Characterization of Black Holes from EHT Images using Deep Learning

Keerthi, K. (2022) Characterization of Black Holes from EHT Images using Deep Learning. Masters thesis, Indian Institute of Science Education and Research Kolkata.

Text (MS dissertation of Keerthi K. (17MS164)) 17MS164_Thesis_file.pdf - Submitted Version Restricted to Repository staff only Download (32MB)

Abstract

According to the no-hair theorem, a black hole can be fully described by 3 externally observable classical parameters: mass, spin and electric charge. Electric charge of a Black hole is negligible. Neural networks and deep learning models can be used to extract parametres of physical relevance and to establish relationship between different components of the data which may otherwise seem impossible. The Event Horizon Telescope (EHT) collaboration has given us 7070 black hole images labelled with its corresponding spin, magnetisation parameter, inclination of the black hole shadow and the frequency of observation, generated using General Relativistic Magneto Hydrodynamic (GRMHD) simulations. These simulations are computationally expensive and therefore sparse. To increase the grid in the data, we need to make use of deep learning models. In this project, Generative Adverserial Network (GAN), a deep learning model is developed to generate fake instances of the data. This is done in three steps: first approach is to develop a GAN that can learn relevant features of the input data distribution which can be used to generate synthetic black hole images similar to the ones EHT has generated using GRMHD simulations. The second step is to condition this GAN to generate images conditioned on the input parameter, i.e., black hole spin. The final step is to validate the results using a Convolutional Neural Network(CNN) classifier. This is done by training and testing the data on real and synthetic images on all the four combinations. The goal of this project is to use this labelled synthetic images along with the real ones as the training set for a CNN model in which 2 frequencies will be used as 2 different channels of the CNN model to compare their performances. This will allow EHT to assess the gain in precision on incorporating a new frequency to the new generation EHT (ngEHT) design, which they are currently building, thereby helping the team to take a billion-dollar worth decision for the ngEHT design. Training the CNN classifier with real data and testing it with synthetic data (TRTS), training with synthetic data testing with real data (TSTR), training with synthetic data testing with synthetic data (TSTS) gave high precision and R2 values, comparable with training on real data and testing on real data (TRTR). This indicates that the synthetic data has successfully incorporated almost all the features of the real data distribution and thus can be used in real world problems that requires larger grid in the training data.

Actions (login required)

View Item