Organized by: Prof. Lu Lu
Unlocking High Redshift Hydrogen Cosmology with Machine Learning (Differentiable Bayesian Forward Models for Detecting Cosmic Dawn)
Date: Thursday, May 1st
Time: 2:30 pm - 3:30 pm
Place: Chamberlin 5280 (Will end in time to get to the Astronomy/Physics Colloquium at 3:30.)
Speaker: Nick Kern, MIT & University of Michigan
Abstract: The frontier of modern astrophysics and cosmology lies at high redshifts, where new generations of experiments are tapping into uncharted astrophysical information to shed light on longstanding problems with LCDM and the birth of the first stars at Cosmic Dawn. In particular, over the next decade, radio telescopes using the 21 cm hyperfine transition of neutral hydrogen will yield unprecedentedly large-volume and statistically-powerful measurements of the high redshift universe. Although widely recognized as a transformative cosmological probe, these radio telescopes face intense foreground contamination and instrumental systematics that have to date precluded a direct measurement of the 21 cm cosmological signal. However, a new era of ML-accelerated analysis is enabling deeper foreground subtraction, faster parameter inference, and improved systematics mitigation that will be key to finally unlocking the potential of 21 cm science. In this talk, I will discuss how current limits of the 21 cm signal at Cosmic Dawn are already providing novel insights into the formation of the first stars and galaxies at z > 6. I will then discuss work on new differentiable Bayesian forward models that will bridge the field into a new era where 21 cm experiments can fully realize their promise as a cosmological probe. Along the way, I'll share insights into how we are addressing systematics-plagued analyses, (very) high-dimensional inference, and how GPUs make all of this work at scale.
Host: Peter Timbie
Add this event to your calendar