Data-driven multi-objective mix design optimization of cementitious underwater repair grout incorporating MICP-related components

Developing cementitious underwater repair grout (CURG) with balanced anti-dispersion performance, workability, and mechanical properties remains challenging because of the strong coupling among polymer stabilization, cement hydration, biological components, and curing conditions. In this study, a CURG incorporating hydroxypropyl methylcellulose (HPMC) and microbially induced carbonate precipitation (MICP)-related components was developed, and a data-driven multi-objective mix-design framework explicitly incorporating MICP-related biological and chemical components was established for CURG. HPMC was used to provide baseline anti-dispersion resistance, with 0.5% selected as the minimum dosage based on preliminary Plunge tests. Diatomite-immobilized bacteria, urea, calcium chloride, and yeast extract were introduced as MICP-related components, and their effects on anti-dispersion performance, fluidity, compressive strength, and splitting tensile strength were experimentally investigated. The experimental results were further used to construct machine learning datasets. Six regression models were trained to predict compressive strength, splitting tensile strength, and fluidity, and Bayesian optimization (BO) was adopted for hyperparameter tuning. Among the models, BO-optimized XGBoost showed the best prediction performance, with testing-set R 2 values greater than 0.96 for all three target properties. Shapley additive explanations (SHAP) analysis indicated that curing age and HPMC were important contributors to mechanical-property predictions, whereas HPMC and the water-cement ratio showed the highest contributions to fluidity prediction. Finally, the BO-XGBoost models were coupled with NSGA-II to optimize CURG mix proportions. Experimental validation of two representative Pareto solutions showed that the relative errors of all performance indicators were within 9.5%, supporting the predictive accuracy and practical feasibility of the proposed BO-XGBoost-NSGA-II framework within the investigated design range.