Location
MC 5501
Speaker
Chris Subich, University of Waterloo
Title
From Navier–Stokes to Neural Networks: Data-Driven Weather Forecasting
Abstract
Numerical weather prediction (NWP) stands as one of the most successful applications of computational fluid dynamics in human history, shaping the development of supercomputing from its infancy. Historically, progress in atmospheric modeling has been driven by a virtuous triplet: advancing computational power, deeper physical understanding of atmospheric processes, and denser global observation networks.
A major byproduct of this classical era is the creation of consistent, high-quality, multi-decadal gridded reanalysis datasets which serve as the "ground truth" for both evaluations and data-driven algorithms. These data-driven algorithms have sparked a profound shift in the meteorological community, as they are increasingly capable of producing highly skillful forecasts with a fraction of the runtime cost required for traditional "physics-based" and PDE-driven prediction systems.
This lecture explores the transition from physics-driven to data-driven weather forecasting. We will examine how generic machine learning architectures such as Graph Neural Networks, Transformers, and Neural Operators view weather as a massive, high-dimensional next-frame prediction problem. We will also address the unique scientific and mathematical challenges at the frontier of AI-NWP: modeling chaotic uncertainty in an ensemble framework, correcting the smoothing behavior imposed by standard loss functions, hybridizing AI with traditional dynamical cores via spectral nudging, and the long-term limits of learning directly from sparse observations.
This event is part of the NSERC CREATE-funded Quantitative Climate Science summer school. More information about this initiative can be found at https://qcs-create2024.github.io/
