Estimating the risk of extreme weather due to climate change, particularly at regional scales, is critically important yet remains one of the most challenging problems to address with traditional physics-based global climate models (GCMs). The computational cost of these models leads to high uncertainty in estimates of the rarest — yet most impactful — extreme weather return periods. AI emulators trained on historical reanalysis data have been shown to reproduce global atmospheric dynamics at greatly reduced computational cost and, in some cases, with decades-long stability. Such emulators can also be combined with coarse-resolution physics-based models to simulate extreme weather events that are absent from the physics-based model alone. I will discuss this in the context of an early reservoir computer-based emulator that successfully simulates sudden stratospheric warming events not captured by its component physics-based model.

However, because emulators are trained and validated on the relatively short historical record, it is difficult to assess whether their predicted extreme event statistics are accurate. Using state-of-the-art emulator architectures, we performed a first-of-its-kind assessment of this capability using 92,000 years of stationary climate data. We find that emulators can generate weather events more extreme than those in the training set that are dynamically similar to those in the long simulation data. However, the accuracy with which emulators reproduce the correct return periods of the most extreme events varies by region, variable, and architecture. These biases are difficult to correct with conventional training methodologies, motivating the development of new sampling methods. One such method, AI-boosted rare event sampling, uses emulator forecasts as a score function in an importance sampling algorithm to more efficiently sample extremes from the physics-based model and produce unbiased return period statistics. I will conclude by discussing future prospects for resampling approaches to improve emulated extreme event statistics.

Alex Wikner (UChicago)

Lieu
École normale supérieure – PSL
24 rue Lhomond – aile Erasme
salle Claude Froidevaux – E314