Abstract ID: 178
Identifying State-Dependent Subseasonal Predictability Bias with Explainable Neural Networks
Lead Author: Kirsten Mayer
National Center for Atmospheric Research, United States of America
Keywords: state-dependent model bias, subseasonal, neural network, explainable artificial intelligence (XAI)
Abstract: Subseasonal timescales are known for their limited predictability. However, this timescale is important for actionable decision-making in many public and private sectors. To improve subseasonal prediction skill, one area of research has explored modes of variability shown to enhance predictability when present, often referred to as “forecasts of opportunity” or state-dependent predictability. Previous work has demonstrated that explainable neural networks can identify these states of enhanced subseasonal predictability in both models and observations. However, Earth system models are known to have biases that can affect the representation of modes of variability and their subsequent impacts, which can hinder the ability to make accurate forecasts. Here we demonstrate a neural network approach to identify biases in Earth System models. In particular, we use explainable neural networks together with transfer learning to examine state-dependent subseasonal predictability biases in a large ensemble of Community Earth System Model version 2 simulations.
Co-authors:
Katherine Dagon (National Center for Atmospheric Research)
Maria J. Molina (University of Maryland, National Center for Atmospheric Research)
