Collaborative Research: CAIG: Framework for Artificial Intelligence-Enhanced Modeling of Wildfire Geohazards (FAIM-WG): Applications for postfire debris flows across the Western US

Wildfires are among the most common and widespread natural hazards that impact landscapes and communities throughout the Western United States. Beyond their immediate effects, wildfires cause vegetation loss that can make watersheds susceptible to postfire debris flows and flash floods for several years, posing risks to downstream communities, properties, and vital infrastructure. Existing early-warning tools for postfire debris flows address specific contexts, such as inland and dry coastal regions, but they are not yet well-adapted to cooler, wetter areas like the coastal Pacific Northwest. This project integrates Artificial Intelligence (AI) and process-based earth surface response models into a Framework of AI-enhanced Modeling of Wildfire Geohazards (FAIM-WG). This framework will enable the identification of rainfall thresholds for triggering debris flows, explore new and missing processes to improve predictions, and develop transferable models for postfire debris flows to aid early-warning and risk mitigation efforts. This project will first create a comprehensive AI-ready, multi-modal dataset that includes topographic, meteorological, and environmental variables for all major wildfires across the Western U.S. This dataset will serve as the foundation for developing probabilistic models to predict postfire debris flow initiation using machine learning methods such as gradient-boosted decision trees, taking into account the unique characteristics of different regions in the Western U.S. These data-driven models, along with process-based initiation mechanisms, will be implemented into the Landlab earth surface modeling toolkit. At several study sites where extensive postfire geomorphic responses have been observed, Landlab models will be established to develop baseline model simulations of eco-hydro-geomorphic dynamics driven by stochastic weather and wildfires. These baseline Landlab models will be used in a discrepancy modeling framework where a new model identification method that involves Sparse Identification of Nonlinear Dynamics with SHallow REcurrent Decoder networks (SINDy-SHRED) will be coupled with Landlab to discover new formulations of postfire watershed response and geomorphic transport laws. The AI-ready curated dataset will be distributed through the AI platform HuggingFace, and models will be shared on Landlab’s GitHub codebase. This project will train two graduate students and a postdoctoral researcher. The research will be disseminated through in-class experiential learning and via asynchronous and open-access teaching materials in the form of YouTube videos and Jupyter notebooks executable on the project's JupyterHub. 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.