El Niño: 20-Month Forecast with an Explainable Neural Network ✛
Context
El Niño forecasting is crucial for agriculture, energy, and managing flood and drought risks worldwide. Classical dynamical models predict the event reliably only 6–12 months ahead, and the spring predictability barrier sharply reduces the accuracy of forecasts made in spring. Deep neural networks have already pushed the horizon to 17–18 months, but most rely solely on sea surface temperature and ignore coupled ocean–atmosphere processes. This work aligns with current efforts to use hybrid architectures and multivariate data for long-term forecasting: neighboring studies have shown that convolutional networks capture spatial patterns and transformers capture long-range dependencies, but combining these approaches with a full set of physical variables remained an open challenge.
Methods
CTEFNet uses a hybrid architecture: three blocks of convolutional layers extract local spatial features, then a transformer with encoder and decoder models temporal dependencies. Inputs are 12 consecutive months of nine variables: sea surface temperature, heat content, mixed layer depth, salinity, sea level pressure, zonal and meridional components of currents and wind stress. All fields are regridded to 1°×2° and normalized. The model was trained on 18 CMIP6 climate models under the SSP370 scenario (2015–2100), validated on ORAS5 and ERA5 reanalyses (1958–1978), and tested on GODAS and ERA5 (1980–2021). The target variable was the Niño 3.4 index for the next 24 months. For interpretation, gradient sensitivity analysis was used: backpropagation computed the contribution of each input point to the index forecast, and statistical significance was selected by t-test at the 95% level.
Results
CTEFNet maintains forecast correlation above 0.7 for all seasons up to 12 months and above 0.6 up to 17 months, with an effective horizon (correlation above 0.5) reaching 20 months. This is 3–5 months farther than dynamical NMME models and 3–4 months farther than the best deep networks (CNN, Geoformer, ResCNN, ResoNet, STPNet). Seasonal analysis shows that the model retains correlation above 0.5 for more than 20 months for forecasts made from June to December, and 16 months for forecasts made in spring, indicating a mitigation of the spring barrier. Gradient analysis for 11 El Niño events revealed a sequence of precursors: 11 months before the peak, positive sensitivity of heat content in the western Pacific indicates warm water accumulation, while negative sensitivity of mixed layer depth indicates suppressed vertical mixing. By 8 months, signals spread to the central and eastern Pacific and the eastern Indian Ocean, corresponding to the onset of Bjerknes feedback. Concurrently, cooling of the northern tropical Atlantic weakens trade winds through atmospheric bridges, favoring El Niño development. By October, westerly winds intensify and warm water is transported eastward, while negative sea surface temperature sensitivity in the Indian Ocean triggers the Matsuno–Gill response, further fueling the event.
Implications
This work differs from neighboring studies by being the first to combine convolutional layers and a transformer in an encoder–decoder with a full set of nine oceanic and atmospheric variables, not just surface temperature. This allows the model to capture coupled interbasin interactions and achieve a forecast horizon of 20 months, 2–3 months farther than previous hybrid approaches. Furthermore, gradient sensitivity analysis is used here as a substitute for adjoint models, which are expensive and linear: backpropagation provides a nonlinear and computationally cheap estimate of each variable's contribution at each time step. This makes interpretation not just a demonstration but a working tool for testing physical mechanisms.
Future development
The model can be extended to distinguish El Niño types: canonical (eastern Pacific) and Modoki (central Pacific), which affect global climate differently. Another direction is further weakening the spring predictability barrier, which would require more detailed modeling of seasonal shifts in winds and currents, possibly using higher-resolution data or additional physical constraints in the loss function. Also promising is using the proposed gradient analysis to study the asymmetry between El Niño and La Niña, which has already emerged in preliminary calculations.
Impact
Improved El Niño forecasts will directly affect seasonal planning in agriculture, water resource management, natural disaster insurance, and energy logistics in regions dependent on rainfall and temperature. The gradient sensitivity method can be transferred to other climate modes, such as the Indian Ocean Dipole or the Atlantic Meridional Overturning Circulation.
