Quantifying Uncertainties for Extreme Weather Events with Prithvi WxC Foundation Model

Abstract

Deep Learning Weather Models have recently emerged as AI emulators for traditional numerical weather prediction (NWP) systems. Though still in the early stages and rapidly advancing, these AI forecast emulators have shown promising performance and accuracy to capture the interest of the meteorological and climate science community. Notably, this progress has occurred alongside the revolution in foundation models. In this talk, we will discuss our progress in modelling weather extremes using the Prithvi WxC foundation model. Through the use of model ensembles, we will illustrate the training and fine-tuning strategies required to capture statistical outliers, as well as the use of physics priors in reducing the amount of data needed for training, a key requirement as data on extreme weather events is inherently limited. Since our model is trained on reanalysis data, we do not expect the introduction of aleatoric uncertainty, which is typically irreducible and originates from the noise in the input data. We focus instead on epistemic uncertainty, and characterize the contrast between physical and AI-based modelling of weather phenomena. We emphasize the necessity of a UQ approach informed by physics for the reliable use of foundation model methods in forecasting extreme weather events. Additionally, we highlight the need for a unified framework to address challenges such as limited, noisy, and heteroscedastic data, as well as the requirement for extrapolation with out-of-distribution data.