Fresh-State Properties and Machine-Learning Prediction of Sulfoaluminate Cement Grout for Water-Rich Tunnels

This study investigated sulfoaluminate cement grout intended for water-rich tunnel applications and developed machine-learning models for predicting its fresh-state properties. The experimental programme included a 31-mixture base series varying the water-to-binder ratio, Class F fly ash, steel slag, and polycarboxylate ether (PCE) superplasticizer dosage; a hydroxyethyl cellulose (HEC)-modified series with a fixed binder composition; and a static-water turbidity series used to evaluate particle dispersion resistance. Flowability, setting time, compressive strength, and turbidity were measured. The results showed that the water-to-binder ratio, fly ash, steel slag, PCE, and HEC affected the balance between workability, setting behaviour, strength development, and static-water dispersion resistance. Increasing the HEC dosage from 0.004 to 0.006 reduced turbidity by 84.1–94.7% under static-water conditions, whereas an HEC dosage of 0.007 reduced the 7 and 28 d compressive strengths by 51.5% and 40.7%, respectively. Machine-learning models trained on the 31-mixture HEC-free dataset provided exploratory predictions of flowability and initial and final setting times. Extra Trees, support vector regression, and Extra Trees gave the lowest nested leave-one-out cross-validation (LOOCV) errors for flowability, initial setting time, and final setting time, with R2 values of 0.819, 0.743, and 0.767, respectively. Shapley Additive Explanations (SHAP) analysis indicated that the water-to-binder ratio dominated flowability prediction, whereas fly ash contributed strongly to setting-time prediction. The results provide complementary experimental and data-driven evidence for comparing sulfoaluminate cement grout formulations within the investigated composition domain.