Next steps
The immediate next step is adapting CTEFNet to forecast indices that distinguish El Niño types and testing on independent data beyond 1980–2021. It is also worth comparing model sensitivity with direct adjoint model calculations for individual events to quantitatively assess the accuracy of the gradient approach.
Key open problems
In climatology, the causes of the spring predictability barrier remain unresolved: why forecasts made in March–May are systematically worse than forecasts from other seasons. CTEFNet mitigates but does not eliminate this barrier, and its gradient analysis may indicate which processes become unstable during that period. Another open problem is the relative role of the three ocean basins in triggering El Niño: the model's sensitivity to the Atlantic and Indian Oceans provides a quantitative basis for testing hypotheses about interbasin teleconnections.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
This paper introduces CTEFNet, a hybrid neural network that combines convolutional layers and a transformer to forecast the Niño 3.4 index up to 24 months ahead. The model was trained on the CMIP6 ensemble and tested on the GODAS and ERA5 reanalyses, achieving a correlation above 0.5 out to 20 months and mitigating the spring predictability barrier. Gradient analysis revealed physically interpretable precursors, such as heat buildup in the western Pacific.
Where its strength is
The strength of this work is that it not only extends the forecast horizon but also uses the full set of physical ocean and atmosphere variables, while interpretability through gradient analysis links the forecast to known mechanisms, paving the way for further physical analysis.
Where it could go next
- It might be interesting to see how the approach of extracting precursors from multivariate data relates to the method of drought forecasting via reverse reading of time series: that approach also uses the idea of identifying early signals, but with a different architecture.based on: How a neural network, reading weather backwards, warns of drought months in advance arXiv:2607.00331
- A related work examines how convolutional networks capture spatial patterns for El Niño forecasting up to one and a half years ahead; comparing with it could help understand which components of CTEFNet provide the main gain.based on: El Niño Can Be Predicted a Year and a Half in Advance arXiv:2601.12251
- Another neighboring work uses lightweight architectures to improve accuracy and save resources; adapting similar techniques could be useful for speeding up CTEFNet training and inference without loss of quality.based on: Lightweight AI for Space Communications: More Accurate and Efficient arXiv:2607.15127
Worth trying
- Test whether the gain in forecast horizon persists if the model is trained only on sea surface temperature and heat content, as in some related works.
- Use the technique of reverse reading of time series to enhance sensitivity to slowly varying processes, such as heat accumulation in the western Pacific.
- Try replacing some convolutional blocks with lighter modules to reduce computational cost and facilitate deployment in operational systems.
Why this may matter
If accuracy can be maintained while simplifying the input data or architecture, this would make long-term El Niño forecasting more accessible for operational use, which is important for early warning of climate risks.
Nearby works in our archive
- El Niño Can Be Predicted a Year and a Half in Advancesimilarity 0.61
- Weather forecast on a single GPUsimilarity 0.52
- Solar Cycles: How Magnetic Swings Sway the Futuresimilarity 0.49
- Space Weather by the Seasonssimilarity 0.48
- How a neural network, reading weather backwards, warns of drought months in advancesimilarity 0.48
- An Underwater Hunter for Earth's Heatsimilarity 0.47
- Lightweight AI for Space Communications: More Accurate and Efficientsimilarity 0.47
The closest work in the archive sits at 0.61 by meaning — the usual level for our corpus.
The 1997–1998 El Niño was so strong that global sea surface temperature rose 0.2°C above the decadal average, and sea level in the eastern Pacific temporarily rose by 20 centimeters due to thermal expansion of water.
In Kim Stanley Robinson's novel 'Forty Signs of Rain', climatologists struggle with the consequences of a sudden change in ocean currents caused by warming; accurate long-term El Niño forecasting could be the tool the characters use to prevent disaster.
Key numbers
- effective forecast horizon: 20 months
- correlation above 0.7: up to 12 months
- correlation above 0.6: up to 17 months
- number of input variables: 9
- number of climate models for training: 18