Relatore
Descrizione
The exponential growth of astronomical datasets demands a paradigm shift in data analysis, with AI emerging as a critical enabler. We present a dedicated local AI infrastructure comprising five NVIDIA GB10 Grace-Blackwell Superchip nodes, each with 128 GB of unified memory, designed to deliver domain-specific AI capabilities while ensuring full data sovereignty. The system runs open-weight Qwen models optimized via Mixture-of-Experts (MoE) architectures for automated code generation, documentation, and agentic workflow orchestration—enabling dynamic pipeline development that keeps pace with scientific discovery. We argue that agentic AI workflows—capable of autonomous task orchestration, failure recovery, and adaptive reasoning—will become the dominant paradigm for complex astronomical pipelines. Key applications include automated flagging, calibration, imaging, and source extraction in radio astronomy, where LLM agents can integrate domain knowledge across heterogeneous data conditions, as required by the data deluge from next-generation observatories like the Square Kilometre Array (SKA). To scale these capabilities, we are planning to expand our infrastructure to support inference on larger LLMs (200B+ parameters) and will leverage this capacity to train a lightweight, domain-specific LLM tailored to astronomical workflows, using curated datasets of pipeline scripts, calibration logs, and peer-reviewed literature. This localized, sovereign approach ensures reproducibility, efficiency, and privacy—critical for next-generation data-intensive science. By embedding AI directly into the scientific workflow, we move beyond tool-assisted analysis toward autonomous, adaptive discovery, establishing a sustainable model for AI-augmented astronomy at scale.
| Sessione | Calcolo, Archivi e Intelligenza Artificiale |
|---|