Topology Optimization for Computational Fabrication

Jun Wu
Depart. of Design Engineering
j.wu-1@tudelft.nl
 Copenhagen
Mechanical Engineering

Delft
Industrial Design Engineering

Munich Computer Science

2
Bone Chair by Joris Laarman

3
 Back

Seat

4
 5
www.jorislaarman.com
Schedule

• Basics of Topology Optimization (45’)

• Break (15’)
• Topology Optimization for Additive Manufacturing (45’)
• Break (15’)
• Exercises and Assignment (45’)

6
Topology Optimization Examples

Frustum Inc. Airbus APWorks, 2016 Qatar national convention

7
Full-scale aircraft wing design

Aage et al., Nature 2017

8
Classes of Structural optimization: Sizing, Shape, Topology

Initial

Optimized

Sizing Shape Topology

9
 A Toy Problem

• Design the stiffest shape, by placing 𝟔𝟎 Lego blocks into a grid of 𝟐𝟎 × 𝟏𝟎

10

20

10
A Toy Problem: Possible Solutions

• Number of possible designs

200!
– 𝐶 200,60 = = 7.04 × 1051
60! 200−60 !

• Which one is the stiffest?

Design C

Design A Design B
11
A Toy Problem: Possible Solutions

• Which one is the stiffest?

Design C

Design A Design B
12
A Toy Problem: Possible Solutions

• Which one is the stiffest?

Design C

Design A Design B
13
Topology Optimization Animation

14
Topology Optimization

1 𝑇 1 1
Minimize: 𝑐= 𝑈 𝐾𝑈 Elastic energy 𝑐= 𝑓𝑢 = 𝑘𝑢2
2 2 2
Subject to: 𝐾𝑈 = 𝐹 Static equation 𝑘𝑢 = 𝑓

𝑘
𝑢
𝑓

15
Topology Optimization

1 𝑇
Minimize: 𝑐= 𝑈 𝐾𝑈 Elastic energy
2
Subject to: 𝐾𝑈 = 𝐹 Static equation

1 (solid)
𝜌𝑖 = , ∀𝑖 Design variables
0 (void)
g = 𝑖 𝜌𝑖 − 𝑉0 ≤ 0 Volume constraint

1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1
1 1 1 1 1
1 1 1 1 1
1 1 1 1
1 1 1 1
1 1 1 1 1
1 1 1 1 1
1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1

16
Topology Optimization

1 𝑇 Compute
Minimize: 𝑐= 𝑈 𝐾𝑈 displacement
2
(KU=F)
Subject to: 𝐾𝑈 = 𝐹
1 (solid)
𝜌𝑖 = , ∀𝑖 Sensitivity
0 (void) analysis
g = 𝑖 𝜌𝑖 − 𝑉0 ≤ 0
Update 𝜌𝑖
1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1
1 1 1 1 1
1 1 1 1 1
1 1 1 1
1 1 1 1 Converged?
1 1 1 1 1 No
1 1 1 1 1
1 1 1 1 1 Yes
1 1 1 1 1 1 1 1 1 1 1 1 1

17
Topology Optimization Animation

18
Relaxation: Discrete to Continuous

1 𝑇
Minimize: 𝑐= 𝑈 𝐾𝑈
2
Subject to: 𝐾𝑈 = 𝐹
1 (solid)
𝜌𝑖 = , ∀𝑖 𝜌𝑖 ∈ [0 , 1]
0 (void)
g = 𝑖 𝜌𝑖 − 𝑉0 ≤ 0

• Motivation: (Difficult) binary problem → (easier) continuous problem

19
Material Interpolation

• Material properties: Young’s modulus 𝐸, and Poisson's ratio 𝜈

• SIMP interpolation (Solid Isotropic Material with Penalization)
– 𝐸𝑖 = 𝜌𝑖 𝑝 𝐸
– 𝑝 ≥ 1, typically 𝑝 = 3

