DeepWaive: A Scalable and Fine-Tunable AI Foundation Model for Probabilistic 2D Inundation Forecasting

Operational flood forecasting and risk management requires high-resolution, spatially and temporally explicit predictions of inundation dynamics alongside robust uncertainty quantification. In practice, forecast skill is strongly constrained by uncertainties in meteorological forcing, boundary conditions, and human controls (e.g., reservoir releases) as well as inherent model uncertainties. While physics-based 2D hydrodynamic models provide physically consistent inundation dynamics, their computational cost make it impractical to generate several of ensemble members needed for uncertainty quantification  and the real-time exploration of “what-if” intervention scenarios.

We present DeepWaive, a physics-informed Foundation Model, that translates precipitation- or discharge-driven boundary conditions and static geospatial input into transient, spatially explicit 2D inundation dynamics within seconds. By leveraging deep-learning architectures trained on synthetic 2D hydrodynamic simulations, DeepWaive achieves zero-shot transferability to previously unseen basins without the need for domain-specific re-training. Crucially, the model architecture maintains the flexibility for optional site-specific fine-tuning, allowing for further optimization using either regional hydrodynamic models or in-situ sensor data to meet localized precision requirements. Benchmarking against classical numerical solvers demonstrates high predictive fidelity, with R² values ranging from 0.85 to 0.97, achieved alongside acceleration factors of 10⁵–10⁶. The model maintains scalability for domains up to 40,000 km² and event durations exceeding 24 hours.

Building on this capability, we develop an ensemble-to-probability workflow that propagates meteorological and hydrological forecast ensembles, and alternative reservoir release scenarios, through DeepWaive to generate probabilistic inundation products (e.g., spatial exceedance probabilities for depth and velocity thresholds) and impact-relevant summary metrics.

Within the Indo-German FLAIR project (Flood Forecasting using AI for Regional Sustainability, funded by BMBF), DeepWaive provides the fast dynamic core required to (i) quantify and communicate forecast uncertainty, (ii) support rapid sensitivity analyses of key uncertainty sources, and (iii) enable tight coupling to consortium modules on EO-derived flood variables, data assimilation, and reservoir operation optimization.