DLS-DMO: Towards High Accuracy DL-Based OPC With Deep Lithography Simulator Guojin Chen supervised by Prof. Yu Bei Department of Computer Science & Engineering The Chinese University of Hong Kong June 17, 2020 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 / 30
Introduction and Background Outline Introduction and Background Previous work DLS-DMO Data Generation DCGAN-HD DCUNet++ Multi D Perceptual Loss DLS DMO Irregular Splitting Algo. Results Our datasets ISPD 2019 datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 / 30
Introduction and Background Background and problem formulation Project backgroud Optical proximity correction (OPC) is a photolithography enhancement technique commonly used to compensate for image errors due to diffraction or process effects. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 / 30
Introduction and Background PRELIMINARIES of OPC: DESIGN, SRAF, MASK, WAFER . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 / 30
Introduction and Background PRELIMINARIES of OPC: Flow, EPE, PVBand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 / 30
Introduction and Background Goal: Using NN to simulate this process And beat one commercial products: Calibre Two main step OPC and Litho Design Mask Wafer DMO DLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 / 30
Introduction and Background Goal: Test our model on the industry data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 / 30
Previous work Outline Introduction and Background Previous work DLS-DMO Data Generation DCGAN-HD DCUNet++ Multi D Perceptual Loss DLS DMO Irregular Splitting Algo. Results Our datasets ISPD 2019 datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 / 30
Previous work cGAN Objective function L cGAN ( G , D ) = E x , y [log D ( x , y )] + E x , z [log( 1 − D ( x , G ( x , z ))] . (1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 / 30
Previous work OPC stage previous work: GAN-OPC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 / 30
Previous work GAN-OPC: shortages ▶ We cannot control the litho simulator. ▶ ILT-based model, come from MOSAIC, small layout. ▶ Only initial solution, bottleneck on the ILT-model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 / 30
Previous work Litho stage previous work: LithoGAN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 / 30
Previous work LithoGAN: shortages 1. Wafer may not have center. Did not make full use of cGAN. 2. The center shift is over design, we just need a powerful generator. 3. Mask must be at the center, one time can only generator one wafer(in the center, few in the dataset.) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 / 30
DLS-DMO Outline Introduction and Background Previous work DLS-DMO Data Generation DCGAN-HD DCUNet++ Multi D Perceptual Loss DLS DMO Irregular Splitting Algo. Results Our datasets ISPD 2019 datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 / 30
DLS-DMO Goal Solution: DLS-DMO Problem ▶ End-to-end mask optimization without ▶ Initial solution need further correction. using traditional model. ▶ One time one via lithography process. ▶ High resolution cGAN model. ▶ Low accuracy and small layout. ▶ Window splitting algorithm for large layout. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 / 30
DLS-DMO Data Generation Generate Training set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 / 30
DLS-DMO Data Generation Self-generated datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 / 30
DLS-DMO DCGAN-HD DCGAN-HD: solution for higher resolution ▶ Generator: DCUNet++ ▶ Discriminator: Multi-discriminator ▶ Perceptual Losses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 / 30
DLS-DMO DCGAN-HD DCUNet++: Generator of DCGAN-HD UNet++ Arch. ▶ UNet++ for low-level information. ▶ Residual blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 / 30
DLS-DMO DCGAN-HD DCUNet++: Generator of DCGAN-HD DCUNet++ UNet++ Backbone Encoder Decoder Arch. Residual Blocks ▶ UNet++ for low-level information. ▶ Residual blocks … Convolution Deconvolution Residual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 / 30
DLS-DMO DCGAN-HD Multi Scale discriminator We design a multi-scale discriminator, different from pix2pixHD using 3 discriminators, our design uses 2 discriminators that have an identical network structure but operate at different image scales, which namedD1,D2. Specially, the discriminators D1,D2are trained to differentiate real and synthesized images at the 2 different scales, 1024 × 1024 and 512 × 512 respectively. As in pix2pixHD claimed, the multi-scale design helps the training of high-resolution model easier. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 / 30
DLS-DMO DCGAN-HD Perceptual Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 / 30
<latexit sha1_base64="7CZ4qKZ6qe92jLDIBbVtlvjZlkY=">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</latexit> DLS-DMO DLS DLS training L cGAN ( G , D k ) + λ 0 L G , Φ ∑ L DLS = L P ( y , ˆ y ) . (2) k = 1 , 2 y D Real Perceptual Loss x D x G Fake z ˆ y . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 / 30
DLS-DMO DMO DMO training L cGAN ( G DMO , ( D DMO ) k ) + λ 1 L G DMO , Φ ∑ L DMO = ( x , ˆ x ) . (3) L P k = 1 , 2 L DLS − OPC = L DMO + L DLS + λ 2 L L 1 (ˆ y , w r ) . (4) DCUNet++ (a) Frozen DMO DLS DCUNet++ Generator Feed-forward Back-propagetion . . . . . . . . . . . . . . . . . . . . (b) . . . . . . . . . . . . . . . . . . . . 21 / 30
DLS-DMO Irregular Splitting Algo. Irregular Splitting Algo: Coarse to Fine, DBSCAN to KMenas Algo. figure Algo. detail 1. DBSCAN then KMeans++ 2. Initialize the number of centroids from 1 to V to run KMeans++. 3. Every cluster contains no more than K via patterns. 4. Every via pattern must be contained in a window. 5. If (3) or (4) is not satisfied, increase the centroid number . SRAF VIA DBSCAN KMeans++ Window . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 / 30
DLS-DMO Irregular Splitting Algo. Main Contribution ▶ DCGAN-HD: we extend cGANs model by redesign the generator and discriminator for high resolution input (1024*1024), combined with a novel window-splitting algorithm, our model can handle input layout of any size with high accuracy. ▶ We build up a deep lithography simulator (DLS) based on our DCGAN-HD. Thanks to the express power of stack convolution layers, DLS is expected to conduct lithography simulation faster with similar contour quality compared to legacy lithography simulation process. ▶ We present DLS-DMO, a unified end-to-end trainable OPC engine that employs both DLS and DMO to conduct mask optimization without further fine-tune with legacy OPC engines. ▶ Experimental results show that the proposed DLS-OPC framework is able to output high quality lithography contours more efficiently than Calibre, which also derives ∼ 4 × speed-up in OPC tasks while generating masks with even better printability. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 / 30
Recommend
More recommend