20
Sensitivity Analysis

• Sensitivity: The derivative of a function with respect to design variables

𝜕𝑐 𝑝 𝑝−1 𝑇 1 𝑇
• = − 𝜌𝑖 𝑢𝑖 𝐾 𝑢𝑖 Minimize: 𝑐= 𝑈 𝐾𝑈
𝜕𝜌𝑖 2 2
– Smaller than zero Subject to: 𝐾𝑈 = 𝐹
𝜕𝑔 𝜌𝑖 ∈ [0 , 1]
• =1
𝜕𝜌𝑖
g = 𝑖 𝜌𝑖 − 𝑉0 ≤ 0

22
Design Update
Compute
• Mathematical programming displacement
– Interior point method (IPOPT package) (KU=F)
– The method of moving asymptotes (MMA)

Sensitivity
• Optimality criterion analysis
𝜕𝑐
– If “− ” is large, increase 𝜌𝑖
𝜕𝜌𝑖
– Otherwise, decrease 𝜌𝑖 Update 𝜌𝑖

– How to determine large or small?

Converged?
– Bisection search for a threshold No

Yes
23
Checkerboard Patterns

24
Sensitivity Filtering by a Convolution Operation
Compute
displacement
(KU=F)

Sensitivity
analysis
Convolution
Update 𝜌𝑖

Converged?
No

Yes
25
Convolution Operation

http://cse19-iiith.vlabs.ac.in/theory.php?exp=neigh
26
Demo

• www.topopt.dtu.dk

27
Topology Optimization
90%
1 𝑇 Compute
Minimize: 𝑐= 𝑈 𝐾𝑈 displacement
2
(KU=F)
Subject to: 𝐾𝑈 = 𝐹
𝜌𝑖 ∈ [0,1], ∀𝑖
Sensitivity
g = 𝑖 𝜌𝑖 − 𝑉0 ≤ 0 analysis

Update 𝜌𝑖
1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1
1 1 1 1 1
1 1 1 1 1
1 1 1 1
1 1 1 1 Converged?
1 1 1 1 1 No
1 1 1 1 1
1 1 1 1 1 Yes
1 1 1 1 1 1 1 1 1 1 1 1 1

28
Geometric Multigrid: Solving 𝐾𝑢 = 𝑓

• Successively compute approximations 𝑢𝑚 to the solution u = lim 𝑢𝑚

𝑚→∞
• Consider the problem on a hierarchy of successively coarser grids to
accelerate convergence

Ωℎ Relax 𝐾 ℎ 𝑢ℎ ≈ 𝑓 ℎ Relax 𝐾 ℎ 𝑢ℎ ≈ 𝑓 ℎ
Residual 𝑟 ℎ = 𝑓 ℎ − 𝐾 ℎ 𝑢ℎ Correct 𝑢ℎ ← 𝑢ℎ + 𝑒 ℎ

Restrict Interpolate
Ω2ℎ 𝑟 2ℎ = 𝑅ℎ2ℎ 𝑟 ℎ ℎ ℎ 2ℎ
𝑒 = 𝐼2ℎ 𝑒

Ω4ℎ

W. Briggs, A multigrid tutorial, 2000

Solve 𝐾 4ℎ 𝑒 4ℎ = 𝑟 4ℎ

⋮ 29
Memory-Efficient Implementation on GPU

• On-the-fly assembly
– Avoid storing matrices on the finest level
• Non-dyadic coarsening (i.e., 4:1 as opposed to 2:1)
– Avoid storing matrices on the second finest level
Ωℎ

Ω2ℎ

Ω4ℎ

Wu et al., TVCG’2016
Dick et al., SMPT’2011
⋮ 30
High-Resolution Design

Resolution: 621×400×1000
#Element 14.2m
Time: 12 minutes

31
33
Negative Poisson's ratio Negative thermal expansion Electric actuator
Larsen et al. 1997 Sigmund &Torquato 1996 Sigmund 2000

