Post on 03-Apr-2022
Confidential
Machine Learning Hotspot Prediction Significantly Improve Capture Rate on Wafer
Wei.YuanICRD/ASML 2020.11.06
SHANGHAI IC R&D CENTER上海集成电路研发中心有限公司
ICRD SINCE 2002 | 商密X级
Confidentialwww.icrd.com.cn
Outline
• Patterning Technology Roadmap
• Hotspot Prediction Background
• Machine Learning Hotspot Prediction
• Wafer Data Verification and Result Analysis
• Summary
2
Confidentialwww.icrd.com.cn
Patterning Technology Roadmap
3
RBOPC
OPC
Illumination
Resist
Integration
Mask
≧ 45nmArF/ArF-iDUV/I-line
32/28nmArF-i
16/14nmArF-i
22/20nmArF-i
Binary
6% PSM
MBOPC
RBAF
PTD
Conventional
Annular
C-Quad
DipoleLEC
SMO
LELE DPT
SAV/SAC
SADP DPT
NTD
Wafer 3D
ILT
MBAF
APF
Tri-Layer
FinFET
OMOG
Mask 3D
10/7nmArF-i/EUV
MPTHigh-T PSM
Resist 3D
Machine learning OPC
Confidentialwww.icrd.com.cn
Hotspot Prediction Background
• OPC verification(LMC) is playing the role after OPC correction to verify OPC quality.
• Key value of OPC verification is to distinguish risky patterns and safe patterns.
4
Killer defect
found in fab
Lithography/
Wafer processing• Problems are caught too late
• Cost of mask re-spin
• Longer time to good yield
Litho related
problems
Re-tune OPC
recipe
Mask Manufacturing
Without LMC
Litho related
problems
Recovers Lost
Cycle Time
LMC
Design RET/OPC
Data
Cycle time - days LMC saves >14 days of lost cycle time
LMC catches yield problems before reticle manufacturing
Design
completion End of Life
Lost
opportunity
$$$
Re
ve
nu
e (
$)
Time
Re-tune OPC
recipe
Printing
defect
predicted by
LMC
1st pass failed
inspection
2nd pass
good
Confidentialwww.icrd.com.cn
Verification become more and more complex
5
CD/EP at NC/PW Less raw inputLess accuracy
More raw inputMore accuracy
Post OPC gdsand model Contour and image
ILS, DOF, MEEF,
PV band…
ML solution
• Traditional verification methods could not capture all mask related defects. Simple CD/EP inspection.
Cross condition inspections, image contrast inspections.
Solution: Construct more KPI; Introduce more accurate models.
• More complex but powerful solutions are required as process keep shrinking on.ML solution?
Mature nodes
Advanced nodes
Confidentialwww.icrd.com.cn
Newron Hotspot Prediction – Training Flow
• Image based defect defection flow constructed on DCNN.
• Makes full use of whole simulated images in order to generate prediction information.
• Multiple projecting routines are supported.To design space. Trained by huge amount simulation data.To behavior space. Trained by limited wafer data.
7
mask and
OPC model
Resist
image
Latency space
vector
CNN
network
Label from
wafer inspection
Design similarity
comparison
DNN
network
Model for behavior
space
Model for
design space
1) Generate resist image around hotspot location by
applying OPC model; Corp and covert whole resist
image to latency space vector via CNN.
2) Prepare the training data; ADI patterns with design
similarity comparison and labels (manually labeled ‘r/s’
for these ADI images by CDSEM value).
3) Feed training data into Newron engine to train two
separate prediction models.
Confidentialwww.icrd.com.cn
Newron Hotspot Prediction – Application Flow
• Generate resist image, convert to latency space vectors for new hotspots.
• Latency space vectors are projected to new feature spaces by prediction models.
• Generate final prediction label by considering the two feature space distribution.
8
Latency
space vector
behavior space
design space
Hotspot candidates
in MTO LMC+ jobs
Verified
Hotspots
Prediction risky
or safe
In design space the Euclidean distance stands for the design similarity of one hotspot candidate to other hotspots in training set.
In behavior space the Euclidean distance stands for the probability of being risky or safe.
Central region of one hotspot image contribute more to the final similarity.
Ref Less similarMore similar
Confidentialwww.icrd.com.cn
• Total 375 unique hotspot candidates verified on wafer. (7 conditions*10shots*375=26250 measurements)
• KPI #1: CD variation in PW conditions. • The top 125 candidates with biggest CD variation --- “Risky”
• The rest 250 candidates with smallest CD variation --- “Safe”
• KPI #2: ADI model error at nominal condition.• The worst 125 candidates with biggest model error --- “Risky”
• The best 250 candidates with smallest model error --- “Safe”
Ground truth labeling based on real wafer data
10
Defocus -45 Defocus 45Nominal Condition
Hotspot A
CD variation 11.4nm
CD variation 1.53nm
Hotspot B
Confidentialwww.icrd.com.cn
• Baseline ADI model has a good performance both on 1D and 2D patterns.
