Optimizing XPS Vacuum Data Drift Detection: A Comparative Study of CUSUM, Kernel CUSUM, and LSTM

Su-Hyoung Ha, Jae Wook Jeon

Abstract: Maintaining an ultra-high vacuum is critical for the stable operation of X-ray sources in X-ray Photoelectron Spectroscopy (XPS) within semiconductor manufacturing. However, challenges in sustaining such vacuums often lead to equipment failures, particularly when vacuum levels drop below 4×10−8 Torr, causing damage to the E-Gun source and extended maintenance periods. Traditional methods like Fault Detection and Classification (FDC) and Time-Based Preventive Maintenance (TBM) struggle to predict these failures accurately. This paper proposes a Kernel CUSUM method for detecting subtle vacuum data drifts, enabling early identification of potential issues before critical damage occurs. By implementing this approach, we enhance maintenance strategies, reduce unplanned downtime, and achieve significant cost savings in maintenance and repairs. Additionally, we compare the general CUSUM, a statistical method, with Kernel CUSUM, which incorporates machine learning techniques, and LSTM, a deep learning method, analyzing the advantages and disadvantages of each. Our findings demonstrate the financial and operational benefits of adopting advanced data analysis techniques for equipment maintenance in semiconductor manufacturing.