I created HydroGraphNet, a physics-informed graph neural network for flood forecasting on unstructured domains. The model was designed to improve the operational usefulness of ML flood surrogates by addressing three persistent weaknesses at once: physical inconsistency, instability over long forecast horizons, and limited interpretability.

Key contributions

• Embedded mass conservation directly into the training loss rather than relying only on purely data-driven prediction objectives
• Used KAN-based encoder components to expose how terrain and hydraulic features such as elevation, curvature, slope, and roughness affect model behavior
• Built an autoregressive encoder-processor-decoder architecture for long-horizon flood forecasting and contributed the work into NVIDIA PhysicsNeMo

Impact

On the White River case study near Muncie, Indiana, HydroGraphNet achieved a 67% reduction in prediction error, near-zero mass balance error, and a 58% improvement in critical success index for major flood events relative to a baseline GNN, while retaining real-time inference characteristics.