Natural convection Fluid flow

34
Alexandersen et al. 2016 Maute & Pingen
A General Formulation

Minimize: 𝑐(𝜌)
Solve state
Subject to: 𝜌𝑖 ∈ [0,1], ∀𝑖 equation
𝑔𝑖 (𝜌) ≤ 0
Sensitivity
analysis

Update 𝜌𝑖
1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1
1 1 1 1 1
1 1 1 1 1
1 1 1 1
1 1 1 1 Converged?
1 1 1 1 1 No
1 1 1 1 1
1 1 1 1 1 Yes
1 1 1 1 1 1 1 1 1 1 1 1 1

35
Outline

• Basics of Topology Optimization

• Topology Optimization for Additive Manufacturing

36
Additive Manufacturing: Complexity is free

TU Delft & MX3D, 2015 Joshua Harker Scott Summit

37
Complexity is free? … Not really!

• Printer resolution: Minimum geometric feature size

• Layer-upon-layer: Supports for overhang region
• Shell-infill composite

Tiny details Supports Infill

Ralph Müller

Concept Laser GmhH mpi.fs.tum.de

Paul Crompton
38
Outline

• Basics of Topology Optimization

• Topology Optimization for Additive Manufacturing

– Geometric feature control by density filters
– Geometric feature control by alternative parameterizations

39
Messerschmidt-Bölkow-Blohm (MBB) beam

40
Messerschmidt-Bölkow-Blohm (MBB) beam

41
Geometric feature control by density filters
(An incomplete list)
Reference

Minimum feature size, Guest’04 Coating structure, Clausen’15

Self-supporting design, Langelaar’16 Porous infill, Wu’16 42

Infill in 3D Printing: Regular Structures

3dplatform.com

www.makerbot.com
43
Infill in Bone: Porous Structures

44
Can we apply the principle of bone to 3D printing?

45
Topology Optimization Applied to Design Infill

No similarity in structure

Topology optimization Infill in the bone

46
Topology Optimization Applied to Design Infill

• Materials accumulate to “important” regions

• The total volume 𝑖 𝜌𝑖 𝑣𝑖 ≤ 𝑉0 does not restrict local material
distribution

Infill by standard Infill in the bone

topology optimization
47
Bone-like Infill in 2D

Cross-section of a human femur

48
Approaching Bone-like Structures: The Idea

• Impose local constraints to avoid fully solid regions

1
Min: c = 𝑈 𝑇 𝐾𝑈
2 𝜌𝑖 = 0.0
s.t. : 𝐾𝑈 = 𝐹
𝜌𝑖 ∈ [0,1], ∀𝑖
𝑖 𝜌𝑖 ≤ 𝑉0
𝜌𝑖 = 0.6
𝜌𝑖 ≤ 𝛼, ∀𝑖

𝑗∈𝛺𝑖 𝜌𝑗
𝜌𝑖 = 1.0
𝜌𝑖 =
𝛺𝑖 𝑗∈𝛺𝑖 1

Local-volume measure 49
 Constraints Aggregation (Reduce the Number of Constraints)

1
𝜌𝑖 ≤ 𝛼, ∀𝑖 max 𝜌𝑖 ≤ 𝛼 lim 𝜌 𝑝 = 𝜌𝑖 𝑝 𝑝 ≤𝛼
𝑖=1,…,𝑛 𝑝→∞ 𝑖

Too many constraints! A single constraint A single constraint

But non-differentiable and differentiable
Approximated with 𝑝 =16

50
Optimization Process: The same as in the standard topopt

• Impose local constraints to avoid fully solid regions

1 𝑇 Compute
Min: c= 𝑈 𝐾𝑈 displacement
2
(KU=F)
s.t. : 𝐾𝑈 = 𝐹
𝜌𝑖 ∈ [0,1], ∀𝑖
Sensitivity
𝑖 𝜌𝑖 ≤ 𝑉0 analysis
𝜌𝑖 ≤ 𝛼, ∀𝑖
Update 𝜌𝑖

