Coupled AI and physics model improves typhoon-wave height forecasting

Typhoons bring sudden, dangerous waves that can threaten ships, offshore platforms, and coastal infrastructure in the northwest Pacific Ocean. Current forecasting methods sometimes miss the largest waves during these storms. Researchers at a Chinese university developed a new model that combines physics guidance with data-driven learning, aiming to improve typhoon-wave height forecasting.

The study, published in Ocean Engineering, uses significant wave height (SWH) as a standard measure of sea-state severity. However, SWH is often underestimated in the Simulating WAves Nearshore (SWAN) model, leading to compounding errors in typhoon forecasts. The issue stems from SWAN's reliance on wind data, which can be difficult to obtain during extreme typhoon conditions. Standard corrections are also challenging when waves fluctuate rapidly.

Previous attempts to improve wave-models have used post-processing, data assimilation, statistical interpolation, and empirical correction methods. However, these techniques may struggle to represent the nonlinear and rapidly changing error structure of typhoon-driven waves. To address this, the researchers designed a hybrid model combining physics-informed neural networks (PINNs) and generative adversarial networks (GANs).

To test their new model, the team ran SWAN simulations for two typhoons in the East China Sea and surrounding waters. They compared modeled waves with observations from 29 buoys and coastal stations. In the first test scenario, the hybrid model reduced root-mean-square error (RMSE) by 34.9% compared to the uncorrected SWAN model during Typhoon Bebinca's decay phase.

In the second test, involving Typhoon Pulasan, the hybrid model reduced error by 11.0%. The hybrid model was most accurate for waves above 3 meters (10 feet), where SWAN tends to underestimate heights.

While the hybrid model shows promise, it still has limitations. It performs less well in complex coastal areas and has been tested only on two typhoons, limiting its generalizability. The researchers plan to refine the model, separate wind-data errors from model-physics errors, and demonstrate real-time operational performance. If successful, the hybrid model could serve as a correction layer to enhance typhoon-wave height forecasting and coastal hazard applications.