Machine Learning for Gravitational Lens Finding

Strong gravitational lenses are rare and scientifically valuable, but finding them in large imaging surveys requires automated methods. Furthermore, simulated data is used due to the lack of available data, but models trained on simulated lensing images suffer a significant performance drop when applied to real observations (the sim-to-real domain gap). This project addresses that gap by combining equivariant neural networks (ENNs), which exploit the rotational and reflectional symmetries of lensing images, with domain adaptation techniques to build a lens-finding pipeline that transfers reliably from simulated training data to real survey images. The project will also develop a hybrid adaptation strategy that chains unsupervised alignment with supervised contaminant rejection to reduce false positives from visually similar objects like spiral galaxies. Through this project, I aim to build a benchmarked ENN+DA pipeline integrated into the DeepLense codebase, a ranked list of lens candidates from an HSC-SSP subregion with catalog cross-matching and property analysis, a cross-survey generalization evaluation, and a final research report with blog post.