The World as a Graph: Neural Network Learns to Predict El Niño
Context
Seasonal forecasting of El Niño has long hit two walls: the chaotic nature of the phenomenon and limited observations. Dynamical models, which faithfully simulate ocean and atmosphere physics, are resource-hungry and not always accurate at horizons of several months. A recent breakthrough with convolutional neural networks (CNNs) showed that climate simulation data can yield forecasts up to a year and a half, but CNNs have inherent limitations: they assume the same pattern matters everywhere on the map and struggle to capture long-range connections. Alongside this work on the preprint server, other studies use neural networks to improve telecommunications, read molecules, or speed up galaxy formation; all of them grapple with the same issues of locality and interpretability. This work stands apart: it brings the idea of graphs to climate, where nodes are ocean points and edges are invisible teleconnections.
Methods
The authors framed ONI forecasting as a regression problem on a graph. Nodes are 1345 ocean grid cells at 5-degree resolution (land excluded); features are sea surface temperature and heat content anomalies over three months. The key novelty is a trainable connectivity module: the adjacency matrix is not predefined but computed from static node attributes (temperature, heat content, latitude, longitude) via two small networks and a sigmoid. To avoid a fully connected graph, only the e=8N strongest edges are kept. A standard graph convolutional network (GCN) with residual connections and jumping knowledge then propagates messages along the learned edges, and the final graph representation is assembled by averaging and summing node embeddings. The model was trained on CMIP5 simulations and SODA reanalysis (about 30,000 examples in total) and tested on the independent GODAS dataset from 1984 to 2017. Unlike the earlier CNN, it did not require 12 separate models for different seasons: one model per forecast horizon sufficed.
Results
Graphiño outperformed the best deep model (CNN) for forecasts 1 to 6 months ahead and the dynamical model SINTEX-F at all horizons. Correlation with observations reached 0.975 at 1 month, 0.917 at 3 months, and 0.780 at 6 months, compared with 0.942, 0.876, and 0.762 for the CNN. For the rare and strong 2015 El Niño, the model achieved a correlation of 0.948 over the year, beating nearly all statistical and most dynamical models. One ensemble member reached a correlation of 0.408 at 23 months ahead, notably higher than the ensemble mean (0.336) and the previous record (0.287). Analysis of the learned adjacency matrix revealed that node importance (eigenvector centrality) shifts with horizon: at 1 month, key points cluster in the ONI region; by 6 months, influence spreads to the western Pacific; and by 23 months, to the Indian and Atlantic Oceans. This aligns with known theories: Kelvin and Rossby waves, the recharge oscillator, and Indian Ocean influence.
Implications
This work brings graph neural networks into seasonal climate forecasting and shows that a learnable connectivity structure can be more than a trick: it can provide physical insight. Unlike neighboring studies where graphs are mostly applied to traffic or molecules, this is the first demonstration on real climate data that a GNN beats a CNN precisely because it captures global, not local, connections. The interpretability of the learned edges, a rare property for deep models, allows testing hypotheses about El Niño mechanisms. This is a step from black box to a tool that can hint to physicists which regions matter for prediction.
Future development
A natural next step is to model time explicitly: currently, three months are simply concatenated into one vector, while spatiotemporal graph networks have already proven themselves in traffic forecasting. Adding winds, pressure, and other variables could boost accuracy at longer horizons. A separate challenge is rare extreme events: the model systematically underestimates the strength of the record El Niños of 1998 and 2016, and specialized loss functions or data augmentation might help. The learned connectivity structure itself deserves study: it may point to new teleconnections not yet described in theories.
Impact
El Niño forecasts directly affect agriculture, insurance, water resource management, and disaster preparedness in the tropics and beyond. A more accurate forecast six months ahead gives time to prepare for drought or flood. The method is also transferable to other climate indices and regions.
Next steps
Immediate next steps include incorporating wind and other atmospheric variables, trying spatiotemporal architectures, and tuning the model for extreme events. It would also be useful to compare the learned connections with independent teleconnection data from climatology.
Key open problems
For climatology, the key unsolved problem is the spring predictability barrier: models sharply lose accuracy when forecasting across spring. The learned connections at 9 months ahead point to the Niño5/Niño6 region, consistent with the western Pacific oscillator theory, and may help identify which observations are needed to overcome this barrier. Another problem is rare super El Niños, which the model underestimates; this is a common weakness of statistical approaches trained on short observational histories.
The model required 12 times fewer neural networks than the previous record: one network per horizon instead of twelve for each season. Training took only 33 to 50 minutes on four GPUs.
In Kim Stanley Robinson's novel 'Forty Signs of Rain', climatologists struggle to predict sudden climate shifts. Graphiño is a step toward making such forecasts routine rather than dramatic.
Key numbers
- correlation at 1 month: 0.975
- correlation at 6 months: 0.780
- correlation at 23 months (best ensemble member): 0.408
- number of graph nodes: 1345
- maximum number of edges: 10760 (8 per node)