Machine learning for earth system science and engineering : critical challenges

Abstract

Predictive characterization of enviro-climate risks at different time scales and the ability to understand their implications on engineering design and operations of lifeline systems downstream is a societal grand challenge. Recent works point to the potential of machine learning (ML), and data-driven (DD) approaches in addressing some of these challenges - hitherto tricky to solve using traditional techniques. At the same time, the suitability of ML/DD methods for problems in real-world domains such as earth and environmental science (EES) is constrained by fundamental limitations. In this thesis, we explore three such challenges of ML methods in the context of EES: 1) big data vs. small data, i.e., limited data (including labels) for domains or regions of interest and the need to go beyond supervised learning for scalability and generalization. 2) inability of standard ML techniques - including deep learning - to quantify process uncertainty in modeling complex nonlinear systems such as the earth's climate. 3) characterizing the information flow and interaction between stressed systems (e.g., urban transportation networks (UTN)) and various hazards (both short-and-long-term) at urban and policy-relevant local scales. We address each of the challenges discussed above through representative research problems.--Author's abstract