[Physics-grads-open] Dissertation Defense, Ming-Feng Ho, 8/12 (Mon) at 1 PM

[Ming-Feng Ho]

Hi all!

I’d like to advertise my PhD thesis defense! Everyone is welcome to join.
P.S., the location is in MRB, not Physics building.

Date/Time: *8/12 (Monday) at 1 PM*
Place: *4th floor, Conference Room, Multidisciplinary Research Building
<https://maps.app.goo.gl/w4Jt6MDZg3Wce9eS6> (MRB)*

Zoom link for remote join:
https://ucr.zoom.us/j/92505382183?pwd=Fqi6yZXr7jTzVXb3HZLytH1WIqtMao.1

ID: 925 0538 2183
Password: 1216

*Title: Model-Driven Cosmology with Bayesian Machine Learning & Population
Inference*

*Abstract*:
In this talk, I will present two research projects during my PhD. In the
first part, I will introduce the concept of “emulation,” a fast-to-compute
regression model that helps researchers utilize slow-to-compute simulations
in data analysis tasks. Emulators have been widely used in cosmology since
the 2010s. During that time, cosmologists began relying on expensive N-body
simulations to explain the observed distributions of galaxies or gases. A
typical Bayesian data analysis for cosmology requires millions of
simulations; however, only hundreds of N-body simulations are computable.
Emulators act as a “surrogate” model for simulations in data analyses by
using a Bayesian interpolator trained on simulations’ inputs and outputs.
Emulators make cosmology data analysis possible for current observations.
However, the curse of dimensionality restricts the development of emulators
from expanding to higher dimensions, which is crucial for exploring beyond
ΛCDM parameters or astrophysical feedback effects, the science goals for
future Stage-V cosmological surveys. To address this problem, we developed
the “multi-fidelity emulation.” Multi-fidelity emulation uses
cheaply-obtained simulations to fill the high-dimensional parameter space
and only a few expensive-and-slow simulations to correct the resolution
differences. I will explain how effective multi-fidelity emulation is
applied in cosmology and discuss potential future improvements by using it.
Then, I will present two applications of multi-fidelity simulations from
Bird’s group: Goku simulations for beyond ΛCDM and PRIYA simulations for
the Lyα forest.

In the second part of the talk, I will discuss gravitational wave
astronomy. LIGO-Virgo’s first detection of gravitational waves from binary
black holes (BBHs) dates back to 2016. By 2021, the number of detected BBH
events had increased to approximately 90. With this number of events,
astronomers can study the formation origins of these BBHs by fitting
Bayesian hierarchical models and inferring the population statistics. The
recent LIGO-Virgo-KAGRA (LVK) population analysis reveals a Gaussian bump
substructure at ~35 M⊙ in the primary mass spectrum, while the rest of the
black holes are mostly distributed in a power-law distribution from ~5 M⊙
to ~80 M⊙. An interesting question is whether this ~35 M⊙ Gaussian
population of black holes is mixed with the power-law distribution or if
they remain separate due to distinct formation channels. We designed a
Bayesian hierarchical model to measure the co-location and separation of
the power-law and Gaussian populations in the LVK catalog using a mixture
model approach. We found that the posterior suggests that the Gaussian bump
black holes prefer to merge within their cohort, indicating a preference
for these two populations to be separate. Current formation channels for
producing the Gaussian bump might also need to consider the separation of
this population.

Thank you and look forward to seeing you!
Ming-Feng

---
Ming-Feng Ho
PhD Candidate (NASA FINESST FI)
Physics and Astronomy
University of California, Riverside
Bird's group
jibancat.github.io
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.ucr.edu/pipermail/physics-grads-open/attachments/20240806/4596545e/attachment.html>