Dependency-aware Interpretable Defect Detection in High-dimensional Semiconductor Production Data

Abstract Semiconductor wafer fabrication relies on multistage process monitoring to maintain yield and identify defective products before downstream processing. Modern fabrication lines generate high-dimensional sensor data from heavy monitoring where machine learning techniques have become prevalent. However, existing approaches commonly either ignore data dependencies or model them with strict assumptions that do not reflect defect data behavior. This problem is compounded by the limited number of defective wafers that are generated, causing class imbalance. This study proposes an end-to-end framework for defective wafer detection from process monitoring data. The framework first applies SHAP-guided feature screening to retain measurements with high predictive relevance, followed by training data undersampling to reduce the effect of imbalance. Principal component analysis is then used to reduce redundancy in the retained feature space. The resulting components are classified using a Defect Conditional Bernstein Bayesian Classifier, which combines class-specific marginal evidence with a sparse set of stable feature dependencies that are more informative for defective wafers. The Bernstein dependence model enables nonlinear and asymmetric feature relationships to be modelled without relying on simplistic assumptions. Experiments on the SECOM and DAIWM benchmark datasets validate our approach, achieving balanced accuracies of 80.16% and 86.65% respectively, outperforming existing approaches with rapid response abilities. The results demonstrate the framework’s efficacy and the model’s compact size enhances scalability for distributed deployment in the industry.