Machine Learning Approaches for Data-Driven Analysis and Forecasting of High-Dimensional Chaotic Dynamical Systems

Abstract

We consider problems in the forecasting of large, complex, spatiotemporal chaotic systems and the possibility that machine learning might be a useful tool for significant improvement of such forecasts. Focusing on weather forecasting as perhaps the most important example of such systems, we note that physics-based weather models have substantial error due to various factors including imperfect modeling of subgrid-scale dynamics and incomplete knowledge of physical processes. In this thesis, we ask if machine learning can potentially correct for such knowledge deficits.

First, we demonstrate the effectiveness of using machine learning for model- free prediction of spatiotemporally chaotic systems of arbitrarily large spatial extent and attractor dimension purely from observations of the system’s past evolution. We present a parallel scheme with an example implementation based on the reservoir computing paradigm and demonstrate the scalability of our scheme using the Kuramoto-Sivashinsky equation as an example of a spatiotemporally chaotic system. We then demonstrate the use of machine learning for inferring fundamental properties of dynamical systems, namely the Lyapunov exponents, purely from observed data. We obtain results of unprecedented fidelity with our novel technique, making it possible to find the Lyapunov exponents of large systems where previously known techniques have failed.

Next, we propose a general method that combines a physics-informed knowledge-based model and a machine learning technique to build a hybrid forecasting scheme. We further extend our hybrid forecasting approach to the difficult case where only partial measurements of the state of the dynamical system are available. For this purpose, we propose a novel technique that combines machine learning with a data assimilation method called an Ensemble Transform Kalman Filter (ETKF).