Tackling the Accuracy–Interpretability Trade-Off in a Hierarchy of Machine Learning Models for the Prediction of Extreme Heatwaves

Abstract When performing predictions that use machine learning (ML), we are mainly interested in performance and interpretability. This generates a natural trade-off, where complex models generally have higher skills but are harder to explain and thus trust. Interpretability is particularly important in the climate community, especially when dealing with extreme weather events, where gaining a physical understanding of the underlying phenomena is crucial to contain impacts. In this paper, we perform probabilistic forecasts of extreme heatwaves over France, using a hierarchy of increasingly complex ML models, which allows us to find the best compromise between accuracy and interpretability. We use models that range from a global Gaussian approximation (GA) to deep convolutional neural networks (CNNs), with the intermediate steps of a simple intrinsically interpretable neural network (IINN) and a model using the scattering network (ScatNet). Our findings reveal that CNNs provide higher accuracy, but their black-box nature severely limits interpretability, even when using state-of-the-art explainable artificial intelligence (XAI) tools. In contrast, the ScatNet achieves similar performance to CNNs while providing greater transparency. Our interpretable models highlight the known drivers of extreme European heatwaves, like persistent anticyclonic anomalies and dry soil, as well as new ones, in the form of subsynoptic geopotential height oscillations. This study underscores the potential of interpretability in ML models for climate science, demonstrating that simpler models can rival the performance of their more complex counterparts, all the while being much easier to understand. This gained interpretability is crucial for building trust in model predictions and uncovering new scientific insights. Significance Statement The purpose of this work is to test increasingly complex machine learning models on the task of forecasting extreme heatwaves and explain their prediction. This is important to quantify what additional information the more complex models are able to capture. We find that answering this question with a black-box model and explainability techniques is not sufficient. Indeed, the explanations are mainly qualitative and do not add much to what was known from the simplest linear models. On the other hand, by using an inherently interpretable architecture design, we are able to identify subsynoptic oscillations in the geopotential height field as the source of additional information, while maintaining the same predictive skill of the black-box model.