Urban airflow prediction by pix2pix trained on FFD

• Published in: Journal of physics. Conference series 2023-11, Vol.2600 (8), p.82009
• Main Authors: Vecchiarelli, Rebekah, Kraus, Michael, Griego, Danielle, Waibel, Christoph
• Format: Article
• Language: English
• Subjects: Air flow; CAD; Computational fluid dynamics; Computer aided design; Design analysis; Divergence; Fields (mathematics); Generative adversarial networks; Mathematical models; Parametric analysis; Physics; Representations
• ISSN: 1742-6588; 1742-6596
• DOI: 10.1088/1742-6596/2600/8/082009

Existing computer-aided design tools render insufficient in their capacity to enable architects and engineers to efficiently evaluate alternative designs during early design phases due to their computationally expensive nature, which is especially the case for computational fluid dynamics (CFD) methods. One of the greatest bottleneck for integrating CFD analysis into early design phases is the limited potential for parametric analysis, where a number of design alternatives need to be quickly generated and evaluated. In this context, the present study investigates the use of the generative deep learning method “pix2pix”, which leverages conditional generative adversarial networks (cGANs) for image-to-image translation, for prediction of airflow characteristics in different representations. The evaluation proposes statistical metrics to judge the fitness of the approach in performing urban airflow prediction. Our study demonstrates that the proposed method to be implemented, trained and validated successfully for different representations of the flow field prediction under parametric city shapes by incorporating building height and vectorial information (either components or magnitudes) into the pix2pix image inputs. The findings of the study reveal that the vortical flow fields can be predicted with a high accuracy in space and magnitude in all model variations tested. Adding building height information to the input images also significantly improves Kullback-Leibler (KL) divergence compared to using uniform building heights as inputs. Using vectorial information in the form of decomposed u, v, w -vector fields during training enabled pix2pix to additionally generate vectorial predictions instead of magnitudes only.