Deep Generative Designs: A Deep Learning Framework for Optimized Shell Structures
Year 2022Â
Author(s)     Stella Pavlidou  Â
ML Tags
Variational Autoencoders
Surrogate Modelling
Gradient Descent Optimization
Topic Tags
Generative Design
Structural Design
> Software & Plug-ins UsedÂ
Rhinoceros, Grasshopper, Karamba 3D for Finite Element Modelling (FEM)Â
Keras (Python library) for training the VAE and the Surrogate ModelÂ
COMPAS for the dataset generationÂ
> Workflow
Proposed methodology. Thesis report page 12,13
> Summary
The objective of this thesis is to optimize the structural performance of shell structures by exploring different topology options through an AI framework. The following steps are taken:
A new dataset is generated, consisting of several mesh tessellations created through a generative workflow.
Next, a Variational Autoencoder (VAE) is trained using this data. To train the VAE, two methods were tested:Â
The first method involves training the VAE using the coordinates of vertices as training data, but this was unsuccessful.Â
The second attempt employs an adjacency matrix as the training data. This approach proved to be more successful.
A third attempt used only a quarter of the adjacency matrix so that the input size would be smaller.
A  fourth attempt used a flattened part of the adjacency matrix, to minimize the size of the samples. This and this method proved to be faster.Â
After training the VAE, a surrogate model was also trained to predict a performance score of decoded designs.
A design is selected and encoded.
The gradient of the design’s performance score with respect to its encoded vector is obtained, and a Gradient Descent Optimizer runs to optimize the initial design.
LIMITATIONS: The dataset was created using specific mesh boundaries and limited pattern exploration. The Variational Autoencoder (VAE) was the exclusive generative model used, and the optimization process focused solely on structural performance.
SPECIAL NOTE: Code parts can be found in appendix of the thesis paper.
> Additional Visuals
Thesis Report Table 4.2: Some of the training samples and their decoded result
Thesis Report Figure 5.6: Various shell structures
Final Presentation Figure: Case study performance scores
> Possible Applications
For ideas on how to implement some of the above mentioned techniques, please see
‘Possible applications for students to try with Generative Deep Learning’