Search for Strong Gravitational Lenses Project Description

Strong gravitational lensing is a powerful tool in exploring various astrophysical questions, including probing the substructure in dark matter haloes of the lensing galaxies. However one of the main limitations of such analysis is the relatively small number of known lens candidates and confirmed lens systems.

Recent works have shown the potential of CNNs in the task of lens finding — classification of images obtained from the telescopes into lensed and non-lensed systems. Since the number of real lenses is insufficient for training a machine learning algorithm, training datasets heavily rely on simulations. However it has been noticed that CNNs perform worse on lens images obtained with the instrument from the one that simulations were tailored to reproduce (for example, different surveys use different color filters and have different resolution).

The goal of this project is to investigate the prospects of using domain adaptation techniques to bridge the gap between simulated data used for training and real images from different surveys (such as HSC-SSP, HST, DES, JWST, and future missions) and explore which type of lenses has a higher risk of being lost during the automated searches.

This repository contain two tests:

1. Common Test I. Multi-Class Classification
2. Specific Test II. Lens Finding

The result for common test I is defined challenge1.ipynb file and result for specific test II are defined in challenge2.ipynb and challenge2_no_da.ipynb file. In specific test II, I use two different approach, the first is using domain adaptation technique and the other one is not using domain adaptation. Summary of the final results for both test are:

1. The trained model weight for common test I is in model_weights_challenge1.pth. AUC score for this model are: 0.97, 0.95, and 0.97 for lensing images with no substructure, subhalo substructure, and vortex substructure respectively.
   
2. The trained model weight for challenge II is in model_weights_challenge2_with_da.pth and model_weights_challenge2_no_da.pth for model with domain adaptation and without domain adaptation respectively. AUC score are 0.95 for model with domain adaptation and 0.94 for model without domain adaptation. Based on this result, we can conclude that domain adaptation can help to increase the accuracy on validation dataset.