Spectroscopic Completeness and Photometric Redshift Performance in Astronomical Foundation Models

Using a combination of self-organized maps (SOM) to perform prototype learning and astronomical foundation models to provide embeddings, we present an analysis of a new photometric redshift model for the DESI Legacy Survey footprint. Using the groupings learned by our SOM, we investigate the role local training-sample density plays into performance. The SOM can be used to flag samples that users request which lie outside the distribution of training data, or those examples which are known to belong to cells where the model under-performs. This flag can help scientists better understand the performance of our model on their specific sample to make educated decisions tailored to their downstream analysis.