Efficient and reliable design of Geosynthetic Reinforced Pile-Supported (GRPS) embankments on soft ground remains challenging due to the complex nonlinear soil–structure interaction and the high computational cost of repeated numerical analyses and optimization procedures. This study presents a hybrid machine-learning-based multi-objective optimization framework for the design and performance enhancement of geosynthetic reinforced pile-supported (GRPS) embankments. A comprehensive database comprising 4, 282 validated 3D numerical simulations was developed using PLAXIS 3D to capture nonlinear interactions between key design and geotechnical parameters. Multiple machine-learning algorithms, including Decision Tree, Random Forest Regression (RFR), Gradient Boosting, XGBoost, and LightGBM, were evaluated. RFR was retained as the most stable overall multi-output surrogate model, with test-set determination coefficients of 0.996 for load efficiency, 0.941 for geosynthetic tension, 0.983 for maximum settlement, 0.973 for differential settlement, and 0.983 for consolidation time. The database is described as a pairwise-informed augmented parametric database rather than a standard balanced orthogonal array. Feature-importance analysis identified embankment height, soft-soil stiffness, pile spacing, pile-cap width, and embankment friction angle as the dominant variables influencing GRPS embankment performance. These variables were used to structure the subsequent optimization stage. The NSGA-II procedure was implemented as a conditional pile-spacing optimization for each fixed combination of soft-soil stiffness, embankment height, pile-cap width, and friction angle, while less influential variables were held constant. The optimization simultaneously minimized the deviation from a target settlement of 5 cm, minimized consolidation time, and maximized load efficiency under serviceability and constructability constraints. The Ideal Point Method was applied using normalized objectives and Euclidean distance to identify the most balanced feasible solution on each Pareto front. Residual diagnostics, convergence assessment, and surrogate-uncertainty sensitivity checks were included to support the interpretation of the optimized design recommendations. The results demonstrate that the proposed ML-NSGA-II framework provides a computationally efficient and physically interpretable approach for identifying practical GRPS embankment configurations with controlled settlement, reduced consolidation time, and enhanced load-transfer efficiency within realistic construction limits.