SADIM: A Pre-trained Foundation AI for Galactic Computing and Autonomous Discovery of Anomalous Stellar Phenomena in Gaia DR3

 

⁠Project SADIM-77M represents a paradigm shift in astronomical data processing, moving from static archival searches to dynamic "Galactic Computing." This framework reimagines the ESA Gaia DR3 archive a collection of trillions of data points as a Living Digital Map.

⁠By employing a pre-trained neural-statistical model with 77 million parameters, SADIM converts raw stellar metadata into a High-Dimensional Neural Vector Space. Each celestial object is encoded into a 512-dimensional latent vector, creating a unique "Digital Fingerprint" that represents the star's statistical essence and its deviation from the cosmic norm.

⁠This approach moves beyond traditional tabular queries, allowing for the autonomous identification of rare stellar phenomena through latent manifold analysis. The system achieves a 99.9% reduction in query latency, proving that the future of discovery lies in the neural representation of the universe. This publication outlines the architecture of the SADIM-77M autoencoder and its capacity to "understand" galactic structures through high-dimensional inference.