Discovering Latent Hydrological Regimes with Interpretable Mixture-of-Experts Surrogates

High-resolution hydrodynamic simulations are essential for assessing urban pluvial flood hazards, but their computational cost limits rapid scenario analysis and operational use. Simulation-based surrogate models provide efficient approximations, yet standard global predictors often hide the heterogeneous structure of simulator-derived flood responses. We investigate Mixture-of-Experts (MoE) surrogates for discovering model-induced latent response regimes in large-scale environmental simulation data. We train a neural-gated mixture of linear experts, where the gate assigns each cell--event instance a graded membership distribution over specialized experts, and interpret the regimes induced by this specialization rather than individual predictions. The approach is evaluated on a large tabular dataset derived from physics-based hydrodynamic simulations of the urban area of Crotone, Italy. We compare the MoE surrogate with linear, tree-based, linear-tree, and neural models, and characterize regimes through routing, tail enrichment, feature profiles, and spatial coherence. The results show that a few experts capture a disproportionate share of high-response cases, while large-coverage experts describe ordinary low-response conditions. The MoE therefore retains competitive flood-response approximation while exposing hydrologically plausible, response-coherent latent regimes.