Applying an Inherently Interpretable Neural Network to Subseasonal-to-Seasonal Climate Prediction

In order to explain the decision-making process of black-box models, new techniques from fields such as computer science have been applied to geoscientific applications. One such technique is to build neural network architectures that are inherently interpretable. Here, we apply such a neural network architecture, named ProtoLNet, in a subseasonal-to-seasonal climate prediction setting. ProtoLNet identifies predictive patterns in the training data that can be used as prototypes to classify the input, while also accounting for the absolute location of the prototype in the input field. In our application, we use data from the Community Earth System Model version 2 (CESM2) pre-industrial control simulation, and we train ProtoLNet to identify prototypes in precipitation anomalies over the north Pacific to forecast 2-meter temperature across the western coast of North America on subseasonal-to-seasonal timescales. These identified CESM2 prototypes are then projected onto fifth-generation ECMWF Reanalysis (ERA5) data to predict temperature anomalies in the observations. We compare the performance of the ProtoLNet with that of other prediction setups, such as convolutional neural networks, and show that the predictions by ProtoLNet have high skill while also being interpretable, sensible, and useful for drawing conclusions about what the model has learned.