TEGLIE: Transformer encoders as strong gravitational lens finders in KiDS. From simulations to surveys

Context. With the current and upcoming generation of surveys, such as the Legacy Survey of Space and Time (LSST) on the Vera C. Rubin Observatory and the Euclid mission, tens of billions of galaxies will be observed, with a significant portion (~10 5 ) exhibiting lensing features. To effectively detect these rare objects amidst the vast number of galaxies, automated techniques such as machine learning are indispensable. Aims. We applied a state-of-the-art transformer algorithm to the 221 deg 2 of the Kilo Degree Survey (KiDS) to search for new strong gravitational lenses (SGLs). Methods. We tested four transformer encoders trained on simulated data from the Strong Lens Finding Challenge on KiDS data. The best performing model was fine-tuned on real images of SGL candidates identified in previous searches. To expand the dataset for fine-tuning, data augmentation techniques were employed, including rotation, flipping, transposition, and white noise injection. The network fine-tuned with rotated, flipped, and transposed images exhibited the best performance and was used to hunt for SGLs in the overlapping region of the Galaxy And Mass Assembly (GAMA) and KiDS surveys on galaxies up to z = 0.8. Candidate SGLs were matched with those from other surveys and examined using GAMA data to identify blended spectra resulting from the signal from multiple objects in a GAMA fiber. Results. Fine-tuning the transformer encoder to the KiDS data reduced the number of false positives by 70%. Additionally, applying the fine-tuned model to a sample of ~5 000 000 galaxies resulted in a list of ~51 000 SGL candidates. Upon visual inspection, this list was narrowed down to 231 candidates. Combined with the SGL candidates identified in the model testing, our final sample comprises 264 candidates, including 71 high-confidence SGLs; of these 71, 44 are new discoveries. Conclusions. We propose fine-tuning via real augmented images as a viable approach to mitigating false positives when transitioning from simulated lenses to real surveys. While our model shows improvement, it still does not achieve the same accuracy as previously proposed models trained directly on galaxy images from KiDS with added simulated lensing arcs. This suggests that a larger fine-tuning set is necessary for a competitive performance. Additionally, we provide a list of 121 false positives that exhibit features similar to lensed objects, which can be used in the training of future machine learning models in this field.