AI-Powered Fire Engineering Design and Smoke Flow Analysis for Complex-shaped Buildings

Yanfu Zeng,Zhe Zheng,Tianhang Zhang,Xinyan Huang,Xinzheng Lu

DOI: https://doi.org/10.1093/jcde/qwae053

2024-06-16

Journal of Computational Design and Engineering

Abstract:This paper aims to automatize the performance-based design of fire engineering and the fire risk assessment of buildings with large open spaces and complex shapes. We first establish a database of high-quality fire simulations for diverse building shapes with heights up to 60 m and complex atriums with volumes up to 22 400 m3. Then, AI models are trained to predict the soot visibility slices for new fire cases in buildings of different atrium shapes, symmetricities, and volumes. Two deep learning models were demonstrated: the pix2pix generative adversarial network (GAN) and image-prompt diffusion model. Compared with high-fidelity CFD fire modelling, the available safe egress time (ASET) predicted by both models shows a high accuracy of 92% for random atrium shapes that are not distinct from the training cases, proving their performance in actual design practices. The diffusion model reproduces more flow details of the smoke visibility profiles than GAN, but it takes a longer computational time to render the fire scene. This work demonstrates the potential of leveraging AI technologies in building fire safety design, offering significant cost and time reductions and optimal solution identification.

computer science, interdisciplinary applications,engineering, multidisciplinary

What problem does this paper attempt to address?

The paper aims to address the issue of performance evaluation in fire engineering design for buildings, particularly those with large open spaces and complex shapes. Specifically, the research goal is to achieve automation in fire engineering design and fire risk assessment for such buildings.

To achieve this, the researchers first established a high-quality fire simulation database, covering various buildings with different shapes, heights up to 60 meters, and atrium volumes up to 22,400 cubic meters. Then, using this database, they trained two artificial intelligence (AI) models to predict smoke visibility slices in new fire cases with different atrium shapes, symmetries, and volumes. These two models are the pix2pix model based on Generative Adversarial Networks (GAN) and the image-prompted diffusion model.

The paper demonstrates the effectiveness of these models by comparing the predicted Available Safe Egress Time (ASET) from the two AI models with the results from high-fidelity Computational Fluid Dynamics (CFD) fire modeling. The results show that for random atrium shapes not significantly different from the training data, the ASET predicted by the two models has an accuracy of up to 92%, showcasing their performance capability in practical design practice.

Although the diffusion model can better reproduce the details of smoke visibility profiles compared to the GAN model, it has a longer computation time. This work demonstrates the potential of using AI technology in building fire safety design, which can significantly reduce costs and time and help find optimal solutions.