• 목록

Diagnosing Out-of-Spec (OOS) anomalies in anonymized semiconductor manufacturing is highly challenging. We propose a Dual-Model Explainable AI framework for post hoc root cause analysis, featuring a Standard Model for global trends and a Tail-Focused Model—optimized via dynamic weighted tail loss—for extreme OOS distributions. By applying SHapley Additive exPlanations (SHAP), this architecture isolates global main effects and localized cross-process interactions. Human-in-the-loop validation by field engineers demonstrated that the Standard Model effectively verified known factors, whereas the Tail-Focused Model significantly reduced OOS prediction error and discovered novel, actionable interactions, offering a robust diagnostic tool for yield enhancement. Experimentally, the Tail-Focused Model reduced OOS RMSE from 0.261 to 0.181 and uncovered 56 previously unknown cross-process interactions validated as physically actionable by domain engineers.