Source-linked AI summary
OceanLight: Efficient Global Ocean Forecasting via Geometry-Adaptive Unstructured Mesh Representation
Wei Wu, Xiang Wang, Hongze Leng, Qingye Min, Junxing Zhu, Junqiang Song
TL;DR
Global ocean forecasting must support climate and societal applications while managing computational demands. OceanLight uses a climatology-adaptive unstructured mesh with a graph neural network to forecast ocean states, achieving strong accuracy and dynamical consistency while reducing computation and memory use.
Problem
Reliable ocean forecasting supports climate monitoring, navigation, fisheries management, and extreme-event early warning, motivating accurate and computationally efficient methods.
Method
OceanLight combines an ocean-climatology-based unstructured mesh with a graph neural network to represent heterogeneous ocean regions efficiently.
Results
73.0% lower FLOPs accompany consistently lower forecast RMSE than GLONET and strong dynamical, geostrophic-balance, and kinetic-energy consistency.
Takeaways & Limitations
The unstructured representation supports scalable ocean forecasting while preserving forecast accuracy and physically consistent ocean dynamics.
Takeaways & Limitations
Large mesh regions exceeding 20% of ocean grid points can aggregate diverse features and reduce spatial resolution, potentially degrading forecast skill.
Abstract
from arXiv · showhide
Reliable global ocean forecasting is critical for climate monitoring, marine navigation, and extreme event early warning. Physics-based ocean forecasting models impose prohibitive computational costs, while existing deep learning approaches predominantly rely on structured-grid architectures, incurring unnecessary computation on masked land cells and enforcing uniform resolution across dynamically heterogeneous ocean regions regardless of local flow complexity. Here we present OceanLight, an efficient global ocean forecasting framework innovatively combining geometry-adaptive unstructured mesh tokenization with a graph neural network (GNN) backbone. OceanLight achieves pointwise forecast accuracy and kinetic energy spectral fidelity exceeding both operational numerical analyses and state-of-the-art AI-based models, while surpassing all AI-based ocean models in geostrophic balance consistency. Furthermore, OceanLight demonstrates reliable mesoscale eddy representation, capturing coherent ocean structures beyond pointwise statistical optimization. These capabilities are delivered with a 62% reduction in GPU memory consumption and 70\% reduction in FLOPs relative to structured-grid baselines. Our unstructured mesh representation establishes a generalizable paradigm for scalable data-driven oceanography.
1 Introduction
OceanLight addresses the computational limitations of structured-grid ocean forecasting by combining climatology-aware unstructured mesh tokenization with a GNN backbone. The framework is motivated by the need for reliable forecasts across climate monitoring, navigation, fisheries management, and extreme-event early warning.
- Motivation: Reliable ocean forecasting supports climate monitoring, marine navigation, fisheries management, and early warning for storm surges and coastal flooding.The ocean’s storage and transport of heat, freshwater, and momentum influence weather patterns, sea level variability, and marine ecosystems.
- Motivation: Operational forecasting uses physics-based general circulation models such as NEMO, MOM6, and HYCOM on structured grids, but high-resolution global simulations require large-scale HPC resources.The passage identifies these computational demands as a central limitation motivating data-driven alternatives.
- OceanLight framework: OceanLight is introduced as a climatology-aware unstructured-mesh framework designed to overcome limitations of structured-grid ocean forecasting.Its inputs include sea surface height, temperature, salinity, and current velocities across 23 vertical depth levels.
- Contributions: OceanLight reportedly surpasses traditional numerical models and the state-of-the-art GLONET baseline in the introduction’s quantitative summary.The introduction states that OceanLight matches numerical simulations in geostrophic balance consistency, leads competitors in kinetic energy spectral fidelity, and reconstructs coherent mesoscale eddies.
- OceanLight framework: OceanLight combines geometry-adaptive unstructured mesh tokenization with an encode–process–decode GNN to forecast global ocean states from 0.25° inputs.The encoder maps ocean variables onto a compact unstructured mesh, an L-layer processor performs message passing, and the decoder returns outputs to the original grid.
2 Results
OceanLight consistently outperforms GLONET and GLO12 and is comparable to or slightly better than GraphCast across multi-variable ocean forecasts, while degrading more slowly at longer lead times. It also improves geostrophic consistency, kinetic-energy fidelity, and computational efficiency relative to competing models.
- Forecast accuracy: OceanLight achieves lower RMSE and higher ACC than GLONET across variables and lead times, while matching or slightly exceeding GraphCast in most cases.Its advantage is especially consistent for sea surface height and surface currents.
