A Novel Multi-Variate Anomaly Detection Method For Chemical Vapor Deposition

Current method of anomaly detection is typically univariate SPC control which false alarm increases proportionally with number of parameters (high dimensional). New anomaly detection methods (ABOD and LOF) were applied to address high-dimensional issues in chemical vapor deposition (CVD) high density plasma oxide (HDPO) for 20 days, but the initial false alarm rate was still 65.6%. Further improvement was achieved in reducing false alarms by applying supervised machine learning approach, which is random forest, on the same tool set, and tested the model against data obtained. The false alarm rate was further down 8.7%, representing a 7.5 times reduction in false alarm rate based on the supervised machine learning model and annotation by module engineering to identify genuine outlier cases.