Layerwise relevance propagation (LRP) is a method for interpreting neural networks that projects the decision pathways of a network back onto the original input dimensions. This enables the analysis of how networks are arriving at their decisions, and more importantly for climate applications, a critical assessment of whether the network is learning physically meaningful patterns. We first show that neural networks interpreted with LRP can extract the dominant modes of variability that characterize common climate patterns such as ENSO and the global markers of anthropogenic climate change. We then discuss how LRP can be used to draw novel scientific conclusions for cases where answers are not known a-priori, such as in the case of multi-year predictability of the climate system. This discussion has broad implications for the climate science community, particularly because it enables the usage of neural networks for physical interpretation of the climate system.
Finally, while we focus on climate applications here, LRP is relevant for interpreting neural networks in all fields of atmospheric science.