Learning and Simulating Building Evacuation Patterns for Enhanced Safety Design Using Generative Models

Jin Han, Zhe Zheng, Yi Gu, Jia-Rui Lin, Xin-Zheng LuarXiv 2025 · DOI · arXiv

Abstract

Traditional building evacuation simulations are computationally expensive and time-consuming, creating a bottleneck in the iterative architectural design process. This paper presents DiffEvac, a generative AI framework leveraging diffusion models to predict evacuation patterns from architectural sketches. By decoupling floor plan features into structural, density, and exit channels, DiffEvac achieves a 16x speedup over industry-standard simulators while significantly outperforming GAN-based baselines in accuracy and robustness across complex geometries.

Introduction

In building design, safety is paramount, but "safety-first" design is often hindered by a massive bottleneck: simulation time. Traditional evacuation tools (like Pathfinder) require detailed 3D modeling and extensive parameter tuning. This makes them nearly impossible to use during the "schematic design" phase, where architects are rapidly iterating through dozens of different floor plans. Furthermore, previous attempts to use Artificial Intelligence—specifically Generative Adversarial Networks (GANs)—to solve this problem struggled with "irregular" floor plans (L-shaped, U-shaped, or T-shaped buildings). There was a clear need for a surrogate model that is fast enough for real-time design and robust enough to handle complex geometry.

Methods

The authors introduce DiffEvac, which shifts the paradigm from traditional physics-based simulation to image-to-image translation using a Diffusion Model. Key technical innovations include a Decoupled Feature Representation. Instead of feeding the model a standard RGB image of a floor plan, the data is decoupled into three distinct channels:

  1. Structural Layout: A binary map of walls and obstacles.
  2. Occupant Density: A matrix representing how many people are in each room based on its function.
  3. Exits: Precise locations of doors and stairs.
Overview of the DiffEvac framework, including dataset development, the decoupled feature diffusion model, and validation.

The model uses these decoupled features as "prompts" to guide the denoising process, gradually turning random noise into a precise evacuation flow heatmap. The core of the model is a U-Net architecture with temporal encoding, designed to capture both the global context of the building and the local details of narrow corridors.

Datasets and Training

The researchers established a custom dataset to train and validate the model using 81 real-world office building floor plans with diverse geometries. To overcome the small sample size, they used "functional augmentation"—randomly re-assigning room functions and densities to generate 399 unique functional layout drawings. Every layout was run through Pathfinder (a gold-standard industry simulator) to generate the ground truth evacuation heatmaps used for training.

Key Results and Insights

DiffEvac demonstrated significant performance gains over previous state-of-the-art approaches like pix2pix:

  • Accuracy: The model achieved a 37.6% improvement in SSIM (Structural Similarity Index) and a 142% improvement in PSNR (Peak Signal-to-Noise Ratio).
  • Efficiency: The total process—from modeling to results—was cut from roughly 20 minutes to 2 minutes. Actual AI inference takes only 74 seconds, representing a 16-fold increase in efficiency.
  • Robustness: Unlike previous models that failed on "complex" or "irregular" layouts, DiffEvac maintained high accuracy regardless of building shape.

Technical analysis revealed that deep learning models perform better when physical constraints are explicitly separated. By giving the model a dedicated "exit channel," it learns the fundamental rule of evacuation—people move toward exits—much faster. Interestingly, the authors found that adding an Attention Mechanism didn't significantly improve results, suggesting that evacuation patterns are driven more by the global relationship between rooms and exits than by local pixel correlations.

Conclusion

DiffEvac provides a viable path toward real-time safety evaluation in architectural design. While currently validated for office buildings, the framework establishes a foundation for future work integrating dynamic factors like fire spread and smoke toxicity. The team aims to extend the model to more building types, such as hospitals and schools, to make simulations even more realistic for emergency management.

Summary updated

Abstract

Evacuation simulation is essential for building safety design, ensuring properly planned evacuation routes. However, traditional evacuation simulation relies heavily on refined modeling with extensive parameters, making it challenging to adopt such methods in a rapid iteration process in early design stages. Thus, this study proposes DiffEvac, a novel method to learn building evacuation patterns based on Generative Models (GMs), for efficient evacuation simulation and enhanced safety design. Initially, a dataset of 399 diverse functional layouts and corresponding evacuation heatmaps of buildings was established. Then, a decoupled feature representation is proposed to embed physical features like layouts and occupant density for GMs. Finally, a diffusion model based on image prompts is proposed to learn evacuation patterns from simulated evacuation heatmaps. Compared to existing research using Conditional GANs with RGB representation, DiffEvac achieves up to a 37.6% improvement in SSIM, 142% in PSNR, and delivers results 16 times faster, thereby cutting simulation time to 2 minutes. Case studies further demonstrate that the proposed method not only significantly enhances the rapid design iteration and adjustment process with efficient evacuation simulation but also offers new insights and technical pathways for future safety optimization in intelligent building design. The research implication is that the approach lowers the modeling burden, enables large-scale what-if exploration, and facilitates coupling with multi-objective design tools.
BibTeX
@misc{han2025learning,
  author = {Jin Han and Zhe Zheng and Yi Gu and Jia-Rui Lin and Xin-Zheng Lu},
  title = {{Learning and Simulating Building Evacuation Patterns for Enhanced Safety Design Using Generative Models}},
  year = {2025},
  eprint = {2510.19623},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.19623},
}
Figure