Abstract
A decoupled edge cyber–physical system architecture was developed in this study for semiconductor manufacturing through chemical vapor deposition. The cyber layer of this system integrates a declarative automated machine learning pipeline that executes data preprocessing, feature selection, and model training, enabling domain experts without extensive programming expertise to control the product life cycle. To preserve bandwidth and support closed-loop cloud–edge collaboration, localized edge nodes perform real-time inference and transmit only filtered feature windows or operational events to the cyber layer for offline model recalibration. To ensure system generalizability across diverse equipment, a dynamic output mapping mechanism separates presentation logic from inference models. Compiled predictive endpoints are deployed onto localized edge nodes by using Docker and a Fast application programming interface, and a fault-tolerant dual-write mechanism is employed for database synchronization. System-level benchmarking indicates that the proposed system can achieve millisecond-level inference latency and a sub-minute context switching overhead, confirming that its decoupled architecture eliminates requirements for manual codebase modifications or service restarts while maintaining strict data sovereignty through onsite deployment. The results of this study validate the feasibility, computational scalability, and reliability of the proposed system; its fab-level yield will be investigated in a future study.