NORMAN, OKLA. – For years, scientists have worked diligently to understand tornadoes to better forecast them. The National Science Foundation has funded a team of OU scientists to take a cutting-edge approach to understanding the life cycle of tornadoes. Over three years, the team, led by Nathan Snook, Ph.D., will use deep learning techniques to better understand how tornadoes form.
The team will create a library of approximately 200 numerical simulations to train a machine-learning model. The simulations use computer models to create a set of grid points containing three-dimensional information on the atmosphere, allowing meteorologists to predict how storms within this simulated cube of atmosphere will evolve over time, all to answer why and how tornadoes form.
According to Snook, a significant benefit of using a machine learning approach in tandem with these simulations will be training the algorithm to predict where a tornado will develop based on not one or two fields but all the information a model offers.
“Many different features have been implicated in the literature as being important to tornadoes,” said Snook, citing variables such as temperature, moisture and wind direction and speed, among many others. “A machine learning model can take all of that information and look at it impartially, and hopefully confirm or refute existing understanding of how tornadoes form and decay.”
Snook plans to take two separate approaches to the machine-learning process. In one approach, researchers will give the machine-learning model information about what scientists believe are the most important features in tornado development. In the second approach, the team will allow the model to learn on its own what features it believes are important, and then human scientists will interpret what the model has learned.
Snook says it is possible that the model will latch on to a new, previously unidentified interaction between variables and features that could help scientists to better understand the ways tornadoes form.
“The machine learning model learns things in a way that is very different from the way a human would, and it may learn things that human scientists would have a blind spot for,” said Snook.
Snook is the director of research and a senior research scientist with the Center for Analysis and Prediction of Storms, or CAPS, at the University of Oklahoma. In addition to Snook, the research team consists of Ming Xue, Ph.D., Amy McGovern, Ph.D. and Andrew Fagg, Ph.D. of OU, and Corey Potvin, Ph.D., a research scientist with the National Oceanic and Atmospheric Administration’s National Several Storms Laboratory.