Physics Guided Machine Learning on Real Gravitational Lensing Images

Strong gravitational lensing is a powerful tool for studying dark matter, but machine learning models trained on simulated data often fail to generalize to real observations due to noise, domain shift, and lack of physical constraints. This project addresses this gap by developing a physics-informed deep learning framework that incorporates the gravitational lensing equation directly into the training process. The proposed approach augments a deep neural network with physically derived deflection field channels and introduces a physics-consistency loss that enforces agreement between learned and predicted physical parameters (e.g., Einstein radius). This enables the model to learn representations that are both data-driven and physically grounded, improving robustness, interpretability, and cross-domain generalization. The project will deliver a modular, open-source library for physics-informed lensing analysis, including baseline and physics-augmented models, reusable physics layers (e.g., SIS/NFW profiles), training and evaluation pipelines, and reproducible benchmarking tools. Additional deliverables include pretrained models, documentation with tutorials, and a systematic evaluation of performance on both simulated and real lensing datasets.