AI for Science
Astrophysics and scientific computing.
- Astrophysics
- Strong Lensing
- Scientific Simulation
- PDEs
- Autonomous Discovery
Open-source research · ML4SCI
DeepLense is an open-source research group studying dark matter with strong gravitational lenses and developing machine-learning methods for scientific data.
Strong gravitational lensing. A foreground galaxy bends light from a distant source into arcs and rings. Dark-matter structure can perturb their shape.
Strong lenses provide a way to study matter that does not emit light. We build models that identify lenses, characterize their morphology, and test how much physical information can be recovered from images.
The work includes simulation, survey-data processing, representation learning, super-resolution, parameter inference, and experimental tools for scientific agents.
How the group worksResearch
We develop methods for astrophysical data and study machine-learning problems that recur across the sciences.
AI for Science
Core AI
Research workflow
The group works across the stages needed to connect controlled simulations with astronomical observations.
Generate controlled lensing systems across candidate dark-matter models.
Identify rare lens candidates in large, noisy astronomical surveys.
Train representations using known symmetries, geometry, and lensing physics.
Measure what the models recover and how reliably results transfer to survey data.
Selected work
A conditional flow-matching model for generating strong-lensing images across physical parameters and dark-matter classes.
A self-supervised model that uses exact lens symmetries and tests when those invariances remove information needed by downstream tasks.
A multi-agent framework for coordinating lens simulation, model search, evaluation, and experiment planning through typed interfaces.
Results above are reported on bounded experimental setups, primarily using simulated data. We distinguish benchmark performance from real-world scientific inference throughout the publication record.
Different measurements require different representations.
A symmetry that improves classification can remove detail needed for reconstruction. A sharper image may not improve a physical estimate. We evaluate models against the scientific task they are intended to support.
News & project notes
The accepted work covers label-efficient lens representations, symmetry-aware field surrogates, and resolution transfer in neural operators.
Explore the papers → CommunityThe projects cover scientific agents, foundation models, quantum learning, simulation, super-resolution, and real-lens finding.
Meet the cohort ↗ ResearchAn ICML workshop study compares learned memory retention with simpler policies under clean and noisy long-horizon experience.
Read the paper ↗Organization
Machine Learning for Science (ML4SCI) is the nonprofit open-source organization that supports DeepLense and related research programs across scientific disciplines.
Explore DeepLense