 |
Center for Nonlinear Studies |
This is the story of how my UCSD graduate student (Azton Wells [1]) hijacked the computational cosmology thesis project I gave him and turned it into an integrated ModSim/AI thesis. Azton wanted to apply what he had learned about machine learning in CS classes to the problem I suggested. The question was: How do the first stars in the universe pollute the universe with heavy elements and how does that affect the properties of the first galaxies forming subsequently? The product of his thesis research is StarNet [2][3][4], a deep learning based surrogate model for primordial star formation and feedback. In this talk I trace how the surrogate model emerged through trial and error, and how the model was trained using data produced by resolved numerical simulations of primordial star formation and feedback in protogalaxies. I discuss the crucial differences between resolved scales, unresolved scales, and inference scales in the context of our model. I then describe how the surrogate model, prototyped in PyTorch, was exported so that it can be used for inference in running Enzo and Enzo-E simulations. I touch on inference mesh data structures incorporated into Enzo-E for this purpose. I then show some results of simulations using StarNet of the formation of the first galaxies in the universe.
Bio: Michael L Norman is Distinguished Professor of Astronomy and Astrophysics at UC San Diego and former director of the San Diego Supercomputer Center. He received his PhD in Engineering and Applied Science from UC Davis while working at the Lawrence Livermore National Laboratory. Subsequently he held research positions at the Max Planck Institute for Astrophysics, Los Alamos National Laboratory, and National Center for Supercomputing Applications, UIUC. His research group, the Laboratory for Computational Astrophysics (est. 1991) develops community application software for astrophysical simulation on supercomputers including ZEUS-2D, ZEUS-3D, ZEUS-MP, Enzo, and Enzo-E. His scientific interests include astrophysical and cosmological fuid dynamics with applications to star formation, interstellar medium, supernova remnants, astrophysical jets, galaxy formation, and X-ray clusters. His technical interests include algorithm development, code development, and parallel computing. In his role as SDSC director, he served as PI for the Gordon, Comet, and Expanse national HPC systems and is currently PI of the CloudBank cloud access project funded by the NSF.
Sponsored by Information Science & Technology Institute (ISTI).
[1] Now an Assistant Computational Scientist at ANL working on AI for Science
[2] Wells, A. I. & Norman, M. L. 2021; Predicting Localized Primordial Star Formation with Deep Convolutional Neural Networks, doi: 10.3847/1538-4365/abfa17
[3] Wells, A. I. & Norman, M. L. 2022a; Connecting Primordial Star-forming Regions and Second-generation Star Formation in the Phoenix Simulations, doi: 10.3847/1538-4357/ac6c87
[4] Wells, A. I. & Norman, M. L. 2022b; The First Galaxies and the Effect of Heterogeneous Enrichment from Primordial Stars, doi:10.48550/arXiv.2210.14805
Join in Microsoft Teams
Meeting ID: 283 676 446 088 9
Passcode: oj7vJ69K
+1 575-323-9652,,436833543# United States, Las Cruces
Host: Patrick Diehl, CCS-7