Graph-Derived and Geometric Drivers of Pollutant Dispersion in Urban Environments: Learning from Water Distribution Networks

Predicting air pollutant dispersion in complex urban environments traditionally relies on computationally intensive fluid dynamics simulations or empirical Gaussian models. This study explores whether topological network properties—originally derived from fluid flow concepts in pipe networks—can directly account for spatial concentration variations in an urban canopy. Using wind-tunnel measurement data from a representative urban setup, we construct a directed flow graph based on local velocity fields and compute key topological metrics, including node degree, betweenness centrality, and topological distances. A multiplicative regression model is then evaluated to assess the predictive capacity of these metrics against measured concentration distributions across multiple emission locations. Results show that graph-derived and geometric predictors capture a substantial portion of concentration variance (R2 up to 0.83, mean 0.57), with geographic distance, node degree, and betweenness centrality emerging as the dominant drivers. Because geographic distance is not itself a topological metric, the reported explanatory power is attributed to the combined set of geometric and graph-derived predictors rather than to graph topology alone. These findings provide preliminary evidence that graph-derived and geometric descriptors of a sensor-space proximity graph are statistically associated with concentration in this dataset. Because geographic distance is a geometric rather than topological predictor, the explanatory power is attributed to the combined predictor set, not to graph topology alone. The results are source-dependent and should not be interpreted as an operational replacement for CFD or as direct design guidance. Further work with distance-only baselines and independent validation is needed.