MM-WAE: Multimodal Wasserstein Autoencoders for Semi-Supervised Wafer Map Defect Recognition

Area of Science:

• Semiconductor Manufacturing
• Artificial Intelligence
• Computer Vision

Background:

• Wafer map defect pattern recognition is crucial for integrated circuit manufacturing yield.
• Existing deep learning methods struggle with scarce data, imbalanced classes, and limited feature representation, impacting minority defect class performance.

Purpose of the Study:

• To develop a robust semi-supervised classification method for wafer maps.
• To address challenges of data scarcity, class imbalance, and complex defect morphologies in defect pattern recognition.

Main Methods:

• Proposed a multimodal Wasserstein autoencoder (MM-WAE) with parallel spatial, frequency, and texture feature branches.
• Employed multi-head attention and gating for adaptive multimodal fusion, characterizing defects comprehensively.
• Utilized Wasserstein autoencoder with MMD loss for latent space regularization and incorporated inverse class-frequency weighted cross-entropy and modality consistency losses.

Main Results:

• The MM-WAE model significantly improved accuracy and robustness in wafer defect classification.
• Demonstrated effective mitigation of performance degradation caused by insufficient labels and class imbalance.
• Showcased comprehensive characterization of defect patterns through fused multi-domain features.

Conclusions:

• The MM-WAE offers a promising solution for industrial wafer defect classification, particularly in data-scarce and imbalanced scenarios.
• The method's ability to leverage unlabeled data enhances its practical applicability.
• Further development holds potential for advancing semiconductor manufacturing yield optimization.