Authors
M. Bernhard, R. Kakooee, P. Bedarf, B. Dillenburger
Abstract
Topology optimization (TO) is a numerical simulation to identify an optimal distribution of solid and void. A more efficient distribution of material means a reduction of natural resources consumption. TO results in branching structures, difficult to manufacture with conventional methods. Advances in additive manufacturing allow the production of components at an unforeseen level of complexity. The computational cost and the need for expert knowledge in setup prevent TO from being part of architects’ set of instruments. Data-driven artificial intelligence (AI) improved not only classification tasks but also spawned various synthetic models. We trained a generative adversarial network (GAN) with the boundary conditions as input and the result of a conventional TO as output. We chose a wall with randomly placed openings as a case study and produced three different training sets. The GAN was able to generate an output in a fraction of a second. The network learned to output structures close to the ground truth and generalized even across data sets. We measured the accuracy of the generated results with different metrics. The accuracy of the results was very encouraging within a few percents of the target value’s deviation. The significant speed improvement is a first and promising indicator of how machine learning could provide real-time feedback to the designer. Integrated into a CAD environment, dynamic updates, even for complex tasks, are invaluable in the conceptual design phase. Such an instrument can help the designer save material and most efficiently layout the building structure.
