Vortex state transitions in deep street canyons enabled by an automated large language model workflow

Understanding airflow dynamics and vortex state transitions in urban street canyons is crucial for evaluating urban microclimates and pollutant dispersion. Computational fluid dynamics (CFD) simulations are a key tool for assessing urban wind flow. However, the traditional manual setup of these simulations is highly time-consuming, creating a critical computational bottleneck for the large-scale parametric studies required to uncover complex physical mechanisms. To address this challenge, we employ a large language model (LLM)-based multi-agent framework, Auto-Fluent, as an enabling methodology to automate CFD workflows. Auto-Fluent accepts natural-language instructions and automatically performs all steps of a simulation in Pointwise and Ansys Fluent, from geometry construction, mesh generation, boundary condition assignment, and solver setup. Using this enabling methodology, we systematically investigate vortex-regime transitions in deep canyons (aspect ratios H/W = 3–5) under varying upstream fetches. Results quantitatively demonstrate that initial upstream canyons exhibit vertically stacked recirculation structures (e.g., up to four vortices for H/W = 5). As fetch increases, these structures undergo a critical vortex-merging process, collapsing into a single dominant clockwise vortex and stabilizing after a critical fetch of five to six canyons. Concurrently, the primary recirculation center descends, significantly improving pedestrian-level ventilation. This transition is driven by the variation in the wind speed above the street canyon resulting from the separating flow at the leading edge of the first building and enables deeper penetration of the primary vortex. Ultimately, these findings provide quantifiable insight into urban canopy aerodynamics, demonstrating LLM-guided workflows as powerful enabling tools for large-scale parametric discoveries in fluid mechanics.