Integrating Machine Learning Techniques for Streamlined Predictive Modeling in Cosmological Applications
Loading...
Author(s)
Sethuram, Snigdaa Sairathi
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
This dissertation advances the integration of machine learning (ML) into computational astrophysics, demonstrating its capacity to streamline cosmological simulations and enhance predictive modeling. The research addresses three interconnected challenges: accelerating radiative transfer (RT) calculations, inferring galaxy properties from early observational data, and emulating stellar feedback in hydrodynamic simulations.
First, we develop ANNgelina, an artificial neural network trained on IllustrisTNG50 and FIREbox simulations, to emulate computationally intensive RT calculations and predict spectral energy distributions (SEDs) of galaxies. By learning the nonlinear relationships between galaxy properties and their SEDs, ANNgelina achieves a median absolute error of 0.06 dex (15%) across UV-to-millimeter wavelengths, bypassing traditional Monte Carlo RT methods while accelerating post-processing by orders of magnitude. Secondly, leveraging early James Webb Space Telescope (JWST) photometric data, we infer stellar masses and star formation histories of high-redshift galaxies. Our models incorporate diverse stellar populations, AGN contributions, and dust attenuation, revealing that stellar-dominated spectra—even with non-standard initial mass functions (IMFs)—best align with observations, yielding stellar masses consistent with initial JWST estimates. Finally, we introduce CosmoConv, a convolutional recurrent network trained on high-resolution SG256 simulation data, to emulate stellar feedback effects in cosmological simulations. By predicting gas density, temperature, and metal enrichment patterns, CosmoConv achieves ~62% accuracy in capturing feedback dynamics, offering a scalable, resolution-independent alternative to subgrid prescriptions.
Collectively, these projects highlight ML’s transformative role in astrophysics, enabling faster exploration of parameter spaces, improved synergy between simulations and observations, and novel insights into galaxy formation and cosmic reionization. The methodologies developed here provide a foundation for future applications in large-scale cosmological modeling, observational data interpretation, and next-generation simulation frameworks.
Sponsor
Date
2025-04-29
Extent
Resource Type
Text
Resource Subtype
Dissertation