Introduction: Artificial Intelligence aided Design and Manufacturing group
1
Design
Process
Performance
Inverse design
Design
Process
Performance
Inverse model
4
Cuttlefish
GrabCAD
Target
Ours
Appearance Fabrication
Design
Performance
3D printer colors
and layer deposition
Appearance of the final product
Sumin et al. Geometry aware scattering compensation for 3D printing. SIGGRAPH 2019.
5
Shape from Release
Design
Performance
The shape of the 3D
printed object
Release curve of the material
Panetta, Mohammadian, Luci, Babaei. Shape from release: inverse design and fabrication of controlled release structures, SIGGRAPH ASIA 2022.
6
Cucerca, Didyk, Seidel, Babaei. Computational image marking on metals via laser induced heating. SIGGRAPH 2020.
Computational image marking on metals via laser induced heating
7
Cucerca, Didyk, Seidel, Babaei. Computational image marking on metals via laser induced heating. SIGGRAPH 2020.
Design
Performance
Laser parameters
Printed image
8
Cucerca, Didyk, Seidel, Babaei. Computational image marking on metals via laser induced heating. SIGGRAPH 2020.
Neural Inverse design
Design
Neural surrogate
Performance
Inverse model
Autoinverse: Uncertainty Aware Inversion of Neural�Networks
Navid Ansari
Hans-Peter Seidel
Nima Vahidi Ferdowsi
Vahid Babaei
10
Native forward process (NFP)
11
Design
Performance
NFP,�e.g., physics simulation
Neural surrogate model
12
Design
Performance
NFP,�e.g., physics simulation
Neural surrogate
Design
Performance
Neural inverse design
13
Design
Performance
NFP,�e.g., physics simulation
Neural surrogate
Design
Performance
Neural Inversion
The gap
14
Design
Performance
Design
Performance
Neural Inversion
NFP Inversion
Inversion methods
15
Train
Neural adjoint (NA)
Ren et. al. Benchmarking deep inverse models over time, and the neural-adjoint method, NeurIPS 2020
Inversion methods
16
Freeze weights and biases
Neural adjoint (NA)
Ren et. al. Benchmarking deep inverse models over time, and the neural-adjoint method, NeurIPS 2020
Inversion methods
17
Freeze weights and biases
Optimize input
Neural adjoint (NA)
Ren et. al. Benchmarking deep inverse models over time, and the neural-adjoint method, NeurIPS 2020
Inversion methods
18
Freeze weights and biases
Optimize input
Targeted value
Inversion methods
19
Freeze weights and biases
Optimize input
Targeted value
Inversion methods
20
Freeze weights and biases
Optimize input
Targeted value
21
NFP
Design
Performance
Native forward process (NFP)
22
Sample the NFP
NFP
Design
Performance
23
Surrogate inversion
Design
Performance
Surrogate
24
Surrogate inversion
Design
Performance
Surrogate
Sparse�sampling
of NFP
25
Surrogate inversion
Performance
Design
Surrogate
Noisy
region
of NFP
Sparse�sampling
of NFP
26
Our solution: find inversions near NFP
Design
Performance
Solution: uncertainty information
27
Design
Performance
Conventional NN
Neural adjoint (NA)
Target
Solution: uncertainty information
28
Design
Bayesian NN
Uncertainty
Autoinverse: neural adjoint optimization with uncertainty
Design
Performance
Conventional NN
Target
Performance
Target
Neural adjoint (NA)
29
Surrogate inversion
Performance
*
*
*
Design
Surrogate
Epistemic
uncertainty
Aleatoric
uncertainty
Spectral printing
30
https://en.wikipedia.org/wiki/Visible_spectrum
Spectral printing
31
Physical reproduction
Original painting
© Azadeh Asadi
Forward model: Ink to color spectrum
32
Ink ratio
Color spectrum
Design
Performance
Inversion: Color spectrum to ink
33
Ink ratio
Color spectrum
Design
Performance
Auto inverse incorporates feasibility
34
-1
0
1
2
Valid design
35
-1
0
1
2
NA
Valid design
Auto inverse incorporates feasibility
Auto inverse incorporates feasibility
36
-1
0
1
2
NA
NA boumdry loss
Valid design
Auto inverse incorporates feasibility
37
-1
0
1
2
NA
UANA
NA boundary loss
Valid design
Inversion: soft robot
38
Sun et. al. Amortized Synthesis of Constrained Configurations Using a Differentiable Surrogate, NeurIPS 2021
Xue et. al. Amortized finite element analysis for fast pde-constrained optimization. In International Conference on Machine Learning, PMLR 2020
