Publication: Synthesis of realistic photo-masks using supervised artificial neural networks
Abstract
Foto maskeler mikro elektronik devrelerin üretiminde önemli bir rol oynarlar ve foto maskelerdeki hataların tespiti ve bulunan hataların düzeltilmesi, mikro ve nano boyutlarda devrelerin doğru çalışması için son derece önemlidir. Maske hata tespiti derin sinir ağları ile gerçekleştirilebilir. Genellikle sinir ağı yapıları, eğitim sürecinde önemli miktarda veriye ihtiyaç duyarlar. Bu sebeple, kusurları bulunan ve bulunmayan foto maske görüntüleri gereklidir. Ancak, kusur bulunduran foto maskeler nadiren ortaya çıkarlar ve bu nedenle hatalı maske görüntülerinin elde edilmesi oldukça güçtür. Bahsedilen sorunu ele almak için bu çalışma, sentetik foto maske görüntüleri üretmeyi amaçlamıştır. Bu doğrultuda sentetik foto maske görüntülerinin üretiminde derin sinir ağı yapıları kullanılmıştır. Foto maske görüntü sentezi için biyomedikal alanda görüntü bölütlemede yaygın olarak kullanılan bir piksel tabanlı sınıflandırma ağı kullanılmıştır. Ardından görüntüden görüntüye dönüşüm ağları olarak da bilinen, pix2pix temelli bir Conditional Generative Adversarial Network yapısı ile sentetik foto maske görüntüleri elde edilmiştir. Devamında, görüntüden görüntüye dönüşüm gerçekleştiren Spatially Adaptive Normalization modelinden yararlanılarak sentetik foto maske görüntüleri oluşturulmuştur. Sentetik veriler oluşturulduktan sonra bu görüntülerin kalitelerini ölçmek için iki yaygın yöntem kullanılmıştır. Bu yöntemler, Fréchet Inception Distance ve Learned Perceptual Image Patch Similarity olarak adlandırılır. Bahsedilenlere ek olarak üretilen sentetik maske görüntüleri, hata tespiti yapan bir yapay sinir ağı modeline beslenerek test edildi. Bu araştırmanın sonuçları, foto maske hata tespitlerinin doğruluğunu artırma potansiyeline, ayrıca yüksek kaliteli mikro elektronik devrelerin üretiminde gerekli maliyeti ve işlem zamanını azaltma imkanına sahiptir
Photo masks play an important role in chip fabrication, and detecting errors in these masks is essential for ensuring correct functionality of micro and nano circuits. Error detection on these masks can be done via a deep neural network. Most of the time deep networks require considerable amount of data. Thus, images of defective and non-defective photo masks are needed. However, defects rarely occur and gathering images of defective photo masks is a challenging task. To address this issue, this study aims to generate synthetic photo-mask images by using deep neural network structures. A pixel-wise classification network which is commonly used for biomedical image segmentation was utilized for photo mask image synthesis. Next, a Conditional Generative Adversarial Network based on pix2pix framework was used to generate synthetic photo-mask images, where this network is known as an image-to-image translation network. Then, Spatially Adaptive Normalization method was utilized with an image translation network to generate synthetic photo-mask images. Quality comparison of the generated synthetic images was done via two common perceptual metrics, namely: Fréchet Inception Distance and Learned Perceptual Image Patch Similarity. In addition to these we have tested the quality for the synthetic images by feeding them into a defect detecting deep network. The results of this research have the potential to significantly improve the accuracy of photo mask error detection and reduce the cost and time required to produce high-quality microelectronic circuits.
Photo masks play an important role in chip fabrication, and detecting errors in these masks is essential for ensuring correct functionality of micro and nano circuits. Error detection on these masks can be done via a deep neural network. Most of the time deep networks require considerable amount of data. Thus, images of defective and non-defective photo masks are needed. However, defects rarely occur and gathering images of defective photo masks is a challenging task. To address this issue, this study aims to generate synthetic photo-mask images by using deep neural network structures. A pixel-wise classification network which is commonly used for biomedical image segmentation was utilized for photo mask image synthesis. Next, a Conditional Generative Adversarial Network based on pix2pix framework was used to generate synthetic photo-mask images, where this network is known as an image-to-image translation network. Then, Spatially Adaptive Normalization method was utilized with an image translation network to generate synthetic photo-mask images. Quality comparison of the generated synthetic images was done via two common perceptual metrics, namely: Fréchet Inception Distance and Learned Perceptual Image Patch Similarity. In addition to these we have tested the quality for the synthetic images by feeding them into a defect detecting deep network. The results of this research have the potential to significantly improve the accuracy of photo mask error detection and reduce the cost and time required to produce high-quality microelectronic circuits.