The semiconductor industry has played a crucial role in societal development over the past several decades. Plasma etching is a key processing step employed in Integrated Circuit (IC) fabrication. In order to improve product yield, Optical Emission Spectroscopy (OES) is widely used to monitor the etching process. OES generates
high-dimensional data, which has a large information capacity but also has significant information redundancy. Based on plasma OES characteristics, two novel data analysis methods are proposed in this thesis: the Internal Information Redundancy Reduction (IIRR) method for dimension and redundancy reduction and Similarity Ratio Analysis (SRA) for fault detection. By identifying peak wavelength emissions and the correlative relationships between them, IIRR outputs a subset of the original variables. Data dimensionality is reduced significantly by IIRR with minimal information loss.
The SRA method is intended for early-stage faultdetection in plasma etching
processes using real-time OES data as input. The SRA method can help to realise a
highly precise control system by detecting abnormal etch-rate faults in real-time during
an etching process, so less energy and materials will be wasted by faulty processing.
Generally, previous research on OES measurements of plasma etching has largely
focused on particular target applications and has used methods that rely on
transforming the original data into an abstract variable space. In contrast, our approach
operates directly in the original variable space allowing a more direct and easier
interpretation of the dimension reduced data.