The Observable Universe working group is currently establishing a "data-to-discovery" pipeline for radio astronomy, specifically designed to address the challenges posed by high-dimensional, low signal-to-noise datasets. Modern radio telescopes generate massive volumes of raw telescope data that must undergo an elaborate and computationally intensive process of calibration and reduction, including removing interference, correcting for atmospheric effects, and varying antenna gains, before it becomes science-ready. This raw data is inherently multidimensional, and once processed, it yields massive hyperspectral cubes. These cubes are three-dimensional data structures where two dimensions represent the sky's coordinates and the third represents frequency (or velocity). As next-generation facilities like the ngVLA and the ALMA WSU come online, our goal is to support astronomical data processing by transitioning from largely manual supervision to foundational AI automation.

The first primary focus of our research is on developing new statistical methods for anomaly detection in raw telescope data. In this context, anomaly detection refers to identifying problematic data—such as Radio-Frequency Interference (RFI), hardware malfunctions, or calibration errors—that can corrupt the final image. We are working toward AI models designed to flag these "anomalies" automatically, pushing the boundaries of what can be recovered from datasets where the signal is nearly buried. By isolating and mitigating these data-quality issues at the source, we aim to fully automate the radio data processing pipeline, ensuring that only high-fidelity signals reach the final archive.

The second primary front focuses on segmentation and identification within these high-dimensional cubes. Using deep learning, we intend to automatically isolate and "mask" regions of scientific interest—such as rotating proto-stellar disks or distant evolving galaxies. Segmentation is the process of partitioning a digital image (or a 3D volume) into multiple segments to simplify its representation. Our objective is to train an AI model to accurately trace the boundaries of these complex structures, providing researchers with precisely defined targets for follow-up analysis without the need for manual inspection.

A significant offshoot of these efforts will be the automated generation of high-quality training data for AI/ML applications by the community. Historically, the "bottleneck" in astronomical AI has been the lack of "ground truth" labels—the human-verified examples needed to train neural networks. As our detection and segmentation algorithms begin to process the archives, they are expected to naturally produce a vast, self-consistent library of labeled data. This secondary outcome will abstract complex raw data into manageable, pre-labeled categories, creating a foundational resource for the broader community to train their own specialized models.

Finally, these research outputs are planned for integration into CosmicCoder, a virtual research assistant. By feeding our future automated labels and segmented data into a large language model (LLM) framework, we aim to allow researchers to query archives using natural language. This would transform the global scientific community's interaction with radio archives from a manual search process into an intuitive, AI-guided exploration.