Double-Patterning Aware DSA Template

Guided Cut Redistribution for Advanced

1-D Gridded Designs
Zhi-Wen Lin and Yao-Wen Chang

National Taiwan University

ISPD 2016

1
 Outline
Introduction

Contributions

Problem Formulation

Algorithm

Experimental Results

Conclusions
2
 Directed Self-Assembly Block Copolymer
․ The directed self-assembly (DSA) technology has
emerged as a promising candidate for sub-10 nm
designs

Self-assembly
+ Topographical
= Smaller and
block copolymer templates denser patterns
Template

Feature

(a) (b) (c) (d)

[Xiao et al., SPIE’13]

3
 1-D Style Layout Design with DSA
․ DSA has applications in
 Contact hole patterning [Wong et al., SPIE’12]
 Cut printing of 1-D gridded designs [Ma et al., SPIE’14]
․ 1-D layouts are popular for their regularity [Lam et al.,
SPIE’11]
․ DSA can be used to pattern dense cuts in 1-D layouts

Real wire Dummy wire (gap)

DSA templates The resulting 1-D layout

4
 Challenge of Cut Printing with DSA
․ The size of the feasible-template library is limited
․ Conflict cost: two close templates might interfere with
each other [Ma et al., SPIE’14]
 # cut pairs with spacing conflicts should be minimized

Complex template
Poor printability

[Ma et al., SPIE’14]

5
 Cut Redistribution
․ Redistribute cuts to reduce conflict costs
․ Wire cost: the total length of extended real wires should
be minimized considering the performance impact
․ Given a layout, minimize its conflict cost and wire cost
 DSA template guided cut redistribution problem (DSA-TCR)

Extended wire

vs.

6
 Double Patterning
․ Double patterning can be adopted to further improve
the image fidelity of templates in cut printing with DSA
[Ma et al., SPIE’14]

(a) Dense templates (b) Double patterning (c) Patterned cuts

Mask 1 decomposition
Mask 2

․ Integrate DSA cut redistribution with double patterning

 Conflict costs and wire costs can be further reduced

7
 Outline
Introduction

Contributions

Problem Formulation

Algorithm

Experimental Results

Conclusion
8
 Main Contributions: Algorithms
․ A linear-time optimal algorithm for the special case with
limited contiguous rows is identified
 Simultaneously minimize conflict costs and wire costs
․ A linear-time algorithm for decomposing a general
problem into subproblems conforming to the special case
 With double patterning, a criterion for non-interfering
subproblems is proposed Subproblem
of special case

Different mask
No conflict Large distance
No conflict
Subproblems with contiguous rows
9
 Main Contributions: Performance
․ The total wire cost of a derived template distribution is
linear to the number of cuts

․ Experimental results show that our algorithm can resolve

all conflicts, with smaller running times, compared with
the related previous works

10
 Outline
Introduction

Contributions

Problem Formulation

Algorithm

Experimental Results

Conclusion
11
 Problem Formulation
․ Given
 A library of templates, design rules, a layout, and original
distribution of cuts
․ Output
 A template distribution patterning required cuts
 A mask assignment of all derived templates
․ Objective
 An ordered optimization on the conflict cost first and then the
wire cost
 Our algorithm can readily apply to the linear combination of
conflict cost and wire cost

12
 Outline
Introduction

Contributions

Problem Formulation

Algorithm

Experimental Results

Conclusion
13
 Overall Flow
․ Propose an algorithm (DPDSA-Core), incorporating
DSA with double patterning
 Divide a general problem into subproblems

Input

Subproblem candidate construction

Solving every subproblem candidate

Subproblem candidate selection

Output
14
Subproblem Candidate Construction

Input

Subproblem candidate construction

Solving every subproblem candidate

Subproblem candidate selection

Output

15
 Special Case with Contiguous Rows
․ Construct subproblems conforming to a special case to
reduce conflicts between two adjacent subproblems
 Each subproblem consists of contiguous rows with # rows ≤ 𝑟𝑀
 Note that 𝑟𝑀 + 1 ≥ the minimum spacing
 Templates in a subproblem are patterned by the same mask
 Adjacent subproblems are patterned by different masks

