A New AI Method Just Made El Niño Predictions Explainable — And Controllable

When a neural network predicts a catastrophic El Niño event 10 months in advance, scientists need more than a number. They need to know why. What oceanic signals, what atmospheric patterns, are driving that forecast? Without an explanation, even the most accurate AI model remains a black box — useful, perhaps, but not trustworthy, not insightful, and certainly not actionable in the way climate science demands.

Now, a new method called average gradient outer product (AGOP) XAI is changing that. In a 2026 study by Yuan Hui, Dorian S. Abbot, and Robert J. Webber, AGOP didn’t just explain AI predictions of El Niño–Southern Oscillation (ENSO) events — it passed three out of four rigorous scientific tests for explainability better than any existing method. More strikingly, when researchers used AGOP’s explanations to nudge a climate model in real time, they could trigger or suppress extreme El Niño and La Niña events 10 months later.

This isn’t just about better AI. It’s about turning machine learning from a predictive tool into a hypothesis-generating engine for climate physics — and potentially, a control knob for Earth’s most powerful climate oscillator.

The Science

The team focused on one of climate science’s hardest problems: forecasting ENSO, the irregular warming and cooling of the tropical Pacific that drives droughts, floods, and heatwaves across the globe. While AI models have recently achieved lead times of up to 1.5 years (Ham et al., 2019; Liu et al., 2023), their inner workings remain opaque. Are they learning real physics, or just statistical coincidences?

To find out, the researchers trained three types of neural networks — a multilayer perceptron (MLP), a convolutional neural network (CNN), and a vision transformer (ViT) — to predict the Niño-3 index (a key ENSO indicator) using data from the Zebiak–Cane model, a well-established intermediate-complexity climate model of the tropical Pacific (Zebiak and Cane, 1987). The input included sea surface temperature (SST), thermocline depth, and zonal and meridional ocean currents across a broad region of the Pacific (

).

They then applied four explainable AI (XAI) methods to interpret the models’ predictions:

• Gradient (GRAD): the direction of steepest change in the prediction.
• Integrated Gradients (IG): accumulates gradients along a path from a baseline state.
• GradientSHAP: a SHAP approximation that averages IG over multiple baselines.
• AGOP XAI: a new method that uses the average gradient outer product across the training data to define a global sensitivity structure, then tailors explanations to individual inputs.

The key innovation of AGOP lies in its use of second-order gradient statistics. While traditional XAI methods rely on local gradients at a single input, AGOP computes:

$$\mathbf{M}=\frac{1}{n}\sum _{j=1}^n\nabla f({\mathbf{z}}_j)\nabla f({\mathbf{z}}_j{)}^{\top }$$

This matrix $\bm{M}$, known as the average gradient outer product, captures directions in input space that consistently produce large changes in the model’s output across many training examples. AGOP then generates an input-specific explanation by transforming the input anomaly through the square root of this matrix:

$$\mathbf{e}(\mathbf{x})=\frac{{\mathbf{M}}^{1/2}(\mathbf{x}-\mathbf{\mu })}{\Vert {\mathbf{M}}^{1/2}(\mathbf{x}-\mathbf{\mu })\Vert }$$

This ensures that the explanation is both globally informed and locally relevant.

To evaluate these methods, the authors introduced four quantitative scores, each scaled to an optimal value of +1:

1. Sensitivity: Does the explanation point in a direction that strongly affects the prediction? (Score: [0,1])
2. Attribution: Can the explanation reproduce the full prediction change from baseline? (Score: [−1,1])
3. Robustness: Does the explanation stay stable under small input changes? (Score: [−1,1])
4. Coherence: Is the explanation spatially smooth, as physical mechanisms tend to be? (Score: [−1,1])

These criteria reflect what climate scientists actually need: explanations that are not just statistically valid, but physically meaningful and stable.

What They Found

Across multiple neural network architectures and forecast lead times, AGOP XAI consistently outperformed the competition — but not in every category.

XAI Method Performance Across Criteria

Label	Value
Attribution	0.85 score
Robustness	0.78 score
Coherence	0.82 score
Sensitivity	0.65 score

AGOP XAI scores are based on CNN architecture at 10-month lead time (Hui et al., 2026).

The results show that AGOP achieved the highest attribution, robustness, and coherence scores, often by a wide margin. For example, in the CNN architecture at a 10-month lead time, AGOP scored 0.85 on attribution, compared to 0.45 for GradientSHAP and 0.38 for IG. On robustness, AGOP reached 0.78, while GradientSHAP and IG hovered around 0.5. Coherence scores were even more striking: AGOP scored 0.82, while Gradient and IG fell below 0.3 due to noisy, patchy patterns.

Only on sensitivity did AGOP fall short — and by design. The gradient method, which defines sensitivity, scored nearly 1.0, as expected. But high sensitivity alone is not enough. As the authors show in a simple linear example (Figure 1), a method can be perfectly sensitive yet fail to reproduce the actual prediction change — a critical flaw for scientific interpretation.

The visualizations of AGOP explanations reveal why it excels.

shows the AGOP explanation for an extreme El Niño event (+4.45°C in Niño-3) 10 months before onset. The CNN-based AGOP explanation highlights a broad equatorial thermocline-depth anomaly, with a deepened thermocline in the central Pacific and a pronounced southeastern Pacific lobe — a pattern consistent with the recharge oscillator theory of ENSO (Jin, 1997; Meinen and McPhaden, 2000), which describes how heat builds up and discharges across the Pacific basin.

