How many design storms are needed to train a machine learning flood depth surrogate model?

Flood depth modelling aims to estimate the spatial distribution and magnitude of flooding under different storm conditions, but physically based hydraulic simulations can be computationally demanding. This study develops a machine learning (ML) surrogate model to emulate the Hydrologic Engineering Center’ s River Analysis System (HEC-RAS 2D) water depths and reduce computational demand. The dataset includes simulations for five flood return periods (20‐1000 years), three storm durations (6, 12, and 24 h), and three antecedent moisture conditions. Because the target water depth field is zero-inflated, with many dry cells and positively skewed values, the surrogate is implemented as an extreme gradient boosting (XGBoost) regression model trained under a Tweedie likelihood. Entire return period and duration combinations were withheld through five leave-out configurations to evaluate interpolation to unseen design storms. The results showed strong performance in the most favourable setting, with coefficient of determination (R 2 ) values up to 0.99, but revealed a clear asymmetry between event descriptors. Assigning the 12 h rainfall duration events to the testing set had only a marginal effect on model performance. Assigning adjacent return periods to the testing set resulted in a substantial reduction in predictive accuracy, with R 2 decreasing from 0.971 to 0.757. Dry antecedent conditions were the most challenging because of the larger proportion of zero-depth cells, while terrain elevation and spatial location dominated the predictions. These findings support the use of ML surrogate models to reproduce unsimulated design storms within the sampled design storm space while reducing the number of computationally expensive hydraulic simulations required.