Abstract: Modern machine learning methods provide exciting new tools for modeling and understanding dynamical systems. Machine learning-based emulators are accelerating simulations of complex physical phenomena and helping create more accurate models by learning directly from data. These emulators enable faster predictions and uncertainty quantification, as well as new approaches for parameter estimation and other inverse problems. A key aspect of building high-quality emulators—and crucial for scientific understanding—is identifying the right variables, coordinates, or parameters that efficiently describe the system and its dynamics. This task is especially challenging when the complexity of the system vastly outweighs our prior knowledge.

In this talk, we will discuss representation learning methods that have the potential to help us automatically discover hidden structure in our data and extract relevant and interpretable variables. The right representation of a dynamical system provides both a solid foundation for building emulators that generalize well and a deeper understanding of the underlying physics.

Bio: Peter Y. Lu is an Eric and Wendy Schmidt AI in Science Postdoctoral Fellow at the University of Chicago working at the intersection of physics and machine learning. He received a Ph.D. in Physics from MIT in 2022, where he was an NDSEG Fellow, and an A.B. in Physics and Mathematics from Harvard in 2016. His research interests include physics-informed machine learning and interpretable representational learning with applications in nonlinear dynamics, condensed matter physics, photonics, fluid dynamics, biophysics, and other areas. Peter aims to develop new computational methods for modeling and understanding physical systems with an emphasis on incorporating physics-informed priors and identifying relevant and interpretable latent representations.