Fang and Chen (2023) established an information-theoretic framework that measures the potential predictability of individual El Niño–Southern Oscillation (ENSO) events as the information gain (relative entropy) between an ensemble forecast probability density function and the observational climatological distribution, computed with a statistically accurate multiscale stochastic conceptual model. The framework was validated only in that cheap conceptual setting, and the paper explicitly left open its transfer to realistic prediction systems. This problem asks for the faithful, event-resolved information-gain predictability assessment of ENSO complexity in intermediate-complexity and coupled general circulation models and operational forecast systems, and for the determination of how practical constraints alter the conceptual-model predictability ceilings: the validity and limits of the Gaussian approximation under small ensemble sizes, the net effect of coupled atmosphere–ocean data assimilation and realistic initialization uncertainty, the quantification and mitigation of model-error contributions to the computed information gain, information-theoretic multi-model forecast combination and model selection, and the conditions under which these ingredients suffice for operational use. All five threads remain open in the published literature, with only partial conceptual-model-level advances on the model-error and calibration thread.

Background

The El Niño–Southern Oscillation (ENSO) is the dominant year-to-year coupled atmosphere–ocean mode of variability in the tropical Pacific, in which sea surface temperature (SST) anomalies in the central and eastern Pacific (El Niño when warm, La Niña when cold) interact with the tropical trade winds through the Bjerknes feedback. "ENSO complexity" refers to the diversity of ENSO events: eastern-Pacific (EP) versus central-Pacific (CP) El Niño spatial patterns, peak intensities ranging from moderate to extreme, and temporal evolutions such as multi-year La Niña events. Conventional forecast-skill measures (anomaly correlation and root-mean-square error of ensemble-mean time series) summarize skill over long records and cannot assign a predictability limit to an individual event; they also lack uncertainty quantification in a chaotic system.

Information theory offers an event-resolved alternative. Given an ensemble forecast started from a specific observed initial condition, the forecast probability density function (PDF) p(t,u)p(t,u) at lead time tt is compared with the observational climatological PDF pclimp^{clim} through the relative entropy (Kullback–Leibler divergence) P(p(t),pclim)=∫p(t,u)log⁡(p(t,u)/pclim(u))duP(p(t),p^{clim})=\int p(t,u)\log\left(p(t,u)/p^{clim}(u)\right)du, called the information gain: the additional information carried by the forecast beyond the already-known climatology. A rescaled quantity E=1−exp⁡(−P)∈[0,1)\mathcal{E}=1-\exp(-P)\in[0,1) is used for presentation. When both PDFs are approximated as Gaussians, PP has a closed form decomposing into a "signal" part (information in the mean, weighted by the inverse climatological variance) and a "dispersion" part (information in the covariance ratio); the accuracy of this Gaussian approximation is central to practical applicability. The framework requires a statistically accurate model: the model must satisfy model fidelity (reproduce the observed climatological PDF) and model memory (reproduce the observed autocorrelation function, ACF), otherwise the computed information gain is dominated by model error rather than intrinsic predictability — for example, a fidelity violation makes the apparent predictability limit infinite.

Fang and Chen (2023) instantiated this framework with a multiscale stochastic conceptual model for ENSO complexity (Chen, Fang, and Yu 2022; Chen and Fang 2023), a low-dimensional coupled stochastic system with conditional Gaussian structure coupling interannual variables (eastern- and central-Pacific SST, western-Pacific thermocline depth, zonal advection), intraseasonal westerly wind bursts, and decadal Walker-circulation modulation, which reproduces observed non-Gaussian PDFs, spectra, and event diversity. Its low cost allows ensembles of thousands to tens of thousands of members (operational practice is 10^3–10^4 members at most, and typically far fewer for coupled general circulation models, CGCMs). Key findings: different ENSO event types have very distinct predictability limits; ensemble spread (the dispersion part) contributes substantially to predictability; the onset of EP El Niño is hard to predict whereas many CP El Niño events are strongly predictable about two years in advance; La Niña events associated with effective discharge are strongly predictable; ignoring wind bursts causes large information loss especially for EP events; and, within this conceptual model, the Gaussian approximation is accurate for computing the information gain, which the authors argue should facilitate applying the framework to more sophisticated models such as intermediate-complexity models (ICMs) and CGCMs, for which only smaller ensemble sizes are affordable.

