This work sits alongside research on the large-scale structure of climate variability, including attempts to predict El Niño 1.5 years ahead, maps of atomic nuclei as an analogy for complex systems, and machine learning models for gravitational maps of galaxies. In this context, the paper stands out because it does not build a physical model of the ocean and atmosphere but instead extracts equations of motion directly from the observed time series. This approach echoes the idea of 'excavating reality' in neural networks, but here, instead of training a network, the search is for an invariant functional relationship in phase space.
They took the MEI.ext index (extended multivariate ENSO index) from January 1871 to December 2005, a total of 1620 monthly values. They filtered out high-frequency noise with periods shorter than a year, leaving the 'slow manifold', a signal retaining 95% of the energy. Then they applied time-delay embedding into a five-dimensional phase space: each state point is defined by five values of the series shifted by τ. They found that the system's trajectory lies remarkably close to a plane in a reduced three-dimensional space, corresponding to the functional relationship $y_{i+2τ}$ = A + B($y_{i+τ}$ − $y_{i−τ}$) + C($y_{i+τ}$ − $y_{i−τ}$) $y_i$. This relationship is equivalent to a third-order nonlinear differential equation. For forecasting, the model was trained on the first 30% of the data (486 points) and tested on the remaining 70%.
The nonlinear model yields an anomaly correlation (ACC) above 0.9 for lead times up to 7 months, and useful forecasts (ACC > 0.6) persist up to 10.5 months. For comparison, the linear model reaches 0.9 only up to 4.5 months, and multimodel ensembles of current climate models have ACC around 0.86 at 6 months. At a 6-month forecast, the nonlinear model shows ACC = 0.96. Strong events (deviation exceeding one standard deviation) are predicted even better. The spring predictability barrier (SPB), a sharp drop in forecast skill when starting from spring months, is greatly weakened in the nonlinear model: the maximum improvement over the linear model occurs precisely in the barrier zone. The model was tested on three ENSO indices: MEI.ext, MEI, and Niño3.4, and in all cases the nonlinear version outperforms the linear one.
The work shows that the observed nonlinearity of ENSO can be described by a simple deterministic equation extracted from a single time series. This differs from neighboring studies that either use physically complete models with high computational costs or linear stochastic models that do not reproduce the asymmetry between El Niño and La Niña. The proposed approach is the first test that an invariant nonlinear relationship can be found directly in ENSO data without a priori assumptions about mechanisms.
The methodology can be transferred to other geophysical time series: the Madden-Julian oscillation, intraseasonal oscillations, and sunspots. The authors have already obtained preliminary good results for the sunspot index. The main unresolved challenge is real-time adaptation: the filtering needed to isolate the slow manifold introduces errors at the end of the series, which spoils forecasts. They are searching for ways to extract the slow component without filtering, similar to what is done for the MJO index.
ENSO forecasting affects agriculture, water resource management, and energy on a global scale. A cheap and fast nonlinear model could become an independent complement to physical models in multimodel ensembles.
Develop a method to extract the slow manifold without filtering to eliminate edge errors. Test the model on other indices and extend it to real-time forecasts.
In climatology, the nature of the ENSO spring predictability barrier remains an open question. This work links it to insufficient accounting for deterministic nonlinearity, which may direct efforts toward improving the representation of nonlinear processes in models.
The model is trained on just 486 points (40.5 years of data) but forecasts 94.5 years ahead. Moreover, the computational cost of forecasting is practically zero, unlike the supercomputer calculations of physical models.