In contrast, GradientSHAP and IG produce fragmented, high-frequency patterns — especially in the ViT and MLP models — with sharp boundaries and patchy artifacts that lack physical plausibility. The CNN’s AGOP explanation, by comparison, is smooth, coherent, and physically interpretable.

For a strong La Niña prediction, the AGOP explanation showed approximately sign-reversed thermocline patterns, consistent with the opposite phase of ENSO. This symmetry further supports its physical relevance.

But the most dramatic finding came not from analysis, but from intervention.

The researchers used the AGOP explanation vector to construct optimized perturbations in the Zebiak–Cane model. Starting from a near-neutral state, nudging the ocean state in the direction of the AGOP El Niño explanation led to a strong El Niño event 10 months later. Conversely, nudging in the opposite direction — suppressing the AGOP pattern — prevented an impending extreme event from developing.

This wasn’t just correlation. It was causation — or at least, model-scale causation. The AGOP explanation didn’t just describe what the AI was seeing; it identified a controllable pathway to extreme ENSO events.

Why This Changes Things

For decades, ENSO prediction has been a game of pattern recognition and statistical modeling. The recharge oscillator, the delayed oscillator, the western Pacific oscillator — these are all theoretical frameworks built from simplified physics. AI has now surpassed them in predictive skill, but without interpretability, it risks becoming a self-contained oracle: accurate, but isolated from scientific progress.

AGOP XAI bridges that gap. It doesn’t just open the black box — it turns the AI into a collaborator.

Consider the southeastern Pacific lobe in the AGOP explanation. This feature isn’t predicted by the simplest recharge oscillator models, which assume symmetry about the equator. Its presence suggests that asymmetries in ocean dynamics or wind feedbacks may play a more important role in extreme El Niño development than previously thought — a hypothesis that can now be tested in full-physics climate models.

Moreover, AGOP’s success across architectures reveals something deeper: not all neural networks learn the same physics. The CNN produced the most coherent explanations, likely because its inductive bias toward spatial continuity aligns with the smooth nature of oceanic processes. The ViT, designed for patch-based vision tasks, generated explanations with artificial boundaries between patches — a reminder that architecture shapes not just accuracy, but interpretability.

This has profound implications for climate modeling. If we want AI to enhance our understanding, not just replace it, we must design models that learn in physically plausible ways. AGOP provides a tool to audit that learning process.

But perhaps the most transformative implication is control. The ability to suppress or trigger ENSO events in a model using AGOP-guided nudging suggests a future where AI doesn’t just forecast climate extremes — it helps us manage them.

Imagine a world where, months before a predicted super-El Niño, we could deploy a fleet of ocean pumps or reflective buoys to nudge the thermocline in the opposite direction, dampening the event’s intensity. Or where early warning systems don’t just alert governments — they suggest interventions calibrated by AI to maximize impact.

This is not geoengineering in the stratospheric aerosol sense. It’s microclimate steering: small, targeted actions guided by deep understanding, aimed at stabilizing a chaotic system at its tipping points.

Of course, we are far from doing this in the real world. The Zebiak–Cane model is simplified. Real oceans are noisier, more complex, and influenced by global feedbacks. But the principle is now demonstrated: explanations can be levers.

What’s Next

AGOP is not a panacea. Its sensitivity score lags behind the gradient method, meaning it may miss some fine-scale, high-impact features. And like all XAI methods, it depends on the quality of the underlying model. If the AI has learned spurious correlations, AGOP will explain them faithfully — but wrongly.

The authors also note that the southeastern Pacific lobe in the AGOP explanation may be specific to the Zebiak–Cane model, not a universal feature of ENSO. Testing AGOP on AI models trained on observational data or high-resolution climate simulations will be crucial.

Future work could extend AGOP to other climate phenomena — the Indian Ocean Dipole, the Atlantic Multidecadal Oscillation, or even abrupt shifts in the Atlantic Meridional Overturning Circulation. Each is a candidate for AI prediction, and each demands physical understanding.

Another frontier is real-time feedback. Could an AGOP-guided system monitor the Pacific in real time, detect emerging ENSO precursors, and recommend interventions? The computational cost of AGOP is low once $\bm{M}$ is computed, making it feasible for operational use.

Finally, there’s the question of uncertainty. AGOP provides a single explanation vector, but climate is probabilistic. Combining AGOP with ensemble methods or Bayesian neural networks could yield explanations with confidence intervals — not just what drives a prediction, but how certain we are about it.

The era of uninterpretable AI in climate science may be ending. With methods like AGOP, we’re not just predicting the future — we’re beginning to understand, and perhaps even shape, the forces that govern it.

As the authors put it: “Beyond diagnosing neural-network behavior, AGOP XAI can generate candidate hypotheses about physical mechanisms.” That’s not just explainability. That’s discovery.

Comparison of XAI Methods

Label	Value
Attribution	0.85 score
Robustness	0.78 score
Coherence	0.82 score
Sensitivity	0.65 score

Scores reflect performance on CNN-based ENSO prediction at 10-month lead (Hui et al., 2026).