Shanghai Astronomical Observatory Astrophysics Colloquium

Title:Machine Learning for Next-Generation Sky Surveys: Photometric Redshift, Image Enhancement, and Galaxy Morphology

Speaker: Zhijian Luo（Shanghai Normal University）

Time: 15:00, Thursday,23 July 2026

Tencent Meeting: 46822606747 password: 6360

Location: Lecture Hall, 3rd floor

Invited by: Fangting Yuan

Abstract：

This talk focuses on the application of machine learning to three key astronomical problems: photometric redshift estimation, galaxy image enhancement, and galaxy morphology classification.

For photometric redshift estimation, we present a series of research progresses, systematically demonstrating how data‑driven models can achieve high‑precision redshift inference using multi‑band photometric data and image information. Specific work includes: optimizing traditional machine learning models to improve the baseline accuracy of photometric redshift estimation; constructing dedicated deep learning photometric redshift templates tailored to the observational characteristics of the Chinese Space Station Survey Telescope (CSST); and developing a novel non‑parametric framework for accurately characterizing the multi‑modal structure of redshift posterior probability distributions. These results collectively enhance the accuracy and reliability of redshift estimation.

In the direction of image enhancement, we propose two deep learning models – the Pix2WGAN hybrid network and the CDDPM conditional diffusion model – aimed at achieving high‑fidelity transformation and data fusion between images from different survey projects. By upgrading ground‑based SDSS and DECaLS images to a resolution and signal‑to‑noise ratio close to those of HSC and the space‑based Euclid telescope, this method effectively overcomes the limitations of atmospheric seeing and significantly improves the detectability of faint, extended peripheral structures such as spiral arms and tidal tails.

For galaxy morphology classification, addressing the high cost of manual labeling in large‑scale survey data, we design the GC‑SWGAN semi‑supervised deep learning model. This model efficiently combines a very small set of labeled samples with massive unlabeled survey data, maintaining robust classification performance while greatly reducing the reliance on manual labels. The model has been validated in practical applications on DESI survey data, successfully detecting a number of rare and evolutionarily significant special galaxy structures, such as polar‑ring galaxies and dust‑lane elliptical galaxies.

CV：Zhijian Luo received his Ph.D. in Astrophysics from the Shanghai Astronomical Observatory, Chinese Academy of Sciences. He is currently a Professor and Doctoral Supervisor at Shanghai Normal University, and also a Researcher at the Shanghai Key Laboratory for Galaxy and Cosmology Semi‑analytic Research. His primary research interests include semi‑analytic modeling of galaxy formation and evolution, the structure and evolution of the Milky Way, and AI‑based processing and mining of astronomical big data. He has led or participated in multiple national and provincial/ministerial‑level research projects. Over the past five years, he has published more than 20 papers in journals such as ApJ, ApJS, A&A, and MNRAS.

Koushare link to the recorded presentation