I developed FloodForecaster, a time-dependent geometry-informed neural operator framework for rapid flood forecasting on irregular river geometries. The work addresses a core deployment challenge in scientific ML: how to transfer a high-performing model to new geographic domains without retraining from scratch or losing performance on the original domain. The resulting work is now represented in a peer-reviewed Journal of Hydrology paper.

Key contributions

• Built a hybrid architecture that combines graph-based operators for irregular terrain with Fourier-style operators for global flood dynamics
• Added a gradient-reversal-based domain adaptation strategy to learn features that transfer across river segments
• Evaluated the method under data-scarce adaptation settings rather than assuming abundant target-domain simulations

Impact

The resulting framework supports real-time flood forecasting, preserves source-domain knowledge during transfer, and reduces target-domain error by roughly 75% relative to standard fine-tuning while performing strongly with as few as 10 target-domain training simulations.