Ryan Lagerquist-Feb 15

Machine learning (ML) has become a widely used tool in meteorological research and often makes better predictions than thecompeting methods. Despite this, many meteorologists are reluctant to adopt ML in day-to-day operations, due to the per- ception thatthe models are “black boxes” (cannot explain their decisions to the user). My recent work in ML interpretation attempts to bridge thisgap. I apply several interpretation methods to a convolutional neural network (CNN) trained for storm-based, 0–1-hour prediction oftornadogenesis. CNNs are a method from deep learning, a subset of ML that specializes in learning from spatiotemporal grids. Thepredictors consist of 3-D storm-centered radar grids and a proximity sounding. To understand what the CNN has learned from thesedata, I use interpretation meth- ods such as class-activation maps, to identify important spatial regions of the storm; saliency maps, toidentify spatial and multivariate patterns that increase or decrease tornado probability; backwards optimization, to create synthetic stormsthat min- imize or maximize tornado probability; novelty detection, to quantify the novelty of each storm and highlight which parts are themost novel; and the permutation importance test, to rank the important of the different radar and sounding variables. Some of thesemethods are also applied to CNNs for front detection and mesocyclone prediction in numerically simulated storms, to emphasize thatthey can be applied to a wide variety of problems. I will include links to Python notebooks that allow other practitioners to apply thesemethods to their own data.