Abstract: During the broaching of tenon grooves in GH4065A high-temperature alloy turbine disks for aero-engines, challenges such as large cutting forces, high susceptibility to machining deformation, and the difficulty in simultaneously achieving high material removal rate and low cutting force in high-speed broaching are encountered. When optimizing and predicting broaching process parameters, problems including small sample size, multi-objective requirements, and strong nonlinearity need to be addressed. Based on finite element simulation and actual broaching experiments, this study collected data (e.g., broaching force, maximum equivalent stress, and material removal rate) during the broaching of GH4065A high-temperature alloy under different broach dimensions and broaching parameters, and constructed an integrated artificial intelligence method named "SMOGN-XGBoost-SHAP-NSGA-II". Firstly, 29 groups of original broaching data were obtained via finite element simulation, and the SMOGN algorithm was applied to expand the sample size to 145 groups, which significantly alleviated the overfitting risk caused by sparse data; subsequently, an XGBoost model with hyperparameters tuned by Bayesian optimization was used to establish high-precision surrogate models for cutting force, cutting temperature, equivalent stress, and material removal rate, with the coefficient of determination (R²) of all models on the test set exceeding 0.90. Through SHAP (Shapley additive explanations) interpretation, broaching linear speed was identified as the primary influencing factor for each response, and the nonlinear coupling effects between parameters were quantified. Finally, the NSGA-II algorithm was employed for three-objective optimization, which achieved convergence within 150 generations, obtained a uniformly distributed Pareto front, and provided engineering-feasible compromise solutions, while comparisons were conducted between real experimental results and model prediction results. Verification experiments show that the error between predicted and simulated measured values under the recommended parameters is less than 5%, and the error between real experimental results and model predicted values is also less than 5%, which realizes the closed-loop optimization of "high precision-interpretability-multi-objective" under small sample conditions and provides a portable decision-making framework for efficient and precise broaching of difficult-to-machine materials.