<p>In the high-pressure die casting (HPDC) process, strong nonlinear couplings exist among process parameters, posing significant challenges for parameter optimization. Although traditional data-driven approaches exhibit certain predictive capabilities, their lack of interpretability and engineering feasibility limits their practical application in manufacturing environments. To address these limitations, this paper proposes an integrated framework that combines surrogate modeling, SHapley Additive exPlanations (SHAP)-based interpretability analysis, and Non-dominated Sorting Genetic Algorithm II (NSGA-II) multi-objective optimization for intelligent and transparent control of the HPDC process. First, high-precision surrogate models for product quality, dosing weight, and cycle time are constructed using the eXtreme Gradient Boosting (XGBoost) algorithm. The SHAP method is then employed to analyze feature importance and reveal the influence mechanisms of key process parameters on multiple performance objectives. Subsequently, a segmented surrogate modeling structure is developed based on the critical feature of “dosing weight” incorporating a kernel density estimation (KDE)-based and principal component interpolation strategy to ensure smooth transitions between segments. A dynamic feasible solution space is further constructed by introducing a quality confidence threshold, enhancing the engineering adaptability and stability of the optimization results. On this basis, a staged NSGA-II optimization strategy is proposed to improve the handling of dependencies among objectives. It should be noted that due to the lack of direct measurements of the net casting weight, this study does not explicitly calculate the material utilization rate (MUR). Instead, it approaches MUR from an optimization perspective, exploring the potential to reasonably reduce the dosing weight while ensuring acceptable product quality. Experimental results demonstrate that the proposed method improves product yield by 0.23, reduces dosing weight by 0.63 kg, and shortens the cycle time by 3.56 seconds, thereby verifying its high adaptability, interpretability, and optimization performance in intelligent die casting applications</p>