Deep Generative Design: A Deep Learning Framework for Optimized Spatial Truss Structures with Stock Constraints

Abstract

Energy use, CO2 emissions, and waste production are all significant causes of environmental issues. The building sector is a major contributor to these problems, specifically the manufacturing of (structural) steel elements. Application of reuse and/or remanufacturing, as done in a circular economy, will reduce these effects. Therefore, these techniques must to be taken into account while designing buildings and structures. However, as actors in the construction industry recognize barriers, such as time delays in the early phases of the design process, these tactics are not yet commonly used. Nonetheless, reusing structural steel components is still favoured by structural engineers, particularly when the aforementioned obstacles can be removed. This is feasible with the use of computational tools.

In this thesis, an AI based deep generative design framework is developed to optimize a 3D spatial truss structure for structural performance, material use and similarity to a defined stock of reusable materials. This workflow consist of a labelled dataset of geometries and performance indicators, a variational autoencoder (VAE), a predictive surrogate model and optimization through gradient descent. The aim of this study is to gain insight into the extent to which such a workflow can be applied in early-stage architectural and structural design exploration, especially with regards to making circular design & reuse more feasible. The case study in this thesis is a spatial truss structure supporting a flat roof. It was found that a surrogate model can be successfully trained to predict the performance of a geometry. For this, various input data representations can be used, including adjacency matrices and edge-vertex matrices. Through training of a VAE, a latent space from which meshes could be generated was successfully constructed. The VAE was able to accurately reconstruct a significant part of the geometry dataset. Through gradient descent, novel meshes were generated. Predicted performance of these meshes was increased. After assessing performance with simulations and calculations, increases in performance often remained. The largest performance increase found was 74.2%. Minor editing of meshes based on user insight demonstrated further increase in mesh performance.