🔬 physics

Njord: A Probabilistic Graph Neural Network for Ensemble Ocean Forecasting

Njord is a novel probabilistic graph neural network that leverages deep latent variables and adaptive K-means cluster meshes to generate efficient, uncertainty-aware ensemble ocean forecasts at both global and regional scales, outperforming deterministic baselines on the OceanBench benchmark.

Original authors: Daniel Holmberg, Joel Oskarsson, Erik Larsson, Fredrik Lindsten, Teemu Roos

Published 2026-05-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Daniel Holmberg, Joel Oskarsson, Erik Larsson, Fredrik Lindsten, Teemu Roos

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 dance floor. The water is constantly swirling, crashing, and shifting in unpredictable ways. For a long time, scientists trying to predict where the water will go next have used two main tools:

The Physics Engine: A super-complex, computer-based simulation that tries to calculate every single drop of water using the laws of physics. It's incredibly accurate but takes hours to run on massive supercomputers.
The Deterministic AI: A newer, faster machine learning model that learns from past data to guess the future. It's fast, but it's like a single person guessing the dance moves: it gives you one specific answer. If the ocean is chaotic, this single answer often misses the wild variations and extreme events, smoothing everything out into a boring, average prediction.

Enter Njord.

The authors introduce Njord (named after a Norse god of the sea), which is like a "super-charged" version of the fast AI. Instead of giving you just one guess, Njord acts like a crowd of fortune tellers. It generates a whole group of possible futures (an "ensemble") in the time it takes a deterministic model to make just one.

Here is how it works, broken down into simple concepts:

1. The "One-Step" Magic Trick

Most advanced AI models that try to predict many possibilities (like weather or ocean waves) are slow. They have to run a long, repetitive process to generate each different scenario.
Njord is different. It uses a special "latent variable" framework. Think of this as a secret ingredient (a random number) that the model injects into its brain at the very start of the prediction.

The Analogy: Imagine baking a batch of cookies. A slow model might bake one cookie, taste it, adjust the recipe, and bake the next. Njord mixes a little bit of "randomness" into the dough once, and then instantly bakes 20 slightly different cookies in a single go. This makes it incredibly fast.

2. The Flexible Map (Graph Neural Networks)

The ocean isn't a perfect grid like a chessboard; it has jagged coastlines, islands, and deep trenches. Traditional models often try to force the ocean into a rigid square grid, which wastes computer power on land (where there is no ocean to predict) or breaks the map in weird places.
Njord uses a Graph Neural Network (GNN).

The Analogy: Instead of a rigid chessboard, Njord uses a fishing net. The knots in the net (the nodes) are placed exactly where the water is. The net stretches and shrinks to fit the shape of the coastline perfectly.
The Innovation: The authors didn't just use a standard net; they used a "K-means clustering" method to place the knots. This ensures the net has knots in narrow straits and bays where the water moves fast, rather than leaving gaps or wasting knots on land. This allows the model to handle the messy, irregular shape of the real ocean.

3. Handling the Ice

The ocean isn't just water; it has sea ice, which has strict rules: it can't be negative, and it can't be more than 100% covered. If a model isn't careful, it might predict "negative ice" or "105% ice," which is physically impossible.
Njord treats ice like a light switch. It predicts two things at once:

Is there ice here? (On/Off)
How thick is it?
If the model predicts "Off," it forces the ice thickness to be exactly zero. This prevents the model from accidentally creating "ghost ice" that slowly accumulates over time and ruins the forecast.

4. The Results: Fast, Accurate, and Honest

The researchers tested Njord in two ways:

Globally: Looking at the entire world's oceans at a resolution of 0.25 degrees (about 25km).
Regionally: Looking specifically at the Baltic Sea at a very high resolution of 2km.

What they found:

Accuracy: Njord is just as good (or better) than the best existing fast AI models and even competes with the slow, heavy physics-based models. It made the most accurate predictions for surface temperature.
Uncertainty: Because it generates a crowd of forecasts, Njord can tell you how confident it is. If all 20 members of the "crowd" agree, the forecast is solid. If they disagree wildly (like in stormy areas or near ice edges), Njord shows you a large "spread" of possibilities. This is crucial for decision-makers who need to know the risk of extreme events.
Speed: It can generate a 20-member forecast in seconds on a single graphics card, whereas traditional methods might take hours.

Summary

Njord is the first machine learning model that can predict the ocean's future fast, accurately, and probabilistically. It doesn't just give you one answer; it gives you a spectrum of possibilities, helping us understand not just what might happen, but how likely it is to happen, all while respecting the messy, irregular shape of the real world.

Technical Summary: Njord – A Probabilistic Graph Neural Network for Ensemble Ocean Forecasting

Problem Statement

Ocean dynamics are inherently chaotic, yet existing machine learning (ML) ocean models predominantly produce deterministic forecasts. These models typically minimize mean squared error (MSE), encouraging predictions toward the conditional mean of future states rather than capturing the full predictive distribution. Consequently, they tend to smooth over fine-scale variance and offer limited insight into the probability of extreme events. For operational decision-making, where risk mitigation relies on understanding forecast confidence and the spectrum of possible scenarios, a single trajectory is often insufficient. While conditional generative models have emerged for probabilistic weather forecasting, no comparable probabilistic data-driven model existed for the ocean domain prior to this work. Furthermore, scaling such models to high-resolution global grids (0.25°) and irregular regional domains (e.g., the Baltic Sea at 2 km) presents significant challenges regarding memory, computational efficiency, and the handling of irregular coastal geometries.

Methodology

The authors introduce Njord, a probabilistic data-driven model for short-range, high-resolution global and regional ocean forecasting. Njord combines a deep latent variable framework with a Graph Neural Network (GNN) architecture, enabling the sampling of each forecast step in a single forward pass.

