CrowdSafeCritic: A Physics-Informed Surrogate for Real-Time Crowd-Safety Evaluation of Architectural Layouts

Public-space disasters, such as the 2010 Love Parade crush, are often traced to a single geometric flaw, a narrow corridor or misplaced obstacle, whose danger is invisible until a crowd actually moves through it. High-fidelity pedestrian simulators can reveal such flaws, but a full safety audit takes on the order of a minute per layout, which becomes hours once a generative or parametric pipeline proposes the dozens to hundreds of candidate designs typical of early-stage exploration. We present CrowdSafeCritic, a physics-informed deep surrogate that predicts dense crowd-density and crowd-pressure heatmaps directly from a floor-plan mask, approximating a JuPedSim simulation in 7.2 ms on GPU (738× faster) or 470 ms on CPU alone (11× faster). The model trains on a generated, morphologically validated dataset of synthetic layouts paired with simulator ground truth. We report a controlled data-scaling study (250 to 10,000 training samples) in which Pearson correlation with simulated ground truth rises from 0.21 to 0.88, and matched ablation studies. An interactive tool lets designers sketch a layout and see its predicted risk profile immediately. We discuss the limitations of the approach and the societal trade-offs of automating safety review.