ENSO forecasting is important for agriculture, energy, and insurance. Dynamical models provide reliable forecasts for 6–12 months, but beyond that their skill drops. Deep neural networks have shown improvements, yet concerns remain about their robustness and interpretability. Surrounding this work are studies where neural networks predict El Niño 18 months ahead, detect particle energy by sound, and warn of droughts months in advance. ResoNet follows this line, applying deep learning to climate data, but stands out with its hybrid architecture and explicit analysis of physical mechanisms.
ResoNet was trained on 3240 monthly samples from 20 CMIP6 models, then fine-tuned on 103 years of ERSST.v5 reanalysis. Input is three months of sea surface temperature anomalies (SSTA); output is the three-month average of the Niño3.4 index. The architecture has three stages: convolutional layers in a Cross-Scale Embedding Layer extract local patterns, while ENSO Mobile Blocks with a transformer capture global interactions. To prevent overfitting, bagging was used: 20 models with different initializations and training subsamples, with the prediction being the ensemble average. Explainability was provided by integrated gradients, which computed the contribution of each pixel of the input map to the forecast.
On all three validation sets (ERSST.v5, ERA5, OISST.v2), anomaly correlation remained above 0.5 for lead times of 19–26 months. This outperforms CNN and dynamical models SINTEX-F and NMME, whose skill drops earlier. With ideal hyperparameter tuning on test data, forecasts reached 26 months, but the authors stress that such an estimate is inflated and cannot be used in real conditions. Integrated gradients analysis revealed that for long lead times the network focuses on the southern Pacific and Indian Ocean, while for short lead times it focuses on the equatorial zone. This aligns with the recharge mechanism, the seasonal footprinting mechanism, and the Indian Ocean capacitor effect. For the first time, asymmetry is shown for neural network ENSO models: for El Niño, the southern Pacific is important 18 months ahead, while for La Niña, the eastern equatorial zone is key.
ResoNet differs from neighboring works in that it not only improves accuracy but also uncovers the physical mechanisms underlying the forecast. The hybrid architecture allows simultaneous viewing of local anomalies and long-range connections between oceans, which pure CNNs or transformers lacked. This reduces skepticism about deep learning in climate science: the model is not a black box but a tool that can be checked against known physics. Demonstrating the asymmetry of El Niño and La Niña is a first step toward neural networks helping discover new patterns, not just reproducing old ones.
The topic could evolve toward including additional variables: wind, ocean heat content, precipitation. This would test which signals are truly necessary for long-term forecasting. More advanced explainability methods may emerge, showing not only where the network looks but also how it combines features. Of particular interest is predicting rare events, such as strong El Niños, and transferring the approach to other climate modes, such as the Indian Ocean or the Atlantic.
Improved ENSO forecasts will help agriculture plan planting, energy companies estimate demand, and insurance companies manage risks. More reliable warnings of droughts and floods will reduce damage from extreme weather events.
The next step is to train ResoNet on multiple variables and see if the forecast horizon grows. It is also worth applying the method to other climate indices and comparing with next-generation ensemble dynamical models.
An open question in climate science is the limit of ENSO predictability: how far ahead can we look using only sea surface temperature? This work shows that deep networks can extract signal for 18–26 months, but it does not answer where the physical limit lies. The connection to the unsolved problem is that the asymmetry of phases and the role of extratropical processes are still poorly represented in dynamical models, and neural networks may help understand which mechanisms are missed.
Bagging, used in ResoNet, comes from 1990s statistics: it was proposed by Leo Breiman, the same author of the famous random forests. And the integrated gradients method is so named because it accumulates the contribution of each pixel along a path from a zero image to the actual one.
In Kim Stanley Robinson's novel 'Forty Signs of Rain', climate scientists desperately try to predict abrupt weather shifts. ResoNet is a step toward making such predictions routine rather than a plot of catastrophe.