Water flow in underwater grouting causes severe slurry dispersion, compromising cement-based material stability. Traditional empirical mix design cannot simultaneously optimize fluidity, anti-dispersion performance, and mechanical strength. To solve this, we propose a multi-scale framework combining molecular dynamics (MD) simulation and machine learning (ML) to link microstructure to macroscopic properties in a Polyacrylamide-Graphene Oxide-U-type Expansive Agent (PAM-GO-UEA) -modified ultrafine cement system. From MD, we extract interfacial interaction energy, radial distribution functions, and ion/molecule diffusion coefficients to quantify interfacial bonding, structural density, and hydration kinetics in the Calcium Silicate Hydrate(C-S-H)–polymer–mineral system. Using experimental fluidity, turbidity, and compressive strength data, we train an XGBoost–SHAP model to identify key microstructural drivers. Five-fold cross-validation gives R 2 values of 0.939, 0.941, and 0.979 for fluidity, turbidity, and strength—improving on response surface modeling by 13.1%, 13.6%, and 48.1%. SHAP analysis shows E int dominates fluidity (0.14) and strength (0.165), while D (N) most strongly influences turbidity (0.17). • Multi-scale MD–ML framework for the design of underwater anti-dispersion grouts is developed. • MD descriptors ( E int , RDF, MSD, D ) provide quantitative microstructural indicators. • XGBoost–SHAP achieves high-accuracy prediction from small MD–experiment datasets. • SHAP reveals dominant microstructural factors governing flowability and strength.
Zhao et al. (Sun,) studied this question.