Interpretable machine learning approaches for identifying and understanding predictable multi-year climate variability
Emily Gordon (Stanford University)
Colloquium
Monday, February 5, 2024, 3:30 pm
Abstract
Identifying predictable climate variability beyond a few weeks is notoriously difficult due to the unpredictable noise in the Earth system. In this talk, I discuss methods for identifying predictable climate variability on interannual to decadal timescales using machine learning, specifically, neural networks. Starting with a simple application for examining predictable Pacific decadal variability, I demonstrate how explainable AI (XAI) can be used to increase trust in neural network predictions, and investigate the mechanisms governing Pacific decadal variability. I formalize this approach by examining predictable sea surface temperatures (SSTs) across the ocean, in model simulations of both pre-industrial and future climate, and how climate change influences the predictability of near-term climate variability. I further demonstrate how thoughtful and creative experimental design, coupled with the power of neural networks to predict non-linear behavior, provides insights into sources of predictable internal variability and how this may manifest under climate change. Finally, I discuss how AI-driven SST predictions can provide constrained estimates of future climate variability of surface temperatures and precipitation, and how these data-driven tools may shape future investigations of climate variability and predictability.