Ensemble Learning for Wafer Classification with Minimum Pre-Labelled Data

Abstract

Data quality during training is crucial for building accurate and reliable machine learning and deep learning models. Obtaining annotations relies heavily on the expertise of domain professionals. Prioritizing manual data labelling is crucial for capturing expert knowledge in machine learning, despite its laborious and time-intensive nature in supervised learning. However, Manual annotation is susceptible to distractions, leading to errors and mislabelling and posing significant disadvantages for data labelling processes in supervised learning. This paper elaborates on the pattern recognition of wafer defect map and the ultimate goal of the paper is to train the convolutional neural network. This deep learning model exhibits confidence in its predictions while requiring minimal pre-labelled data for accurately classifying wafer patterns. It is also capable of pseudo-labelling the wafer maps. For this process, deep model is used to perform the proposed method which involves a series of CNN layers. The paper presents hybrid ensembling as a distinctive method to improve wafer classification efficiency. This approach aims to minimize the need for manual annotation compared to existing methodologies.