile strain capacity via improving crack-bridging effects. Alkali equivalent (AAD) and alkali activator modulus (AAM) both affect all three performance indices. Three optimization objectives are implemented by the integration of CatBoost and NSGA-III, i.e., 1) simultaneous maximization of all three target performances, 2) targeted optimization focusing on maximum tensile strain, and 3) precise matching of predefined target performance values. From the set of Pareto-optimal solutions, the optimal mix proportion is definitively selected using the TOPSIS ranking method for subsequent experimental validation. The validation results show that the errors between predicted and measured values for the three selected mix proportions are all within 10%, with the minimum error being only 5.3%, confirming the reliability and engineering accuracy of the proposed intelligent design framework.

This work could establish an intelligent design framework for EGC that integrates INN prediction, SHAP interpretability analysis, and CatBoost-NSGA-III multi-objective optimization, achieving a full-process closed-loop design from performance prediction to targeted optimization. The major findings were 1) The INN model demonstrated a superior predictive accuracy for the three key performance indicators of EGC, providing a robust foundation for forward performance prediction; 2) SHAP analysis quantitatively elucidated the mechanisms of key mix parameters on EGC performance, offering a theoretical support for precise material design; and 3) The CatBoost and NSGA-III integration effectively enablec\d both multi-objective collaborative optimization and single-objective precise targeting, validating the feasibility of the reverse design from desired performance targets to optimized mix proportions through experimental verification. This intelligent design strategy effectively resolved the conventional strength-toughness conflict, thereby transcending the limitations of "black-box" and experience-driven methods and propelling a field toward a paradigm of intelligent, data-driven design for next-generation geopolymer composites.