A Masked Autoencoder -Based Approach for Defect Classification in Semiconductor Manufacturing

In semiconductor manufacturing, automatic defect classification is of paramount importance. Even the slightest defects can compromise chip performance or lead to complete failure, subsequently impacting chip yield rates. Currently, defect classification still heavily relies on manual processes, often leading to a significant number of misclassifications. In this paper, we propose a method based on MAE (Masked Autoencoder) for automatic defect classification in chip manufacturing. The core concept of MAE involves applying a high-proportion random mask to images, creating a challenging image reconstruction task. Using the unmasked image patches, the model predicts the masked patches for self-supervised pretraining. When applied to downstream tasks, this methodology enhances the model's generalization and feature representation capabilities. In a task-agnostic way, we conduct self-supervised pretraining on a large number of SEM (Scanning Electron Microscope) images without the necessity of any labels. In a task-specific way, we fine-tune the network using a limited amount of highly reliable labels. Experimental results suggest that our method is capable of accurately classifying defects with minimal labeled data, greatly reducing labor costs.