A Complex Systems Perspective on Machine Learning for Earth Science

Explore a complex systems approach to applying machine learning in Earth science through this 41-minute conference presentation. Discover how the traditional reductionist paradigm of constructionism breaks down when studying complex systems due to intractable nonlinear interactions, making computational methods and numerical approximations essential. Learn how modern computational models have become so sophisticated that their outputs match the complexity of natural systems they simulate, and understand how machine learning provides new nonlinear tools to parse this complexity using data from model outputs and observational measurements. Examine the breakdown of constructionism in understanding extreme weather events and see how unsupervised discovery methods can bridge this gap. Address the "black box" nature of data-driven models and understand how optimal predictive models implicitly learn underlying physics, with particular focus on probability distributions over possible futures conditioned on past observations. Investigate the role of these predictive distributions in stochastic data-driven models, including ensemble weather forecasting applications. Delve into how data-driven causal discovery techniques can help untangle the complex web of nonlinear interactions in Earth systems, with practical applications to weather and climate modeling throughout the discussion.