CAIG: Advancing AI atmospheric emulators through rigorous theoretical evaluation and uncertainty quantification

Accurately predicting how regional weather extremes like heat waves and droughts will evolve in the coming decades is essential for national preparedness and resilience. Traditional physics-based Earth system models (ESMs) exhibit discrepancies when compared to available observations, especially in regional trends, limiting their usefulness. Artificial intelligence (AI) models trained directly on observational data offer a promising alternative, having already shown an ability to predict regional extremes more skillfully than traditional models in some cases. However, before these AI tools can be trusted to forecast future conditions — particularly conditions never seen in the historical record — scientists need rigorous ways to evaluate their reliability and quantify their uncertainty. This project brings together geoscientists, statisticians, and applied mathematicians to rigorously test and improve AI-based atmospheric models, making them more trustworthy for predicting future regional conditions. The resulting tools and open-source software will benefit researchers, policymakers, and communities seeking to understand and prepare for evolving regional weather hazards, while also training the next generation of scientists in this interdisciplinary field. This project will advance AI-based atmospheric emulators through three interconnected research thrusts: (1) rigorous theoretical evaluation of how well AI models capture extremes relative to physics-based models and observations; (2) development of robust methods, including conformal prediction, for quantifying aleatoric and epistemic uncertainty in multi-decadal AI emulator predictions; and (3) construction of a hierarchy of physically consistent, coupled atmosphere-land AI emulators that incorporate physical coupling constraints. The work will clarify the roles of extrapolation versus translocation in regional discrepancies and produce emulators with quantified uncertainty suitable for out-of-sample and out-of-distribution regional prediction. The interdisciplinary team will train PhD students and undergraduate researchers, contribute to the Data Science Institute's AI+Science summer school, and release code and emulators through open-access repositories. 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.