𝑟𝑀 + 1
Different mask
No conflict No conflict

16
 Solutions of Partitions
․ Subproblems with fewer rows can have less internal
wire cost and conflict cost

․ Also consider the subproblems with fewer rows when a

subproblem partition with fewer inter-costs exists

Subproblem with
fewer rows

․ A linear-time partition algorithm is proposed

17
Solving Subproblem Candidates

Input

Subproblem candidate construction

Solving every subproblem candidate

Subproblem candidate selection

Output

18
 Overview
․ Given a fixed rM, solve a subproblem optimally in linear
time to the number of gaps in the subproblem
…

…
Divide a sub-problem
vh,j vh+1,j vh+2,j into multiple columns
vh,j+1 vh+1,j+1 vh+2,j+1
vh,j+2 vh+1,j+2 vh+2,j+2
…

…
For every column,
h h+1 h+2 generate feasible
gap solutions

g1 Distribution Template
g2 specifier dh,j specifier th,j
h
․ Find a path of gap solutions with minimized conflict
costs and wire costs
19
 Data Structure of Column Solutions
․ Template specifier specifies the topologies of cuts in
the current column and restrict the templates next to it
․ To consider cut conflict completely, store the
distribution of previous cuts in a column solution

Distribution Template
specifier dh,j specifier th,j

Previous Horizontal
cut column
distribution index
․ For every column solution, find the best compatible
previous column solution

20
 Column-wise Searching
․ Construct column solutions from left to right
․ For every pair of a column solution and its predecessor,
evaluate the current conflict costs and wire costs
Add 1 conflict Add 1 grid wire cost

g1
g2 Right cut of g2 hasn’t
addressed yet
Predecessor Current solution
․ Then trace back according to predecessors for a path

21
 Reduced Column-Space
․ Prune redundant columns to speed up the process
․ For ever gap, calculate the remaining columns within it
 First, calculate the possible range for cuts
 Scan from left to right, and use the information of possible
range to get the remaining columns

Remaining
Possible range columns
․ Restrict the complexity of the number of remaining
columns and the total wire cost to be linear
 Thus the running time grows linearly

22
Subproblem Candidate Selection

Input

Subproblem candidate construction

Solving every subproblem candidate

Subproblem candidate selection

Output

23
 Overview of Partition Algorithm
․ Divide the whole layout into regions
Subproblem candidate selection
Region construction R0
R1
Region solution R2
graph construction
R0
Optimal region solution
path determination
R1
Template distribution
construction R2
v1 v2 v3 v4 v5

24
 Region Solutions
․ Every region solution is a set of solved subproblems
Subproblem candidate selection
Region construction
R1 1
Region solution
graph construction R2 2

Optimal region solution R1 1

path determination
R2 2
Template distribution
construction

25
 Region Solution Graph
․ An edge exists between two region solutions when
they share a subproblem

Subproblem candidate selection R1

Region construction

Region solution R2
graph construction

Optimal region solution R1

path determination

Template distribution R2
construction

․ Inter-subproblem conflicts can occur only between

adjacent regions for the large enough height of regions
26
 Partition from Region Solution
․ Get subproblems by selected region solutions
․ Construct the whole solution by selected subproblems
Subproblem candidate selection 1 1 1
Region construction

Region solution 2 2 2
graph construction

Optimal region solution

path determination 3 3 3

Template distribution
construction 4 4 4

5 5 5
27
 Extension
․ A framework to relax the same-mask constraint in a row

Input

Connected component Terminal condition

construction
Not satisfied Satisfied
DPDSA-Core for every Conflicting component
conflicting component selection and merging

