NSF Award
2425658
Thunderstorms produce many disruptive, impactful hazards, including heavy rainfall, strong winds, large hail, and tornadoes. Accurate prediction of thunderstorm hazards is limited by our ability to predict exactly where storms will initially form. Current state-of-the-art storm-permitting numerical models perform best in predicting storm evolution only after storms have formed. Newer AI models can learn space and time patterns across different scales but require deeper physical validation to understand what patterns they are learning and how they translate into improved predictions. To address these challenges, this project will develop AI models to predict storm evolution. The project is expected to enable significant breakthroughs in storm initiation understanding through advancing explainable AI techniques and sophisticated uncertainty quantification methods. The project would also generate a convection initiation benchmark dataset and train students to conduct cross-disciplinary collaborative research.<br/> <br/>This project will advance geosciences and AI impact through collaborative development of a prognostic, diagnostic, and generative deep learning nowcasting tool for convection initiation, quantifying the predictability and uncertainty of predictions and working to unify uncertainty quantification paradigms from ensemble NWP and DA with those from statistical and evidential ML. Explainable AI methods adapted to link realistic perturbations in physical processes to changes in predictions can improve physical understanding of CI through exploration of specific scientific hypotheses. New uncertainty quantification approaches can generate well-tuned ensembles of ML predictions, providing insights to the practical predictability of CI. Synergy with data assimilation can also be attained by using AI for the ensemble forecast model and background error covariance, to anticipate and initiate convection in models, and to aid process studies. In particular, this research will: (1) design and optimize a state-of-the-art combined prognostic, diagnostic, and generative deep learning convection nowcasting system using artificial intelligence techniques, high resolution satellite and radar data, and surface weather fields; (2) assess the predictability and quantify the uncertainty associated with the AI system predictions, and compare to traditional ensemble NWP; and (3) discover the physical and dynamical mechanisms that control the predictability of convection through physically-informed explainable AI. The project is expected to enable significant breakthroughs in research questions surrounding convection initiation through advancing explainable AI techniques and sophisticated uncertainty quantification methods to overcome challenges associated with convection initiation predictability and explainability. The project will enable insights of the fundamental dynamical and physical processes behind convection initiation from a data-driven perspective, and how these physical processes affect predictive skill.<br/><br/>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.
