Coupled AI and physics model improves typhoon-wave height forecasting
Typhoons pose significant threats, with sudden, dangerous waves that endanger ships, offshore platforms and coastal infrastructure in the northwest Pacific Ocean. Current typhoon forecasting methods sometimes underestimate the largest waves during typhoons, but researchers in China have designed a new model that may offer improvements. Their new study, published in Ocean Engineering, describes how the model combines physics guidance with data-driven learning to improve typhoon forecasting.
Predicting typhoon wave dynamics
Significant wave height (SWH) is used as a standard measure of sea-state severity and is central to marine warnings, but this parameter is often underestimated in the commonly used Simulating WAves Nearshore (SWAN) model. This leads to compounding errors that throw off typhoon forecasts. The underestimation partially results from the model's dependence on wind data, which can be challenging to collect during the extreme conditions induced by typhoons. Furthermore, standard corrections can be difficult when waves grow and fade quickly.
Researchers have attempted to improve methods but have faced challenges. The authors of the new study write, "To reduce wave-model biases under tropical cyclones, a range of post-processing and correction strategies have been explored. Data assimilation can improve wave state estimates when observations are available, but it is computationally expensive and its effectiveness is often limited during extreme events by data sparsity and the strong sensitivity of wave dynamics to forcing uncertainties. Statistical interpolation or empirical correction methods are relatively lightweight, yet they may struggle to represent the nonlinear and rapidly evolving error structure of typhoon-driven waves."
A physics-guided hybrid correction framework
There is a clear need for a practical add-on feature capable of improving existing wave models without rebuilding the entire forecast system. Past studies indicated that physics-informed neural networks (PINNs) can produce steadier, more physically plausible results but may smooth away extreme peaks. Other studies showed that generative adversarial networks (GANs) are better at learning rare, extreme patterns but can produce unrealistic fluctuations without physical safeguards. This motivated the researchers to design a combination of the two as an add-on to SWAN models.
To test their new design, the team ran SWAN simulations for two September 2024 typhoons in the East China Sea and nearby waters and compared modeled waves with observations from 29 buoys and coastal stations. Their AI learned to correct the gap between SWAN outputs and observed wave heights using wave, wind, location, depth and energy-related information.
The team used two different scenarios (A and B) to test model performance in different dynamic marine environments with GAN-only, PINN-only and hybrid combined models. Test A was representative of a typhoon decay period, while Test B represented a full typhoon life cycle.
In Test A, during the fading phase of Typhoon Bebinca in 2024, the hybrid model reduced root-mean-square error by 34.9% compared with the uncorrected SWAN model. In Test B, involving Typhoon Pulasan, the hybrid model reduced error by 11.0%. The physics-guided model alone transferred better in Test B than the GAN alone, suggesting physical rules help when conditions differ from training data. However, the combined model was most reliable for waves above 3 meters (10 feet), where standard SWAN showed a strong tendency to underestimate heights.
"The results show that correction skill varies significantly with test scenario. Quantitatively, in the within-typhoon decay scenario (Test A), the GAN achieves a 31.77% RMSE improvement relative to SWAN, outperforming the PINN (21.75%). In the cross-typhoon extrapolation scenario (Test B), the PINN (9.21%) outperforms the GAN (2.94%), demonstrating the superiority of physical constraints for out-of-distribution generalization," the study authors explain.
Combining the PINN-GAN model with SWAN
Although the hybrid model offers impressive improvements for SWAN models, it still has some limitations that need to be addressed before practical use. The hybrid model showed some weaknesses near complex coastlines, slightly worsening predictions in one nearshore test group. Also, this study covered only two typhoons, which is insufficient to prove performance across many storms or ocean regions. Real-time operational performance also needs to be demonstrated because the tests were performed on past events.
However, the researchers have plans for further improvements to the model. They plan to refine and test the system with more wave-model settings, separate wind-data errors from model-physics errors and address other limitations in the model. They note that the model can act as a correction layer, correcting errors and enhancing robustness during typhoon extremes. They say it will offer a scalable pathway for operational forecasting and coastal hazard applications.
Written for you by our author Krystal Kasal, edited by Gaby Clark, and fact-checked and reviewed by Robert Egan—this article is the result of careful human work. We rely on readers like you to keep independent science journalism alive. If this reporting matters to you, please consider a donation (especially monthly). You'll get an ad-free account as a thank-you.
More information
Yinhe Cheng et al, A PINN–GAN coupled model for physics-guided correction of SWAN significant wave height under typhoon-driven sea states, Ocean Engineering (2026). DOI: 10.1016/j.oceaneng.2026.127299
Who's behind this story?
Freelance science writer with Master's in physics. Five years clinical research and physics education experience. Science communicator. Full profile →
MA in English, copy editor since 2021 with experience in higher education and health content. Dedicated to trustworthy science news. Full profile →
Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile →
© 2026 Science X Network
Citation: Coupled AI and physics model improves typhoon-wave height forecasting (2026, August 6) retrieved 6 August 2026 from https://phys.org/news/2026-08-coupled-ai-physics-typhoon-height.html
This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.