Quantum Kernel-Aligned Regression for Data-Efficient Predictive Modeling in Semiconductor Manufacturing

Semiconductor manufacturing analytics face a persistent imbalance between abundant process telemetry and scarce, high-cost metrology labels, limiting the practical effectiveness of predictive models during product ramps, recipe adjustments, and post-maintenance requalification. This study proposes Quantum Kernel-Aligned Regression (QKAR), a deployment-oriented pipeline that combines manufacturing-aware feature engineering, quantum-kernel similarity estimation, and supervised kernel alignment to improve regression performance under constrained labels. Across time-respecting evaluations, QKAR demonstrates consistent data-efficiency gains over strong classical and modern baselines, including classical kernel ridge regression with an RBF kernel, gradient-boosted trees, and tabular deep models. Learning-curve results show that QKAR achieves RMSE 0.603 at 80 labeled wafers/lots, outperforming the best baseline (RBF-KRR) at RMSE 0.652 under the same budget, with simultaneous improvements in MAE (0.468 vs 0.507) and R² (0.636 vs 0.592). At higher budgets, QKAR maintains its advantage, reaching RMSE 0.468 and R² 0.781 at 300 labels, compared with RMSE 0.503 and R² 0.742 for the best baseline. Drift-oriented evaluation across production weeks indicates smaller post-maintenance degradation and faster recovery, with error increases after maintenance events remaining more contained for QKAR than for classical kernels. Operational utility is further validated through metrology prioritization experiments: an uncertainty-plus-diversity policy guided by QKAR achieves a high-risk capture rate of 0.66 for top excursion cases, versus 0.41 under random sampling, while reducing the labeling requirement to reach RMSE ≤ 0.55 by 46 wafers at fixed weekly budget. These results indicate that aligning quantum kernel geometry to label structure can materially improve early generalization and robustness in non-stationary manufacturing environments, enabling faster decision readiness with fewer metrology actions.