Abstract

Spectroscopic surveys are essential for measuring galaxy redshifts and probing the physical processes driving galaxy evolution. As datasets grow in scale and complexity, traditional analysis methods face increasing limitations, motivating the development of scalable, data-driven alternatives. This thesis presents SpecPT, a transformer-based deep learning framework for general-purpose spectroscopic analysis, with a focus on redshift estimation. The model is first trained on DESI Early Data Release spectra from the Bright Galaxy and Emission Line Galaxy samples, jointly performing spectral reconstruction and redshift regression while learning a latent representation that captures the intrinsic properties of galaxies. SpecPT is then extended to a unified model trained across BGS, ELG, and LRG samples, achieving robust performance across diverse galaxy types and redshifts without class-specific supervision. Finally, the DESI-trained model is fine-tuned on a small set of Keck/DEIMOS spectra from the COSMOS field, demonstrating strong transfer learning capabilities and accurate redshift predictions even in low-data regimes. SpecPT establishes a scalable and adaptable foundation model for spectroscopy capable of robust inference across instruments, redshifts, and data quality. The results lay the groundwork for future applications to space-based grism spectroscopy and downstream physical property estimation.

Library of Congress Subject Headings

Spectrum analysis--Data processing; Deep learning (Machine learning); Red shift--Measurement

Publication Date

8-15-2025

Document Type

Dissertation

Student Type

Graduate

Degree Name

Astrophysical Sciences and Technology (Ph.D.)

Department, Program, or Center

Physics and Astronomy, School of

College

College of Science

Advisor

Jeyhan Kartaltepe

Advisor/Committee Member

Linwei Wang

Advisor/Committee Member

Jonathan Trump

Campus

RIT – Main Campus

Plan Codes

ASTP-PHD

Download

Share

COinS