This ongoing project explores a neural-operator domain-decomposition framework for accelerating two-dimensional shallow-water flood forecasting. The approach combines learned patch-level surrogates with physics-driven interface coupling so the overall system remains computationally efficient while preserving physically meaningful global behavior.

Key directions

• Using neural operators to replace expensive patch-level solves in local-to-global flood simulation workflows
• Combining learned patch interiors with conservative interface coupling and stabilization across subdomains
• Supporting fast deterministic forecasting and efficient uncertainty quantification under limited boundary information

Why it matters

Operational flood forecasting often depends on repeated high-fidelity simulations that are too slow for real-time use. This work is aimed at building surrogate methods that are faster than full simulation while still retaining stronger physical structure than purely black-box emulators.