Where accurate graph neural network surrogates fail in urban drainage systems: Hidden local failures under overflow transitions.

Graph neural network (GNN) surrogates can support rapid, repeated urban drainage analyses. Yet a model that is fast and accurate on average may still fail at critical nodes and hydraulic transitions. We developed a reliability-oriented diagnostic framework to identify hidden local failures in a purely data-driven GNN surrogate. We used a calibrated Storm Water Management Model (SWMM) reference and eight Chicago design storms with return periods from 1 to 100 years. The framework examined where errors showed hydraulically consistent pairwise linkages, when they intensified during overflow evolution, and which information made the remaining errors predictable. The unified cross-return-period model was more stable than return-period-specific models, but produced hydraulic boundary exceedances of 0.208-0.243 m lasting 6-21 min. Within the pairwise screening domain, 0.59%-1.78% of upstream-downstream pairs met the correlation, event-matching, and positive-lag criteria, with a pooled rate of 0.87%. Removing the late stage of renewed overflow and recession reduced the median spectral residual-fluctuation index (SpecE) by 5.81%-39.17%, while MAE and RMSE changed less consistently. Information scope contributed more to residual predictability than ordered sequence modeling. Using the current-state inputs, gradient boosting and temporal convolution reduced pooled MAE by 27.0% and 23.8%, respectively. Adding prior residual observations and reference flooding states increased these reductions to 48.9% and 53.5%, but required observation-assisted information. These results are specific to the tested case and support evaluating drainage surrogates in terms of where and when failures occur and which information makes the remaining errors predictable, alongside pooled accuracy.