[CS] TODAY 5/19 Elena Orlova Dissertation Defense

[via cs]

This is an announcement of Elena Orlova's Dissertation Defense.
===============================================
Candidate: Elena Orlova

Date: Monday, May 19, 2025

Time:  3 pm CST


Location: JCL 298

Title: Machine Learning for Complex Physical Systems: Applications to Climate Forecasting and Dynamical Systems Simulation

Abstract: This dissertation investigates machine learning (ML) approaches for modeling complex physical systems across multiple domains. We examine three interconnected research areas: subseasonal climate forecasting, chaotic dynamical systems, and quantum mechanical simulations. Across these diverse domains, we address common challenges in the application of ML to physical systems: effectively encoding information to capture important features of spatiotemporal dynamics, developing robust emulators of physical systems that preserve critical invariant properties such as chaotic attractors, and systematic incorporation of known physical constraints into neural architectures and training schemes to simultaneously enhance prediction accuracy and computational efficiency. These methodological considerations form the conceptual framework for our various applications. For subseasonal climate forecasting, we develop frameworks that use lagged ensemble members and observational data to significantly improve temperature and precipitation forecasts. In chaotic dynamical systems, we propose neural operator training approaches that preserve invariant measures through optimal transport distance minimization and contrastive learning, maintaining statistical fidelity in long-term simulations. For quantum mechanics, we introduce Deep Stochastic Mechanics (DSM), a framework inspired by stochastic mechanics and generative diffusion models that may have far lower computational complexity in higher dimensions compared to traditional numerical methods by exploiting wave function latent structure. Our DSM simulations of bosonic systems outperform conventional approaches in both accuracy and computational efficiency. Furthermore, we extend DSM to effectively model fermionic systems, demonstrating its capabilities through a hydrogen molecule time dynamics simulation. Throughout these applications, we integrate domain-specific physics with advanced learning techniques to enable more accurate, efficient, and physically consistent simulations.


Advisors: Rebecca Willett

Committee: Yuxin Chen, Ian Foster, Rebecca Willett