🔬 physics

Sampling sea state using a diffusion model

This paper introduces a diffusion-based generative model that conditions on five days of global wind forcing to efficiently sample probabilistic sea states, offering a computationally accelerated and calibrated alternative to traditional spectral wave models for both bulk and derived ocean variables.

Original authors: Jiarong Wu, Bertrand Chapron, Laure Zanna

Published 2026-06-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiarong Wu, Bertrand Chapron, Laure Zanna

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the ocean as a giant, chaotic kitchen where the wind is the chef. Sometimes the chef stirs gently, creating small ripples. Other times, the chef goes wild, whipping up massive waves that travel across the entire ocean. Predicting exactly what this "sea state" looks like at any given moment is crucial for sailors, climate scientists, and anyone who cares about the weather.

Traditionally, scientists have tried to predict these waves using complex math equations that act like a super-detailed recipe. They calculate every single ripple and wave interaction. The problem? This recipe is so complicated and computationally heavy that it takes a long time to cook. It's like trying to bake a cake by measuring every single grain of flour and sugar with a microscope—it's accurate, but you can't do it fast enough to help with real-time decisions or to run thousands of different scenarios at once.

The New Approach: A "Guess-and-Refine" Artist

This paper introduces a new way to predict the ocean using a type of Artificial Intelligence called a Diffusion Model. Think of this model not as a chef following a strict recipe, but as a talented artist who has spent 30 years studying thousands of photos of the ocean.

Here is how the process works, using a simple analogy:

The Input (The Wind History): The artist doesn't just look at the wind blowing right now. Instead, they look at a 5-day "movie" of the wind blowing across the entire globe. This is important because waves, especially the long, rolling ones called "swells," take time to travel. Just like a wave you see today might have been started by a storm three days ago in a different part of the world, the artist needs to see the whole history to understand the picture.
The Starting Point (The Noise): The model starts with a blank canvas covered in static noise—like a TV screen with no signal. It's a complete mess.
The Process (Denoising): The model then starts to "clean up" the noise. It asks itself, "If the wind was like this for the last 5 days, what would the ocean look like?" It slowly removes the static, refining the image step-by-step until a clear picture of the sea emerges.
The Result (Sampling): Because the ocean is chaotic, there isn't just one possible answer. The model can generate many different, slightly different versions of the ocean for the same wind history. This is like the artist painting five different versions of the same scene to show the range of possibilities.

What Makes This Special?

Most previous AI models for weather were like a "domino effect." They predicted the next hour based on the last hour, then the next based on that, and so on. If they made a small mistake early on, the error would pile up, like a snowball rolling downhill getting bigger and bigger.

This new model skips the dominoes. It looks at the wind history and jumps straight to the result. This means:

It's incredibly fast: It runs more than 20 times faster than the traditional math-heavy models.
It doesn't make mistakes pile up: Since it doesn't rely on the previous hour's prediction, it doesn't accumulate errors over time.
It sees the details: While other AI models mostly guess the "average" wave height (like saying "the waves are 2 meters high"), this model can guess much more specific details. It can tell you about:
- Crossing Seas: When two different sets of waves crash into each other from different directions (a dangerous situation for ships).
- Stokes Drift: How the water itself moves in a circle, which is important for how oil spills or plastic trash drift.
- Surface Roughness: How "choppy" the water is, which affects how gases like oxygen move between the air and the sea.

The "Ensemble" Advantage

The paper highlights that this model is "probabilistic." Imagine you are trying to guess the weather for next week. A deterministic model gives you one answer: "It will rain." A probabilistic model gives you a range: "There's a 70% chance of rain, a 20% chance of sun, and a 10% chance of a storm."

This AI model generates a whole "ensemble" of possible ocean states. If the model is unsure, the different versions it generates will look very different from each other (high spread). If it is very confident, they will all look almost the same. This helps scientists know when to trust the prediction and when to be cautious.

In Summary

The researchers trained this AI on 30 years of historical ocean data. They found that by looking at the wind's 5-day history and using this "denoising" technique, they could recreate the ocean's state with high accuracy, but in a fraction of the time it takes traditional computers.

This doesn't just help predict if a ship will get wet; it provides a fast, flexible way to understand the complex, multi-layered nature of the ocean, which is essential for improving climate models and keeping maritime activities safe. It turns a slow, rigid calculation into a fast, flexible, and probabilistic glimpse into the future of the sea.

Technical Summary: Sampling Sea State Using a Diffusion Model

Problem Statement
Accurate sea state prediction is critical for operational maritime safety and coupled earth system modeling, particularly for representing air-sea exchanges (e.g., Langmuir turbulence, mixing). However, current state-of-the-art spectral wave models, such as WAVEWATCH-III, are computationally prohibitive for many applications, including online coupling with climate simulations and generating probabilistic ensemble forecasts. While deep learning has advanced weather forecasting, existing AI-based wave models are predominantly deterministic, focusing on bulk variables (e.g., significant wave height) via autoregressive time-stepping. These approaches fail to capture the complex, multi-modal nature of sea states (e.g., crossing seas) and do not provide calibrated probabilistic estimates or derived quantities necessary for advanced parameterizations. Furthermore, the lack of probabilistic wave forecasting limits the application of advanced data assimilation techniques.

