In Earth sciences, neural networks help find patterns in short and noisy observational records, but their opacity undermines trust in their conclusions. Echo-state networks (ESNs) are a rare type of recurrent network that trains in seconds and hardly overfits, yet they have remained a black box. This work sits alongside studies that apply machine learning to the large-scale structure of climate and ocean, but it differs by opening the internals of an echo-state network for the first time using an explainable AI tool.
The authors took monthly sea surface temperature anomaly maps from 1880 to 2021 and turned each map into a sequence of 180 columns, feeding them into the echo-state reservoir as a time series. A reservoir of 300 randomly connected neurons processed the data, and only the output layer was trained, a total of 301 parameters instead of 87,993 for a conventional fully connected network. They then applied layer-wise relevance propagation, tracing backward from the prediction through all 180 steps to see which parts of the map contributed most to the answer.
The model achieved 100% accuracy on training data and 99% on validation data, recognizing El Niño and La Niña from temperature anomalies. Relevance maps showed that the network itself focused on the Niño 3.4 region in the tropical Pacific, exactly where the phenomenon originates. An experiment with shuffled map columns confirmed that the network does not rely on spatial structure but extracts the needed features from any data order. When the leaking rate was increased from 0.01 to 0.4, accuracy dropped to 58%, and relevance maps showed that the network forgot early columns and looked only at the end of the map.
This work shows for the first time that echo-state networks are not only fast but also explainable: their decisions can be decomposed across input data as clearly as those of deep networks. Unlike neighboring studies that applied explainability to multilayer perceptrons or convolutional networks, here the black box of reservoir computing is opened. This removes the main barrier to climate and ocean scientists trusting such models.
The method can be extended to other Earth science tasks: predicting climate indices, finding teleconnections between distant regions, and analyzing observational time series. Because echo-state networks train quickly and require little data, they can be applied to rare events where deep networks overfit. Further steps include adapting other explainable AI tools, such as sensitivity maps and inverse optimization.
The results will influence climatology, oceanography, and meteorology, where fast and transparent models are needed for short data records. The method will also be useful in geophysics and remote sensing of the Earth.
Apply the proposed approach to unsolved problems, such as detecting atmospheric teleconnections or forecasting rare climate anomalies. Test the method on other geospatial datasets and compare it with alternative explainability methods.
For climatology, the key problem is trust in machine learning models with limited data. This work directly addresses it by showing that even compact reservoir networks can be fully interpretable, opening the way to their use in operational forecasting and scientific analysis.
An echo-state network with 300 reservoir neurons required only 301 trainable parameters, whereas a conventional neural network for the same task needed almost 88,000. Yet the accuracy in recognizing El Niño was the same.