Oisin Boyle, Joseph Brindley, Patricia Killen, Benoit Daniel, Gencoa Ltd, Liverpool, United Kingdom
Vacuum processes are highly complex processes whose performance can be difficult to predict by analytical modelling alone. Residual gas analysers such as remote plasma optical emission spectroscopy or quadrupole mass analysers can be used to monitor vacuum processes but often need expert analysis of the data produced to determine the process performance. Furthermore, vacuum processes are being equipment with a greater number of sensors, producing “operator overload” in terms of being able to use the data produced effectively. As such, the success of the process is often only known upon completion. This can be costly both in terms of time and money when production failures are not identified early.
This work presents a method of using an artificial intelligence algorithm that learns from sensor data to identify the condition of a vacuum process in real-time. Whilst one can use simplistic classifications like off-the-shelf SVMs to categorise process data as an indicator of a successful/unsuccessful process, these methods fail to capture the complexity of the process. It is necessary therefore combine this with a deep learning approach to better create a model representing vacuum processes.
This presentation will demonstrate a tool developed using deep learning methods to identify the state of the process. Examples are shown for with monitoring of vacuum quality, magnetron power condition and a sputter coating process deposition quality.