Real-Time End Point Detection via Automated Dynamic Mode Decomposition

In semiconductor manufacturing, an integral process is plasma etching, where endpoint detection (EPD), the moment the film has cleared to the underlying layer, is crucial. Optical emission spectrum (OES) is traditionally used for EPD due to its non-invasive nature and sensitivity to chemical changes in the plasma chamber, with plasma species’ emission lines undergoing significant changes at the endpoint. As etching processes progress, monitoring a single emission line becomes insufficient due to limited exposed areas, leading to low signal-to-noise ratios. Manual observation risks human error, jeopardizing wafer integrity and EPD sensitivity [1]. Therefore, automated real-time EPD algorithms are essential for timely operation cessation. Traditional methods like linear regression, PCA, and Spectral Clustering applied to OES data are slow, impeding real-time control.