SpecPT (Spectroscopy Pre-trained Transformer) Model for Extragalactic Spectroscopy: I. Architecture and Automated Redshift Measurement

TL;DR

The paper presents SpecPT, a transformer-based model for analyzing extragalactic spectra, achieving accurate spectrum reconstruction and redshift measurement across diverse datasets, paving the way for a versatile foundation in spectroscopic analysis.

Contribution

Introduction of SpecPT, a novel transformer-based model that effectively reconstructs spectra and measures redshift, demonstrating robustness and versatility across multiple extragalactic datasets.

Findings

Accurate spectrum reconstruction with noise reduction.

Redshift prediction with NMAD of 0.0006 and 0.0008.

Consistent performance across the redshift range 0-1.6.

Abstract

We introduce the Spectroscopy Pre-trained Transformer (SpecPT), a transformer-based model designed to analyze spectroscopic data, with applications in spectrum reconstruction and redshift measurement. Using the Early Data Release (EDR) of the DESI survey, we evaluate SpecPT's performance on two distinct datasets: the Bright Galaxy Survey (BGS) and Emission Line Galaxy (ELG) samples. SpecPT successfully reconstructs spectra, accurately capturing emission lines, absorption features, and continuum shapes while effectively reducing noise. For redshift prediction, SpecPT achieves competitive accuracy, with Normalized Median Absolute Deviation (NMAD) values of 0.0006 and 0.0008, and catastrophic outlier fractions of 0.20% and 0.80% for BGS and ELG, respectively. Notably, SpecPT performs consistently well across the full redshift range ($0 < z < 1.6$), demonstrating its versatility and robustness. By leveraging its learned latent representations, SpecPT lays the groundwork for a foundational spectroscopic model, with potential applications in outlier detection, interstellar medium (ISM) property estimation, and transfer learning to other datasets. This work represents a first step in building a generalized framework for spectroscopic analysis, capable of scaling to the full DESI dataset and beyond.