Design to Acoustics mittels Deep Learning

Prof. Dr.-Ing. Sabine C. Langer (Braunschweig)
Dr. Timo Lüddecke (Göttingen)

Motivation / Background

  • Design assistants are desirable in many areas of engineering
  • Large vibroacoustic models (e.g. aircraft):
    • Numerical methods are the standard approach (e.g. finite element models)
    • Extensive parameter samplings are computationally expensive and hardly possible to consider during early design stages
  • Deep Learning:
    • Increasingly used in scientific and engineering applications
    • Recent success in generative language and image modelling
Figure: Symmetric finite element model of an aircraft fuselage
Symmetric finite element model of an aircraft fuselage

Approach: Combine both

Hypotheses

  • Neural networks can learn to represent vibroacoustic problems (Replacement of classic surrogates)
  • Neural network architectures are crucial for generalization to new scenes
  • Generative modeling is transferable to acoustic design spaces
  • Generative Adversarial Networks (GANs) enable fast one-shot proposals

Goals

  • Design of vibroacoustic systems by generative neural networks
  • Transfer of insights and methods to other engineering disciplines

Milestones

1.Representation of vibroacoustic scenes

  • Fit a neural network to numerical solutions
    of vibroacoustic problems

2.Generalization to novel vibroacoustic scenes

  • Apply neural networks on unseen vibroacoustic problems
    (e.g. new frequencies, new mechanical input parameters)

3.Generation of designs (e.g. geometries, material properties/distributions)

  • Direct, gradient-based design space optimization
  • GAN-based one-shot proposal generation
Figure: Neural network for sound pressure level predictions
Neural network for sound pressure level predictions

Contacts

 

Prof. Dr.-Ing. Sabine C. Langer

Universitätspl. 2, 38106 Braunschweig, Langer Kamp 19, Raum 104, 1.OG

 

Dr. Timo Lüddecke

To the top of the page