TEGLIE: Transformer encoders as strong gravitational lens finders in KiDS

Margherita Grespan,Hareesh Thuruthipilly,Agnieszka Pollo,Michelle Lochner,Marek Biesiada,Verlon Etsebeth

2024-05-20

Abstract:We apply a state-of-the-art transformer algorithm to 221 deg$^2$ of the Kilo Degree Survey (KiDS) to search for new strong gravitational lenses (SGL). We test four transformer encoders trained on simulated data from the Strong Lens Finding Challenge on KiDS survey data. The best performing model is fine-tuned on real images of SGL candidates identified in previous searches. To expand the dataset for fine-tuning, data augmentation techniques are employed, including rotation, flipping, transposition, and white noise injection. The network fine-tuned with rotated, flipped, and transposed images exhibited the best performance and is used to hunt for SGL in the overlapping region of the Galaxy And Mass Assembly (GAMA) and KiDS surveys on galaxies up to $z$=0.8. Candidate SGLs are matched with those from other surveys and examined using GAMA data to identify blended spectra resulting from the signal from multiple objects in a fiber. We observe that fine-tuning the transformer encoder to the KiDS data reduces the number of false positives by 70%. Additionally, applying the fine-tuned model to a sample of $\sim$ 5,000,000 galaxies results in a list of $\sim$ 51,000 SGL candidates. Upon visual inspection, this list is narrowed down to 231 candidates. Combined with the SGL candidates identified in the model testing, our final sample includes 264 candidates, with 71 high-confidence SGLs of which 44 are new discoveries. We propose fine-tuning via real augmented images as a viable approach to mitigating false positives when transitioning from simulated lenses to real surveys. Additionally, we provide a list of 121 false positives that exhibit features similar to lensed objects, which can benefit the training of future machine learning models in this field.

Astrophysics of Galaxies,Instrumentation and Methods for Astrophysics

What problem does this paper attempt to address?

This paper mainly explores how to use deep learning techniques, especially the Transformer encoder, to automatically detect Strong Gravitational Lens (SGL) phenomena in large-scale astronomical observations. With the progress of new observation projects such as the Legacy Survey of Space and Time (LSST) and the Euclid mission, billions of galaxy images will be generated, which contain a large number of potential SGL events. Due to the rarity of these events and the inefficiency of manual identification, automated methods are needed.

The researchers applied the Transformer algorithm to search for new SGLs in a 221 square degree area of the Kilo Degree Survey (KiDS). They first trained four Transformer encoders on simulated data and fine-tuned them on real images of previously identified SGL candidates. By using data augmentation techniques to expand the training set, the number of false positives was reduced. The fine-tuned models filtered approximately 51,000 SGL candidates out of around 5 million galaxy samples, and 231 candidates, including 71 high-confidence SGLs and 44 new discoveries, were ultimately confirmed through visual inspection.

The paper points out that although the Transformer model has been improved, its performance is still inferior to models trained directly on KiDS galaxy images. The authors suggest fine-tuning with augmented data from real images to reduce false positives, and provide 121 false positive instances that may be helpful for training future machine learning models.

In conclusion, this paper aims to address the effective automatic identification of rare strong gravitational lensing events in large-scale astronomical observation data. By applying and optimizing the Transformer encoder, the accuracy and efficiency of automatic detection have been improved.