Speaker
Description
We introduce OJALA (Optimizing J-PAS Astronomy for Large-scale Analysis), a Transformer-based foundation model specifically designed to analyze narrow-band photometry from the Javalambre Physics of the Accelerating Universe Astrophysical Survey (J-PAS). The model is pre-trained on synthetic photometry derived from 19.6 million spectra from the Dark Energy Spectroscopic Instrument (DESI). Leveraging DESI value-added catalogs, OJALA is trained to predict stellar masses and photometric redshifts for galaxies, equivalent widths (EWs) of emission lines, and black hole masses for quasars. Additionally, it estimates stellar parameters, including effective temperature, surface gravity, metallicity ([Fe/H]), and alpha-enhancement ([α/Fe]).
A key advantage of our approach is its unified architecture, which performs classification and regression tasks simultaneously. This eliminates the need for multiple specialized pipelines and significantly accelerates inference compared to traditional techniques. Furthermore, the Transformer architecture naturally handles incomplete data, facilitating the integration of multi-wavelength datasets. We also introduce a domain adaptation framework to bridge the gap between simulated J-PAS fluxes and real observations—a crucial step for future foundation models trained on hybrid datasets.
We demonstrate that OJALA's embeddings can be fine-tuned with minimal training to predict physical properties not seen during pre-training. Finally, beyond standard predictive tasks, we illustrate the model's versatility in performing similarity searches, which we leverage to generate galaxy segmentation maps for resolved galaxies in J-PAS.
| Stream | Science or Engineering |
|---|