Methodology
The authors propose a Denoising Diffusion Probabilistic Model (DDPM) to sample the complex conditional distribution of global sea states. Unlike autoregressive models that step forward in time, this framework conditions on a static, relatively long history of global wind forcing to generate sea state snapshots directly.

Input Conditioning: The model conditions on a 5-day history of global zonal and meridional wind fields ( $U, V$ ), along with auxiliary static and dynamic fields: sea ice fraction and ocean depth. The 5-day window is chosen based on the approximate propagation time of swells across ocean basins.
Architecture: The core is a conditional U-Net with approximately 90 million parameters. It features three encoder/decoder resolution levels (320 $\to$ 160 $\to$ 80) and uses sinusoidal embeddings for noise levels. The conditioning fields are concatenated with the noisy input. A wet mask is applied to ensure training and loss calculation occur only over ice-free ocean regions.
Training and Sampling: The model is trained on a 30-year global hindcast (1993–2023) generated by WAVEWATCH-III forced by ERA5 winds. The training objective minimizes the squared error between the predicted and actual noise. Sampling involves a reverse diffusion process (linear-log schedule) starting from Gaussian white noise to generate a sample $x^*$ from the distribution $p(x|f)$ .
Target Variables: The model is designed to sample a diverse array of variables beyond bulk parameters:
- Bulk Variables: Significant wave height ( $H_s$ ), mean wave period ( $T_m$ ), and mean direction ( $\theta_m$ ).
- Partition-Related Variables: Characteristics of the two most energetic wave systems (partitions), including their heights, periods, and directions, as well as a binary Crossing Sea Index (CI) indicating the presence of misaligned, co-existing wave systems.
- Derived Variables: Surface Stokes drift ( $U_{ss}, V_{ss}$ ) and Mean Square Slope (MSS), which are higher-order spectral moments used in air-sea gas transfer and mixing parameterizations.

Key Results
Evaluated on the year 2004 (held-out test set), the diffusion model demonstrates the following performance:

Computational Efficiency: The model achieves a substantial acceleration, offering over a 20x wall-clock speedup compared to numerical spectral models.
Bulk Variable Skill: The model achieves low Root-Mean-Square Error (RMSE) and high Anomaly Correlation Coefficient (ACC) for bulk variables. For significant wave height ( $H_s$ ), the RMSE is 0.27 m and ACC is 0.96. While the RMSE is slightly higher than short-lead autoregressive models (as expected due to the lack of prior wave state information), it remains comparable to long-lead (360-hour) autoregressive forecasts.
Probabilistic Calibration: The ensemble spread is well-calibrated. The Spread-Skill Ratio (SSR) for bulk variables falls between 0.9 and 1.1, indicating the ensemble variance accurately reflects forecast uncertainty. Regions of high error (e.g., complex directional intersections) correspond to high ensemble spread.
Partition and Crossing Sea Prediction: The model successfully predicts the Crossing Sea Index with a Brier Skill Score of ~0.4, demonstrating an ability to distinguish multi-modal sea states. It also captures partition-level variables (e.g., swell-dominated periods), though with higher RMSE and slight under-dispersion (SSR < 1) compared to bulk variables.
Derived Variables: The model effectively samples Stokes drift and MSS. Interestingly, including the 5-day wind history improves Stokes drift (a third-moment statistic) but slightly degrades MSS (a fourth-moment statistic), suggesting that higher-order moments may be less dependent on long-term wind history or more sensitive to local forcing.
Physical Consistency: The joint distribution of non-dimensional wave height and period preserves physical scaling laws (e.g., Toba's law) observed in the hindcast data.

Significance and Claims
The paper claims that diffusion-based sea state sampling offers a promising path toward probabilistic wave forecasting and the efficient coupling of sea state information into broader earth system models.

Key contributions highlighted by the authors include:

Beyond Bulk Variables: Unlike prior AI approaches limited to bulk parameters, this framework naturally extends to partition-related and derived variables, enabling the estimation of complex sea state features (like crossing seas) and quantities required for physical parameterizations (Stokes drift, MSS).
Probabilistic Framework: By sampling from a conditional distribution rather than predicting a deterministic mean, the model provides calibrated ensemble spreads, addressing a major gap in wave forecasting and enabling future data assimilation of wave measurements beyond just significant wave height.
Non-Autoregressive Efficiency: The approach avoids error accumulation inherent in autoregressive time-stepping and allows for flexible querying of sea state variables at any temporal interval without re-running a time-stepping loop.
Foundation for Assimilation: The ability to generate low-cost, probabilistic sea state estimates with associated variance estimates facilitates the development of data assimilation algorithms that can leverage detailed wave measurements from modern satellite platforms (e.g., SAR).

The authors maintain a modest stance, acknowledging that while the current results are skillful, the metrics can likely be improved through further exploration of model architecture (e.g., latent spaces), sampling strategies, and additional conditioning fields like ocean currents. They also note that the current model is trained on hindcast data but is applicable to reanalysis or observations.

Technical Summary: Sampling Sea State Using a Diffusion Model

More like this