[Physics-grads-open] Physics & Astronomy Student Seminar (PASS), October 14 (Fri) at 4 pm: Virtual

[Ming-Feng Ho]

Hi Physics and Astronomy Grads!

I am happy to announce we will resume the *P*hysics & *A*stronomy *S*tudent
*S*eminar (PASS) *this Friday (Oct 14)*.

PASS is a student-run and -oriented seminar, providing graduate students a
safe place to practice their job talks and qualify exams. PASS is also a
safe place to practice asking questions and being asked questions.

We have been running PASS since 2022 Spring, and have already helped
several students on practicing their job talks. The past program could be
found here
<https://jibancat.github.io/ucr-student-seminars/program/index.html>.

This quarter, I am happy to be the first presenter:


*Title: MF-Box: Multi-fidelity emulation for simulations from different box
sizes using a graphical Gaussian process*


*Ming-Feng Ho (University of California, Riverside)*

*Abstract: *Bayesian inference is necessary for any scientific discipline,
especially for explaining cosmological and astrophysical observations.
However, Bayesian inference also introduces the computational burden of
>10^6 likelihood evaluations on the theoretical model. Emulation is the
current only way to perform Bayesian inference using accurate numerical
simulations. By approximating the input/output relationship of numerical
simulations using a machine learning model, emulators work as cheap
surrogates, allowing fast likelihood evaluations suitable for inference
using Markov Chain Monte Carlo (MCMC). Nevertheless, as the experiments
become highly precise in the future, the training data for an accurate
emulator will also become impossible to generate, making the emulator
approach impractical. In this talk, I will present MF-Box, a multi-fidelity
emulation framework that allows us to combine cosmological simulations from
different resolutions and different box sizes. Multi-fidelity emulation
augments expensive high-resolution simulations with many cheap
low-resolution simulations, minimizing the computational budget. MF-Box
extends our previous multi-fidelity emulator, allowing the machine learning
model to use simulations from different box sizes to accelerate the
training via transfer learning. I will also present a theoretical approach
to analyzing the emulator's performance and determining the optimal budget
allocation.


*Friday, October 14, 2022 at 4 pmVirtual via Zoom:*
https://ucr.zoom.us/j/94967485304?pwd=WFczZnpVcWpkR3ZVZVBETmdTNFI0UT09



PS, We need more speakers for Fall! *If you want to recommend or volunteer,
write the speaker's name in this sheet
<https://docs.google.com/spreadsheets/d/13Y7qZKoB8Qzy44ZukZWKcU_nF9U6MnLAIOj51bEcVP4/edit?usp=sharing>.*
We will contact the speakers to confirm the date.

Hope to see you there!
Ming-Feng Ho
on behave of PASS organizing team

---
Ming-Feng Ho
PhD Student
Physics and Astronomy, UCR
NASA FINESST FI
jibancat.github.io
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.ucr.edu/pipermail/physics-grads-open/attachments/20221012/065fd923/attachment.html>