• 목록
• 게시판 리스트 옵션
  • 검색

[Abstract]

Defect monitoring in semiconductor manufacturing is constrained by sparse inspection and the inherent process-to-inspection interval, during which many processed wafers remain unverified. This creates a quality blind spot because conventional prioritization relies mainly on information available after inspection. This study proposes a three-step framework that uses uninspected wafer volume as an equipment-level exposure indicator, enabling proactive prioritization from the time of production. The framework first predicts future wafer inspection volume from equipment throughput and operational sampling information. It then estimates future defect levels and prediction intervals using a horizon-conditioned, Tweedie-based forecasting model. Finally, Monte Carlo scenarios and the volume-adjusted risk score (VARS) are used to derive equipment priorities, while the risk prioritization change score (RPCS) measures deviations from the historical baseline. Using industrial data from eight inspection steps, the CatBoost-based inspection-volume model achieved a coefficient of determination of 0.816. The selected Tweedie-based XGBoost model achieved an average prediction interval coverage probability (PICP) of 89.8% and a prediction interval normalized average width (PINAW) of 3.31. In the final prioritization evaluation, VARS outperformed the mean-based baseline in 12 of 19 defect trends, reduced the mean ranking RMSE from 2.56 to 1.57, and improved the overall mean win rate from 57.64% to 76.88%. These results show that throughput-informed reprioritization can identify operationally relevant suspected equipment before inspection feedback becomes available, supporting earlier engineering action and helping to mitigate latent quality risk under sparse and irregular defect-monitoring conditions.