Observable Universe Working Group

Part 1: The Observable Universe Working Group: progress and prospects
Speaker: Brian Mason (Scientist, NSF NRAO)

Part 2: SIFARI: Self-Supervised Neural Networks for High-Resolution Radio Imaging
Speaker: Shunyuan Mao (CosmicAI Postdoctoral Fellow at the National Radio Astronomy Observatory)

Part 3: LabelScope: Context-Aware Discovery of Label Outliers in Radio Astronomy Signals
Speaker: El Kindi Rezig (Assistant Professor, University of Utah’s Kahlert School of Computing)

Zoom & in person at NSF NRAO (Room 230)
Link to Zoom
Meeting ID: 886 3021 8251 / Passcode: 684859

Part 1 Abstract: The Observable Universe Working Group of the NSF-Simons AI Institute for Cosmic Origins brings together astronomers and AI researchers to advance the state of the art in the operation of world-leading observatories like ALMA and ngVLA. In this talk, I will provide an overview of the working group's activities, highlighting our goals, ongoing projects, and progress to date. I will also discuss emerging opportunities and where we are going next.

Part 2 Abstract: Radio telescopes such as ALMA do not take pictures directly. Instead, they measure samples of the sky’s Fourier transform, leaving astronomers to reconstruct an image from incomplete and noisy data. How can neural networks help recover the underlying sky—and assess which features are reliable?

In this talk, I will introduce SIFARI (Self-Supervised Interferometric Fitting for Astronomical Radio Imaging), a framework that represents sky brightness as a continuous function using a coordinate-based neural network. Each observation provides its own training signal: the network is fitted directly to the measured interferometric data, without an external image training set. I will explain how the network’s representation guides reconstruction and how sampling its weights provides approximate, spatially resolved uncertainty estimates.

Using synthetic ALMA observations, I will demonstrate improvements over CLEAN in resolving compact sources, recovering extended emission, and reconstructing complex structures under the tested conditions. I will then present an application to the planet-forming disks PDS 70. Throughout, I will discuss how the observations and the model’s assumptions jointly shape the reconstructed image.

Part 3 Abstract: The NSF National Radio Astronomy Observatory (NRAO) operates one of the largest radio telescopes in the world. As part of its data processing pipeline, it produces thousands of bandpass calibration signals per observation cycle, each capturing how an antenna's frequency response varies across a spectral window—essentially a sequence of amplitude measurements over frequency channels. To ensure data quality, automated detectors assign anomaly scores to these signals, and human experts review flagged cases to assign labels. In practice, scores and labels can disagree, and expert review cannot scale to every mismatch. We present LabelScope, an interactive system developed in collaboration with NRAO to address this problem. LabelScope groups signals using contextual metadata and feature augmentation, identifies minority-label instances within locally coherent groups, and generates predicate-based explanations for why they are interesting disagreement cases. Through its interface, users explore candidate label outliers and determine whether score-label disagreements reflect labeling errors, meaningful anomalies, or limitations of the scoring method. We demonstrate LabelScope on real astronomy signals, enabling NRAO scientists to efficiently direct limited reviewer attention toward the most informative disagreements.

Speaker Bios:
Brian Mason is a staff scientist at NSF NRAO and a co-lead of the Observable Universe working group. His observatory work has included instrumentation development, telescope operations, and data processing, with a current focus on ALMA and ngVLA development. His research interests include the cosmic microwave background, galaxy clusters, and applications of machine learning to radio astronomy.

Shunyuan Mao is a CosmicAI Postdoctoral Fellow at the NSF National Radio Astronomy Observatory in Charlottesville, Virginia. His research combines astronomy, scientific machine learning, and radio interferometry. He developed SIFARI, a self-supervised neural framework for reconstructing radio images directly from interferometric measurements. He also develops physics-informed and operator-learning methods for astrophysical fluid dynamics, with a focus on planet formation and protoplanetary disks.

El Kindi Rezig is an Assistant Professor at the University of Utah’s Kahlert School of Computing. His research focuses on data management and data quality, building systems that simplify data discovery, preparation, and debugging. Previously, he was a research scientist and postdoctoral associate at MIT CSAIL in the Data Systems Group. He earned his Ph.D. in Computer Science from Purdue University.