Layout Hotspot Detection with Feature Tensor Generation and Deep - - PowerPoint PPT Presentation

Layout Hotspot Detection with Feature Tensor Generation and Deep Biased Learning

Haoyu Yang1, Jing Su2, Yi Zou2, Bei Yu1, Evangeline F. Y. Young1

1The Chinese University of Hong Kong 2ASML Brion Inc.

1 / 15

Outline

Introduction Feature Tensor Generation Biased Learning Experimental Results

2 / 15

Outline

Introduction Feature Tensor Generation Biased Learning Experimental Results

3 / 15

Lithography Hotspot Detection

◮ 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 / 15

Pattern Matching based Hotspot Detection

library'

hotspot&

Pa)ern' matching'

hotspot& hotspot&

4 / 15

Pattern Matching based Hotspot Detection

library'

hotspot&

Pa)ern' matching'

hotspot& hotspot&

detected

hotspot&

undetected detected

Cannot&detect& hotspots&not&in& the&library&

◮ Fast and accurate ◮ [Yu+,ICCAD’14] [Nosato+,JM3’14] [Su+,TCAD’15] ◮ Fuzzy pattern matching [Wen+,TCAD’14] ◮ Hard to detect non-seen pattern

4 / 15

Machine Learning based Hotspot Detection

Hotspot& detec*on& model&

Classifica*on& Extract&layout& features&

5 / 15

Machine Learning based Hotspot Detection

Non$ Hotspot Hotspot

Hotspot& detec*on& model&

Classifica*on& Extract&layout& features& Hard,to,trade$off, accuracy,and,false, alarms,

◮ Predict new patterns ◮ Decision-tree, ANN, SVM, Boosting, Bayesian, ... ◮ [Ding+,TCAD’12][Yu+,JM3’15][Matsunawa+,SPIE’15][Yu+,TCAD’15][Zhang+,ICCAD’16][Wen+,TCAD’14] ◮ Feature reliability and model scalability

5 / 15

Why Deep Learning?

• 1. Feature Crafting v.s. Feature Learning

◮ Manually designed feature–> Inevitable information loss ◮ Learned feature–> Reliable

• 2. Scalability

◮ More pattern types ◮ More complicated patterns ◮ Hard to fit millions of data with simple ML model

• 3. Mature Libraries

◮ Caffe [Jia+,ACMMM’14] ◮ Tensorflow [Martin+,TR’15]

6 / 15

Special Issues for Layout Hotspot Detection

Layout image size is large (≈ 1000 × 1000)

◮ Compared to ImageNet (≈ 200 × 200) ◮ Associated CNN model is large ◮ Not storage and computational efficient

Hotspot detection accuracy is more important

◮ Hotspot –> Circuit Failure ◮ False Alarm –> Runtime Overhead ◮ Consider methods for better trade-off between

accuracy and falsealarm

Layout clip with 1nm precision has resolution 1200 × 1200

7 / 15

Outline

Introduction Feature Tensor Generation Biased Learning Experimental Results

8 / 15

Feature Tensor Generation

◮ Clip Partition ◮ Discrete Cosine Transform ◮ Discarding High Frequency Components ◮ Feature Tensor

Division

8 / 15

Feature Tensor Generation

◮ Clip Partition ◮ Discrete Cosine Transform ◮ Discarding High Frequency Components ◮ Feature Tensor

Division DCT

8 / 15

Feature Tensor Generation

◮ Clip Partition ◮ Discrete Cosine Transform ◮ Discarding High Frequency Components ◮ Feature Tensor

Division DCT

2 6 6 6 4 C11,1 C12,1 C13,1 . . . C1n,1 C21,1 C22,1 C23,1 . . . C2n,1 . . . . . . . . . ... . . . Cn1,1 Cn2,1 Cn3,1 . . . Cnn,1 3 7 7 7 5 2 6 6 6 4 C11,k C12,k C13,k . . . C1n,k C21,k C22,k C23,k . . . C2n,k . . . . . . . . . ... . . . Cn1,k Cn2,k Cn3,k . . . Cnn,k 3 7 7 7 5

