Lunch talk on Nov. 20, 2023

Advancing Astrophysics: Machine Learning Solutions for Large-Scale Sky Surveys and Data Challenges

Speaker:Zechang Sun (Tsinghua University)

Venue:SWIFAR Building 2111

Time:12:30 PM, Monday, Nov. 20, 2023

Abstract:

The ongoing and upcoming large-scale sky survey projects such as DESI and Euclid offer valuable opportunities to unravel the holy grails of astrophysics: dark matter and dark energy. However, the vast amount of data also poses significant challenges in data processing and analysis. Machine learning is currently considered a promising statistical method with great potential for the future, but its application is limited by incomplete datasets and a lack of interpretability. In this talk, I will share my research experience on how to build robust and interpretable machine learning systems to serve large-scale survey projects, addressing various dataset imperfections and interpretability. When datasets are relatively easy to prepare, we use CNNs to detect damped Lya absorbers in DESI quasar spectra; in the absence of labeled data, we use unsupervised learning to generatively model quasar spectra, separating intrinsic continua and absorption components. The model is highly robust and interpretable, suitable for various downstream tasks. In the case of anisotropy in photometric redshift training sets in weak lensing analysis, we combine mixed density estimation with normalizing flow to provide an additional quality flag for photometric redshift estimates in addition to the estimation itself. If time permits, I will also discuss how large language models can help us understand the shift in hotspots of astronomical research and assist astronomers in their work as a member of UniverseTBD.

Report PPT:SWIFAR_Zechang Sun.pptx