Distilling the Ocean: How to Compress Thousands of El Niño Models into Twelve Pure Drops ✛ ✎

Original: "Distillation and Interpretability of Ensemble Forecasts of ENSO Phase using Entropic Learning"

· Michael Groom, Davide Bassetti, Illia Horenko, Terence J. O'Kane

arXiv:2602.16857 · 2026-02-15 · arXiv non-exclusive · independent analysis · 2 min

Computational Physics physics.data-an Geophysics Statistical ML

From thousands of forecasting models, only the successful ones are selected, merged into twelve transparent states, and yield a two-year forecast that reads like a map.

Thousands of models simmer in a single cauldron, and the steam rises thick and impenetrable. Each whispers something about the future of El Niño, but in the general din you cannot tell who is right and who is just boiling over. Scientists have devised a still: they select only those models that have already guessed the phase of the event and distill from them twelve pure prototype drops. These drops are the superclusters, the typical states toward which successful predictors converge.

A supercluster is not a single model but a composite portrait: dozens and hundreds of models that share similar internal representations are drawn to it.

The distillation is cleverly arranged. First, an ensemble of fifty sparse probabilistic approximations is brewed, trained on satellite-era data: sea surface temperature, vertical temperature gradient, wind stress. For each lead time from one to twenty-four months, there is a separate pot. Then comes the tasting: a model that erred in phase is poured away, while one that guessed correctly is kept. The remaining models are clustered, but not just anyhow: the clustering accounts for which features each model considers important. This yields twelve superclusters for each lead time.

Next, the drops are strung into a chain. Transitions between superclusters over time form a Markov chain: the next state depends only on the current one. The mean relaxation time of this chain is six months. Almost like the Niño-3.4 index itself, whose characteristic decay time is 6.1 months. This is no coincidence: the distillate has preserved the ocean's flavor.

The effective dimensionality of the feature space peaks when the forecast must cross the spring predictability barrier, a period when El Niño forecasts sharply lose accuracy.

Feature importance maps extracted from the distillate reveal known precursors, such as the North Pacific Meridional Mode. The case study of the 2015/16 El Niño shows how the model traces the path from extratropical anomalies to the mature event. Previously, such insights could only be gleaned from gradient sensitivity maps, which are applied post hoc and are often unstable. Here, interpretability is built into the structure itself.

The cauldron no longer belches black-box smoke. Twelve drops for each lead time hold all the aroma of the forecast, but now they can be examined under a microscope. And if a drop turns cloudy, you can see which feature was stirred up.

Whole graph →

An ensemble of 50 models for each of the 24 forecast lead times and 252 start dates yields 302,400 models analyzed in the study. Distillation compresses them to just 12 superclusters per lead time, a reduction by a factor of 25,200.

D_{\mathrm{eff}} = \exp\left(-\sum_{d=1}^{D} W_d \log W_d\right)

Entropic measure of the number of significant features: equals 1 if only one feature is important, and D if all features are equally likely.

r = \frac{1}{1 - |\lambda_2|}

Characteristic decay time of correlations in the chain of transitions between superclusters; determined by the second largest absolute eigenvalue of the transition matrix.

Concepts

ensemble forecast spring predictability barrier importance sampling nested sampling El Niño index Multivariate ENSO Index epidemic model ensemble average probabilistic forecast IRI parameter estimation

I am the author

Distilling the Ocean: How to Compress Thousands of El Niño Models into Twelve Pure Drops ✛ ✎