Upcoming Event: NSF-Simons CosmicAI Hybrid Seminar Series

Emulators for Asteroseismology & Physics Based Deep Learning for Stellar Structure Equations

Speaker 1: Earl Bellinger Speaker 2: Manuel Ballester, Speaker 1: Assistant Professor, Yale University Speaker 2: Wyant College of Optical Sciences, University of Arizona

Abstract

Speaker 1: Asteroseismology---the study of stellar oscillations---has revolutionized the field of stellar astrophysics by enabling ultra precise inferences of internal stellar structure. This in turn helps other fields of astrophysics, such as exoplanets and galactic archaeology, by delivering ages and distances for stars throughout the Galaxy. However, accurately interpreting the observations requires detailed stellar evolution and pulsation models, which are too expensive to calculate on a star-by-star basis. I will present our work in developing neural network emulators trained on numerical simulations that enable rapid theoretical characterization of pulsating stars. I will also discuss our efforts toward building Bayesian hierarchical models in order to study the uncertain physics underpinning the simulations.

Speaker 2: Large-scale stellar population studies require accurate stellar models across potentially billions of stars. Traditional stellar structure calculations rely on numerical methods such as finite differences to solve a coupled system of stiff, nonlinear differential equations. While highly accurate, these approaches can become computationally expensive when repeatedly evaluating large stellar populations or evolving complex stellar systems.   We present a self-supervised physics-informed neural network (PINN) for solving the equations of stellar structure in hydrostatic and thermal equilibrium. Rather than learning from precomputed stellar models, the network is trained directly from the governing physics. The four stellar structure equations are incorporated into the loss function through their differential residuals, while the stellar boundary conditions are imposed as hard constraints through the neural-network architecture. The resulting model learns continuous profiles of pressure, radius, luminosity, and temperature as functions of enclosed mass.   Applying PINNs to stellar structure introduces several challenges, including the stiffness and nonlinearity of the governing equations, spectral bias, and the selection of collocation points. We show how recent advances in scientific machine learning can address these difficulties and enable accurate neural solutions of the stellar structure equations. This work demonstrates the potential of physics-informed neural networks as mesh-free stellar structure solvers and provides a foundation for developing accelerated methods for stellar evolution and, ultimately, large-scale single and binary stellar population modeling.

Biography

Speaker 1:  Earl Bellinger is an Assistant Professor in the Department of Astronomy and the Institute for the Foundations for Data Science at Yale University. He did his PhD in Computer Science and Astrophysics at the Max Planck Institute, followed by postdoctoral fellowships in Denmark, Australia, and Germany. He leads the Yale Astro Machine Learning Group with support from the DOE, NSF, and NASA. He is a core developer of the MESA stellar evolution code and the mission pipeline for the forthcoming PLATO satellite. His work focuses on the theory and observation of all kinds of pulsating stars, and using these stars to study physics at large scales and early times.

Speaker 2: Manuel Ballester is a postdoctoral researcher at the University of Arizona’s Wyant College of Optical Sciences. He received his Ph.D. in Computer Science from Northwestern University, where his research focused on physics-informed and data-driven methods for computational physics. His research spans scientific machine learning, astrophysics, and computational imaging, with a particular interest in combining physical models with machine learning to accelerate scientific simulations.