Core Architecture

Njord extends the Graph-EFM latent variable model (originally designed for weather) with specific modifications for ocean physics:

Latent Variable Framework: The model maps initial ocean states $X_{-p+1:0}$ and forcing inputs $F$ to a distribution of future states. It employs a hierarchical GNN where stochasticity is introduced via a latent variable $Z_t$ sampled from a Gaussian distribution $N(\mu_t, I)$ . The prior mean $\mu_t$ is learned via a separate GNN encoder.
Residual Prediction: The network predicts residuals $\hat{r}_t$ which are added to the previous state $X_{t-1}$ to produce the forecast $\hat{X}_t$ .
Graph Adaptation (K-means Cluster Meshes): Unlike global weather models that use icosahedral meshes, Njord utilizes a K-means cluster-based graph construction. Nodes are placed based on the density of ocean grid points using spherical K-means clustering with latitude-based area weights. This approach:
- Adapts to the irregular geometry of the ocean surface (bays, straits, canals) without breaking graph connectivity.
- Allows for a continuous selection of node counts, avoiding the discrete resolution jumps inherent in icosahedral refinement.
- Filters edges crossing land masses, ensuring the graph operates only over ocean points.
Sea Ice Modeling: To handle physical constraints (Sea Ice Concentration bounded to [0, 1]; Sea Ice Thickness $\ge$ 0), Njord employs a density channel mechanism combined with soft clamping. A binary density channel is predicted alongside other variables; if the predicted density falls below a threshold, ice variables are set to zero in the next state. This prevents the accumulation of unphysical residual values during autoregressive rollouts.
Scalability: To manage the memory footprint of large global grids (0.25°), the authors modify the GNN layers to use separate dimensionalities for grid-level embeddings, edge representations, and graph node representations. They also employ gradient checkpointing and a two-stage training schedule (pretraining on 1° data, finetuning on 0.25° data).

Regional Extension (Njord-Baltic)

The framework is adapted for regional modeling by introducing boundary forcing inputs ( $B_t$ ) derived from a global ocean model. This allows the regional model to condition on the surrounding ocean state without relying on data from the system it aims to replace at the boundaries.

Training Objective

Njord is trained using a combination of the Evidence Lower Bound (ELBO) and the Continuous Ranked Probability Score (CRPS). The loss function includes masking for bathymetric structure and weighting based on grid cell area, depth level, and the inverse variance of daily changes. A curriculum learning strategy is used, gradually activating the KL divergence term, introducing multi-step rollout training, and finally enabling the CRPS loss.

Key Contributions

First Probabilistic Global Ocean Emulator: Njord is presented as the first generative ensemble forecasting model for global ocean physics, operating at 0.25° resolution.
Novel Graph Layout: The introduction of a clustering-based graph layout that better conforms to the irregular geometry of the ocean surface compared to traditional icosahedral or quadrilateral meshes.
Sea Ice Integration: The incorporation of sea ice variables with explicit physical constraints (density channels and clamping) to ensure physically realistic fields, a feature often missing in high-resolution ML ocean models.
Comprehensive Evaluation: The model is evaluated against the OceanBench benchmark, demonstrating competitive performance against state-of-the-art deterministic models (GLONET, WenHai, XiHe) and operational physics-based systems (GLO12), while providing calibrated uncertainty estimates.
Regional Adaptability: The successful application of the framework to the Baltic Sea (Njord-Baltic) at 2 km resolution, achieving errors comparable to deterministic baselines while providing probabilistic forecasts.

Results

Global Performance (0.25°)

Accuracy: On the OceanBench benchmark, Njord achieves the lowest errors on average across upper-ocean variables when evaluated against real-world observations (IV-TT CLASS-4 dataset). It shows the largest improvements in surface temperature (SST) prediction, outperforming both operational baselines and other ML emulators.
Ensemble Calibration: The 20-member ensemble exhibits a Spread-Skill Ratio (SSR) that increases steadily toward 1 over the 10-day forecast horizon, indicating good calibration and reliable uncertainty representation.
Spatial Dynamics: The ensemble mean accurately captures global thermal structures and sharp gradients (e.g., Gulf Stream, Kuroshio). The ensemble spread correctly highlights regions of increased uncertainty in dynamically active areas.
Efficiency: Njord samples a single next-step ensemble member in 3 seconds on an AMD MI250X GPU.

Regional Performance (Baltic Sea, 2 km)

Accuracy: Njord-Baltic achieves RMSE values comparable to the deterministic baseline (SeaCast) while providing probabilistic forecasts.
Calibration: Fine-tuning on operational analysis data leads to consistent improvements in probabilistic calibration (higher SSR).
Uncertainty: The ensemble spread effectively highlights regions of high variability, such as the Danish Straits (Kattegat and Skagerrak) and marginal ice zones.

Significance and Claims

The paper claims that Njord represents a significant step forward in operational ocean forecasting by addressing the "deterministic gap" in current ML models. By providing well-calibrated uncertainty estimates alongside high-accuracy forecasts, Njord enables better risk mitigation and decision-making for applications ranging from maritime navigation to environmental monitoring.

The authors emphasize that while the current vertical resolution is sufficient for benchmarking, the framework is designed to scale to more vertical levels, additional variables, and coupled ocean-atmosphere systems in future work. The significance lies not only in the accuracy of the point forecasts but in the ability to generate ensembles that capture dynamic variability and quantify uncertainty at resolutions previously unattainable for probabilistic data-driven ocean models. The work demonstrates that probabilistic generative modeling can be successfully scaled to large, irregular ocean grids, offering a viable alternative to computationally expensive numerical simulations.