Machine Learning Prediction Model for Automation Equipment Control in Semiconductor Manufacturing

The conventional control systems in semiconductor manufacturing today typically depend on humans to monitor the performance of their equipment. Not only is this method performed manually, but the way of working is a passive approach where investigation analysis is initiated when failure has occurred. This method, on the other hand, has the potential to produce bad electronic-chip and escapee to consumers in general. Today, Industry 4.0 and automotive applications have set higher standards for requirements toward the semiconductor manufacturing industry. One way of solving this is by implementing advanced automation control with machine learning capability to enable the system to predict equipment failures. There are various machine learning techniques, namely Decision Tree (DT), Random Forest (RF), Super Vector Machine (SVM), K-Nearest Neighbors (KNN), Logistic Regression (LR), and Naïve Bayes (NB), that will be evaluated to develop the prediction model. The results indicate the Decision Tree (DT) algorithm has the best performance with a higher prediction capability of 93% as compared to the other classifiers. Therefore, with the known prediction accuracy of the model, the system is equipped with the capability of forecasting expected failures where affected equipment should be down for countermeasure. The scope of this research will be focused on back-end semiconductor manufacturing, specifically the wire bond process. Here the most common equipment errors, such as First Bond Non-Stick and Second Bond Non-Stick, are considered as independent variables while the dependent variable is based on the quality of the electronic-chip produced. However, this initial study can be further expanded and implemented to include all other equipment errors and processes as well.