The paper's conclusion states the unresolved practical issues that define this problem: (i) the event-resolved information-theoretic assessment has not been extended to ICMs and CGCMs, because these models cannot afford the very large ensembles needed to fit unbiased non-Gaussian forecast PDFs — the reasonability and limits of the Gaussian approximation across more complex model classes must be quantified; (ii) the effect of data assimilation (the estimation of the initial model state from sparse observations, an essential step toward practical forecasting) and realistic initialization uncertainty on the information-gain predictability is unknown, since coupled atmosphere–ocean assimilation could either shrink or broaden event-specific E\mathcal{E} lifetimes; (iii) model-error contributions to the computed information gain require systematic quantification and mitigation, e.g. via calibrated stochastic parameterizations or statistical closures whose calibration minimizes PDF and ACF mismatches; (iv) multi-model forecast combination and model-selection methods founded on information-theoretic criteria remain to be developed and tested for event-specific ENSO forecasting; and (v) whether the Gaussian approximation, calibrated stochastic closures, and reduced-ensemble protocols suffice to transfer the conceptual-model findings into operational forecast systems (ICMs/CGCMs) is explicitly identified as future work.

Problem Statement

The common objective is to obtain a faithful, event-resolved, information-gain-based potential-predictability assessment of ENSO complexity in realistic coupled prediction systems — intermediate-complexity models, coupled general circulation models, and operational forecast systems — and thereby to determine how the conceptual-model predictability ceilings (distinct per-event limits, with substantial dispersion contributions) are altered by the practical constraints of those systems. The specific, jointly linked questions stated by Fang and Chen (2023) are: (1) Extending the event-resolved information-theoretic assessment to ICMs and CGCMs, where the very large ensemble sizes (order 10^3–10^4 members) needed to fit unbiased non-Gaussian forecast PDFs are unaffordable — which requires quantifying the reasonability and the limits of the Gaussian approximation for computing the information gain as a function of model class, ensemble size, event type, and lead time, so that reduced-ensemble protocols can be justified. (2) Determining the net effect of data assimilation and realistic initialization uncertainty on the information-gain predictability: coupled atmosphere–ocean assimilation alters initial-condition uncertainty and could either shrink or broaden the event-specific E\mathcal{E} lifetimes, and the direction and magnitude of this effect is unknown. (3) Systematically quantifying and mitigating model-error contributions to the computed information gain — including determining calibration methods for stochastic parameterizations and statistical closures that minimize PDF and autocorrelation-function mismatches in the models used. (4) Developing and testing multi-model forecast combination and model-selection methods founded on information-theoretic criteria for event-specific ENSO forecasting. (5) Assessing whether the Gaussian approximation, calibrated stochastic closures, and reduced-ensemble protocols together suffice to transfer the conceptual-model findings into operational forecast systems. These threads jointly support one objective — a validated event-resolved predictability ceiling for ENSO complexity in realistic systems and its exploitation for forecast improvement. A full resolution of the problem requires addressing all five threads jointly; a result on any single thread is a partial contribution to the objective rather than a complete resolution. Each of threads (1)-(4) is an independently refutable sub-claim, addressed only by the concrete deliverable its own statement specifies — for (1) the information-gain error of the Gaussian approximation resolved over model class, ensemble size, event type, and lead time; for (2) the sign and magnitude of the change in the event-specific information-gain lifetimes from data assimilation and initialization uncertainty; for (3) a calibration procedure with a measured reduction of climatological-PDF and autocorrelation-function mismatch; and for (4) a multi-model combination and selection method assessed on event-specific forecasts — while thread (5) is not a further independent deliverable but the synthesis of (1) and (3).

