SHAP Analysis for Diagnosing Anomalies in Semiconductor Manufacturing

We consider the problem of predicting the quality of semiconductor devices and, in case of low quality, diagnosing the anomaly occurred during production. A multi-branch neural network is developed for quality prediction based on multimodal data. Specifically, a dedicated autoencoder is trained for each data modality; then, the latent representations provided by the encoders are concatenated and a regression layer is added for quality prediction. Shapley Additive exPlanation (SHAP) is used to quantify the contribution of each data modality to the quality outcome. Since different data modalities contain information about different production stages, the causes of the production anomaly can be identified. The developed method is demonstrated using a synthetic case study, which mimics the complexity of semiconductor manufacturing. Wafer map (images) and signal measurements (time series) from a production machine are the two considered data modalities. The method is shown able to effectively predict the quality of semiconductor devices and diagnose anomalies occurred at different stages of production.

Keywords: Semiconductor manufacturing, Quality, Multimodality data, Multi-branch neural network, SHAP.