Baseline ADI model performance
11
14nm M1 NTD Model PerformanceSEM Prediction
TypeIn Spec
ratioOverall
RMSAnchor error
1D 96.7% 1.05
2D 99.8% 1.4
Total 1.24 -0.09
Confidentialwww.icrd.com.cn
Use traditional detectors as baseline prediction solutions
• PVB-Focus has the most relevant physical meaning with through-focus band in wafer. Thus PVB detector shows best prediction accuracy.
12
Hotspot A
Hotspot B
0
0.25
0.5
0.75
Nuisance Rate Capture Rate F1
Prediction on KPI #1
PVB CD ILS NILS AI MEEF
Hotspot A
PredictionRisky Safe
Ground Truth
RiskyTrue Positive
(TP)False Negative
(FN)
SafeFalse Positive
(FP)True Negative
(TN)
Nuisance = FP
TP+FP, the lower the better
Capture = TP
TP+FN, the higher the better
F1= 2
Capture−1+(1−Nuisance)−1, the higher the better
Hotspot B
(simulation CD)
(w
afe
r C
D)
Confidentialwww.icrd.com.cn
Newron hotspot prediction result comparison
• Risky and safe hotspots show clear separation in feature space.
• Final F1 score get ~47% improvement in KPI#1.
13
Although some safe hotspot are mixed with risky hotspot in test set, the overall distribution matched well to train set.
Hotspot A
Hotspot B 0.31
0.43
0.53
0.33
0.91
0.78
0
0.25
0.5
0.75
1
Nuisance Capture F1
Prediction on KPI #1
PVB
Newron
+47%+111%
Confidentialwww.icrd.com.cn
Newron can handle more complex predictions
• Traditional detectors do not have the capability to predict the model error because all the detectors use the information simulated by the model itself.
14
Model error
There's no obvious correlation between NC model error and PVB-Focus, MEEF or ILS.
Confidentialwww.icrd.com.cn
Newron can handle more complex predictions
• Resist image contain sufficient info for such kind of KPI;
• Machine learning solution can achieve good enough prediction.
15
Newron solution result on model error at NC.
Capture the model error by ML method.
0.12
0.68
0.78
0
0.25
0.5
0.75
1
Nuisance Capture F1
Prediction on KPI #2
Model error
Confidentialwww.icrd.com.cn
Summary
• ICRD and ASML-Brion co-worked to evaluate an innovative machine learning hotspot
prediction method on 14nm metal layer process.
• Traditional detectors are insufficient to predict wafer hotspots for advanced tech node.
• To predict the ground truth labeled by CD variation in PW conditions, Newron Hotspot
improves the prediction F1 score by 47% compared with baseline detectors.
• Newron Hotspot achieved a breakthrough in predicting model error with 68% capture
rate and 12% nuisance rate while there is no traditional solution yet.
• It has been proved that Newron Hotspot flow can be flexibly adjusted to handle
various prediction targets.
16
Confidentialwww.icrd.com.cn
More Study on ML
17
ML SRAF(IWAPS 2019)
ML OPC(Continued )
ML OPC(SPIE 2019)
Machine Learning
ML hotspot Prediction
(IWAPS 2020)
ML ADI model(SPIE 2019)
ML etch model(SPIE 2020)
ML ADI model
ML AEI model ML OPC
ML SRAF
• Machine Learning technologies have high potential to be applied in OPC area and will bring significant benefit to improve the OPC performance.
Confidentialwww.icrd.com.cn
Acknowledgement
• Thank ICRD OPC team and ASML-Brion LMC/MKT team’s great support and co-work to finish this study.
18
Wei Yuan
Yifei Lu
Ming Li
Bingyang Pan
Ying Gao
Yu Tian
Zhi-qin Li
Liang Ji
Ying Huang
Hao Chen
Yueliang Yao
Yanjun Xiao
Confidential
ICRD is a registered trademark of SHANGHAI IC R&D CENTER.
Names of ICRD products and services are the registered trademarks and/or trademarks of SHANGHAI IC R&D CENTER or its Group companies.
Other company names and product names are registered trademarks and/or trademarks of the respective companies.
INNOVATION EFFICIENCY PERSISTENCE EXCEED
Thank you.