Discovering interaction mechanisms in crowds via deep generative surrogate experiments.

Area of Science:

• Active matter physics
• Urban infrastructure design
• Complex systems analysis

Background:

• Pedestrian crowd dynamics are complex, driven by social interactions often modeled as simple forces.
• Quantitative studies face a trade-off between controlled lab experiments and large-scale real-world data.
• Understanding these dynamics is vital for urban planning and safety.

Purpose of the Study:

• To bridge the gap between controlled experiments and real-world data resolution in crowd dynamics.
• To develop a virtual surrogate experimentation paradigm for studying pedestrian behavior.
• To uncover fundamental interaction structures in complex social systems.

Main Methods:

• Utilized a generative simulation model based on graph neural networks (GNNs).
• Trained the GNN model on real-world pedestrian tracking data.
• Validated the model against statistical properties of crowd dynamics and known experimental results.

Main Results:

• Successfully reproduced known experimental results on collision avoidance.
• Revealed that N-body interactions in crowds are primarily topological.
• Identified that individuals react to a limited number of neighbors within a narrow field of view.

Conclusions:

• Data-driven virtual surrogate experiments can provide laboratory-like control with real-world statistical resolution.
• Topological interactions, rather than simple distance-based forces, govern pedestrian crowd behavior.
• This approach offers new avenues for scientific discovery in complex systems where direct experimentation is challenging.