Learning to tune: AI-enabled virtual commissioning for CNC interpolator parameter optimization

CNC interpolator commissioning still relies heavily on trial-and-error manual tuning, limiting repeatability and efficiency. This study presents an AI-enabled virtual commissioning framework that integrates a back-propagation neural network (BPNN) surrogate with CrossQ reinforcement learning for interpolator parameter optimization. Trained on 729 full-factorial design-of-experiments samples, the surrogate is evaluated using ten-fold out-of-fold validation and 600 independent off-grid samples, enabling continuous parameter search without repeated on-machine trials. Pearson correlation and SHAP analyses identify influential interpolator parameters. A range-normalized reward balances cycle time against soft penalties for contouring error and jerk. Sensitivity and ablation analyses across ten random seeds demonstrate that stronger jerk penalization reduces jerk violations with less than a 0.5% increase in cycle time. Compared with BP-GA and BP-TD3, BP-CrossQ is the only surrogate-assisted method satisfying both contouring accuracy and jerk targets on the KAKINO benchmark under semi-finish and finishing. A frozen, condition-specific policy attains 100% surrogate-predicted feasibility from 50 initial settings per condition without retraining, showing consistent reusable parameter refinement. Transferring the optimized settings to a FANUC test mold on the same machine reduces maximum contouring error by 30.97% and 12.61% and cycle time by 1.20% and 1.43%, respectively, relative to manual tuning. Total commissioning time decreases from 24 h to about 11 h, supporting efficient, reusable virtual commissioning and cross-trajectory performance improvement within the evaluated machine and conditions.