Semiconductor fabrication facilities often suffer from undesired particle introduction into process chambers in vacuum systems. Ideally, it is unusual to observe particles formed in the exhaust pumping line inside the chamber, but non-volatile compound products at relatively low temperatures jeopardize the vacuum pumping system, gas scrubber and the wafer-in-process. This study proposes a monitoring system for constructing a complete condition-based maintenance system for diagnosing the powder build-up within exhaust pipes used in the semiconductor manufacturing industry. This system includes ultrasonic sensors and machine learning.
Employing ultrasonic sensors, physical and data-driven models are established. The time- or frequency-domain data acquired by the monitoring system are converted into cepstrums for modeling the powder layer thickness using machine learning.
The algorithms used in the proposed system successfully classified the thicknesses with an average accuracy of above 97%, and feature importance analysis identified the quefrency that varied with the thickness of the powder layer.
The limitation of this research lies within the lab environment. It is unfortunate that the suggested method has not been evaluated in actual semiconductor manufacturing facilities, as powder build-up may take more than a few months to be called the facility maintenance. However, the submitted paper is still valid in academic and engineering aspects to be utilized in industry.
We modeled the system using data acquired by an ultrasonic sensor, and we constructed a data-driven model that was trained using cepstral data to replace the physical models that monitor thickness. We are the first to use ultrasound and machine learning to estimate the thickness of powder in the exhaust vacuum pumping line.
