Using Deep Learning to Enhance Prediction and Understanding of Weather Phenomena

Deep learning has recently been used to solve many challenging problems -- including galaxy classification, beating the world champion in Go, and creating "deep fakes" -- that involve the processing of gridded data.  Motivated by these successes, deep learning has recently gained popularity in the atmospheric and Earth sciences as well.  I will present two novel applications of deep learning from my doctoral work: synoptic-scale front detection and next-hour tornado prediction.  For front detection, a convolutional neural network (CNN) is used to detect warm and cold fronts in 2-D reanalyses of temperature, humidity, and wind velocity.  The CNN appears to outperform the previous state of the art and is currently being used to investigate the climatology and variability of fronts in the United States over the last 40 years.  For tornado prediction, a CNN is used to transform 2-D radar images, 3-D radar images, and proximity soundings into a next-hour tornado probability for each storm.  Several interpretation methods are used to understand what the CNN has learned.  Efforts are currently underway to compare interpretation outputs with human knowledge, allowing us to identify whether or not CNN has learned anything new.  In the future we envision a tighter coupling between data science and physical science, where data science is used to identify new hypotheses for physical scientists to explore.