High-resolution climate information is essential for local climate risk assessment, infrastructure planning, and decision support, yet conventional statistical and dynamical downscaling workflows remain difficult to scale, reproduce, and adapt across regions, variables, and stakeholder needs. This project will integrate agentic AI into the downscaling workflow using NASA’s Prithvi WxC foundation model as a core modeling component. Prithvi WxC will be fine-tuned to learn relationships between coarse-resolution atmospheric predictors and high-resolution target fields, enabling efficient downscaling of climate and weather variables relevant to hazards such as extreme precipitation, heat, drought, wildfire conditions, and air quality.
The proposed agentic workflow will automate key stages of the downscaling pipeline, including dataset discovery, preprocessing, experiment configuration, model fine-tuning, inference, evaluation, uncertainty characterization, and generation of stakeholder-ready outputs. By coupling Prithvi WxC with tool-using AI agents, the system will be able to select appropriate predictors, launch reproducible experiments, compare model performance against baseline statistical methods, diagnose failure modes, and summarize results in plain language. The workflow will also support iterative human-in-the-loop refinement, allowing scientists and decision-makers to specify domains, variables, metrics, and application-relevant constraints through natural language interfaces.
This integration will reduce technical barriers to applying foundation models for regional climate analytics while improving transparency, reproducibility, and scalability. The resulting capability will provide a pathway toward AI-assisted production of next-generation downscaled climate datasets and decision-support information, strengthening NASA’s ability to translate Earth system observations and model projections into actionable insights for state, local, and regional resilience planning.