VLSI Mask Optimization: From Shallow To Deep Learning Haoyu Yang 1 , - - PowerPoint PPT Presentation

VLSI Mask Optimization: From Shallow To Deep Learning

Haoyu Yang1, Wei Zhong2, Yuzhe Ma1, Hao Geng1, Ran Chen1, Wanli Chen1, Bei Yu1

1The Chinese University of Hong Kong 2Dalian University of Technology

1 / 22

Moore’s Law to Extreme Scaling

1940 1950 1960 1970 1980 1990 2000 2010 2020 10,000,000,000 1 10 100 1,000 10,000 100,000 1,000,000 10,000,000 100,000,000 1,000,000,000

Invention of the Transistor

10 1 0.1 0.01 Year

Number of Transistors per Integrated Circuit

Moore’s Law

Process Technology (µm

)

4004 8086 286 386 486 Pentium Pentium II Pentium 4 Core 2 Duo Core i7

Doubles every 2.1 yrs

A7 A10 A11 A12

Intel Microprocessors Apple Microprocessors

2 / 22

Challenge 1: Failure (Hotspot) Detection

Pre-OPC Layout Post-OPC Mask Hotspot on Wafer

◮ RET: OPC, SRAF, MPL ◮ Still hotspot: low fidelity patterns ◮ Simulations: extremely CPU intensive

Ra#o%of%lithography%simula#on%#me% (normalized%by%40nm%node)% Technology%node

Required(computa/onal( /me(reduc/on!

3 / 22

Challenge 2: Optical Proximity Correction (OPC)

Design target Mask Wafer

without OPC with OPC

4 / 22

Why Deep Learning?

◮ Feature Crafting v.s. Feature Learning

Although prior knowledge is considered during manually feature design, information loss is inevitable. Feature learned from mass dataset is more reliable.

◮ Scalability

With shrinking down circuit feature size, mask layout becomes more complicated. Deep learning has the potential to handle ultra-large-scale instances while traditional machine learning may suffer from performance degradation.

◮ Mature Libraries

5 / 22

Outline

Hotspot Detection via Machine Learning OPC via Machine Learning Heterogeneous OPC

6 / 22

Outline

Hotspot Detection via Machine Learning OPC via Machine Learning Heterogeneous OPC

7 / 22

Hotspot Detection Hierarchy

Increasing verification accuracy

Sampling Hotspot Detection Lithography Simulation

(Relative) CPU runtime at each level

◮ Sampling (DRC Checking):

scan and rule check each region

◮ Hotspot Detection:

verify the sampled regions and report potential hotspots

◮ Lithography Simulation:

final verification on the reported hotspots

7 / 22

Early Study of DNN-based Hotspot Detector∗

◮ Total 21 layers with 13 convolution layers and 5 pooling layers. ◮ A ReLU is applied after each convolution layer.

… …

Hotspot Non-Hotspot

512x512x4 256x256x4 256x256x8 128x128x16 128x128x8 64x64x16 64x64x32 32x32x32 32x32x32 16x16x32 2048 512 C1 P1 C2-1C2-2 C2-3 P2 C3-1 C3-2 C4-1 P3 C3-3 C4-2 C4-3 C5-1 C5-2 C5-3 P4 P5

∗Haoyu Yang, Luyang Luo, et al. (2017). “Imbalance aware lithography hotspot detection: a deep learning

approach”. In: JM3 16.3, p. 033504.

8 / 22

What Does Deep Learning Learn?

Origin Pool1 Pool2 Pool3 Pool4 Pool5

9 / 22

The Biased Learning Algorithm [DAC’17]†

Training Set Update ε yh=[0,1] yn=[1-ε, ε] MGD: end-to-end training Stop Criteria Trained Model Yes No

80 82 84 86 88 90 2,000 3,000 4,000

Accuracy (%) False Alarm Shift-Boundary Bias

†Haoyu Yang, Jing Su, et al. (2017). “Layout Hotspot Detection with Feature Tensor Generation and Deep

Biased Learning”. In: Proc. DAC, 62:1–62:6.

10 / 22

Optimizing AUC [ASPDAC’19]‡

The AUC objective:

LΦ(f) =

1 N+N−

N+

i=1

N−

j=1 Φ

• f
• x+

i

• − f
• x−

j

• .

Approximation candidates: PSL ΦPSL(z) = (1 − z)2 PHL ΦPHL(z) = max(1 − z, 0) PLL ΦPLL(z) = log(1 + exp(−βz)) R ΦR∗(z) =

−(z − γ)p,

if z > γ

0,

• therwise

‡Wei Ye et al. (2019). “LithoROC: lithography hotspot detection with explicit ROC optimization”. In:

• Proc. ASPDAC, pp. 292–298.

