Forecasting El Niño–Southern Oscillation (ENSO) is critical for agriculture, energy, and disaster preparedness worldwide. In recent decades, El Niño Modoki events have become more frequent, where the temperature anomaly is centered in the central Pacific rather than off the coast of South America. The physical mechanisms driving these events are not fully understood, and dynamical models often fail at lead times beyond six months. This work sits alongside studies that use machine learning to improve climate index forecasts, but it stands out by deliberately separating the contributions of the western and eastern ocean regions rather than just chasing accuracy.
The authors built a simple convolutional neural network (CNN) with two convolutional layers, two pooling layers, and one fully connected layer. Input was a set of sea surface temperature anomaly maps from the previous 12 months, and the output was a forecast of the NINO3.4 index for the next 12 months. Three experiments were run: CTRL (entire equatorial Pacific from 120°E to 90°W), EAST (only the eastern part from 160°W to 90°W), and WEST (only the western part from 120°E to 160°W). For each experiment, an ensemble of 5000 models was created with random initialization of convolutional filter weights to assess the spread of skill. Models were trained on data from 1950–2001 and tested on 2002–2021. The primary dataset was sea surface skin temperature from NCEP reanalysis, with additional checks on ERSST and COBE SST.
The 5000-model ensemble showed that forecast skill (correlation between forecast and observations) for the WEST experiment was nearly as high as for CTRL, while EAST was notably worse. At lead times of 2–5 months, WEST even achieved higher correlation than CTRL. For example, at 6 months, correlation for CTRL and WEST was around 0.5–0.6, while EAST dropped below 0.5 by 5 months. Root-mean-square error also grew faster for EAST. The skill distribution across the ensemble was Gaussian, with spread increasing at longer lead times. Validation on independent ERSST and COBE datasets confirmed WEST's advantage over EAST. The model struggled with extreme peaks, such as the 2015–2016 event, and accumulated a phase shift over time.
This work shows that the spatial structure of sea surface temperature in the western Pacific carries more information about the future state of the central Pacific than the eastern part does. This aligns with the delayed oscillator theory: Rossby waves reflected off the western boundary return as Kelvin waves and alter temperatures in the center. Unlike related studies that use complex architectures and many variables for maximum accuracy, this deliberately simple model reveals the physical source of predictability. The result underscores the asymmetry of ENSO dynamics and the importance of including the western ocean in forecasting systems.
A natural next step is to incorporate subsurface ocean temperature, especially thermocline depth, which is a key ENSO predictor. One could also explore how the relative roles of west and east change for different El Niño types (eastern, central, Modoki) and across phases of decadal variability. It would be interesting to test whether the western advantage persists with deeper architectures such as U-Net or transformers, and at longer lead times.
These findings are useful for climate forecast centers developing statistical or hybrid ENSO models. They also provide guidance for physicists studying ocean–atmosphere interactions in the tropics.
Test how adding subsurface data changes the relative roles of the western and eastern Pacific. Run similar experiments for forecasting El Niño Modoki indices and other regional indices.
An open question in climate science is what causes the diversity of El Niño events and why some develop near South America while others form in the central ocean. This work links spatial temperature gradients to predictability but does not fully reveal the physical mechanism. The spring predictability barrier, where model skill drops sharply in certain seasons, also remains unresolved.
The authors created 5000 versions of the same neural network, changing only the initial random filter weights. The spread in forecast accuracy among these versions was comparable to the difference between different architectures, echoing the sensitivity of chaotic systems to initial conditions.