Conceptual
Browse Instruct Curate

Pretrained Transformer for Galaxy Spectrum Reconstruction and Redshift Measurement

A self-supervised transformer model that encodes galaxy spectra into a latent representation, using an autoencoder plus transformer encoder to reconstruct spectra (denoising while preserving emission lines, absorption features, and continuum) and to predict spectroscopic redshift as a continuous regression rather than by binning. The learned latent space is intended as a foundational representation transferable to other spectroscopic analysis tasks such as galaxy-property estimation and outlier detection.