Gravitational Lensing with Generative Adversarial Networks

Session Number

Project ID: PHYS 17

Advisor(s)

Dr. Brian Nord; Fermilab, University of Chicago

Discipline

Physical Science

Start Date

22-4-2020 10:25 AM

End Date

22-4-2020 10:40 AM

Abstract

We present a new method to simulate gravitational lensing using model-assisted generative adversarial networks (MAGAN) developed by Alonso-Monsalve and Whitehead (2018). The MAGAN is trained to emulate Lenstronomy simulations created by Birrer et. al (2015). The network model is used to save time in generating large datasets of gravitational lensing. MAGANs are neural networks with parameter inputs to target specific features that the network should generate. Our research shows the feasibility of this method and an analysis of the accuracy of our MAGAN after certain training steps, comparing its training and run time to Lenstronomy. The majority of our training time is spent simulating images in order to train the neural network, and lies in the ray tracing step of the Lenstronomy package, placing a bottleneck on accuracy with lower training iterations possible in a given amount of time. The trade-off in training time does result in progressively more accurate images produced by the MAGAN, and at a faster runtime than Lesntronomy.

Share

COinS
 
Apr 22nd, 10:25 AM Apr 22nd, 10:40 AM

Gravitational Lensing with Generative Adversarial Networks

We present a new method to simulate gravitational lensing using model-assisted generative adversarial networks (MAGAN) developed by Alonso-Monsalve and Whitehead (2018). The MAGAN is trained to emulate Lenstronomy simulations created by Birrer et. al (2015). The network model is used to save time in generating large datasets of gravitational lensing. MAGANs are neural networks with parameter inputs to target specific features that the network should generate. Our research shows the feasibility of this method and an analysis of the accuracy of our MAGAN after certain training steps, comparing its training and run time to Lenstronomy. The majority of our training time is spent simulating images in order to train the neural network, and lies in the ray tracing step of the Lenstronomy package, placing a bottleneck on accuracy with lower training iterations possible in a given amount of time. The trade-off in training time does result in progressively more accurate images produced by the MAGAN, and at a faster runtime than Lesntronomy.