UT DA GAN-SRAF: Sub-Resolution Assist Feature Generation using Generative Adversarial Networks Mohamed Baker Alawieh , Yibo Lin, Zaiwei Zhang, Meng Li, Qixing Huang and David Z. Pan The University of Texas at Austin Work funded in part by NSF 1
Motivation t With the IC technology scaling, resolution enhancement techniques are becoming indispensable t Sub-Resolution Assist Feature (SRAF) generation is used to improve the lithographic process window of target patterns https://slideplayer.com/slide/9416386/ 2
Conventional Approaches t Rule-Based approaches: › Work well for simple designs with regular patterns › Cannot handle complex shapes t Model-Based (MB) approaches: › Achieve high quality results › Suffer from exorbitant computational cost t Machine Learning (ML) Based approach: › Achieves results quality similar to MB › Results in 10X reduction in runtime 3
Conventional Approaches t Rule-Based approaches: › Work well for simple designs with regular patterns › Cannot handle complex shapes t Model-Based (MB) approaches: › Achieve high quality results › Suffer from exorbitant computational cost t Machine Learning (ML) Based approach: › Achieves results quality similar to MB › Results in 10X reduction in runtime Can we do better?! 4
ML Based Approach t Proposes local sampling scheme with a classification model t On a 2D grid, the classifier predicts the presence of SRAF in each grid 0 1 2 N%1 sub%sampling0point SRAF%label:%0 SRAF%label:%1 Target pattern SRAF SRAF box OPC region SRAF region [Xu et al, ISPD’16, TCAD’17] 5
ML Based Approach t Proposes local sampling scheme with a classification model t On a 2D grid, the classifier predicts the presence of SRAF in each grid t While achieving 10X runtime improvement, this approach has large room for further enhancement [Xu et al, ISPD’16, TCAD’17] 6
ML Based Approach t Proposes local sampling scheme with a classification model t On a 2D grid, the classifier predicts the presence of SRAF in each grid t While achieving 10X runtime improvement, this approach has large room for further enhancement › Do we need a 2D grid and local sampling? [Xu et al, ISPD’16, TCAD’17] 7
ML Based Approach t Proposes local sampling scheme with a classification model t On a 2D grid, the classifier predicts the presence of SRAF in each grid t While achieving 10X runtime improvement, this approach has large room for further enhancement › Do we need a 2D grid and local sampling? › Can we avoid the feature extraction step? [Xu et al, ISPD’16, TCAD’17] 8
ML Based Approach t Proposes local sampling scheme with a classification model t On a 2D grid, the classifier predicts the presence of SRAF in each grid t While achieving 10X runtime improvement, this approach has large room for further enhancement › Do we need a 2D grid and local sampling? › Can we avoid the feature extraction step? › Most importantly, with all advancements in Computer Vision, can we recast this problem to leverage these advancement? [Xu et al, ISPD’16, TCAD’17] 9
CGAN for Image Translation t GANs have been proposed to produce images similar to those in training data set t CGAN, takes as an input a picture in one domain and translates it to a new one › During training it sees pairs of matched images [Isola et al, CVPR’18] 10
SRAF Generation & Image Translation t What does SRAF generation have to do with Image translation?! 11
SRAF Generation & Image Translation t What does SRAF generation have to do with Image translation?! t Can we define the problem as translating images from the Target Domain (D T ) to the SRAF Domain (D S )? 12
Challenges t Layout images have sharp edges which pose a challenge to GANs › Model is not guaranteed to generate polygon SRAF shapes › Sharp edges can complicate gradient propagation t Generated images need ultimately be changed to layout format › Images cannot be directly mapped to ‘GDS’ format › Post-processing step should not be time consuming 13
