Seminar by Balu Nadiga from Los Alamos National Lab, USA
02 September 2024
KST 14:00
The Seminar is being held in Room 1010 (Jasmin) – Integrated mechanical engineering building. Click here for the campus map.
We focus on the use of probabilistic and ensemble machine learning (ML) methods to facilitate the modeling of Earth system processes. The techniques we consider range from generative ML models such as conditional versions of generative adversarial networks (GAN) and variational autoencoders (VAE) to transformers and Bayesian neural networks and ensemble versions of reservoir computing, Given the nonlinear and multiscale nature of various Earth system processes, in one application, we focus on inferring a stochastic closure to represent the effect of unobserved (unresolved) small scale processes on observed (resolved) large scale processes conditioned on the latter. In a second application, we focus on the large scales of the climate system: Predictability typically arises from deterministic dynamics, dynamical symmetries and their associated invariants, and various linear, nonlinear and emergent oscillations/patterns (e.g., the North Atlantic Oscillation, the El Nino Southern Oscillation, the Atlantic Meridional Overturning Circulation, etc.). However, the inevitable problem of model bias prevents comprehensive first-principles based climate models from being able to model such variability skillfully. As such, we focus on data-driven predictive, spatiotemporal modeling of such natural variability using probabilistic and generative ML techniques.