Collaborative Research: CAIG: Data-Centric Geoscience Foundation Model for Extreme Events

This project will transform how floods and droughts are predicted by treating each event as a coherent, trackable object rather than isolated data points. Current models struggle with extreme events because they miss how storms and dry spells evolve over time. The team will develop an artificial intelligence framework that learns these natural patterns, enabling earlier and more accurate warnings for dangerous conditions. The resulting tools will help emergency managers and communities better prepare for extreme-weather related disasters. All software and data will be publicly released, and the project will train students at the intersection of computing and Earth science while producing educational materials for university classrooms. The project will construct a Geoscience Foundation Model using an event-manifold learning framework, encoding each extreme event as a latent trajectory with spatial footprint and contextual drivers. Lightweight adapter layers enable specialization to extremes without retraining large models. Physics-regularized decoding enforces conservation laws to prevent unphysical predictions, while generative modeling provides uncertainty estimates. The framework will be demonstrated on flash-flood and drought applications and benchmarked against operational numerical models. 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.