(

k

Encoding

8 / 15

CNN Architecture

Feature Tensor

◮ k-channel hyper-image ◮ Compatible with CNN ◮ Storage and computional efficiency Layer Kernel Size Stride Output Node # conv1-1 3 1 12 × 12 × 16 conv1-2 3 1 12 × 12 × 16 maxpooling1 2 2 6 × 6 × 16 conv2-1 3 1 6 × 6 × 32 conv2-2 3 1 6 × 6 × 32 maxpooling2 2 2 3 × 3 × 32 fc1 N/A N/A 250 fc2 N/A N/A 2

… Hotspot Non-Hotspot Convolution + ReLU Layer Max Pooling Layer Full Connected Node

2 6 6 6 4 C11,1 C12,1 C13,1 . . . C1n,1 C21,1 C22,1 C23,1 . . . C2n,1 . . . . . . . . . ... . . . Cn1,1 Cn2,1 Cn3,1 . . . Cnn,1 3 7 7 7 5 2 6 6 6 4 C11,k C12,k C13,k . . . C1n,k C21,k C22,k C23,k . . . C2n,k . . . . . . . . . ... . . . Cn1,k Cn2,k Cn3,k . . . Cnn,k 3 7 7 7 5

(

k

9 / 15

Outline

Introduction Feature Tensor Generation Biased Learning Experimental Results

10 / 15

Recall The Training Procedure

◮ Minimize difference with ground truths

y∗

n = [1, 0], y∗ h = [0, 1].

(1)

F ∈ N,

if y(0) > 0.5

H,

if y(1) > 0.5 (2)

10 / 15

Recall The Training Procedure

◮ Minimize difference with ground truths

y∗

n = [1, 0], y∗ h = [0, 1].

(1)

F ∈ N,

if y(0) > 0.5

H,

if y(1) > 0.5 (2)

◮ Shifting decision boundary

F ∈ N,

if y(0) > 0.5 + λ

H,

if y(1) > 0.5 − λ (3)

10 / 15

Recall The Training Procedure

◮ Minimize difference with ground truths

y∗

n = [1, 0], y∗ h = [0, 1].

(1)

F ∈ N,

if y(0) > 0.5

H,

if y(1) > 0.5 (2)

◮ Shifting decision boundary (✗)

F ∈ N,

if y(0) > 0.5 + λ

H,

if y(1) > 0.5 − λ (3)

◮ Biased ground truth

y∗

n = [1 − ǫ, ǫ]

(4)

10 / 15

The Biased Learning Algorithm

Training Set Update ε yh=[0,1] yn=[1-ε, ε] MGD: end-to-end training Stop Criteria Trained Model Yes No Biased Learning v.s. Shift Boundary

80 85 90 2,000 3,000 4,000

Accuracy (%) False Alarm Shift-Boundary Bias

11 / 15

Outline

Introduction Feature Tensor Generation Biased Learning Experimental Results

12 / 15

Comparison with Two Hotspot Detectors

◮ Detection accuracy improved from 89.6% to 95.5%

ICCAD Industry1 Industry2 Industry3 Average

40 60 80 100

Accuracy (%) SPIE’15 ICCAD’16 Ours

12 / 15

Comparison with Two Hotspot Detectors

◮ Comparable false alarm penalty

ICCAD Industry1 Industry2 Industry3 Average

2,000 4,000 6,000 8,000

False Alarm SPIE’15 ICCAD’16 Ours

13 / 15

Comparison with Two Hotspot Detectors

◮ Comparable testing runtime

ICCAD Industry1 Industry2 Industry3 Average

500 1,000 1,500 2,000 2,500

Runtime (s) SPIE’15 ICCAD’16 Ours

14 / 15

Thank You

15 / 15