- Forecast accuracy: OceanLight shows a slower accuracy-degradation trend with increasing lead time, especially for dynamic ocean-current variables, indicating strong multi-variable generalization.All models lose accuracy as lead time increases under the IV-TT evaluation framework.
- Geostrophic consistency: OceanLight has the lowest geostrophic-current RMSE and the highest zonal and meridional correlation coefficients, indicating stronger internal geostrophic balance.The especially strong meridional improvement reflects better capture of two-dimensional SSH-gradient and mesoscale dynamical structure.
- Kinetic-energy fidelity: OceanLight achieves the lowest RMSE for KE, MKE, and EKE, with most variance, amplitude, and area-integral ratios closer to 1 than competing models.GLONET instead substantially overestimates energy intensity and spatial variability, with ratio metrics significantly above 1.
- Computational efficiency: Compared with GraphCast, OceanLight reduces mesh nodes by 97.0%, mesh edges by 99.9%, FLOPs by 73.0%, and encoder, processor, decoder, and backpropagation training times by 40.3%, 92.2%, 61.4%, and 71.2%.It also reduces grid→mesh edges by 55.7% and mesh→grid edges by 76.8%.
3 Methods
OceanLight constructs a geometry-adaptive unstructured mesh from ocean climatology and applies an encode-process-decode GNN to forecast global ocean states at 0.25° resolution. The method combines feature-guided region segmentation, mesh–grid message passing, and a sparse graph topology to reduce computational cost while preserving physical information transmission.
- Graph architecture: The model uses 1,212 climatology-based regions and 4,120 mesh edges instead of uniform icosahedral refinement, yielding a substantially sparser graph.Its architecture follows the encode-process-decode paradigm, with 16 unshared message-passing layers and residual connections for nodes and edges.
- Feature-guided segmentation: PCA compresses multivariate temperature, salinity, velocity, and sea-surface-height features into a scalar field whose Sobel gradient guides seed-point selection.The input climatology includes ocean variables across 23 discrete depth levels.
- Adaptive region construction: Threshold-based region growing partitions spatially connected areas from minimum-gradient seeds, but regions exceeding 20% of the ocean grid are refined with HDBSCAN and connected-component decomposition.The refinement prevents large regions from aggregating diverse ocean features and degrading spatial heterogeneity representation.
- Grid–mesh coupling: The final graph combines first-order and 2-hop mesh adjacency with bidirectional edges linking each 0.25° grid node to its assigned mesh region.An Interaction Network propagates information across grid and mesh nodes through directed message passing.
- Forecast decoding: The grid–mesh topology reduces edge count while preserving essential physical information pathways, providing computational efficiency without compromising forecasting performance.The decoder maps mesh information back to grid nodes using MLP embeddings and GNN-based message passing.
4 Discussion
OceanLight’s climatology-based unstructured mesh reduces computational and training demands while adapting resolution to oceanic geometry and variability. Its representation is architecture-agnostic, but current meshes do not capture transient daily dynamics.
- Mesh design: Climatology-driven region growing and HDBSCAN clustering create coarse meshes in stable open oceans and finer meshes near land-sea boundaries and other rapidly changing regions.
- Efficiency: OceanLight trains on a single 40GB GPU, whereas GLONET requires multiple 40GB GPUs and model-parallel training, while improved time efficiency further reduces training cost.
- Limitations: The climatology-based mesh reflects long-term averaged conditions and therefore omits transient short-term processes relevant to high-resolution forecasting.
- Generalizability: The sparse, irregular unstructured mesh is proposed as an architecture-agnostic spatial representation whose cell tokens interact only with geometrically local neighborhoods.
Supplementary
The supplementary material defines diagnostics for geostrophic balance, pointwise and anomaly accuracy, energy-field agreement, and mesoscale eddy forecast quality. It also provides regional, seasonal, vertical-profile, eddy-structure, and annual mean-field comparisons.
- Evaluation metrics: Geostrophic currents are derived from predicted sea surface height using geostrophic balance, and Pearson correlations compare derived and predicted zonal and meridional currents.The correlation is computed separately for the zonal and meridional velocity components.
- Evaluation metrics: RMSE measures pointwise prediction accuracy, while ACC measures spatial correlation relative to climatological anomalies.
- Energy diagnostics: Variance, amplitude, and area-integral ratios evaluate agreement between modeled and observed KE, MKE, and EKE fields, with values closer to 1 indicating better agreement.These metrics assess spatial variability amplitude, mean energy intensity, and total integrated energy magnitude, respectively.
- Mesoscale eddy evaluation: Bias, CSI, FAR, and POD quantify mesoscale eddy forecast accuracy using categorical hit, false-alarm, and miss outcomes.