Identifying Strongly Lensed Galaxies with Prospector

We present a new method for identifying strongly lensed galaxies in astronomical surveys. Combining data from ~10,000 galaxies chosen at random from the COSMOS field, which are assumed to be unlensed, along with ~1,000 known lensed galaxies from the Master Lens Database and Huang et al. 2021, we train a machine learning algorithm to detect lensed galaxies in any data set given the same initial parameters. The training data set consists of photometry from the DESI Legacy Survey DR9 (filters g, r, z, w1, and w2) and fitted parameters from the Prospector stellar modeling package (mass, metallicity, star formation history, etc.) for each galaxy. Theoretically, lensed galaxies will have subtly different parameters when they are fitted with Prospector, even if the lensing is not resolved by the telescope/instrument. As lensed galaxies are very rare in nature, we augment this set to match the number of unlensed galaxies and train a Random Forest Classifier with balanced class weight, manually shifting the decision boundary to minimize false positives. The final algorithm is 92% accurate, with a False Positive Rate of 0.09%. This opens up the prospect for a much improved allocation of follow-up resources to specifically target supernovae in potential strong-lensing systems in order to carry out future measurements of the Hubble-Lemaître constant.