AI Learns to Read the Ocean: Physically Interpretable Deep Learning Reshapes ENSO Prediction

The El Niño-Southern Oscillation, the vast sloshing of warm water and shifting winds across the tropical Pacific that reshapes weather from Peru to Ethiopia, has long defied the best efforts of forecasters. Now a systematic review published in Earth Science Informatics argues that the field is crossing a decisive threshold: deep learning models are shedding their black-box reputation and becoming physically interpretable machines that forecasters can actually trust. The review, led by Haolin Jiang of Anhui Jianzhu University with colleagues at Nanjing University of Information Science and Technology, King’s College London and the Chinese Academy of Meteorological Sciences, maps the journey from opaque neural networks of the late 2010s to today’s physics-consistent architectures, and lays out a roadmap for the next generation of intelligent climate prediction systems.

Why does ENSO prediction matter so much? The oscillation is the dominant year-to-year climate signal on Earth, and its warm El Niño and cold La Niña phases trigger droughts, floods, and shifts in monsoon timing that affect disaster mitigation and sustainable development worldwide. Yet the system is ferociously nonlinear, with multi-layer interactions between the ocean surface, the subsurface heat reservoir, and the atmosphere above. Traditional statistical tools such as Empirical Orthogonal Function analysis and Canonical Correlation Analysis compress these interactions into a handful of linear patterns, which works reasonably for short lead times but collapses when the system departs from linearity. Dynamical coupled models, which solve simplified equations of motion, carry physical logic but are expensive to run and still stumble over the notorious spring predictability barrier, the window around April to June when forecast skill drops sharply because small errors amplify explosively.

The first wave of deep learning attacks on ENSO arrived with a bang. In 2019, researchers showed in Nature that a convolutional neural network trained on sea surface temperature and heat content fields could deliver multi-year ENSO forecasts with skill exceeding that of conventional dynamical models. Convolutional Neural Networks treat climate fields as images and learn spatial pattern recognition, while Long Short-Term Memory networks track how ocean states evolve through time. Subsequent work added temporal convolutional networks, deep residual architectures with dropout and transfer learning, autoencoder-LSTM hybrids, and attention-based models that process three-dimensional multivariate ocean states. But the review is blunt about the flaw in these early triumphs: the models lacked physical consistency and suffered from opaque decision-making logic. A network might nail the Niño3.4 index for the right reasons or for the wrong ones, and nothing in the architecture could tell you which.

The heart of the review is a three-tier technical framework for physically interpretable deep learning, organized around the input, model, and output layers. At the input layer, the authors advocate multimodal data fusion oriented toward physical mechanisms, combining classes of variables that dynamical theory identifies as essential: sea surface temperature, upper-ocean heat content, and surface wind stress. Feeding a network the variables that the recharge-discharge oscillator and other ENSO theories single out is not just data augmentation; it orients the model toward the causal channels that actually drive the oscillation, rather than letting it latch onto spurious correlations in the historical record.

At the model layer, the review analyzes how physically constrained loss functions and mechanism-embedded architectures reinforce dynamical conservation. This is the fast-growing field of physics-informed machine learning, in which neural networks are penalized during training for violating analytic constraints such as energy balance or known equations of motion. Architectures can also embed mechanisms directly: models such as ENSO-ASC with its multivariate air-sea coupler and ENSO-GTC with its global spatial-temporal teleconnection coupler build the coupling between ocean and atmosphere into the network’s wiring, while graph neural networks represent the tropical Pacific as a network of interacting regions whose edges mirror real ocean-atmosphere connections. The payoff is dual, the authors argue: enforcing physics extends prediction lead times and deepens mechanistic understanding, because a model that respects conservation laws cannot drift into physically impossible states.

At the output layer, the review examines visualization techniques that verify the model’s physical logic. Saliency maps highlight which regions of the input fields most influence the forecast, Layer-wise Relevance Propagation traces the forecast back through the network’s layers to assign relevance to individual input pixels, and attention mechanisms reveal which spatial and temporal features the model weights most heavily. When such tools show that a forecast of an El Niño event draws its relevance from the western Pacific warm pool and the subsurface heat content along the equator, forecasters gain confidence that the network has learned genuine dynamics. When the relevance lands on irrelevant regions, the model can be diagnosed and retrained. Explainability thus becomes a quality-control instrument, not just a cosmetic feature.

The review also documents how interpretable models have begun to deliver scientific discoveries alongside forecasts. Explainable deep learning analyses have connected El Niño predictability to climate mode interactions, identified convection-sensitive oceanic regions that weaken the spring predictability barrier, and revealed enhanced ENSO predictability under historical anthropogenic forcing. Transformer-based models with integrated gradient methods, such as the 3D-Geoformer, have been applied to tropical basin interactions, showing that cross-basin exchanges reduce the spring barrier in deep learning forecasts. In effect, the network becomes a numerical laboratory: by interrogating what the model has learned, scientists extract hypotheses about which physical processes carry the predictability, hypotheses that can then be tested against observations and dynamical theory.

Serious challenges remain, and the review does not minimize them. Climate non-stationarity means the statistical relationships a network learns during training may shift as the background climate warms; ENSO itself is changing under anthropogenic forcing, with altered diversity and teleconnections documented in recent assessments. Small sample sizes compound the problem: the instrumental record contains only a few dozen ENSO events, a tiny dataset by deep learning standards, raising acute overfitting risks. Models trained on one climate regime may generalize poorly to another, and purely data-driven approaches cannot extrapolate reliably beyond the conditions they have seen. The review proposes future directions that address these gaps head-on: large-scale meteorological foundation models, such as ClimaX and the skillful medium-range global forecasters demonstrated in recent years, pretrained on vast reanalysis datasets and then fine-tuned for ENSO; physics-data driven hybrid models that couple neural components to dynamical cores; and generative approaches, including diffusion models, that bridge deep learning and physics while learning climate variability from scarce data.

The trajectory the review describes is striking: in barely half a decade, ENSO forecasting has moved from linear statistical patterns through black-box neural networks to physically consistent, explainable systems that extend lead times beyond two years in some studies while illuminating the underlying dynamics. The authors argue that this physically consistent reference framework provides a reliable foundation for next-generation intelligent ENSO prediction systems, ones that disaster planners, water managers, and agricultural forecasters can interrogate rather than merely obey. As artificial intelligence reshapes weather and climate science at large, the ENSO case may prove the template: the machines that earn trust are not the ones that predict best in the dark, but the ones that can show their work in the light of physical law.

,

Subject of Research: Application of physically interpretable deep learning in ENSO prediction: status, challenges, and future directions

Article Title: Application of physically interpretable deep learning in ENSO prediction: status, challenges, and future directions

Article References: Jiang, H., Fang, W., Li, Y., & Wang, Z. (2026). Application of physically interpretable deep learning in ENSO prediction: status, challenges, and future directions. Earth Science Informatics, 19(12), Article 215. https://doi.org/10.1007/s12145-026-02273-w

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02273-w

Keywords: Application, physically, interpretable, deep, learning, ENSO, prediction, status, challenges, future, directions, scientific research