AstroAgent Projects¶
AstroAgent is an initiative exploring how to use AI agents to help manage and develop complex control software in astronomy and other scientific projects. The idea is to fine-tune large language models (LLMs) on specific project documentation and code, so they can understand the context, interact with external APIs, and communicate in natural language. These AI agents act like smart assistants that can handle tasks such as generating code or configuring systems, taking some load off human experts. In short, it’s about letting an AI helper handle part of the heavy lifting in running sophisticated instruments and software.
Publications & Prototypes¶
AI Agents for Ground-Based Gamma Astronomy¶
An introduction to using LLM-based agents in gamma-ray astronomy, featuring two prototypes integrated with the Cherenkov Telescope Array Observatory (CTAO) pipelines.
Agent-based code generation for the Gammapy framework¶
Focuses on an AI agent that writes, executes, and validates code for the Gammapy library. It describes a code-writing assistant running in a sandbox that checks its own output, and it includes a minimal web demo (interactive prototype) and a benchmarking suite.
Enhancing the development of Cherenkov Telescope Array control software with LLMs¶
Discusses integrating LLM-driven agents into the CTA observatory’s Control and Data Acquisition (ACADA) software. It introduces “CTAgent” for helping with control software development.
AI-Based Assistance System for Control Rooms in Large-Scale Infrastructures: Multi-agent architecture and verifiable workflows¶
An AI-based approach to supporting control rooms in large-scale infrastructures is presented. Distributed data sources, unclear documentation, and complex system dependencies make rapid and reliable decision-making difficult in such environments. The developed assistance system consolidates knowledge from operating manuals, experiential expertise, and real-time data, and makes it accessible through a natural language interface. Technically, the system is based on a locally operated multi-agent architecture that integrates data from monitoring and control software. A verifiable workflow with fixed feedback loops ensures that inputs and outputs remain traceable and stable. This makes it possible to translate probabilistic models into comprehensible and reproducible action steps—an essential aspect for deployment in safety-critical environments. The development and testing take place in scientific facilities being established at DESY Zeuthen. The paper describes initial results, challenges related to data quality and integration, and the potential contributions of the approach to resilient, transparent, and sustainable AI support systems for control rooms.
Partners¶
