Constraining Weather Extremes Uncertainty Using AI Climate Models

Heat extremes, heavy precipitation, drought, and wind patterns have changed markedly at regional scales due to human influence, creating cascading socioeconomic risks. Predicting future changes in these events remains a major challenge for physics-based climate models. Recently, AI emulators trained on observation-derived reanalysis data have emerged and shown great promise, demonstrating skill in multi-decadal prediction of extremes and, in some cases, outperforming physics-based models. If these emulators are going to be used for any kind of future projections—critical inputs for climate economics and public health—it is essential to understand their strengths and limitations. This project brings together geoscientists, statisticians, and applied mathematicians to deliver trustworthy multi-decadal predictions of extremes through two objectives: (1) rigorous theoretical evaluation of extreme-event behavior in AI models, and (2) robust methods for uncertainty quantification. The outcome will be more trustworthy emulators with quantified uncertainties, enabling improved regional predictions, including for out-of-distribution extremes lacking observational validation.

“We’re addressing a critical barrier to the responsible use of AI in climate science by rigorously testing whether AI climate models accurately capture extreme weather and by quantifying the uncertainty in their predictions. By establishing a stronger scientific foundation for trustworthy AI-based climate projections, we aim to enable more reliable assessments of future climate risks that can inform adaptation, public health, and policy.”

Tiffany Shaw, Professor, Geophysical Sciences