Generative Simulation Benchmarking for coastal climate resilience planning during mission-critical recovery windows

Description and use

A generative surrogate pipeline delivers flood maps in seconds, enabling emergency managers to evaluate evacuation routes, generator placement, and sandbag distribution within a 72‑hour hurricane recovery window.

Problem

Municipal emergency managers had 72 hours to decide evacuation corridors, generator placement, and sandbag distribution, but running high-fidelity hydrodynamic models would consume most of that window.

Outcome

The surrogate pipeline produced flood maps within seconds, allowing emergency managers to evaluate multiple scenarios and make informed decisions before the hurricane landfall.

Workflow

• 1. Generated training data with a reduced‑order shallow‑water solver in JAX and validated against NOAA surge hindcasts.
• 2. Trained generative models (conditional GANs, diffusion models, neural operators, latent‑variable surrogates) using a curriculum that started with mild storms and progressed to severe ones.
• 3. Developed a deadline‑aware, risk‑weighted scoring function to benchmark surrogate outputs against ground truth within the time budget.
• 4. Implemented an agentic orchestration layer to decide when to use the surrogate versus defer to the slow physics model.
• 5. Deployed the pipeline to evaluate evacuation corridors, generator pre‑positioning, and sandbag prioritization during the 72‑hour window.

Evidence

Running their existing hydrodynamic models for even a handful of scenarios would eat most of the recovery window before they'd made a single decision.

Source

I trained on mild storms first, then progressively harder ones, mirroring how the surrogate would be queried during a real event (calm pre‑landfall, then rapidly intensifying).

Source

The alpha term encodes what emergency managers told me directly: a missed inundation zone is a catastrophic failure, while a conservative over‑prediction is an inconvenience.

Source