Speaker
Description
Active Galactic Nuclei are powered by accretion onto supermassive black holes and exhibit a wide range of observed properties. While the Unified Model explains part of this diversity through orientation effects, growing observational evidence suggests that additional factors-such as evolutionary stage and host galaxy properties-also play a significant role.
In this context, we adopt a data-driven approach to investigate whether physical and evolutionary patterns can be identified across different AGN types. Our goal is to explore the relationships between optical spectral features, X-ray, and host galaxy properties, and to assess whether a unified representation can capture the intrinsic properties of AGN beyond traditional classification schemes.
To this end, we develop a multimodal machine learning framework integrating optical spectra, imaging, and tabular X-ray data. Each modality is modeled through a dedicated encoder, pre-trained independently on large, domain-specific datasets.
The encoders are then jointly fine-tuned on the eROSITA eFEDS dataset, enabling the construction of a shared latent space where different data modalities provide a consistent representation of the same underlying physical properties of AGN.
We analyze the learned representations using dimensionality reduction and clustering techniques, identifying structures, sub-populations, and potential evolutionary trends, along with correlations with known physical parameters.
Beyond its exploratory nature, the proposed framework provides a basis for downstream tasks such as AGN classification, redshift estimation, and the prediction of physical properties.