Center for AstroPhysical Surveys

View Full Calendar

CAPS Seminar: Photometric Identification of Tidal Disruption Events using Neural Processes

Event Type
Seminar/Symposium
Sponsor
Center for AstroPhysical Surveys
Location
NCSA - 1205 W Clark St, Urbana - RM 1040
Virtual
wifi event
Date
Mar 1, 2024   12:00 - 12:30 pm  
Speaker
Nicholas Earl, 2023 - 2024 Graduate Student Fellow
Contact
Cynthia Trendafilova
E-Mail
ctrendaf@illinois.edu
Views
16

Speaker: Nicholas Earl, 2023 - 2024 Graduate Student Fellow 
Date/Time: March 1, 2024 / 12 noon central.
Location: NCSA, 1040.
Zoom: https://illinois.zoom.us/j/82318062756?pwd=M3g1MFF6cytsOWFEbmU0UW1XWVoxQT09
Title/Abstract: 

Photometric Identification of Tidal Disruption Events using Neural Processes

Tidal disruption events (TDEs) are unique astrophysical phenomena that occur when the orbit of stars brings them close enough to super-massive black holes (SMBHs) for their self-gravity to be overpowered by tidal forces. This disruption and subsequent accretion onto SMBHs produces a bright flare of radiation, providing a means to probe quiescent black holes lying at the center of most galaxies which would otherwise remain observationally inaccessible. Current neural network classification systems for TDEs are limited by the lack of a unifying physical model for the expected behavior of TDEs, are unable to adapt to new observations once a network has been trained, and struggle in the regime of small data where good uncertainty estimation is needed. Alternatives like Gaussian processes take a probabilistic approach by using stochastic processes and are often preferred for their ability to describe the uncertainty of the function domain. However, Gaussian processes scale poorly with the amount of data and/or dimensionality of the problem, and are sensitive to the chosen kernel. A new approach has emerged in recent years to address both these shortcomings: Neural Processes (NPs). NPs are stochastic processes parameterized by neural networks. Similar to GPs, NPs learn distributions over functions, but differ by implicitly learning the kernel function through observed data. NPs use a neural network to parameterize and learn a mapping from the observed data to posterior predictive distributions. Given the lack of conclusive physical models from which to draw explicit kernel functions, and the often sparse data accompanying TDEs, NPs are a prime framework for establishing a predictive model for quickly identifying potential TDEs in survey observations. I will discuss the status of an on-going project to apply NPs to TDEs and develop an adaptive solution to identifying TDEs for use in current and upcoming astronomical surveys.

link for robots only