Sea surface temperature (SST) drives global weather and climate, and its anomalies, such as El Niño, cause droughts, floods, and wildfires worldwide. Existing dynamical models of the coupled ocean-atmosphere system provide El Niño forecasts only 3–4 seasons ahead, while regional anomalies like marine heatwaves are even harder to predict. In recent years, machine learning has shown promise in improving forecasts, but most models target a single index, such as Niño3.4, and do not provide a complete spatial picture. This work sits at the intersection of oceanography and deep learning: instead of using climate model outputs, it directly uses reanalysis data to predict the two-dimensional temperature field across the globe.
The authors used ERA5 reanalysis data: monthly mean SST and 2-meter air temperature from January 1950 to May 2021, totaling 857 months. The original 0.25° grid was averaged to 1° and then bilinearly interpolated to a 64×128 grid (latitude from -64° to 62°, full longitude circle). For training, they used the Unet-LSTM architecture, a convolutional encoder-decoder with long short-term memory layers, which had previously succeeded in precipitation forecasting. The model took 12 months of temperature fields as input and predicted the next 2 months; for long-term forecasts, they used an autoregressive approach, feeding predictions back as input. Training ran on 760 time steps (up to April 2013) on four Nvidia V100 GPUs with a batch size of 4 and the Adam optimizer (learning rate 0.003).
The model keeps the root mean square error below 0.75 °C for all 24 forecast months, and in July 2010 the error grew from 0.43 °C at month 1 to 0.65 °C at month 24. Correlation with actual SST fields remains significant up to 12 months ahead, especially in the tropical, northeastern, and southern Pacific, as well as in the high latitudes of the North Atlantic. The model successfully predicted the El Niño events of 2009–2010 and 2015–2016 with lead times up to 24 months, including the strong La Niña of 2010–2011, although the peak of the extreme 2015–2016 El Niño was underestimated. For the Blob index (northeastern Pacific), the model accurately reproduced the warm events of 2014, 2015, 2019, and 2020, as well as the cooling of winter 2016–2017. The forecast of Ningaloo Niño off the western coast of Australia showed limited accuracy, which the authors attribute to a lack of similar events in the training sample.
This work demonstrates that a purely data-driven approach, trained only on surface temperature reanalysis, can capture the physics of large-scale ocean variability on horizons up to two years. Unlike neighboring studies that predict a single index or use climate model outputs, this model reproduces the full two-dimensional field, allowing any index to be extracted post hoc. This is a step toward more flexible and accurate forecasts of regional anomalies, such as marine heatwaves, which are poorly captured by traditional models.
The authors plan to increase the model's resolution to the original 0.25° of ERA5 to study regional SST variations in more detail, and to implement an ensemble approach to quantify forecast uncertainty. Another direction is moving to three-dimensional fields using ConvLSTM3D, which would allow incorporating wind and other variables. A hybrid path is also promising: combining observational data with climate model outputs through transfer learning or physics-informed neural networks to better predict rare events, such as consecutive El Niños.
Accurate long-term SST forecasts would help provide early warnings for droughts, floods, wildfires, and extreme marine heatwaves, which affect fisheries, aquaculture, and coastal ecosystems. The method could also be transferred to forecasting other geophysical fields, such as precipitation or geopotential height.
Immediate next steps include increasing spatial resolution to 0.25°, ensemble forecasts to assess uncertainty, and testing the model on other indices, such as the Indian Ocean Dipole. They also plan to investigate how including additional variables (wind, upper-ocean heat content) improves predictions of rare events.
A key unresolved problem in oceanography is the limited predictability of regional marine heatwaves and rare climate events due to their diversity and the scarcity of historical examples. This work shows that deep learning can partially overcome this barrier, but for rare phenomena like Ningaloo Niño, richer training data or hybrid approaches with physical models are still needed.
The model was able to predict the Blob marine heatwave 24 months before its peak, even though this phenomenon was considered elusive due to its irregular seasonal timing. However, with Ningaloo Niño off the western coast of Australia, the neural network did worse: there were too few similar cases in the training data.
The idea of predicting climate anomalies with machine learning echoes Kim Stanley Robinson's novel 'The Ministry for the Future', where algorithms help humanity respond to extreme heat and climate disasters. Here, the neural network acts as an early warning system that sees the onset of El Niño two years in advance.