Physics-Informed Super-Resolution for Strong Lensing

DeepLense-SR: Advancing Analysis for LSST & Euclid Surveys
ML4Sci GSoC 2026 Preparation

Try Interactive Demo View Analysis Physics-Informed Scientific Showcase

Why Super-Resolution?

Upcoming surveys like Euclid and LSST are predicted to discover over 100,000 strong lenses. However, many of these systems will have small Einstein radii ($\theta_E < 1"$) that are barely resolved at survey resolution.

DeepLense-SR applies physics-informed machine learning to enhance these images, recovering sub-pixel details critical for:

  • Detecting dark matter substructure (subhalos).
  • Accurately modeling lens mass profiles.
  • Maximizing scientific yield from ground-based surveys.
"Euclid wide survey will find ~100,000 galaxy-scale lenses."
— Collett, ApJ 2015

Interactive Viewer

Select a sample from the left to explore different reconstruction methods.

Select a sample to view

Metrics

Arc Sharpness --
Ring Contrast --

Description

Select a sample to see details about the lens configuration.

* Metrics compared to Ground Truth HR.

Results Analysis

Method Comparison (DeepLense Model I-compatible Mock Validation)

Comparison of average performance across the validation set.

Method Arc Sharpness Score Ring Contrast Ratio Training Signal
Bicubic Interpolation 221.96 5.08 N/A
SR Baseline (Supervised) 268.50 4.97 HR labels
SR Hybrid (Physics-Informed) 267.41 5.39 Sup + Physics. Highest contrast.
Unsupervised (Physics-Only) 249.40 4.83 Physics only. TV Regularized.
HR Ground Truth 266.39 5.14 Target

Visual Comparison Grid

Qualitative comparison showing arc sharpness and ring continuity.

Comparison Grid: LR vs Methods vs HR

Left to Right: Low Res, Bicubic, SR Baseline, SR Hybrid, High Res (Ground Truth)
*(Note: Unsupervised not shown in standard validation grid)*