Estimating dynamical parameters of two interacting galaxies using Deep Learning

TL;DR

This paper employs deep learning, specifically CNNs, to estimate dynamical parameters of galaxy interactions from simulation and real images, achieving high accuracy and demonstrating potential for astrophysical analysis.

Contribution

The study introduces a CNN-based method to accurately predict continuous dynamical parameters of galaxy mergers, bridging simulation data and real observations.

Findings

Achieved 93.63% classification accuracy for galaxy interaction parameters.

Attained 99.86% R-squared in regression for continuous parameter prediction.

Successfully tested the model on real SDSS galaxy images.

Abstract

The science behind galaxy interaction and mergers has a fundamental role and gives us an insight into galaxy formation and its evolution. Fluctuating angular momentum is responsible for extraordinary events like polar rings, tidal tails, and ripples. To study different phenomena related to galaxy interactions, various parameters like the mass ratio of the interacting galaxy, orbital parameters, mass distribution, morphologies are required. Convolutional Neural Networks (CNN) are widely used to classify image data. Thus, we used CNN as our approach to the problem. In this work, we will be using data from state-of-the-art magneto-hydrodynamic simulations of galaxy mergers from the GalMer database at different dynamical parameters using image snapshots of merging pairs of galaxies and feeding them to our Deep Learning model (ResNet). The dynamical parameters we are aiming for; would be spin, relative inclination ($i$), viewing angle ($\theta$), and azimuthal angle ($\phi$). We aim to download bulk data using the web scraping method. The first approach is to create different combinations of these parameters to form 60 classes. Feeding the data into the model, we achieved 93.63% accuracy. As we received good results in minute classification, we moved to our second approach, regression. Here the model can predict the continuous and exact values of the dynamical parameters. We have achieved a 99.86% R-squared value and the mean squared error of 0.0833 on testing data. In the end, we used data from Sloan Digital Sky Survey to test our trained model on some real images.