> Summary

In this thesis, the writer develops a workflow for early-stage design of 3D spatial truss structures. This is achieved by optimizing the structure based on structural performance, material use, and similarity to a defined stock of reusable materials (library) through a deep generative design framework. 

The deep generative design framework has 4 components: 

• Material library (defined stock) 

• Generated dataset for training 

• Structural evaluation of the dataset 

• Variational Autoencoders (VAE), surrogate model & gradient descent algorithm 

As a case study, the non-curved triangular spatial truss of the Orange Hall and Model Hall of the Architecture and the Built Environment Faculty of TU Delft is used. 

To create this workflow, the following steps were followed: 

1. Creating an input dataset that contains different configurations of the spatial truss structure 

2. Measure the performance of the variations in the dataset through a simulation based on 

   
   

   1. Structural performance 

      
      

   2. Material use 

   3. Similarity to the stock library 

3. Train a surrogate model with the simulations 

4. To train the VAE 

   1. Generate a new data point (design) using the VAE 

      
      

5. Find optimal solutions though a gradient descent (GD) algorithm: 

   
   

   1. Evaluate the structural performance of the new design with the trained surrogate model 

   2. Update the VAE based on the analysis 

   3. Continue training until the GD can identify the optimum solution from new generated geometries  

Although not implemented in this thesis, an option to try to solve the same problem is as follows. A generative adversarial network (GAN) is another generative deep learning method that could be applied to the problem by using a combination of a GAN and a surrogate model to replicate the workflow, replacing the VAE with a GAN to generate and evaluate the meshes. 

LIMITATIONS: Most meshes produced by this workflow perform better than their original input under the same degree of design freedom, even though better outputs are predicted by the surrogate model for all meshes. 

SPECIAL NOTE: Full Grasshopper and Python scripts available in Apendix II of the thesis.