functional generative design an evolutionary approach to
play

Functional Generative Design: An Evolutionary Approach to - PowerPoint PPT Presentation

Functional Generative Design: An Evolutionary Approach to 3D-Printing GECCO 2018 Kyoto, Japan Overview Motivation Motivation (FDM) (kinematic) Motivation Printed supports Fused


  1. Functional Generative Design: An Evolutionary Approach to 3D-Printing GECCO 2018 Kyoto, Japan

  2. Overview • • – – – – • – • – – •

  3. Motivation

  4. Motivation (FDM) (kinematic)

  5. Motivation Printed supports Fused Deposition Modeling Process

  6. Motivation Printed supports Fused Deposition Modeling Process requires post-processing

  7. Motivation •

  8. Motivation •

  9. Motivation •

  10. Motivation • Supports! (Grey)

  11. Motivation • Supports! (Grey)

  12. 3D Printing Functional Parts Slicing (Output: GCODE ) 3D gear design 3D Printing model (Continuum model) (Discrete model)

  13. Motivation Slicing (Output: GCODE ) 3D gear design 3D gear design 3D Printing model (Continuum model) (Discrete model)

  14. Motivation Slicing (Output: GCODE ) 3D gear design 3D gear design 3D Printing model (Continuum model) (Discrete model)

  15. Motivation Slicing (Output: GCODE ) 3D gear design 3D Printing model

  16. 3D Printing Functional Parts Slicing (Output: GCODE ) 3D gear design 3D Printing model (Continuum model) (Discrete model) • – – –

  17. Motivation • Launcher Rails Ruler Car

  18. Motivation •

  19. Motivation • SPRING

  20. Motivation • SPRING

  21. Available Methods ● ○ ■ ○ ■ ● ○ ○ ■ ■

  22. Methodology • ➢ ➢ ➢ ➢ •

  23. VAE • → Latent Variable Space AE Training Loss = VAE Recons.Error + KL-divergence Kingma and Welling, 2014 Forces LVs to follow a unit Gaussian distribution •

  24. VAE Interpolation in LV-space

  25. Noisy (Regressing) Kriging • → • Correlation between two points: Prediction at new point x*: f Predicted (re-interpolation) error (for EGO):

  26. EGO • → • Initial Sample Set Expensive Fitness Evaluations Build Surrogate Add Update Search for Improvement Update? Best Design

  27. EGO • → • Initial Sample Set Expensive Fitness Evaluations Build Surrogate Add Update Search for Improvement Update? Best Design

  28. rGA • → • ▪ ▪ α ▪ • ▪ ▪ •

  29. Integrated Method Exp-1 : Random Sampling in LV-space ( 12-D ) Initial Designs Exp-2: Uniform Sampling 2-D / 3-D Conversion in LV-space ( 12-D ) Fitness Evaluations Build Kriging rGA VAE-Encoder (13-D) Add Update (New design in VAE-Decoder LV-space) EI: Search for an rGA update (12-D) Stop? Best Design

  30. Experiments • • • – – • •

  31. Exp-1 Results

  32. Exp-2 Results

  33. Discussion • → • → • •

  34. Discussion • Continuous Gap (red region)

  35. Future Work • VAE EG Conv. NN

  36. Future Work • VAE EG Conv. NN • vs.

  37. Future Work • VAE EG Conv. NN • vs. •

  38. Future Work • VAE EG Conv. NN • vs. • •

  39. Conclusions • • → • – – – ▪ ▪ – – •

Recommend


More recommend