星系特别兴趣小组研讨会，2026年9月8日

学习看得更清晰：用于天文数据增强的深度学习

Learning to See Sharper: Deep Learning for Astronomical Data Enhancement

演讲者

Speaker

Shoubaneh Hemmati (加州理工学院/红外处理与分析中心 IPAC)

Shoubaneh Hemmati (Caltech/IPAC)

摘要

Abstract

天文巡天在本质上需要在覆盖范围、深度和分辨率之间进行权衡，这构成了一个定义不佳的逆问题（inverse problem），使得传统的反卷积方法难以取得进一步突破。通过利用巡天重叠区域作为经验先验，深度学习模型能够学习从低分辨率到高分辨率的直接映射，从而增强广域档案数据。我将介绍针对不同成像和光谱挑战的专业神经网络架构，包括用于分离拥挤视场和对红外观测进行超分辨率重建的生成模型、用于弱引力透镜效应（weak lensing）的保形残差网络、用于生成逼真巡天模拟的扩散模型，以及能够解析混合发射线（emission line）的物理信息模型。将合适规模的架构与特定的科学测量需求相匹配，展示了学习到的先验知识如何能够跨越下一代巡天项目实现数据增强。

Astronomical surveys inherently trade off coverage, depth, and resolution, presenting an ill-posed inverse problem where classical deconvolution methods plateau. By leveraging survey overlaps as empirical priors, deep learning models learn direct low- to high-resolution mappings to enhance wide-area archives. I will present specialized neural architectures tailored to distinct imaging and spectroscopic challenges, from generative models that deblend crowded fields and super-resolve infrared observations, to shape-preserving residual networks for weak lensing, diffusion models for realistic survey simulations, and physics-informed models that resolve blended spectral emission lines. Matching right-sized architectures to specific scientific measurements shows how learned priors can scale data enhancement across next-generation surveys.