Item

Abstract

Typical earth system processes are nonlinear and multiscale in nature. Furthermore, when there is no clear separation of scales, simple closures are insufficiently accurate in capturing the complexity of interactions and feedbacks across scales. In this setting the inability to resolve the full range of scales involved in modeling such processes due to limited computational resources is likely to bring in a stochastic component to the input-output relationship characterizing the behavior of such processes. As such, it is important to avoid a false sense of confidence that arises from a modeling perspective that ignores the stochastic nature of such processes. This is also the case when machine learning is used for data-driven modeling of such processes. We discuss the application of a variety of probabilistic machine learning techniques ranging from reservoir computing to generative models (conditional generative adversarial networks and variational autoencoders) to Bayesian neural networks to model two example systems, one in which the data comes from a popular Earth System Model (Community Earth System Model ver. 2; CESM2) and another which uses reanalysis data.