Output
Double-patterning for templates

28
 Outline
Introduction

Contributions

Problem Formulation

Algorithm

Experimental Results

Conclusion
29
 Experimental Setting
․ Platform: 2.13 GHz Linux workstation with 16 GB memory
․ Programming language: C++
․ Resource of benchmarks: the authors of [Xiao et al.,
SPIE’13]
․ Compared with
 [Xiao et al., SPIE’13] [Ou et al., GLSVLSI’15]
 Extend these works by a general high-level framework [Badr et al.,
SPIE’15] for DSA-multiple patterning hybrid lithography
 The framework consists of the two stages GP-MP (GM) and MP-
GP (MG)
 GP: group cuts into templates
 MP: Assign masks

30
 Conflict Costs
․ Our algorithm can resolve all conflicts
 [Xiao et al., SPIE’13]: S13
 [Ou et al., SPIE’15]: S15 (ILP-based algorithm), S15sp (ILP-
based speed-up heuristic), SC (set-cover-based algorithm)
 “-”: running time ≥ an hour

Conflict cost
Number Initial
GP-MP MP-GP
of cuts conflict
S13 S15sp S15 SC S13 S15sp S15 SC ours
50 12 0 0 0 0 0 0 0 1 0
100 46 0 0 0 0 0 1 0 0 0
200 83 0 0 0 0 3 0 0 4 0
500 223 0 0 - 0 4 0 0 5 0
1000 421 2 0 - 1 7 3 0 8 0
2000 855 0 0 - 2 17 2 0 11 0
Average 0.33 0 - 0.50 5.17 1.00 0 4.83 0
no conflict!!

31
 Wire Cost
․ Our wire cost is 6X smaller than S15sp with GP-MP
․ Except pure-ILP works, our wire costs are consistently
smaller than others with the same numbers of conflicts

Wire cost
Number of
GP-MP MP-GP
cuts
S13 S15sp S15 SC S13 S15sp S15 SC ours
50 48 8 6 3 13 1 1 0 1
100 109 29 24 30 48 5 2 6 6
200 272 58 48 63 150 27 11 23 10
500 665 153 - 166 235 33 14 29 27
1000 1466 307 - 289 542 48 32 45 49
2000 2647 593 - 576 1055 89 65 76 96
Average 867.83 191.33 - 187.83 340.50 33.83 20.83 29.83 31.50
Normalized
27.55 6.07 0.00 5.96 10.81 1.07 0.66 0.95 1.00
average

32
 Running Time
․ Our algorithm has better scalability than others

Running time (seconds)

Number
GP-MP MP-GP
of cuts
S13 S15sp S15 SC S13 S15sp S15 SC ours
50 0.01 0.09 0.22 0.01 0.01 0.08 0.04 0.01 0.01 Better
100 0.02 0.45 9.78 0.03 0.02 0.16 0.07 0.03 0.02 scalability
200 0.04 0.53 609 0.09 0.04 0.33 0.45 0.07 0.03
500 0.12 1.95 - 0.34 0.12 0.84 1.06 0.2 0.08
1000 0.31 3.33 - 1.22 0.32 1.84 6.12 0.64 0.15
2000 0.91 7.99 - 4.38 0.92 4.49 55.1 2.45 0.3
Average 0.24 2.39 - 1.01 0.24 1.29 10.47 0.57 0.10
Normalized
2.39 24.31 - 10.29 2.42 13.12 106.51 5.76 1.00
average

33
 Outline
Introduction

Contributions

Problem Formulation

Algorithm

Experimental Results

Conclusion
34
 Conclusions
․ A linear-time optimal algorithm for the special case of
DSA-TCR

․ A linear-time algorithm for decomposing general double

patterning DSA-TCR problems and reduce inter-
subproblem conflicts cost

․ Better experimental results compared with the previous

work

․ Future Work
 Incorporate DSA with multiple patterning technology

35
Thank You!

36