A Morphology-Guided Conditional Generative Adversarial Network for Rapid Prediction of Hazard Gas Dispersion Field in...

Area of Science:

• Environmental Science
• Computational Fluid Dynamics
• Artificial Intelligence

Background:

• Accurate hazard gas dispersion prediction in urban areas is crucial for emergency response and risk assessment.
• Traditional Computational Fluid Dynamics (CFD) is computationally expensive for real-time applications.
• Simplified models like Gaussian plume lack fidelity in complex urban environments with building obstructions.

Purpose of the Study:

• To develop a real-time gas dispersion modeling framework for urban environments using a morphology-guided conditional Generative Adversarial Network (cGAN).
• To overcome the computational limitations of CFD and the fidelity issues of Gaussian models.
• To enable rapid risk assessment and guide emergency sensor deployment in complex urban settings.

Main Methods:

• Urban areas discretized into grid cells with morphological parameters, categorized using K-means clustering.
• High-fidelity dispersion datasets generated using Lattice Boltzmann Method (LBM) simulations for each morphological type.
• Two cGAN architectures (Pix2Pix and Pix2PixHD) trained to map urban morphology, release conditions, and time to dispersion fields.

Main Results:

• The Pix2PixHD model achieved 92.5% prediction accuracy (FAC2) and 97.6% spatial pattern fidelity (SSIM).
• The framework demonstrated a computational speed-up of approximately 18,000 times compared to traditional CFD.
• Inference time was reduced to under 0.1 seconds per scenario, enabling real-time capabilities.

Conclusions:

• The morphology-guided cGAN framework provides a computationally feasible and physically informed solution for real-time urban gas dispersion modeling.
• This approach significantly enhances emergency management capabilities through rapid concentration field estimation.
• The model's speed and accuracy support potential applications in sensor network planning and gas source localization.