Improved Yield Prediction and Failure Analysis in Semiconductor Manufacturing with XGBoost and Shapley Additive exPlanations Models

In the domain of semiconductor manufacturing, the accurate prediction of wafer yield and subsequent failure analysis (FA) are of critical importance with respect to the assurance of product quality and the optimization of production processes. However, traditional methods frequently exhibit a high dependency on engineering expertise, modest accuracy, and an inability to adapt swiftly to manufacturing environment changes. To address these issues, this study introduces an advanced machine learning (ML) model that combines Particle Swarm Optimization (PSO) with XGBoost. This innovative approach is adept at capturing the complex non-linear relationships between wafer acceptance test (WAT) data and wafer yield. The integration of PSO enables effective WAT feature selection while fine-tuning the hyperparameters of the XGBoost, thereby resulting in a significant improvement in model performance. The proposed PSO-XGBoost model exhibits remarkable accuracy (96.4%), AUC-score (91.4%), F1-measure (97.9%), precision (97.2%), and recall (98.6%). Moreover, the integration of the Shapley Additive exPlanations (SHAP) method facilitates the interpretability of the model, elucidating the critical WAT parameters that influence yield on both global and local scales. This study not only provides a cost-effective methodology that reduces reliance on human expertise but also paves the way for the expanded role of artificial intelligence (AI) in enhancing semiconductor manufacturing processes. This represents a new era in the field of smart manufacturing and quality assurance.