Plenary Day3 - Navigating Uncertainty in the AI Era of Weather Forecasting - Dr. Christopher Subich

Speaker: Dr. Christopher Subich (Meteorological Research Division of Environment and Climate Change Canada (ECCC)) Abstract: The emergence of artificial intelligence in numerical weather prediction has fundamentally disrupted atmospheric science, promising rapid and even unprecedented gains in both computational efficiency and forecast accuracy. However, this revolution also presents challenges at nearly every stage of national weather centres' forecast production chains. Speaking from the perspective of medium-range global atmospheric forecasting, this plenary explores how the meteorological community and Environment and Climate Change Canada (ECCC) are navigating these uncharted waters across three distinct fronts: scientific, operational, and institutional. Scientifically, we face the challenge of predicting a chaotic atmosphere using unexplainable models. Early AI systems learned to minimize error by "hedging their bets" and smoothing out extreme events. Researchers are now actively working to characterize and mitigate this structural uncertainty. We will explore how native ensemble generation, adjusted loss functions that target realism, and physically informed neural architectures like PARADIS force AI models to generate realistic forecasts that capture dangerous extremes. Operationally, life-safety organizations cannot simply deploy "black boxes." We examine ECCC's pragmatic approach to safe AI adoption, utilizing spectral nudging to pilot traditional physics-based systems with AI guidance while retaining the essential, high-resolution, and physically consistent outputs that clients and the public depend upon. Finally, we address the institutional and computational uncertainty looming over the field. AI inverts the traditional compute paradigm, shifting the "heavy lifting" from steady, CPU-based operational runs to massive, "bursty," GPU-heavy model training. As tech giants leverage immense capital to hyperscale their models, we will discuss the "resolution ratchet," the hard limits of historical training data, and the future of sovereign modelling in the AI era.