11 / 22

Conventional Clip based Solution

Region

…

Conventional Hotspot Detector Hotspot Non- Hotspot Clips

◮ A binary classification problem. ◮ Scan over whole region. ◮ Single stage detector. ◮ Scanning is time consuming and single stage is not robust to false alarm.

12 / 22

Region based approach [DAC’19]

Region

Hotspot Core

Region-based Hotspot Detector Feature Extraction Clip Proposal Network Refinement

◮ Learning what and where is hotspot at same time. ◮ Classification Problem -> Classification & Regression Problem.

Ran Chen et al. (2019). “Faster Region-based Hotspot Detection”. In: Proc. DAC, 146:1–146:6.

13 / 22

Outline

Hotspot Detection via Machine Learning OPC via Machine Learning Heterogeneous OPC

14 / 22

OPC Previous Work

Classic OPC

◮ Model/Rule-based OPC

[Cobb+,SPIE’02][Kuang+,DATE’15] [Awad+,DAC’16][Su+,ICCAD’16]

• 1. Fragmentation of shape edges;
• 2. Move fragments for better printability.

◮ Inverse Lithography

[Pang+,SPIE’05][Gao+,DAC’14] [Poonawala+,TIP’07][Ma+,ICCAD’17]

• 1. Efficient model that maps mask to

aerial image;

• 2. Continuously update mask through

descending the gradient of contour error.

Machine Learning OPC

[Matsunawa+,JM3’16][Choi+,SPIE’16] [Xu+,ISPD’16][Shim+,APCCAS’16]

• 1. Edge fragmentation;
• 2. Feature extraction;
• 3. Model training.

14 / 22

Machine Learning-based SRAF Insertion

SRAF Insertion with Machine Learning [ISPD’16]¶

Label: 1 Label: 0 0 1 2 N%1 sub%sampling0point

Tackling Robustness with Dictionary Learning [ASPDAC’19]‖

n + s + 1

{

N

{

{

{

≈

s

{

{

@ D √αA √ βW 1 A

xt

s

N

n + s + 1

(yt

>, √αqt >,

p βht)>

¶Xiaoqing Xu et al. (2016). “A machine learning based framework for sub-resolution assist feature

generation”. In: Proc. ISPD, pp. 161–168.

‖Hao Geng et al. (2019). “SRAF Insertion via Supervised Dictionary Learning”. In: Proc. ASPDAC,

• pp. 406–411.

15 / 22

Machine Learning Assists Model-based OPC [ASPDAC’19]∗∗

◮ Replace lithography simulation (slow) with machine learning-based EPE predictor

(fast) in OPC iterations.

∗∗Bentian Jiang et al. (2019). “A fast machine learning-based mask printability predictor for OPC

acceleration”. In: Proc. ASPDAC, pp. 412–419.

16 / 22

GAN-OPC [DAC’18]††

… 0.2 0.8 Bad Mask Good Mask Discriminator Generator … Encoder Decoder Target Mask Target & Mask

◮ Better starting points for legacy OPC engine and reduce iteration count.

††Haoyu Yang, Shuhe Li, et al. (2018). “GAN-OPC: Mask Optimization with Lithography-guided Generative

Adversarial Nets”. In: Proc. DAC, 131:1–131:6.

17 / 22

Outline

Hotspot Detection via Machine Learning OPC via Machine Learning Heterogeneous OPC

18 / 22

An Observation of Previous OPC Solutions

Machine learning solutions rely on legacy OPC engines

Generator

Real Fake

Generator

(a) (b)

Feed-forward Back-propagetion

Discriminator Litho- Simulator

Legacy OPC engines exhibit different performance on different designs

1 2 3 4 5 6 7 8 9 10

2 4 6 8 ·104

Design ID MSE MB-OPC ILT

18 / 22

A Design of Heterogeneous OPC Framework

Classification Model ILT MB-OPC Mask Mask Design Designs OPC-1 OPC-N Masks Masks … …

We design a classification model that can determine the best OPC engine for a given design at trivial cost.

19 / 22

Training on Artificial Designs

◮ Training data comes from GAN-OPC and is labeled according to results of MB-OPC

and ILT.

◮ Test on 10 designs from ICCAD 2013 CAD Contest. 20 40 60 80 100 0.00 20.00 40.00 60.00 80.00 100.00 Epoch Accuracy (%) Training Testing

20 / 22

Experimental Results

1 2 3 4 5 6 7 8 9 10

2 4 6 8 ·104

Design ID MSE MB-OPC ILT HOPC Several Benefits

◮ Does not require extremely high prediction accuracy of the classification model. ◮ Take advantages of different OPC solutions on different designs.

21 / 22

Conclusion and Discussion

So Far:

◮ Recent progress of deterministic machine learning model for hotspot detection ◮ State-of-the-art machine learning solutions for OPC and SRAF insertion ◮ A heterogeneous OPC framework guided by a classification engine

Future:

◮ Manufacturability issues. ◮ Classification challenge when more than two OPC engines are available.

22 / 22