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

Advancements in AI have propelled multi-modal models to the forefront of astronomical research, particularly in time-domain astronomy. These models integrate diverse data types—images, light curves, spectral data, and metadata—to enhance analysis and prediction of dynamic celestial phenomena like supernovae, variable stars, X-ray bursts, and tidal disruption events. The Time Domain Astronomical Information Center (TDIC) science platform of the National Astronomical Data Center (NADC) provides a data management and analysis pipeline service for the Einstein Probe (EP). This talk will explore how to use multimodal models to deeply mine the X-ray transient source data generated by the EP under the technical framework of the National Astronomical Data Center (NADC).

This talk focuses on our X-ray multi-modal model for rapid preliminary classification of EP’s real-time data. The model employs a two-stage classification scheme. In the first stage, the input observational data are classified into true sources, cosmic rays, and arm (a unique phenomenon produced by the scientific payload of EP). In the second stage, the true sources are further subjected to preliminary classification of celestial types. We created a multi-modal labeled dataset using the observational data from EP WXT after calibration for model training (as of July 10, 2025, with more than 230,000 observational data entries). During the experiments, we found that the image data in the multimodal data effectively helped the model distinguish the interference of arm in true source identification. The X-ray multimodal classification model will be integrated into TDIC to conduct rapid identification and classification at the initial stage of EP observational data generation. The EP X-ray multimodal classification model, as an attempt and application of NADC in AI for Astronomy, is expected to be applied to more large-scale time-domain projects in the future to help researchers better explore and mine data.