Gravitational Lens Detection with CNN & ResNet

Gravitational lensing — the bending of light by massive bodies — is invaluable in astrophysics for probing dark matter and refining cosmological parameters. But detecting lensing events in survey data is hard: low signal-to-noise, varied morphologies, and image artifacts (PSF variability, cosmic rays, foreground stars) all confound classical pipelines.

This project benchmarks deep learning approaches end-to-end. Models are trained on simulated lens images and evaluated against real HST cutouts (CASTLES positives, COSMOS negatives) to test real-world generalization. Two tasks: binary classification, and Einstein-radius regression as a direct physical quantity prediction.

Key findings: deeper ResNet-18 models significantly outperform shallow CNNs on real test data, and training on realistic (PSF / noise / cosmic-ray augmented) simulations improves precision and recall substantially over idealized training. Grad-CAM saliency reveals that better-performing models attend to actual arc structures rather than background artifacts.