GANs for Data Augmentation in semiconductor manufacturing is the use of Generative Adversarial Networks to generate realistic synthetic training data — addressing the chronic shortage of labeled defect images, rare process conditions, and imbalanced datasets that limit ML model performance.

GAN Architectures for Fab Data

• DCGAN: Deep convolutional GAN for basic image generation.
• Conditional GAN (cGAN): Generates specific defect types conditioned on class labels.
• WGAN-GP: Wasserstein GAN with gradient penalty for stable training.
• CycleGAN: Translates between domains (e.g., optical ↔ SEM images) without paired data.

Why It Matters

• Data Scarcity: Real defect images are scarce and expensive to label — GANs multiply the dataset.
• Improved Accuracy: GAN-augmented training typically improves classifier accuracy by 5-15%.
• Balanced Training: Generate minority-class samples to balance severely imbalanced datasets.

GANs for Augmentation are the data multiplier — using adversarial generation to create realistic synthetic fab data that improves ML model training and robustness.