The 35th annual conference on Astronomical Data Analysis Software & Systems 2025 in Görlitz, Germany

To address the challenge of manually analyzing the approximately 50 daily X-ray transient candidates from the Einstein Probe (EP) satellite—a process that can take 10-30 minutes per source—we have developed an AI-driven Real-Time Transient Identification Assistance System. Built upon the AI Agent framework and leveraging Large Language Models, the system is designed to emulate an experienced researcher by automatically retrieving multi-source data, analyzing alerts, and providing fully supported identification conclusions.
The system's intelligence is founded on a multi-agent architecture powered by a comprehensive Time-domain Astronomy Knowledge Base containing over 14,000 scientific papers and 200,000 EP observation records. It is continuously refined through multi-stage learning with Human-in-the-Loop feedback from EP Duty Scientists.
Upon data downlink, the agents automatically ingest transient candidates and begin a multi-layer distributed decision process. The first step is vetting, using specialized models to distinguish true X-ray sources from cosmic rays and instrumental artifacts. If a candidate is deemed a real source with significant flux changes compared to historical data, the system initiates a broad multi-band analysis, retrieving data from Simbad, NED, Gaia, AllWise. Concurrently, it cross-matches the candidate with time-domain alerts from GCN/ATel/TNS to check for associations with previously transients. By synthesizing this multi-modal information, the agents generate a identification conclusion with a detailed reasoning process.
The effectiveness of this identification process has been confirmed through verification against manual Identification records, demonstrating the system's strong real-world performance. It achieves high accuracy rates for identifying specific source types, including cosmic_ray (99.3%), instrumental artifacts (97.3%), ordinary known source (94.1%), stellar_flare (95.5%) and transient (92.3%).
Furthermore, for high-value candidates, the agent intelligently submits follow-up observation plans to multiple telescopes. This initiates a closed-loop process where the newly acquired observation data is fed back into the system, enabling the agents to refine their initial assessment and provide a more accurate identification.