📄 Paper:
Lupin-Jimenez et al. (2025)
"Simultaneous Emulation and Downscaling..."
doi.org/10.1029/2025JH000851

💾 Code & data:
zenodo.org/record/14607130

We’d love to hear from collaborators in ocean ML, emulation, and climate AI 🌊🤝

🚀 Why it’s exciting:
✅ Physically grounded
✅ 10x–1000x faster than ROMS
✅ Enables regional “digital twins”
✅ Sets up for coupled ocean–atmosphere emulation
✅ Works across different reanalysis sources

AI meets ocean science.

📊 Results:

Beats interpolation

Matches or outperforms ROMS in short-term accuracy

Stays stable & realistic over 10 years

Captures mean state and eddy variability

Preserves spectral energy across scales

No exploding gradients here 💥

🧠 What’s different:

We don't just super-resolve existing data.
We downscale from an emulator that predicts ocean dynamics.

Plus: our downscaler learns to correct both model bias and physical mismatch (GLORYS → CNAPS). That’s new.

⚙️ Our framework (FCDS):

An FNO emulator predicts SSH, SSU, SSV, SSKE daily at 8 km

A UNet + PatchGAN-VAE downscales to 4 km & corrects bias

Spectral loss + online fine-tuning ensures physical consistency

Together: speed, structure, and stability.

🌍 Why this matters:

Regional ocean models like the Gulf of Mexico are hard—complex coastlines, eddies, Loop Current, chaotic boundary forcing.

Physics models = accurate but slow.
ML = fast, but unstable after a few weeks. We wanted the best of both.

Led by @baskinengineering.bsky.social PhD students Niloofar & Lenny with Tianning Wu & Roy He @ncstate.bsky.social

If you're working on GenAI for Earth systems, let’s connect — curious to hear your thoughts!
#GenAI #ClimateAI #OceanML #FNO #DDPM #DataAssimilation

Our method is:
⚡️ One-shot
🌀 Physics-consistent
🌐 Scalable

It captures high-wavenumber, fine-scale structures other ML baselines miss. Spectral diagnostics & vorticity metrics confirm this. (4/5)

🧠 The framework combines:
• FNO (Fourier Neural Operator)
• DDPM (Denoising Diffusion Probabilistic Model)

✅ Reconstructs high-resolution states from 1%–0.1% data
✅ Works on synthetic turbulence, GLORYS reanalysis & real satellite altimetry
✅ No forward solver required (3/5)

Ocean observations are often sparse, noisy, and Lagrangian (they move with the flow).
This makes reconstructing fine-scale ocean dynamics like eddies and fronts very hard — especially for forecasting.
We tackle this using a diffusion model conditioned on a neural operator. (2/5)

A key takeaway is that both a priori and a posteriori performance of ML-based parameterization (stability, accuracy, etc) can be derived from insights embedded in the spectral representation of neural networks. Take a look at some of our older work if interested. academic.oup.com/pnasnexus/ar...

We find an interesting distribution of Gabor filters and low-pass filters before and after fine-tuning and predictable spectral dynamics of the hidden layers both during training and fine-tuning phase.

The key idea lies in analyzing the network in spectral space during training, inference, and fine-tuning. Interestingly, more often than not, generalizing to a new system means generalizing to a new shape of the Fourier spectrum and that is a key indicator of model performance a priori.

We have released the codes and the framework as a part of the pre-print. Do check it out if you are interested or work in this space. The framework is adaptable to other areas of geophysics and generally Earth system modeling beyond just the ocean and atmosphere.

The 8Km emulated ocean is then downscaled to a 4KM reanalysis product with a generative model. The coupled emulator + downscaling framework is long-term stable, demonstrates accurate kinetic energy spectrum, and has the right mean and variability over decadal time scales.

Some the key ideas in the work involves building a framework where instead of costly reanalysis products or forecasts which are downscaled, we built an ocean emulator at 8Km over the Gulf of Mexico which is long-stable, does not drift, and remain physically consistent.