<p>Ultrasonic-assisted milling is recognized as an effective method for mitigating tool wear when machining difficult-to-cut materials such as austenitic cast steel. Nevertheless, high-precision prediction of tool wear is hindered by nonlinear behavior and uncertainties arising from multi-parameter coupling. To this end, an Improved Particle Swarm Optimization–based Least Squares Support Vector Regression ensemble (IPSO-LSSVR-AdaBoost) model was proposed for tool wear prediction. By integrating multiple models, accurate prediction of ceramic tool flank wear under small-sample conditions was achieved. In the present study, flank wear data were collected from 32 longitudinal–torsional ultrasonic-assisted milling (LTUAM) experiments, and spindle speed, vibration amplitude, and cutting depth were identified as the dominant influencing factors through Pearson correlation analysis and Lasso regression. An asymmetric learning factor (ALF) was incorporated into the PSO algorithm to enhance its optimization capability, and the algorithm was subsequently embedded within the AdaBoost framework to further improve prediction accuracy and generalization. The experimental results indicated that the RMSE of the proposed IPSO-LSSVR-AdaBoost model was reduced by 50.1% and 29.7% compared with LSSVR and PSO-SVM on the test set. MAE was decreased by 65.2% and 42.1%, while R² was increased by 9.5% and 5.5%. The effectiveness and superiority of the proposed model in ceramic tool wear prediction were thereby fully verified.</p>