Machine Learning in Computational Lithography
Yu Cao
ASML-Brion
San Jose, February 2019
Public
This is really a no-brainer …
Public
Slide 2
What’s Happening and What’s the Meaning:
ASML introduces AI to its product portfolio: This is really a no-brainer. That said, the
problem that most equipment companies have is finding good applications for it, as they find all
they have is little data to feed the NN (more on that below). Anyway, this will be a good test for
DLNNs as to whether engineers will accept results without knowing what’s in the ‘blackbox,’
which is a classic barrier to this technology. I believe they will because comparing results to
input consistency are pretty easy to test out in this case. Especially since ASML led the way into
computational lithography, albeit with plenty of customer pull.
Source: The Chip Insider®
Public
Machine learning brings revolution to many applications!
Public
Slide 3
Can machine learning be the moonshot for us?
Public
We have been doing machine learning for a long time …
Public
with manual feature engineering Slide 4
Mask layout Optical model design Optical kernels Aerial image
1: E. Abbe, H. H. Hopkins Aerial image Resist model design Resist kernels Resist contour
2: F. H. Dill, C. Mack Few hidden layers = Shallow neural network Public
Machine Learning evolved to Deep Learning
Less human intervention in growing data volume analysis Public
Slide 5
Machine Learning
95% cat
5% dog
Input Feature extraction Classification Output
Feature extraction & classification
95% cat
5% dog
Input Deep Learning
Feature extraction + Classification Output
Based on: Jagreet Kaur Gill, “Log Analytics With Deep Learning And Machine Learning” April 28, 2017 Public
Massive metrology data & deep learning models further
Public
improve OPC accuracy in customer case Slide 6
• Big data improve pattern coverage & enhance model accuracy
• Deep Learning Model has more benefits with big data vs Traditional Model
Model accuracy improvement
with big data
(1 / Model Error, 3-sigma, nm-1)
0.4
Model Accuracy
0.3
Deep Learning Model
Traditional Model
0.2
~16,000 verification gauges
Regular hole patterns
0.1
0
0 600 800 1000 2000 8000 16000
Source: Jeff Dean, Google, “Trends and developments in deep learning”, Jan’17 Calibration gauge number
Public
High speed e-beam metrology and large field of view
Public
Slide 7
Excellent precision across large field of view
#1 #2 #3
eP5 Field of View
(12um x 12um)
Throughput advantage over CD-SEM
#4 #5 #6
CD-SEM CD-SEM eP5 68
Resolution 1 nm 1 nm
Field of View
Metrology Throughput
Current 8 pA 250 pA
(1um x 1um) Scan Rate 16 MHz 100 MHz
Field of view 1 um 12 um
12
#7 #8 #9 1
CD-SEM eP5 eP5
(1um FOV) (1um FOV) (12um FOV)
Public
Improving OPC model accuracy by 2X on DRAM Public
enabled by deep learning, fast e-beam, and metrology processing Slide 8
Model prediction error (RMS) eP5-MXP metrology
validated with >150k Edge Placement (EP) Gauges Improving 2D metrology accuracy
with contour based measurement
-32%
• Systematic error removal
• Random noise reduction
• Improved pattern coverage
-18% 1.3
nm
• Capture unknown Irregular shape Center shift
physical effects Systematic error reduction EP gauges capture
without shape fitting pattern shift
Deep learning resist model
Improve OPC model accuracy
Baseline metrology eP5-MXP eP5-MXP
w/o deep learning w/o deep learning
Error
with deep learning 0
(CD gauges) 3.3x (CD+EP gauges) 3.3x (CD+EP gauges)
~18 hours ~3 hours
Metrology collection time Metrology collection time Gauges
3x images
OPC Model accuracy study using high volume contour based gauges and deep learning on memory device, Young-Seok Kim et al., SPIE 2019, 10959-37 Public
Better accuracy of lithography models by deep learning
Public
Enabled by fast e-beam metrology and physical based models Slide 9
Model Prediction Accuracy
Physical driven training using (RMS in nm)
physics based lithography models EUV Cases: 7 nm and 5 nm logic
Stability
1.0
0.7
0.6
ASML Deep 0.4
0.5 0.7
Learning model 0.34 0.32
Physical Resist Data expansion
1D 2D 1D 2D
Resist surface through simulated
Shrinkage stress contours DUV Cases: 7 nm and 5 nm logic
1.2 1.3
Data-driven training based on fitting 1.1 1.1
1.0
0.8
spec and wafer measurements 0.7
0.4
Accuracy
EP SET1 EP SET2 1D 2D
with Deep Learning
without Deep Learning
Large volume wafer metrology data,
further enhanced by fast e-beam
“EP SET1” Edge Placement Gauge Set 1
Public
“Moore’s Law” of Computational Lithography
Runtime and cost of OPC increases node-by-node Public
Slide 10
Technology Node 28nm 22nm 14nm 10nm 7nm 5nm
Production start 2011 2013 2014 2016 2017 2019
Average transistor density (billion/cm2) 1.17 1.63 2.34 3.75 6.25 10.71
Number of critical layer masks 18 24 33 37 47 66
