DeepLense: Lens Finding for LSST Images

In GSoC 2025, I prepared a novel physics-informed Swin Transformer to classify between lens models in simulated gravitational lensing images. The model overtook its contemporaries in sparse datasets. However, the core challenge often is that models trained on perfect simulation data often fail when applied to noisy, complex, and unlabeled observational data, like the Legacy Survey of Space and Time corpus. Thus, I am proposing an application to that model, through a real-life LSST dataset. The strategy that will be used will be a hybrid of “conventional” substructure classification using vision transformers, and a masked autoencoder, as the dataset itself is largely unlabeled. This will allow us to apply the novel architecture defined in the last project into real-world unlabeled datasets like the LSST corpus. The deliverable will be a foundational model with its core being the HEAL-Swin model, having both the strengths of a a physics-informed relativistic inference, and applicability to real-world unlabeled datasets. The aim will be to deliver publishable results by the end of the program.