Challenges t Layout images have sharp edges which pose a challenge to GANs › Model is not guaranteed to generate polygon SRAF shapes › Sharp edges can complicate gradient propagation t Generated images need ultimately be changed to layout format › Images cannot be directly mapped to ‘GDS’ format › Post-processing step should not be time consuming t Hence, a proper encoding is needed to address these challenges! 14
Multi-Channel Heatmap Encoding t Key Idea: encode each type of object on a separate channel in the image › Channel index carries object description (type, size,...) › Excitations on the channel carry objects location Original Layout Encoded Layout 15
Challenges Revisited t Layout images have sharp edges which pose a challenge to GANs › Model is not guaranteed to generate polygon SRAF shapes › Polygon shapes are not needed, the objective of model is to predict locations on different channels › Sharp edges can complicate gradient propagation › No sharp edges in encoded image t Generated images need ultimately be changed to layout format › Images cannot be directly mapped to ‘GDS’ format › Decoding is straight forward, it suffices to detect excitation location on each channel to get full GDS information 16
CGAN Approach t Generator: Fake/Real › Trained to produce images in D S based on input from D T Discriminator › Tries to fool the Discriminator Real Diff t Discriminator: › Trained to detect ‘fakes’ generated by the Generator Encoder Decoder t The two networks are jointly trained until convergence Fake Generator Input 17
CGAN Approach t Generator: Fake/Real › Encoder: Downsampling › Decoder: Upsampling Discriminator Real Diff Encoder Decoder Fake Generator Input 18
CGAN Approach t Generator: Fake/Real › Encoder: Downsampling › Decoder: Upsampling Discriminator Real t Discriminator: Diff › CNN trained as a classifier t After training, only the generator is used Encoder Decoder Fake Generator Input 19
Results Decoding t Decoding the generated layout images consists of two steps: › Thresholding & Excitation detection 20
Results Decoding t Decoding the generated layout images consists of two steps: › Thresholding & Excitation detection 21
Results Decoding t Decoding the generated layout images consists of two steps: › Thresholding & Excitation detection 22
Results Decoding t Decoding the generated layout images consists of two steps: › Thresholding & Excitation detection SRAF location Isolated pixel 23
Results Decoding t Decoding the generated layout images consists of two steps: › Thresholding & Excitation detection Decoding scheme is fast è GPU accelerated SRAF location Isolated pixel 24
• LS_SVM: Xu et al, ISPD’16, TCAD’17 Sample Results • MB: Model-Based Approach - Calibre MB 25
• LS_SVM: Xu et al, ISPD’16, TCAD’17 Sample Results • MB: Model-Based Approach - Calibre A post processing legalization step is applied GAN-SRAF MB 26
• LS_SVM: Xu et al, ISPD’16, TCAD’17 Sample Results • MB: Model-Based Approach - Calibre LS_SVM GAN-SRAF MB Target Pattern MB SRAF LS_SVM Final 27 LS_SVM Prediction CGAN Final CGAN Prediction
Lithography Compliance Checks MB CGAN LS SVM NO SRAF 3000 2000 Histogram of PV ( u m 2 ) 1000 0 • LS SVM: Xu et al, ISPD’16, TCAD’17 0 . 0020 0 . 0025 0 . 0030 0 . 0035 0 . 0040 • MB: Model-Based Approach - Calibre MB CGAN LS SVM NO SRAF 6000 Histogram of EPE (nm) 4000 2000 0 − 6 − 4 − 2 0 2 28
Comparison Summary No SRAF MB LS_SVM CGAN PV Band ( u m 2 ) 0.00335 0.002845 0.00301 0.00291 EPE (nm) 3.9287 0.5270 0.5066 0.541 Run time (s) - 6910 700 48 t The proposed CGAN based approach can achieve comparable results with LS_SVM and MB with 14.6X and 144X reduction in runtime 29
Conclusions t GAN-SRAF, a novel SRAF generation framework, is presented featuring: › Novel problem formulation as image translation › Smart heatmap encoding scheme and GPU accelerated decoding t Results demonstrate significant speedup when compared to ML and MB › While achieving comparable lithography performance 30
Thank You! 31
Recommend
More recommend