This project addresses a recurring bottleneck in flood modeling: producing detailed hazard maps over large domains without paying the cost of repeated high-resolution hydrodynamic simulations. I developed a multi-fidelity graph neural network (MFGNN) framework that combines coarse and fine simulation data directly on unstructured meshes and later matured into a peer-reviewed publication in Environmental Modelling & Software.

Key contributions

• Designed a two-stage workflow in which a low-fidelity model captures broad flood behavior and a second model refines the prediction at higher resolution
• Worked directly on irregular mesh representations rather than converting the problem into regular grids
• Framed the model as a practical surrogate for repeated hazard analysis and scenario generation

Impact

The resulting workflow is aimed at scalable hazard mapping, scenario analysis, and operational forecasting settings where conventional solvers are often too slow to support rapid iteration.