| Abstract : | Measuring a galaxy's stellar mass is essential for comparing it to simulations of galaxy formation and evolution. The most common method involves dynamical techniques, where we observe the velocity of tracer objects, like globular clusters, to recover the gravitational potential and thus the mass.
My goal is connecting observed and intrinsic galaxy properties using Machine Learning techniques. We seek an explicit expression relating stellar mass to observables: u, g, r, i, z band magnitudes, redshift (z), and color (g-r). The challenge is creating an expression that works across broad range of redshifts. I test these methods on simulated universe generated using the GALFORM code, consisting of 475412 galaxies. Once we identify a model predicting stellar mass effectively based on observables, we will then apply it to real data for performance assessment.
I use ML methods to predict mass values. Initially, a straightforward ML method provides stellar mass predictions as black-box process. We use this as a benchmark and subsequently employ symbolic methods to derive an explicit equation. The galaxies exhibits bi-modality in observer frame color, indicating differences in star formation histories and dust extinction properties for red and blue galaxies.
I apply SISSO, a Machine Learning approach used in Material Sciences, for the first time in astronomy. It automates feature engineering and selects a multi-dimensional expression with high covariance with the target variable, i.e. stellar mass, providing a simple feature-based equation with physical interpretation. I also explore techniques like Principal Component Analysis, Linear, Polynomial and Non-linear regression. We analyze scatter and bias plots between predicted and true stellar mass for these ML methods. So far, Polynomial Regression has produced the least scatter for red galaxies.
Our findings provide insights into the complex relationship between galaxy properties and have the potential to establish a correlation between observed and intrinsic properties. |
