Making the Black Box More Transparent: Understanding the Physical Implications of Machine Learning (Core Science Keynote)

This paper introduces and synthesizes multiple methods for machine learning (ML) model interpretation focusing on meteorological applications.
ML has recently exploded in popularity in many domains. In meteorology, applications include nowcasting (0-2 hours) and short-term forecasting (24-48 hours) of severe weather, forecasting of renewable-energy production and rainfall, hydrometeor classification, and object identification, among others. Although ML has been successful in meteorology, it has yet to be widely accepted primarily due to the perception that ML models are ``black boxes'', meaning the ML methods take inputs and provide outputs but do not yield physically interpretable information to the user. This paper bridges the gap by analyzing and comparing several model-interpretation, applied to meteorological data. Specifically, we discuss permutation-based variable importance, forward and backward selection, saliency maps, class-activation maps, and activation maximization, among others. The last three methods are made possible by deep learning, a subfield of ML that has recently solved many difficult problems. Interpretation methods can find important variables and spatiotemporal patterns for the model in general, for a particular subset of cases, or for a single case. They can also be used to generate synthetic data, such as prototypical tornadic storms, hailstorms, freezing-rain soundings, etc. We apply these methods at multiple spatiotemporal scales to tornado, hail, precipitation-type, and storm-mode prediction. By analyzing on such a wide variety of applications, we intend for this work to demystify the black box of ML, offer insight in applying model-interpretation techniques, and serve as a model-interpretation toolbox for meteorologists and other physical scientists.