Generative design of underground station pedestrian facilities: Evolutionary optimization with graph-based agent acceleration

Underground metro stations are high-occupancy public infrastructure where pedestrian facilities (e.g., gates, escalators, and stairs) form the primary circulation system. Designing these systems involves mixed discrete–continuous decisions, multiple conflicting objectives, and strong spatial dependencies among facilities, which together make realistic performance evaluation expensive. As a result, early-stage pedestrian facility design often relies on limited manual iteration and yields configurations that are hard to compare across objectives. To address this problem, this paper proposes a surrogate-assisted multi-objective generative design approach for station-scale pedestrian facility systems. An enhanced evolutionary algorithm is used to generate Pareto-optimal designs that balance system complexity, facility waiting time, and walking distance. For the expensive-evaluation bottleneck, a physics-informed graph neural network surrogate whose architecture mirrors the underlying queueing network is introduced to predict facility-level waiting times and flow distributions on an abstracted facility graph, replacing slow simulations in the optimization loop. Benchmark tests show that this surrogate achieves an R² above 0.99 on both objectives while reducing evaluation time from approximately 1800 s to 2.3 s per 100 candidates. A lightweight online correction mechanism then interleaves SimEvents re-evaluation and FEEC fine-tuning into the search to improve the surrogate's performance in the regions explored by the optimizer. Experiments on two station scenarios and a Tai Koo Station case study across three peak-hour demand levels show that the framework enables efficient Pareto exploration across varying geometries and demand conditions and can reduce system complexity by 19–34% relative to the existing layout while keeping operational performance comparable.