𝑗∈𝛺𝑖 𝜌𝑗
𝜌𝑖 = Converged?
𝛺𝑖 𝑗∈𝛺𝑖 1 No
Local-volume measure Yes 51
A Test Example

52
 Effects of Filter Radius and Local Volume Upper Bound

R=6

𝛼, 𝑐 = (0.6, 76.9) (0.5, 96.0) 0.4, 130.0

R=12

(0.6, 73.9) (0.5, 91.2) 0.4, 119.8

53
 Local + Global Volume Constraints

R=6

𝛼, 𝛼𝑡𝑜𝑡𝑎𝑙 , 𝑐 = (0.6, 0.56, 76.9) (0.6, 0.50, 79.1) 0.6, 0.40, 94.0

54
Result: 2D Animation

55
Result: 2D Animation

56
Robustness wrt. Force Variations

• Porous structures are significantly stiffer (126%) in case of force variations

Total volume constraint Local volume constraints

c = 30.54 c = 36.72
c’= 45.83 c’ =36.23
57
Robustness wrt. Material Deficiency

• Porous structures are significantly stiffer (180%) in case of material deficiency

Total volume constraint Local volume constraints

c = 76.83 c = 93.48
c’ =242.77 c’= 134.84

58
Bone-like Infill in 3D

Infill in the bone Optimized bone-like infill

Wu et al., TVCG’2017
59
FDM Prints

60
Chair

61
Video

63
Geometric feature control by density filters
(An incomplete list)
Reference

Minimum feature size, Guest’04 Coating structure, Clausen’15

Self-supporting design, Langelaar’16 Porous infill, Wu’16 64

Concurrent Shell-Infill Optimization

Wu et al., CMAME 2017

65
Outline

• Basics of Topology Optimization

• Topology Optimization for Additive Manufacturing

– Geometric feature control by density filters
– Geometric feature control by alternative parameterizations

66
 Geometric feature control by alternative parameterizations
(An incomplete list)

11111111 11111
11 1 11
11 1 11
11
1 1
1 11
11
Voronoi cells, Lu’14
1 1 11
11 1 11
11 1 11
11 1 11
11111111 11111 Skin-frame, Wang’13
Reference: Voxel discretization

Offset surfaces, Musialski’15 Ray representation, Wu’16 Adaptive rhombic,67Wu’16

Overhang in Additive Manufacturing

• Support structures are needed beneath overhang surfaces

https://www.protolabs.com/blog/tag/direct-
metal-laser-sintering/ 69
Support Structures in Cavities

• Post-processing of inner supports is problematic

Inner supports

Print Outer supports

direction

70
Infill & Optimization Shall Integrate

Solid, Optimized, With infill,

Unbalanced Balanced Unbalanced
71
The Idea

• Rhombic cell: to ensure self-supporting

• Adaptive subdivision: as design variable in optimization

Print
direction

Rhombic cell Adaptive subdivision

72
Self-Supporting Rhombic Infill: Workflow

0.4X
Compute
Initialization displacement
(KU=F)

Sensitivity
analysis
Optimization
Update
subdivision

Converged?
No
Yes 73
Self-Supporting Rhombic Infill: Results

• Optimized mechanical properties, compared to regular infill

• No additional inner supports needed

Optimization process Reference Print

Wu et al., CAD’2016
74
 Under same force (62 N) Under same displacement (3.0 mm)
Mechanical Tests

Dis. Force
2.11 mm 90 N

Dis. Force
4.08 mm 58 N
75
Outline

• Basics of Topology Optimization

• Topology Optimization for Additive Manufacturing
– Geometric feature control by density filters
– Geometric feature control by alternative parameterizations

76
Bone-inspired infill Self-supporting infill
Topology Optimization
• Lightweight
• Free-form shape
• Customization
• Mechanically optimized

Additive Manufacturing
• Customization
• Geometric complexity

77
Thank you for your attention!

Jun Wu
Depart. of Design Engineering, TU Delft
www.jun-wu.net
j.wu-1@tudelft.nl