Long-term prediction of ENSO with physics-guided Deep Echo State Networks

TL;DR

This study tackles the challenge of long-lead ENSO predictability by integrating physically motivated climate modes from the Extended Recharge Oscillator into a Deep Echo State Network (DESN). The approach yields skillful Niño3.4 forecasts up to $16$–$20$ months with $ACC$ benchmarks comparable to or better than state-of-the-art deep learning and dynamical models, while maintaining CPU-scale training efficiency and interpretability. Mechanistic analyses reveal that long-lead predictability is driven by nonlinear, seasonally modulated cross-basin couplings, especially WWV interactions with inter-basin modes such as NPMM, and that higher-order nonlinearities beyond the XRO framework are important. Error-growth analyses indicate an intrinsic predictability horizon near $30$ months, with DESN approaching this barrier due to its physics-guided structure and multiscale dynamics. Overall, physics-guided reservoir computing provides a tractable, interpretable framework for diagnosing and extending ENSO predictability and offers a blueprint for applying similar approaches to other coupled climate phenomena.

Abstract

The El Niño-Southern Oscillation (ENSO) is a dominant mode of interannual climate variability, yet the mechanisms limiting its long-lead predictability remain unclear. Here we develop a physics-guided Deep Echo State Network (DESN) that operates on physically interpretable climate modes selected from the extended recharge oscillator (XRO) framework. DESN achieves skillful Niño3.4 predictions up to 16-20 months ahead with minimal computational cost. Mechanistic experiments show that extended predictability arises from nonlinear coupling between warm water volume and inter-basin climate modes. Error-growth analysis further indicates a finite ENSO predictability horizon of approximately 30 months. These results demonstrate that physics-guided reservoir computing provides an efficient and interpretable framework for diagnosing and predicting ENSO at long lead times.

Figures (4)

• Figure 1: Regions of interest and architecture of the DESN model for ENSO prediction.a, Standard deviation of sea surface temperature anomalies (SSTA) from the ORAS5 reanalysis (1979–2023). Colored boxes denote regions used to compute area-averaged climate indices for ENSO and associated modes, following the XRO framework zhao2024. b, Schematic of the Deep Echo State Network (DESN) architecture. The input layer comprises multiple climate indices and seasonal cycles as periodic bootstrap sequences ($\mathbf{PBS_t}$). Inputs are passed through a hierarchy of $n_l$ recurrent reservoirs, each generating a neuron state vector $\mathbf{r}_t^l$. Additive white noise $\xi_{\text{rc}}^l$ is injected into each reservoir to emulate stochastic variability. Internal weights consist of input matrices $\mathbf{W}_{\text{in}}^l$ and reservoir matrices $\mathbf{W}_{\text{res}}^l$. Reservoir outputs are concatenated and mapped to next-month targets via a linear readout matrix $\mathbf{W}_{\text{out}}$. Multi-step forecasts are obtained recursively using rolling inputs. When $n_l = 1$, the model reduces to a standard ESN. The reservoir transition function at the bottom illustrates neuron dynamics with activation function $g(\cdot)$ and leakage rate $\alpha$. See Methods for definitions and implementation details.
• Figure 2: Forecast performance of ESN and DESN models for ENSO prediction.a, Forecast correlation skill (ACC) of the 3-month running mean Niño3.4 index as a function of lead time, evaluated out-of-sample for 2002–2023. Shown are DESN (red), ESN (magenta), minimal-ESN (dark blue), XRO (black; trained on 1958–1999), 3D-Geoformer (light blue), and the IRI operational ensemble forecasts—ensemble means of dynamical models (dark purple) and statistical models (dark cyan). b, Same as panel a, but evaluated in-sample over 1979-2023, including the ensemble mean of dynamical models from the North American Multi-Model Ensemble (NMME). Individual NMME model forecasts (1981–2021) are shown in various colors. Forecast periods for the 3D-Geoformer and NMME correspond to their respective training spans. c-e, Target-month-dependent ACC of Niño3.4 forecasts for the minimal-ESN (c), ESN (d), and DESN (e). Colors denote correlation skill as a function of target month (vertical axis) and lead time (horizontal axis). Black contours delineate regions where ACC exceeds 0.5, highlighting combinations of target months and lead times with high predictive skill.
• Figure 3: Role of WWV and cross-basin modes in long-lead ENSO predictability.a,b, Mode-decoupling experiments for XRO (a) and DESN (b). Each curve shows the anomaly correlation coefficient (ACC) of Niño 3.4 forecasts as a function of lead time when one climate mode is removed from the full predictor set while all remaining inputs are retained. The black curve denotes the full model, and colored curves correspond to decoupling individual modes (WWV, NPMM, SPMM, IOB, IOD, SIOD, TNA, ATL3, and SASD). Removing WWV leads to the strongest degradation of long-lead skill in both models, while removing other modes generally reduces skill without eliminating extended predictability. The horizontal dashed line marks ACC = 0.5. c,d, Mode-addition experiments for XRO (c) and DESN (d). Forecast skill is shown for a baseline configuration consisting of Niño 3.4 and WWV (black dashed curve), with one additional climate mode added at a time (colored curves). Extended forecast skill emerges only when WWV is coupled with additional inter-basin modes, with the North Pacific Meridional Mode (NPMM) yielding particularly robust improvements. The dashed horizontal line again indicates ACC = 0.5. e,f, Increasing input-dimensionality experiments for WWV-selected sub-models in XRO (e) and DESN (f). Curves show the average forecast skill across ensembles of sub-models with progressively increasing numbers of climate modes coupled to Niño 3.4 and WWV (from 2 to 9 dimensions). Forecast skill increases with input dimensionality, indicating the cumulative contribution of cross-mode interactions.
• Figure 4: Error-growth analysis and nonlinear predictability limits of ENSO models.a, Evolution of absolute forecast error between perturbed and unperturbed trajectories for the XRO (black), ESN (green), and DESN (red) models, shown as $\ln(\delta_i)$ as a function of lead time. Dashed vertical lines mark the predictability limits for XRO (22 months), ESN (18 months), and DESN (34 months), defined as the time when absolute error reaches 95% of the XRO saturation value. b, Absolute error growth in DESN for a range of initial perturbation magnitudes $\delta_0$. While smaller initial perturbations delay the approach to saturation, the long-term saturation level and the inferred predictability limit remain largely unchanged once $\delta_0 < 0.1$