A Machine Learning Model for Interpretable PECVD Deposition Rate Prediction

Plasma‐enhanced chemical vapor deposition (PECVD) is a key process in manufacturing thin films for various semiconductor applications. Accurate and interpretable predictions of film properties are crucial for optimizing manufacturing processes and quality control. In this study, we develop six machine learning (ML) models: k‐nearest neighbors (KNN), support vector regression (SVR), decision tree, random forest, CatBoost, and backpropagation neural network (BPNN), to predict the SiN x deposition rate during PECVD based on a hybrid dataset of production and simulation data. Bayesian optimization is used to fine‐tune hyperparameters, and the best model is selected based on multiple evaluation metrics. The SHapley Additive exPlanations (SHAP) method analyzes global feature importance, individual feature contributions, and interaction between process parameters on the outcome. The results show that the CatBoost model outperforms other ML models, with high accuracy on both simulation data (R 2 = 0.990) and production data (R 2 = 0.934). SHAP analysis identifies the gas flow rate as the most influential factor and reveals how interactions between process parameters affect deposition rate. By attributing contributions to individual samples, the model provides insights into the underlying decision‐making process of its predictions. We propose an interpretable, ML‐based framework for advanced semiconductor manufacturing, enabling data‐driven and efficient process optimization.