Deep Learning-Driven Simultaneous Layout Decomposition and Mask - - PowerPoint PPT Presentation

Deep Learning-Driven Simultaneous Layout Decomposition and Mask Optimization

Wei Zhong1,2, Shuxiang Hu1,2, Yuzhe Ma3, Haoyu Yang3, Xiuyuan Ma1,2, Bei Yu3

1Information Science & Engineering, Dalian University of Technology 2Key Lab for Ubiquitous Network & Service Software of Liaoning Province 3CSE Department, Chinese University of Hong Kong

1 / 24

Biography Shuxiang Hu

Dalian University of Technology vsxhoo@mail.dlut.edu.cn He is now studying for M.Sc. degree at the International School of Information Science and Engineering, Dalian University of Technology, under the supervision of Prof. Wei Zhong.

2 / 24

Outline

Introduction Algorithm Experimental Results Conclusion

3 / 24

Outline

Introduction Algorithm Experimental Results Conclusion

4 / 24

Optical Proximity Effect

Target Result

◮ Resolution enhancement Technologies (RETs):

• OPC
• MPL

4 / 24

Layout Decomposition for Mask Optimization

◮ Different decomposition results converge to divergent printability

Target LD MO Printed Image

#EPE Violation = 3 #EPE Violation = 1

5 / 24

Option for Decomposition Selection

◮ Solution: Collaboration of LD and MO in a unified framework [Ma+,ICCAD’17].

Output Optimized Masks Numerical Layout Optimization Discrete Layout Optimization Input Layout

6 / 24

Issues

◮ Not Accurate: Greedy pruning. ◮ Not Efficient: OPC suffers from large computational complexity. 10 20 30 10 20 30

#Iterations EPE#

DECMP#1 DECMP#2 DECMP#3

Decomposition convergence of EPE

DS 59.1% MO 40.9%

Runtime break down

7 / 24

Motivation

◮ Powerful convolutional neural network (CNN)

• Build mapping relationship automatically.
• Large amount of data required.

◮ CNN application in EDA field:

• Routing predicting [Xie+,ICCAD’18]
• Hotspot detection [Yang+,TCAD’18]
• Resist modeling [Lin+,TCAD’18]

◮ How about integrating CNN for decomposition selection?

8 / 24

Outline

Introduction Algorithm Experimental Results Conclusion

9 / 24

Forward Optimization Flow

Decomposition Generation Print Violation Detected ? Printability Prediction ILT Optimization Optimized Masks Input Layout N Y

Printability Predictor

ILT Prediction

9 / 24

Decomposition Generation

10 / 24

Decomposition Generation

◮ Classify patterns & build minimal spanning tree E ∈      SP, if d ≤ nmin, VP, if nmin < d ≤ nmax, NP, if nmax < d.

A B E D C F H I J

Component 1 Component 2

75 78 60 76 60 G K

SP

NP

VP

Component 1 Component 2

75 60 76 60

11 / 24

Decomposition Generation

◮ n-wise arrays

• SP and VP with three-wise
• NP with two-wise

A B E D C F H I J

Component 1 Component 2

75 78 60 76 60 G K

SP

NP

VP

Arrs1 SP VP

B F H I #1 1 1 #2 1 1 1 1 #3 1 1 #4 #5 1 1 #6 1 1 #7 1 1 #8 1 1

Arrs2 NP

G J K #1 1 #2 1 #3 1 1 1 #4 1

12 / 24

Decomposition Generation

◮ n-wise arrays

• Relax combination strength
• Complete combination of n factors

Three-wise arrays

13 / 24

Printability Prediction & Mask Optimization

14 / 24

Printability Prediction & Mask Optimization

◮ Select the best decomposition candidate for OPC engine

Printability score Conv1 Layer1 Layer2 Layer3 Layer4 224 64 64 128 256 512 1000 112 56 28 14 7 Decomposition candidate

Printability score = α × #EPE + β × L2 Error + γ × #Print Violation

15 / 24

How to Sample Data?

◮ Sample typical data for train

16 / 24

How to Sample Data?

◮ Layout sampling ◮ Decomposition sampling

• Similar to decomposition generation stage

◮ Get printability score

17 / 24

Layout Sampling

◮ Calaulate point distance

• Match points
• Euclidean distance as matched points distance

◮ Calculate layout distance

• Sum up matched poinnts as layout distance

◮ Cluster layouts

18 / 24

Outline

Introduction Algorithm Experimental Results Conclusion

19 / 24

Comparision on EPE violations

◮ Outperform state-of-the-art. ◮ Reduce 68.0% EPE violations on average.

NOR2_X1 OAI211_X1 NAND4_X1 NAND3_X2

2 4 6 8 10

EPE Violations ISQED’13+DAC’14 ICCAD’13+DAC’14 ICCAD’17 Ours

19 / 24

Comparision on Runtime

◮ About 4X speed up.

NOR2_X1 OAI211_X1 NAND4_X1 NAND3_X2

500 1,000 1,500 2,000

Runtime ISQED’13+DAC’14 ICCAD’13+DAC’14 ICCAD’17 Ours

20 / 24

Optimization results

ICCAD’17 Ours AOI211_X1 NAND3_X2 BUF_X1

21 / 24

Comparision with Random Sampling

◮ Reduce half of EPE violations.

EPE# Runtime

1 2

Ratio Random Sampling Ours

22 / 24

Outline

Introduction Algorithm Experimental Results Conclusion

23 / 24

Conclusion

◮ Deep learning based layout decomposition and mask optimization framework.

• Decomposition generation approach.
• Decomposition printability estimation.

◮ A set of sampling strategies. ◮ Experimental results demonstrate the effectiveness and efficiency.

23 / 24

Thank You

Wei Zhong (zhongwei@dlut.edu.cn) Shuxiang Hu (vsxhoo@mail.dlut.edu.cn) Yuzhe Ma (yzma@cse.cuhk.edu.hk) Bei Yu (byu@cse.cuhk.edu.hk)

24 / 24