In semiconductor manufacturing, Chemical Vapor Deposition (CVD) is a crucial process for wafer production, which requires extensive machine qualification testing before production. Traditionally, engineers must fine-tune five adjustable process parameters to meet three critical output specifications—film thickness, boron concentration, and phosphorus concentration—within strict tolerance limits for quality control after a CVD equipment has undergone preventive maintenance (PM). This qualification process will rely on engineer experience and trial-and-error adjustments, resulting in low success rates, repeated testing cycles, and reduced utilization efficiency. This study proposes an integrated advanced process control (APC) system that combines predictive modeling with multi-objective optimization to overcome these challenges. In the prediction phase, we developed a novel Siamese-LGBM model, which merges the shared-weight architecture of Siamese Networks with LightGBM predictors. This design addresses the limitations of historical data, mitigates overfitting in small-sample scenarios, and reduces computation time by 20% compared to traditional Siamese Network architectures. In the optimization phase, we implemented a NSGA-II-based algorithm that employs a comprehensive feasible-solution scoring mechanism that integrates five weighted evaluation metrics to assess the quality of feasible solutions. Besides, a fast-convergence mechanism shortens computation time from over 20 minutes to approximately 1 minute by employing early-termination strategies once high-quality solutions are detected. Additionally, a fixed-parameter search mode allows flexible operational constraints according to the engineer's requirements. Implementation results demonstrate substantial improvement in qualification success rates and a marked reduction in testing iterations and equipment downtime.

Keywords

Semiconductor Manufacturing, Advanced Process Control, Siamese Neural Network, NSGA-II Algorithm, Multi-objective Optimization