Transformers for Dark Matter Morphology with Strong Gravitational Lensing

Since Dark Matter was discovered, physicists have been trying to understand its composition. In practice, the best method to detect substructure is from strong gravitational lensing images. Given the huge success of deep learning algorithms in computer vision across various fields, it seems a promising path to use these algorithms on gravitational lensing images to classify the Dark Matter substructure. In recent years, attention-based networks like Vision Transformers have shown immense promise in Computer Vision tasks. Hence, the intention is to use them for classifying Dark Matter substructure from gravitational lensing images. DeepLense is a deep learning pipeline for particle dark matter searches with strong gravitational lensing. This project will focus on implementing Vision Transformers in the DeepLense pipeline, which is expected to immensely boost the current performance of the system. The overall deliverables of this project are: - A Python module implementing Vision Transformers (ViT). - Training/evaluation scripts to train this ViT model on gravitational lensing dataset. - Pre-trained, ready-to-use ViT models for other pipelines using lensing datasets.