Structural investigation of regime-dependent surrogate representations under controlled hydrodynamic conditions: A Mixture-of-Experts approach

Rapid flood mapping requires surrogate models capable of approximating computationally expensive two-dimensional hydrodynamic simulations, but monolithic networks may smooth regime-dependent wetting–drying and floodplain-activation responses. This study evaluates a conditional Mixture-of-Experts (MoE) surrogate that decomposes discharge–inundation mappings into specialized expert responses through learned gating. Using 43 synthetic TELEMAC-2D flood simulations of the Lower Piura River Basin (Peru), sampled at 5-minute temporal resolution and exported to 10-m raster grids, the MoE was tested under a scenario-disjoint partition against a single-expert baseline. On unseen hydrographs, the four-expert MoE improved CSI from 0.7915 to 0.8323, reduced depth loss from 0.0127 to 0.0103 m, and reduced relative volume error from 0.1209 to 0.1024. Gate trajectories suggested hydrograph-phase specialization, while ablations indicated that peak performance required combining conditional gating, inundation-state supervision, and physics-regularized learning.