In this study, a hybrid data-driven optimization framework is proposed to simultaneously minimize surface roughness (Ra) and maximize material removal rate (MRR) in the Powder-Mixed Electrical Discharge Machining (PMEDM) of hardened 90CrSi tool steel. A series of experiments was conducted based on a Taguchi L18 orthogonal array using SiC nanopowder suspended in a dielectric fluid. Gaussian Process Regression (GPR) models were trained on the experimental data to serve as surrogate models for Ra and MRR prediction. These surrogate models were then integrated into a Non-dominated Sorting Genetic Algorithm II (NSGA-II) to obtain a Pareto-optimal front of solutions that balance surface finish and machining efficiency. The results demonstrated the effectiveness of the proposed GPR-assisted NSGA-II framework in capturing the trade-offs between Ra and MRR, while significantly reducing computational cost compared to traditional evolutionary algorithms. The best compromise solution was further validated using Analytic Hierarchy Process (AHP), providing a practical decision-making approach for selecting optimal process parameters in PMEDM applications.