Research map / Time series / Causal and generative time series

Njord: A Probabilistic Graph Neural Network for Ensemble Ocean Forecasting

NeurIPS 2026 · Time series › Causal and generative time series · approach: Probabilistic and conformal forecasting · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Njord is a probabilistic graph-neural-network latent-variable ensemble model for global (0.25°) and Baltic Sea (2 km) ocean forecasting, using K-means cluster meshes adapted to the ocean geometry.

Problem

Existing machine learning ocean forecasting models are deterministic and trained with MSE, so they cannot quantify uncertainty or capture the full predictive distribution of chaotic ocean dynamics at high resolution.

Approach

Combines a hierarchical GNN conditional latent variable model (extended from Graph-EFM) with a K-means cluster-based mesh that places nodes by ocean grid density, trained with ELBO plus almost-fair CRPS, enabling single-forward-pass ensemble sampling for ocean forecasting; also adds sea ice with a density channel for physical constraints.

Key results

• On OceanBench, Njord achieves the lowest errors on average across upper-ocean variables against observations, with largest gains in surface temperature.
• Njord achieves the lowest RMSE for 0-5 m temperature across all lead times on the observation track, outperforming GLO12 and ML emulators.
• Njord-Baltic achieves RMSE comparable to SeaCast while providing calibrated 5-member ensemble forecasts; a single next-step member samples in 3 s (global) and 1 s (Baltic) on one MI250X GPU.

Surprising findings

• Machine learning models trained only on reanalysis can outperform the physical GLO12 system on several reanalysis fields.

Limitations

• Only 6 depth levels (global) and 5 (Baltic) modeled, limiting vertical coverage.
• Ensemble generation costs more than deterministic models; minimum-error mean needs multiple members.
• Visual artifacts in Baltic SLA standard deviation fields, possibly from noisy satellite SLA targets.

Experimental setup

Datasets: GLORYS12 GLO12 ERA5 IFS Baltic Sea Physics Reanalysis OceanBench IV-TT CLASS-4 NEMO analysis

Benchmarks: OceanBench IV-TT CLASS-4 Baltic Sea NEMO analysis

Metrics: RMSE SSR CRPS

Baselines: GLONET WenHai XiHe GLO12 SeaCast Persistence

Builds on: Graph-EFM SeaCast GraphCast OceanBench

Closest papers by meaning

• OceanCBM: A Concept Bottleneck Model for Mechanistic Interpretability in Ocean Forecasting (NeurIPS 2026 · similarity 0.86)
• Probabilistic Weather Forecasting with Hierarchical Graph Neural Networks (NeurIPS 2024 · similarity 0.86)
• FuXi-Ocean: A Global Ocean Forecasting System with Sub-Daily Resolution (NeurIPS 2025 · similarity 0.86)
• Common Task Framework For a Critical Evaluation of Scientific Machine Learning Algorithms (NeurIPS 2025 · similarity 0.84)
• Tyche: One Step Flow for Efficient Probabilistic Weather Forecasting (NeurIPS 2026 · similarity 0.84)
• Mesh Interpolation Graph Network for Dynamic and Spatially Irregular Global Weather Forecasting (NeurIPS 2025 · similarity 0.82)