1 of 68

Introduction: Artificial Intelligence aided Design and Manufacturing group

1

2 of 68

Design

Process

Performance

3 of 68

Inverse design

Design

Process

Performance

Inverse model

4 of 68

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 of 68

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 of 68

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 of 68

7

Cucerca, Didyk, Seidel, Babaei. Computational image marking on metals via laser induced heating. SIGGRAPH 2020.

Design

Performance

Laser parameters

Printed image

8 of 68

8

Cucerca, Didyk, Seidel, Babaei. Computational image marking on metals via laser induced heating. SIGGRAPH 2020.

9 of 68

Neural Inverse design

Design

Neural surrogate

Performance

Inverse model

10 of 68

Autoinverse: Uncertainty Aware Inversion of NeuralNetworks

Navid Ansari

Hans-Peter Seidel

Nima Vahidi Ferdowsi

Vahid Babaei

10

11 of 68

Native forward process (NFP)

11

Design

Performance

NFP,�e.g., physics simulation

12 of 68

Neural surrogate model

12

Design

Performance

NFP,�e.g., physics simulation

Neural surrogate

Design

Performance

13 of 68

Neural inverse design

13

Design

Performance

NFP,�e.g., physics simulation

Neural surrogate

Design

Performance

Neural Inversion

14 of 68

The gap

14

Design

Performance

Design

Performance

Neural Inversion

NFP Inversion

15 of 68

Inversion methods

15

Train

Neural adjoint (NA)

Ren et. al. Benchmarking deep inverse models over time, and the neural-adjoint method, NeurIPS 2020

16 of 68

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

17 of 68

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

18 of 68

Inversion methods

18

Freeze weights and biases

Optimize input

Targeted value

19 of 68

Inversion methods

19

Freeze weights and biases

Optimize input

Targeted value

20 of 68

Inversion methods

20

Freeze weights and biases

Optimize input

Targeted value

21 of 68

21

NFP

Design

Performance

Native forward process (NFP)

22 of 68

22

Sample the NFP

NFP

Design

Performance

23 of 68

23

Surrogate inversion

Design

Performance

Surrogate

24 of 68

24

Surrogate inversion

Design

Performance

Surrogate

Sparse�sampling

of NFP

25 of 68

25

Surrogate inversion

Performance

Design

Surrogate

Noisy

region

of NFP

Sparse�sampling

of NFP

26 of 68

26

Our solution: find inversions near NFP

Design

Performance

27 of 68

Solution: uncertainty information

27

Design

Performance

Conventional NN

Neural adjoint (NA)

Target

28 of 68

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 of 68

29

Surrogate inversion

Performance

*

*

*

Design

Surrogate

Epistemic

uncertainty

Aleatoric

uncertainty

30 of 68

Spectral printing

30

https://en.wikipedia.org/wiki/Visible_spectrum

31 of 68

Spectral printing

31

Physical reproduction

Original painting

 

© Azadeh Asadi

32 of 68

Forward model: Ink to color spectrum

32

Ink ratio

Color spectrum

Design

Performance

33 of 68

Inversion: Color spectrum to ink

33

Ink ratio

Color spectrum

Design

Performance

34 of 68

Auto inverse incorporates feasibility

34

-1

0

1

2

Valid design

35 of 68

35

-1

0

1

2

NA

Valid design

Auto inverse incorporates feasibility

36 of 68

Auto inverse incorporates feasibility

36

-1

0

1

2

NA

NA boumdry loss

Valid design

37 of 68

Auto inverse incorporates feasibility

37

-1

0

1

2

NA

UANA

NA boundary loss

Valid design

38 of 68

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

39 of 68

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

40 of 68

Inversion: soft robot

40

Design

Performance

Soft robot

shape

Controllable edge

Target

41 of 68

Inversion: soft robot

41

Design

Performance

Soft robot

shape

Controllable edge

Target

42 of 68

Inversion: soft robot

42

Design

Performance

Soft robot

shape

Controllable edge

Target

43 of 68

Inversion: soft robot

43

Design

Performance

Soft robot

shape

Controllable edge

Target

44 of 68

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

45 of 68

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

46 of 68

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

47 of 68

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

48 of 68

Autoinverse: Initialization free

48

NA with correct

initialization

49 of 68

49

Autoinverse: Initialization free

NA without correct

initialization

50 of 68

Autoinverse: Initialization free

50

Autoinverse with

correct initialization

Autoinverse without

correct initialization

51 of 68

Limitations

  • The cost of calculating the uncertainty

52 of 68

Limitations

  • Quality of the uncertainty

53 of 68

54 of 68

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

?

Design

e.g., photonic metasurface

Performance

e.g., scattering cross-section

Inverse model

?

Mixed-integer neural inverse design

55 of 68

Target

Reproduced

Design

e.g., photonic metasurface

Performance

e.g., scattering cross-section

Inverse model

?

56 of 68

Target

Reproduced

Design

e.g., photonic metasurface

Performance

e.g., scattering cross-section

Inverse model

?

57 of 68

Target

Reproduced

Design

e.g., photonic metasurface

Performance

e.g., scattering cross-section

Inverse model

?

58 of 68

Target

Reproduced

Design

e.g., photonic metasurface

Performance

e.g., scattering cross-section

Inverse model

?

59 of 68

Multi-joint robot

59

Target

Design:

Performance:

Position of the tip

Angle of the joints

60 of 68

Multi-joint robot

60

Target

Design:

Performance:

Position of the tip

Angle of the joints

61 of 68

Multi-joint robot

61

Target

Design:

Performance:

Position of the tip

Angle of the joints

62 of 68

Multi-joint robot

62

Target

Design:

Performance:

Position of the tip

Angle of the joints

63 of 68

Multi-joint robot

63

Target

Design:

Performance:

Position of the tip

Angle of the joints

64 of 68

Multi-joint robot

64

Target

Design:

Performance:

Position of the tip

Angle of the joints

65 of 68

Multi-joint robot

65

Target

Design:

Performance:

Position of the tip

Angle of the joints

66 of 68

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

67 of 68

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

68 of 68

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