From Simulations to Surveys: Fine-Tuning Transformers for Strong Lens Detection in KiDS

With the upcoming Legacy Survey of Space and Time (LSST) poised to reveal tens of billions of galaxies, including a significant fraction exhibiting gravitational lensing features (10^5). We explore the potential of Transformer Encoders to detect strong gravitational lenses (SGLs) across wide-area surveys. Leveraging 221 square degrees of the Kilo Degree Survey (KiDS) as a testing ground, we first train our model on simulated data from the Bologna Lens Challenge, closely resembling real KiDS observations. Subsequently, we fine-tune the model using actual KiDS data, including images of data augmented SGL candidates from previous searches and non-lenses examples.

Fine-tuning results in a substantial 70% reduction in false positives. This search leads to a catalog of 263 SGL candidates, including 43 newly identified high-confidence SGLs. Our approach demonstrates the effectiveness of fine-tuning transformer encoders with real augmented images. This refinement enhances SGL identification precision but also offers a crucial methodology for transitioning from simulations to real surveys. Furthermore, we present a catalog of 118 false positives, mimicking lens-like features, enriching the training resources for future machine learning models in the intricate domain of lens detection.