Abstract ID: 256
Interpretable Machine Learning for S2D Prediction and Discovery: data-driven approaches to the method of analogs
Lead Author: Elizabeth Barnes
Colorado State University, United States of America
Keywords: explainable machine learning, method of analogs, forecasts of opportunity, subseasonal-to-decadal prediction, sources of predictability
Abstract: Machine learning is increasingly being used to identify sources of subseasonal-to-decadal (S2D) predictability within the climate system. In recent years, our group has worked toward building AI models that mimic scientific human reasoning to improve intrinsic interpretability. Here, we propose a methodology to combine the age-old “method of analogs” with machine learning to identify sources of predictability within the climate system. We do this in an inherently interpretable way, training a neural network to identify the most important regions of the globe for the specific S2D prediction task. Our approach is competitive with a range of baselines, including a fully, “black box” neural network, as well as a naive analog approach. Furthermore, we are able to learn sources of predictability from what the network learns. Finally, taking an analog approach allows for scientists to leverage vast libraries of imperfect climate model output to improve S2S prediction of the real world.
Jamin Rader (Colorado State University)
Randal J Barnes (University of Minnesota – Twin Cities)