CAIG: Investigating Tornadogenesis via Explainable Deep Learning

Each year, tornadoes cause significant loss of life, injuries, and economic damages. In 2023 alone, two separate U.S. tornado outbreaks caused >$1 billion in damage. Improving our understanding of the atmospheric processes leading to tornado formation is thus of vital importance.

In recent years, artificial intelligence and machine learning (AI/ML) have advanced rapidly. This project will develop ML models to investigate how and why tornadoes form. An improved understanding of the processes by which tornadoes form will enable weather forecasters to produce higher quality tornado watches and warnings, which ultimately will improve the nation’s ability to protect life and property during tornadic storm events.

Advancements in AI/ML applications for very-high resolution weather models in this project will inform the development of future AI/ML systems for use by weather forecasters. The methods developed in this project for science discovery using AI/ML models could, in the future, be applied to other high-impact weather hazards, such as hurricanes, floods, or winter storms.

This project seeks to advance our understanding of the dynamic and thermodynamic processes within the atmosphere which contribute to tornadogenesis (tornado formation). Deep learning ML models will be trained using a library of 200 tornado-resolving numerical simulations of supercell thunderstorms to investigate modes of tornadogenesis and precursors of tornado formation.

Specifically, using interpretable ML methods, the project team will investigate what meteorological features are most related to tornadogenesis, how skillful ML models are in predicting tornadogenesis, and how tornadogenesis modes identified by ML compare to those identified by human researchers. A key goal of the project is enabling fundamental science discovery using ML systems; to this end, a ML encoder capable of identifying coherent meteorological features interpretable by domain scientists (meteorologists) will be developed for use with the ML models in this project.

Tornado-resolving numerical simulations will be performed using the CM1 cloud model, with configurations selected to ensure that dynamic and thermodynamic processes important to tornadogenesis are represented accurately. Training of ML models using mechanistic explainable AI will simultaneously optimize for predictive skill and quality/utility of the generated explanations for understanding tornadogenesis.

The validity of currently-accepted theories regarding modes of tornadogenesis, the initial sources of vorticity in an incipient tornado, and the methods by which this initial vorticity is intensified into a tornado will be investigated.

This award by the Division of Research, Innovation, Synergies, and Education within the Directorate for Geosciences is jointly supported by the Division of Information and Intelligent Systems within the Directorate for Computer and Information Science and Engineering.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.