Cumincad : CUMINCAD Papers : Search Results

CumInCAD is a Cumulative Index about publications in Computer Aided Architectural Design
supported by the sibling associations ACADIA, CAADRIA, eCAADe, SIGraDi, ASCAAD and CAAD futures

PDF papers
References

Hits 1 to 2 of 2

Reformat results as: short short into frame detailed detailed into frame

_id	ecaade2024_234
id	ecaade2024_234
authors	Mueller, Lisa-Marie; Andriotis, Charalampos; Turrin, Michela
year	2024
title	Using Generative Adversarial Networks to Create 3D Building Geometries
doi	https://doi.org/10.52842/conf.ecaade.2024.1.479
source	Kontovourkis, O, Phocas, MC and Wurzer, G (eds.), Data-Driven Intelligence - Proceedings of the 42nd Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe 2024), Nicosia, 11-13 September 2024, Volume 1, pp. 479–488
summary	Generative Artificial Intelligence (AI) promises to make a vast impact across disciplines, including transforming the architectural design process by autonomously generating full building geometries. One form of generative deep learning that has been used to create 2D and 3D representations of objects is Generative Adversarial Networks (GANs). Existing literature, however, has limited applications that utilize 3D data for building geometry generation, with previous studies focused on low-scale 3D geometries suitable for objects such as chairs or cars. This paper develops a new GAN architecture to produce high-resolution feasible building geometry. The training dataset used is a selection of 3D models of single-family homes from an existing database, pre-processed for the specific application. State-of-the-art GAN models are initially tested to establish baseline performance and applicability potential. Then, a systematic study is performed to identify the structure and hyperparameters necessary to successfully fit a GAN to this design task. The successful architecture, named 3DBuildingGAN, uses a combination of Wasserstein loss with gradient penalty, leaky rectified linear units for neuron activation in the generator and the critic, and the root mean squared propagation optimizer with a fixed learning rate. The proposed model generates outputs similar in size, shape, and proportion to the training data with minimal noise in the output. Evaluation of memorization properties indicates open research directions, such as incorporating memorization rejection and training on larger data sets. Finally, the study reflects on how AI algorithms can reshape creativity through data-driven design solutions.
keywords	3D Generative Adversarial Networks, Deep Learning, Artificial Intelligence
series	eCAADe
email
last changed	2024/11/17 22:05

_id	ecaade2024_235
id	ecaade2024_235
authors	Mueller, Lisa-Marie; Andriotis, Charalampos; Turrin, Michela
year	2024
title	Data and Parameterization Requirements for 3D Generative Deep Learning Models
doi	https://doi.org/10.52842/conf.ecaade.2024.1.615
source	Kontovourkis, O, Phocas, MC and Wurzer, G (eds.), Data-Driven Intelligence - Proceedings of the 42nd Conference on Education and Research in Computer Aided Architectural Design in Europe (eCAADe 2024), Nicosia, 11-13 September 2024, Volume 1, pp. 615–624
summary	It is now within reach to use generative artificial intelligence (AI) to autonomously generate full building geometries. However, existing literature utilizing 3D data has focused to a limited degree on architecture and engineering disciplines. A critical first step to expanding the use of generative deep learning models in generative design research is making training data available. This study investigates 3D building model data characteristics that make it suitable for generative AI applications. Key data set attributes are identified through a systematic review of the object-containing datasets currently used to train state-of-the-art 3D GANs. These requirements are then compared to attributes of existing available building datasets. This comparison shows that publicly available data sets of 3D building models lack essential characteristics for generative deep learning. Features that make these building models inadequate for the task include but are not limited to, their mesh formats, low resolution and levels of detail, and inclusion of irrelevant geometry. To achieve the desired properties in this work, necessary transformations of the data are incorporated into a tailored preprocessing pipeline. The pipeline is applied to an existing dataset that contains 3D models of single-family homes. The transformed dataset is tested within state-of-the-art GAN models to assess training performance and document future data requirements for applying deep generative design to buildings. Our experiments show promise for the impact that architectural datasets can make on deep learning applications within the discipline. It also highlights the need for additional 3D building model data to increase the diversity and robustness of new designs.
keywords	Generative Deep Learning, Data Sets, Generative Adversarial Networks
series	eCAADe
email
last changed	2024/11/17 22:05

No more hits.