Normalized OPC runtime per layer per unit area 1 1.4 2 2.7 4 5.6
Node-on-node OPC runtime trends
6.0
OPC runtime per layer (normalized)
5.0 Total computational litho
capacity worldwide on the
4.0
order of 10 petaflops
3.0
2.0
1.0
0.0
28nm 22nm 14nm 10nm 7nm 5nm Public
Deep learning inverse model speeds up full-chip OPC
Public
by providing a good starting point Slide 11
Design target Inverse mask image
Selected clips
Complex, Iterative Deep Learning Inverse Model
Optimization
Training on clips Training input Training input
Inference on full chip Design target Inverse mask image
Full chip layout Full chip layout (Post-OPC)
Deep Learning Inverse Model
Public
Improving training pattern coverage
Public
using machine-learning-based pattern selection Slide 12
Full-chip Layout Pattern Library Machine-learning-based pattern selection Training Pattern Set
Pattern Representative
… Feature Clustering Selection
Pattern Generation
Collector
…
…
Normalized RMS between prediction
and ground-truth (inverse solution) Critical PV-band (>15% CD error ) Comparison
This Manual This Manual
Random selections Random selections
method selection method selection
Full-chip application of machine learning SRAFs on DRAM case using auto pattern selection K. Chen et al., SPIE 2019, 10961-37 Public
Deep learning SRAF improves full-chip DoF by 24%
for DRAM contact hole layer, validated on wafer Slide 13
Newron SRAF places more accurate assist Newron SRAF wafer validation
features to remove the process window limiter shows 24% DoF improvement
Off-nominal
OPC Mask
(-40nm defocus)
9
Exposure Latitude (%)
8
7
6
5 24%
Baseline solution
4
(Rule-Based SRAF) 3
2
1
0
-10 40 90 140
Depth Of Focus (nm)
Newron SRAF
Baseline Newron
(Deep Learning)
(RB-SRAF) SRAF
DOF @ 5%
102 nm 126 nm
EL
Full-chip application of machine learning SRAFs on DRAM case using auto pattern selection K. Chen et al., SPIE 2019, 10961-37 Public
Leverage confluence of new technologies to meet OPC
Public
technology and cost requirements Slide 14
2016 2017 2018 2019 2020 2021
Inverse OPC Inverse OPC (CTM+) Inverse with phase control
(CTM) Intel DL Boost Deep Learning Inverse Hardware Accel. (tentative)
14 10 nm
Skylake Cascade Lake Cooper Lake Ice Lake
Spring Crest
Pascal Volta Turing Next Gen?
Mask making
Mask writer Mask infrastructure is ready
& inspection inspection for inverse OPC &
available curvi-linear masks
Multi-beam Mask Writer available Public
Context-aware control extends holistic solutions
Public
Slide 15
Lithography Step N-2
Etch
Dep
CMP
Lithography Step N-1
Etch
Dep
CMP
Lithography Step N
Context Data Corrections
Litho InSight and Pattern Fidelity Enhancements
Leveling, Alignment, Metrology Overlay, Focus, pattern defect control and scanner/etcher co-optimization
Metrology design & Model, sampling & Monitoring &
setup control setup analytics Public
Leverage machine learning to address wafer-to-wafer
Public
variation induced by different wafer process routes Slide 16
16
Overlay Variations
Wafer to Wafer Variation
Context based Overlay Control results
ASML Machine
Learning Model
Every month > 100,000 wafers
exposed per scanner
Different Process Route
Correlate wafer-to-wafer variation
Number of Wafer Routes
Variation source
100,000,000 to process context and apply run-
10,000,000
1,000,000
to-run control with context-based
100,000 grouping
10,000
1,000
100
10
A novel patterning control strategy based on
1
1context 2contexts 3contexts … All real-time fingerprint recognition and adaptive
contexts
wafer level scanner optimization
H. E. Hakli et al., SPIE 2018, 10585:105851N
Public
Predict dense alignment from dense leveling data
hybrid metrology enabled by machine learning Public
Slide 17
Known input Type 1
• Leveling &
Alignment all wafers
• Overlay or alignment
sampling from same
wafer (blue points)
Machine Learning
Prediction
Unknown (orange points)
• Overlay or alignment from
same wafer at different Pairing wafer leveling metrology from a lithographic
coordinate locations apparatus with deep learning to enable cost effective dense
wafer alignment metrology
E. Schmitt-Weaver & K. Bhattacharyya, SPIE 2019, 10961-7 Public
Holistic Lithography delivering significant customer value
Public
Lithography scanner with advanced control capability Slide 18
Etch and deposition tools
Applications
Patterning Control
Algorithms Data
Physical Models, Scatterometry, SEM,
Optimization, & other fab
Machine Learning equipment
Computational Optical and e-beam metrology
lithography and metrology
Public
Major trends in semiconductor-enabled computing
Public
Autonomous decisions Slide 19
Immersive experience
Applications
Moore’s Law
Performance Connectivity
Data Value
Algorithms Cost Data Real-time
Deep Learning
Volume
Public
Public
