Internal defect depths prediction technique by combining light-scattering imaging and machine learning

Fine polishing is crucial for achieving nano-scale flatness in high-performance wafers. With increasing demand for high-precision processes, the emergence of latent flaws during polishing has become a significant challenge. Detecting latent flaws is crucial; however, acquiring precise depth information regarding these defects is equally essential to determine the appropriate removal allowance. This study focuses on developing a technique for estimating the depth of latent flaws in glass substrates from a single light-scattering image by integrating optical observations with either filter processing or deep learning techniques. Defects were introduced into a glass substrate through laser processing and then observed using the light-scattering method at various focal points. The data from these observations were used to develop two defect depth prediction models based on different theories. The models based on filtering techniques lack versatility; however, they offer the advantage of being straightforward to interpret and explain. In contrast, models based on deep learning exhibit high versatility and hold the potential for improved accuracy as additional data is acquired. Our approach enables highly sensitive and efficient analysis of latent flaw depths, contributing to improving precision and reliability in the semiconductor manufacturing industry.