Continued LSST-style surveying using ML: A survey-native foundation model for the faint targeting crowded end scene-level detection, deblending, and classification learned from Rubin's own unlabeled pixels

Continued deep surveying drives most detections into the blended, faint regime where catalog-first pipelines strain, yet vetted labels scale with spectroscopic follow-up rather than with imaged area or depth — leaving roughly 90% of an extended survey's footprint label-poor. This Rubin 2036 Group A white paper proposes a survey-native foundation model that learns scene-level detection, deblending, and star/galaxy classification directly from Rubin's own unlabeled ugrizy coadds through self-supervised pretraining, anchored on the small labeled overlap with high-resolution imaging and spectroscopy. In the first application of the DeepDISC scene-level framework to real LSST imaging — 1,458 six-band tiles over ECDFS+EDFS, 288,277 objects, validated against HST/CANDELS — the model works directly from pixels even when catalog entries or bands are missing, reaches near-parity with a catalog-based Random Forest on star/galaxy classification, and produces cleaner deblending than the Scarlet-based production pipeline (94.6% vs. 77.6% purity on a held-out HST footprint). The central figure of merit is label efficiency: matching fully supervised detection and classification from only 5–15% of the labels, the threshold that would make a scene-level catalog deployable across a deepening survey. The proposal requires no new hardware and no additional telescope time — only sustained compute, a deep-field truth program, and an integration path to release scene-level ML catalogs as supported value-added data products.