Modern El Niño forecasts increasingly rely on ensembles of machine learning models that achieve high accuracy but remain "black boxes." Existing approaches around this work explain deep network predictions through gradient-based sensitivity maps, yet these are post hoc and often unstable. This study aligns with the search for interpretable methods but differs by embedding interpretability into the model structure itself rather than adding it externally.
The authors train an ensemble of 50 entropy-optimal sparse probabilistic approximations (eSPA) on satellite-era data (SST, vertical temperature gradient, wind stresses) for each forecast lead time from 1 to 24 months. They then select only those models that correctly predicted the phase and cluster their internal states into 12 "superclusters" using k-means with feature importance weights. For each lead time, transition matrices between clusters over time and across lead times are constructed, allowing reconstruction of typical event development pathways.
Distilled models retain the predictive accuracy of the original ensemble: the ranked probability skill score (RPSS) remains above zero for all lead times except 6 months and is significantly positive for most lead times. The mean relaxation time of the Markov chain of superclusters is 6.0 months, close to the e-folding time of the Niño-3.4 index (6.1 months). The effective dimensionality of the feature space peaks when the forecast must cross the spring predictability barrier. Spatial importance maps highlight known precursors such as the North Pacific Meridional Mode, and the case study of the 2015/16 El Niño shows how the model traces the path from extratropical anomalies to a mature event.
This work differs from neighboring studies in that interpretability is not imposed post hoc on a deep network but is extracted directly from the structure of an entropy model ensemble. This yields more stable and physically consistent explanations of forecasts than gradient methods and allows constructing causal-like narratives of event development. The resulting superclusters and transition matrices serve as a compact diagnostic tool for analyzing long-term El Niño predictability.
The method can be extended to continuous target variables, such as the relative Niño-3.4 index, or to multi-index formulations including the Indian Ocean Dipole. A shift to cyclostationary superclusters, specific to each season, is possible to account for seasonal inhomogeneity in dynamics. Also promising is replacing verification-based filtering with a meta-model that would select successful ensemble members in real time.
The results could improve operational seasonal climate forecasts by making them more transparent and trustworthy. They are also useful for climatologists studying the physical mechanisms of El Niño, as they provide quantitative precursor maps.
The next step will be developing a meta-model for real-time selection of successful ensemble members, enabling distillation in operational mode. Plans also include incorporating categorical features into the affiliation calculation to improve short-term forecasts.
The work links distillation to the fundamental problem of climate predictability: the spring predictability barrier of El Niño. It shows that the complexity of input data required for a successful forecast is maximal precisely when crossing this barrier, consistent with physical notions of weak precursor signals.
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
The authors distill an ensemble of 50 entropy-optimal sparse probabilistic approximations (eSPA) to forecast the El Niño phase up to 24 months ahead. They select skillful models, cluster their states into 12 superclusters, and build transition matrices, which allows reconstructing typical event pathways. The distilled models retain the ensemble's accuracy and yield physically meaningful feature importance maps.
Where its strength is
Interpretability here is built into the model structure itself, not added post hoc: the superclusters and transition matrices provide stable and physically consistent explanations of the forecasts, suitable for causal analysis.
Where it could go next
Worth trying
Why this may matter
The distilled models can serve as a compact diagnostic tool for operational forecasting centers, enabling real-time tracking of El Niño evolution and a better understanding of predictability sources beyond the spring barrier.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.50, against a usual 0.64. It looks like few have drilled nearby.
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.