Generative artificial intelligence for fire scenario analysis in complex building design layouts

Abstract

Performance-based fire safety design requires thoroughly evaluating building fire scenarios to ensure comprehensive fire safety. However, conventional Computational Fluid Dynamics (CFD) fire simulations are computationally intensive and time-consuming, limiting the number of scenarios that can be practically analyzed. This study addresses these challenges by using generative artificial intelligence (AI) to predict fire scenes in realistic multi-room building layouts, characterized by complex shapes and intricate wall partitions. Three generative AI models for image generation are employed for this purpose: GAN-based pix2pix and pix2pixHD, as well as the diffusion model. These models were trained on an extensive dataset of CFD fire simulations to generate near-ceiling smoke movement and temperature distribution outcomes. When tested on new unseen building layouts, these models demonstrated remarkable accuracy and provided near real-time assessments. The diffusion model achieved the highest accuracy (>94 %) while requiring the more computational time. The high performance of these models highlights the potential of using generative AI to enhance fire safety engineering by enabling faster and more comprehensive fire risk assessments.