2025 workshop program
Super-Resolving Low-Resolution Galaxy Spectra with Diffusion Models
The work
Abstract
High-resolution spectroscopy is essential for studying galaxy properties, but acquiring such data is observationally expensive. We present a novel application of diffusion-based deep learning models to enhance the spectral resolution of galaxy spectra, increasing the resolving power from R~100 to R~1000. Our approach utilizes JWST/NIRSpec observations from the JADES program, where we reconstruct high-resolution spectra from the low-resolution PRISM dataset (R~100). To achieve this, we combine three medium-resolution datasets—F070LP/G140M, F170LP/G235M, and F290LP/G395M—each covering different wavelength ranges, into a unified high-resolution reference spectrum. The model is trained on 80% of the data, with the remaining 20% used for validation. Validation is performed by assessing the accuracy of recovered atomic line wavelengths and flux ratios (e.g., Hα/Hβ). By leveraging learned priors from high-resolution data, this method enables superior deblending of emission lines, such as [NII] and Hα, compared to traditional stellar mass scaling approaches, leading to more precise redshift measurements for galaxy clustering among other things. Our method has broad implications for future deep-field surveys, including those from the Roman Space Telescope, where low-resolution grism spectra could be computationally transformed into high-resolution datasets, unlocking new insights into galaxy formation and evolution.