Global Location-Invariant Peak Storm Surge Prediction

1. Problem Statement

Storm surges pose a significant threat to coastal communities, causing loss of life and billions in economic damage. While high-fidelity physics-based models (like ADCIRC) are accurate, they are computationally expensive, often requiring supercomputing resources, which limits their use in real-time emergency management or probabilistic risk assessment.

Data-driven surrogate models have emerged as a cost-effective alternative. However, existing surrogate models suffer from two critical limitations:

1. Data Scarcity: Training datasets are often small (typically <1,000 storms).
2. Regional Specificity: Most models are trained on narrow geographical regions (e.g., only the US or China). Because traditional models map inputs to a fixed set of output points, they implicitly encode specific geographies. Consequently, a model trained on one region cannot be applied to another, necessitating the creation of entirely new datasets and models for every new location. This defeats the purpose of a "cheap" surrogate model.

2. Methodology

A. Data Generation: The Global Dataset

The authors constructed the largest dataset of its kind to date, featuring over 15,000 synthetic tropical cyclones distributed globally.

• Source: Synthetic storms were derived from the STORM IBTrACS dataset (10,000 years of simulated cyclones).
• Physics Model: The ADCIRC (ADvanced CIRCulation) model was used to simulate storm surges on a high-resolution global mesh (12.8 million nodes, 24.9 million elements).
• Meteorological Forcing: Parametric Holland models were used to generate wind and pressure fields based on storm tracks, rather than expensive high-resolution meteorological reanalysis.
• Optimization (Storm Packing): To reduce computational costs, the authors modified the ADCIRC code to run multiple storms (one per ocean basin) simultaneously in a single simulation. This reduced the required number of runs from ~15,000 to 3,000, cutting costs by a factor of five without affecting peak surge accuracy.
• Initial Conditions: 120 unique 7-day "tidal spin-up" runs were performed to ensure realistic initial conditions and avoid bias toward specific tidal phases.

B. Feature Engineering

To enable global applicability, the authors moved away from fixed-point outputs to a grid-based approach:

• Grid Transformation: For each storm, irregular mesh data was interpolated onto a regular 128x128 lat-lon grid centered on the storm's landfall location.
• Input Channels (41 total):
  • Meteorological: 13 temporal snapshots (24h pre-landfall to 12h post-landfall) of sea-level pressure and two wind components (39 channels).
  • Geophysical: Bathymetry (1 channel) and a binary land mask (1 channel).
• Output: A 128x128 grid representing the peak storm surge (maximum water elevation).

C. Model Architecture

The authors employed a U-Net architecture, a fully convolutional encoder-decoder network with skip connections, originally designed for image segmentation.

• Encoder: Extracts hierarchical features through 5 stages (doubling channels, halving spatial resolution).
• Decoder: Reconstructs spatial resolution using transposed convolutions.
• Skip Connections: Crucial for preserving fine-grained spatial details, allowing the model to precisely locate surge maxima near complex coastlines.
• Training: Minimized Mean Squared Error (MSE) using the AdamW optimizer.

3. Key Contributions

1. Unprecedented Global Dataset: The creation of a high-resolution dataset covering >15,000 storms across six ocean basins, an order of magnitude larger than previous datasets.
2. Location-Invariant Surrogate Model: A novel formulation that predicts peak surge on a relative grid centered on landfall. This allows the model to be applied to any location on Earth, including regions with no prior training data, solving the "regional specificity" bottleneck.
3. Storm Packing Technique: A methodological innovation allowing multiple independent storms to be simulated in a single ADCIRC run, drastically reducing computational overhead.
4. Demonstration of Transfer Learning: Evidence that training on a global, heterogeneous dataset improves performance in specific regions compared to training on local data alone.

4. Results

• Architecture Performance: The UNet-5 (5-stage U-Net) outperformed other architectures (CNN, SegNet, UNet-4).
  • RMSE (Near Coast): 0.51 meters.
  • RMSE (All Points): 0.33 meters.
• Global vs. Local:
  • The global model consistently outperformed basin-specific local models. For example, in the North Indian (NI) basin, the global model reduced error by 28% compared to the local model.
  • This confirms the model learns general storm surge mechanics rather than memorizing specific coastline geometries.
• Validation on Historical Storms:
  • Tested against 67 historical North Atlantic hurricanes (2003–2023) with 188 observational data points from NOAA tidal gauges.
  • ADCIRC (Physics-based): RMSE = 0.22m (weighted by data points) / 0.48m (weighted by storm).
  • UNet-5 (ML-based): RMSE = 0.37m (weighted by data points) / 0.52m (weighted by storm).
  • The ML model achieved accuracy comparable to the high-fidelity physics model but at a fraction of the computational cost.

5. Significance and Future Work

• Paradigm Shift: This work challenges the traditional region-centric approach to storm surge modeling. It proves that a single, globally trained model can generalize to disparate geographical regions, making high-fidelity surge prediction accessible for any coastal area.
• Operational Potential: The low computational cost of the surrogate model enables real-time probabilistic forecasting for thousands of potential storm tracks, aiding emergency management and infrastructure planning.
• Future Directions:
  • Incorporating physical constraints (e.g., conservation laws) into the loss function, though difficult for peak surge which lacks direct governing equations.
  • Better integration of tidal effects.
  • Operationalizing the model for real-time forecasting of tropical cyclones.

In summary, this paper presents a breakthrough in storm surge emulation by combining a massive, globally diverse dataset with a computer vision-based architecture, successfully creating a location-invariant surrogate model that bridges the gap between computational efficiency and physical accuracy.