Searching for Primordial Black Holes with Machine Learning
Advisor(s)
Dr. James Annis, Fermilab
Dr. Brian Nord, Fermilab
Location
Room B110-1
Start Date
26-4-2019 10:05 AM
End Date
26-4-2019 10:20 AM
Abstract
The idea that Primordial Black Holes (PBHs) constitute the majority of dark matter was revived in 2015 by LIGO’s detection of the merger of two ~30 solar mass black holes. We can search for PBHs via gravitational microlensing, a phenomenon which occurs when a PBH passes in front of a star, forming an Einstein ring and increasing the apparent brightness of the star. When an observed star undergoes a microlensing event, the apparent magnitude data graphed over time form a Paczynski curve. We are participating in a Dark Energy Survey (DES) project to detect microlensing in the DES data. PBHs of 10-100 solar masses have microlensing events of time duration t ~> 2.5 years and can be observed in the DES. We employ both the traditional astrophysics toolkit as well as machine learning (ML) methods including RNNs, LSTMs, and DNNs. By testing different ML structures, we can find the ideal structure for this novel application. The structures are trained and tested with generated Paczynski curves of varying complexity. They will then be tested on supernova data with minimal noise before being used on DES data. If our observed number of microlensing events is far less than the expected number of microlensing events, we can rule out PBHs as dark matter.
Searching for Primordial Black Holes with Machine Learning
Room B110-1
The idea that Primordial Black Holes (PBHs) constitute the majority of dark matter was revived in 2015 by LIGO’s detection of the merger of two ~30 solar mass black holes. We can search for PBHs via gravitational microlensing, a phenomenon which occurs when a PBH passes in front of a star, forming an Einstein ring and increasing the apparent brightness of the star. When an observed star undergoes a microlensing event, the apparent magnitude data graphed over time form a Paczynski curve. We are participating in a Dark Energy Survey (DES) project to detect microlensing in the DES data. PBHs of 10-100 solar masses have microlensing events of time duration t ~> 2.5 years and can be observed in the DES. We employ both the traditional astrophysics toolkit as well as machine learning (ML) methods including RNNs, LSTMs, and DNNs. By testing different ML structures, we can find the ideal structure for this novel application. The structures are trained and tested with generated Paczynski curves of varying complexity. They will then be tested on supernova data with minimal noise before being used on DES data. If our observed number of microlensing events is far less than the expected number of microlensing events, we can rule out PBHs as dark matter.