Physics Master of Science Thesis Defense by Tyson George
When: Wednesday,
August 30, 2023
2:00 PM
-
4:00 PM
Where: > See description for location
Description: Physics Master of Science Thesis Defense
by Tyson George
Date: Wednesday August 30, 2023
Time: 2:00 pm
Topic: Building Numerical Relativity Surrogate Models with Neural Networks
Zoom Link: https://umassd.zoom.us/j/93334680459?pwd=c1RkdW9yRTdnYk9EcU9uaTVGSFg2QT09
Abstract:
With the detection of gravitational waves and use of numerical relativity (NR), we are able to study the properties of chaotic astrophysical events. However, due to the complexities of solving Einstein's equations, other techniques, such as surrogate modeling, have come into fruition. Surrogate models offer comparable accuracy to that of their NR counter-part, while making waveform evaluations in a fraction of the time. For tests of general relativity, the use of surrogate models is ideal, however these too can fall short when performing parameter estimations, which can have upwards of millions of waveform evaluations, ramping up total run-time. To mitigate this, we employ the use of GPU-accelerated neural networks, which offer a significant speed improvement. Using neural networks, we have built a 1D model trained from the hybrid surrogate NRHybSur3dq8 where we target the dominant $l=m=2$ mode. Building over the mass-ratio range $q\in[1,10]$, we analyze the overall accuracy in waveform generation and compare total run times. Further work towards building a 3D model, and attaining feature-parity with the surrogate is also explored.
ADVISOR(s):
Dr. Scott Field, Department of Mathematics (sfield@umassd.edu, 508.999.8318)
COMMITTEE MEMBERS:
Dr. Collin Capano, CSCVR
Dr. Sarah Caudill, Department of Physics
NOTE: All PHY Graduate Students are ENCOURAGED to attend.
by Tyson George
Date: Wednesday August 30, 2023
Time: 2:00 pm
Topic: Building Numerical Relativity Surrogate Models with Neural Networks
Zoom Link: https://umassd.zoom.us/j/93334680459?pwd=c1RkdW9yRTdnYk9EcU9uaTVGSFg2QT09
Abstract:
With the detection of gravitational waves and use of numerical relativity (NR), we are able to study the properties of chaotic astrophysical events. However, due to the complexities of solving Einstein's equations, other techniques, such as surrogate modeling, have come into fruition. Surrogate models offer comparable accuracy to that of their NR counter-part, while making waveform evaluations in a fraction of the time. For tests of general relativity, the use of surrogate models is ideal, however these too can fall short when performing parameter estimations, which can have upwards of millions of waveform evaluations, ramping up total run-time. To mitigate this, we employ the use of GPU-accelerated neural networks, which offer a significant speed improvement. Using neural networks, we have built a 1D model trained from the hybrid surrogate NRHybSur3dq8 where we target the dominant $l=m=2$ mode. Building over the mass-ratio range $q\in[1,10]$, we analyze the overall accuracy in waveform generation and compare total run times. Further work towards building a 3D model, and attaining feature-parity with the surrogate is also explored.
ADVISOR(s):
Dr. Scott Field, Department of Mathematics (sfield@umassd.edu, 508.999.8318)
COMMITTEE MEMBERS:
Dr. Collin Capano, CSCVR
Dr. Sarah Caudill, Department of Physics
NOTE: All PHY Graduate Students are ENCOURAGED to attend.
Contact: > See Description for contact information
Topical Areas: Faculty, Students, Students, Graduate, Students, Undergraduate, Bioengineering, Civil and Environmental Engineering, College of Engineering, Computer and Information Science, Co-op Program, Electrical and Computer Engineering, Mechanical Engineering, Physics