Scalable Photometric Quantity Retrieval (SPQR)
| Program ID | 19088 |
|---|---|
| Science Category | Galaxies |
| Program Type | Analysis |
| Category | Small |
| Principal Investigator | Charles Steinhardt |
| PI Institution | University of Missouri |
| Co-Investigators |
|
| Abstract | We propose SPQR (Scalable Photometric Quantity Retrieval), a machine learning framework for rapid, survey-scale photometric redshift inference from Roman photometry. Existing template-fitting methods are either computationally prohibitive or prone to catastrophic errors for rare or unusual sources. SPQR emulates the mapping from photometry to galaxy parameters, reproducing complex spectral synthesis calculations orders of magnitude faster than these techniques, allowing it to scale to the size of Roman surveys. Because only a small fraction of Roman galaxies can be targeted for follow-up observations, accurate identification of high-redshift galaxies is critical, since these objects act as signposts for early structure formation and rare evolutionary stages. To identify rare or poorly modeled galaxies, we also propose VESTA (Variational Encoder for Spectral Type Anomalies), a variational autoencoder that flags photometric outliers, reducing catastrophic redshift errors. SPQR and VESTA will first start from existing multi-wavelength survey catalogs, which provide a well-studied training set. From this initial model, two fine-tuned versions will be released taking advantage of the efficiency of transfer learning : a mock catalog-based version for early deployment on Roman photometry and a refined version trained with the first Roman observations to capture survey-specific systematics. Because this fine-tuning is far less computationally intensive than training from scratch, the framework can readily incorporate alternative template sets generated by the community, enabling scalable, flexible inference and the accurate identification of rare and high-redshift targets without additional high-performance computing. |