References

  1. Fang, X., and N. Chen, 2023: Quantifying the Predictability of ENSO Complexity Using a Statistically Accurate Multiscale Stochastic Model and Information Theory. Journal of Climate, 36, 2681–2702. doi:10.1175/JCLI-D-22-0151.1; arXiv:2203.02657. https://doi.org/10.1175/JCLI-D-22-0151.1
  2. Chen, N., X. Fang, and J.-Y. Yu, 2022: A multiscale model for El Niño complexity. npj Climate and Atmospheric Science, 5, 16. doi:10.1038/s41612-022-00241-x. https://doi.org/10.1038/s41612-022-00241-x
  3. Chen, N., and X. Fang, 2023: A Simple Multiscale Intermediate Coupled Stochastic Model for El Niño Diversity and Complexity. Journal of Advances in Modeling Earth Systems, 15, e2022MS003469. doi:10.1029/2022MS003469. https://doi.org/10.1029/2022MS003469
  4. Andreou, M., and N. Chen, 2024: Statistical Response of ENSO Complexity to Initial Condition and Model Parameter Perturbations. Journal of Climate, 37, 5629–5651. doi:10.1175/JCLI-D-24-0017.1. https://doi.org/10.1175/JCLI-D-24-0017.1
  5. Zhang, Y., N. Chen, and C. Moser, 2025: A Simple Intermediate Coupled MJO-ENSO Model: Multiscale Interactions and ENSO Complexity. Journal of Advances in Modeling Earth Systems, 17, e2025MS005374. doi:10.1029/2025MS005374; arXiv:2507.14395. https://doi.org/10.1029/2025MS005374
  6. Zhang, Y., N. Chen, J. Vialard, and X. Fang, 2024: A Physics-Informed Auto-Learning Framework for Developing Stochastic Conceptual Models for ENSO Diversity. arXiv:2402.04585. https://arxiv.org/abs/2402.04585
  7. Wang, J., X. Fang, N. Chen, B. Qin, M. Mu, and C. Ji, 2025: A dual-core model for ENSO diversity: unifying model hierarchies for realistic simulations. npj Climate and Atmospheric Science, 8, 269. doi:10.1038/s41612-025-01164-z. https://doi.org/10.1038/s41612-025-01164-z
  8. Chen, C., N. Chen, Y. Zhang, and J.-L. Wu, 2025: CGKN: A deep learning framework for modeling complex dynamical systems and efficient data assimilation. Journal of Computational Physics, 532, 113950. doi:10.1016/j.jcp.2025.113950; arXiv:2410.20072. https://doi.org/10.1016/j.jcp.2025.113950
  9. Behnoudfar, P., C. Moser, M. Bocquet, S. Cheng, and N. Chen, 2026: Bridging idealized and operational models: an explainable AI framework for Earth system emulators. npj Climate and Atmospheric Science, 9, 65. doi:10.1038/s41612-026-01334-7. https://doi.org/10.1038/s41612-026-01334-7
  10. Qin, B., Z. Yang, M. Mu, Y. Wei, Y. Cui, X. Fang, G. Dai, and S. Yuan, 2024: The first kind of predictability problem of El Niño predictions in a multivariate coupled data-driven model. Quarterly Journal of the Royal Meteorological Society, 150, 5452–5471. doi:10.1002/qj.4882. https://doi.org/10.1002/qj.4882
  11. Zhang, J., S. Hu, W. Duan, J. Peng, and M. Hou, 2024: Consistent Initial Error Modes Causing the Largest Prediction Errors and the Strongest Predictability Barrier for Two Types of El Niño Events in CMIP6. Journal of Geophysical Research: Oceans, 129, e2024JC021633. doi:10.1029/2024JC021633. https://doi.org/10.1029/2024JC021633
  12. Ji, C., M. Mu, X. Fang, and L. Tao, 2023: Improving the Forecasting of El Niño Amplitude Based on an Ensemble Forecast Strategy for Westerly Wind Bursts. Journal of Climate, 36, 8675–8694. doi:10.1175/JCLI-D-23-0233.1. https://doi.org/10.1175/JCLI-D-23-0233.1
  13. Ji, C., M. Mu, B. Qin, T. Lian, S. Yuan, J. Feng, X. Song, Y. Wei, G. Dai, J. Wang, and X. Fang, 2025: Toward skillful forecasting of super El Niño events using a diffusion-based westerly wind burst parameterization. npj Climate and Atmospheric Science, 8, 273. doi:10.1038/s41612-025-01158-x. https://doi.org/10.1038/s41612-025-01158-x
  14. Tang, W., X. Fang, and J. Wang, 2025: Optimal precursors of El Niño: spring zonal wind stress as a non-negligible indicator of event types. Environmental Research Letters, 20, 094027. doi:10.1088/1748-9326/adf906. https://doi.org/10.1088/1748-9326/adf906
  15. Vialard, J., F.-F. Jin, M. J. McPhaden, A. Fedorov, W. Cai, S.-I. An, D. Dommenget, X. Fang, M. F. Stuecker, C. Wang, A. Wittenberg, S. Zhao, F. Liu, S.-K. Kim, Y. Planton, T. Geng, M. Lengaigne, A. Capotondi, N. Chen, L. Geng, S. Hu, T. Izumo, J.-S. Kug, J.-J. Luo, S. McGregor, B. Pagli, P. Priya, S. Stevenson, and S. Thual, 2025: The El Niño Southern Oscillation (ENSO) Recharge Oscillator Conceptual Model: Achievements and Future Prospects. Reviews of Geophysics, 63, e2024RG000843. doi:10.1029/2024RG000843. https://doi.org/10.1029/2024RG000843