Inversion: soft robot
39
40 Controllable edges
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
20
19
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
40
39
Sun et. al. Amortized Synthesis of Constrained Configurations Using a Differentiable Surrogate, NeurIPS 2021
Xue et. al. Amortized finite element analysis for fast pde-constrained optimization. In International Conference on Machine Learning, PMLR 2020
Inversion: soft robot
40
Design
Performance
Soft robot
shape
Controllable edge
Target
Inversion: soft robot
41
Design
Performance
Soft robot
shape
Controllable edge
Target
Inversion: soft robot
42
Design
Performance
Soft robot
shape
Controllable edge
Target
Inversion: soft robot
43
Design
Performance
Soft robot
shape
Controllable edge
Target
Autoinverse: Imperfect data set
44
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
20
19
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
40
39
0
0
0
0
0
0
0
Autoinverse: Imperfect data set
45
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
20
19
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
40
39
0
0
0
0
0
0
0
0
Autoinverse: Imperfect data set
46
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
20
19
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
40
39
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
Autoinverse: Imperfect data set
47
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
20
19
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
40
39
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
Autoinverse: Initialization free
48
NA with correct
initialization
49
Autoinverse: Initialization free
NA without correct
initialization
Autoinverse: Initialization free
50
Autoinverse with
correct initialization
Autoinverse without
correct initialization
Limitations
…
Limitations
…
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
?
Design
e.g., photonic metasurface
Performance
e.g., scattering cross-section
Inverse model
?
Mixed-integer neural inverse design
Target
Reproduced
Design
e.g., photonic metasurface
Performance
e.g., scattering cross-section
Inverse model
?
Target
Reproduced
Design
e.g., photonic metasurface
Performance
e.g., scattering cross-section
Inverse model
?
Target
Reproduced
Design
e.g., photonic metasurface
Performance
e.g., scattering cross-section
Inverse model
?
Target
Reproduced
Design
e.g., photonic metasurface
Performance
e.g., scattering cross-section
Inverse model
?
Multi-joint robot
59
Target
Design:
Performance:
Position of the tip
Angle of the joints
Multi-joint robot
60
Target
Design:
Performance:
Position of the tip
Angle of the joints
Multi-joint robot
61
Target
Design:
Performance:
Position of the tip
Angle of the joints
Multi-joint robot
62
Target
Design:
Performance:
Position of the tip
Angle of the joints
Multi-joint robot
63
Target
Design:
Performance:
Position of the tip
Angle of the joints
Multi-joint robot
64
Target
Design:
Performance:
Position of the tip
Angle of the joints
Multi-joint robot
65
Target
Design:
Performance:
Position of the tip
Angle of the joints
Multi-joint robot
66
Error type | Neural adjoint method (NA) | NA with Autoinverse (UANA) | Invertible neural networks |
Surrogate | (1.99 ± 0.05) × 10-8 | (9.13 ± 6.08) × 10-7 | (2.04 ± 0.017) × 10-13 |
Ardizzone et. al. Analyzing Inverse Problems with Invertible Neural Networks, ICLR 2019
Multi-joint robot
67
Error type | Neural adjoint method (NA) | NA with Autoinverse (UANA) | Invertible neural networks |
Surrogate | (1.99 ± 0.05) × 10-8 | (9.13 ± 6.08) × 10-7 | (2.04 ± 0.017) × 10-13 |
Smaller surrogate error for NA and INN
Ardizzone et. al. Analyzing Inverse Problems with Invertible Neural Networks, ICLR 2019
Multi-joint robot
68
Error type | Neural adjoint method (NA) | NA with Autoinverse (UANA) | Invertible neural networks |
Surrogate | (1.99 ± 0.05) × 10-8 | (9.13 ± 6.08) × 10-7 | (2.04 ± 0.017) × 10-13 |
NFP | (3.24 ± 0.51) × 10-4 | (3.21 ± 1.48) × 10-6 | (9.48 ± 0.021) × 10-3 |
Ardizzone et. al. Analyzing Inverse Problems with Invertible Neural Networks, ICLR 2019