AI for Science
Learning from the universe.
- Astrophysics
- Strong Lensing
- Scientific Simulation
- PDEs
- Autonomous Discovery
Open-source research · ML4SCI
Building intelligent systems for scientific discovery.
DeepLense explores fundamental problems in artificial intelligence and applies them to one of science’s hardest questions: what is dark matter made of?
01
A cosmic lens. A foreground galaxy bends light from a distant source. Small dark-matter structures leave subtle signatures in the resulting arcs.
What is DeepLense?
We make the invisible measurable.
DeepLense is an open-source research group developing physics-aware and data-efficient machine-learning methods to study dark matter through strong gravitational lensing.
Our work spans the full scientific pipeline—from realistic simulation and lens discovery to representation learning, super-resolution, inference, and auditable scientific agents.
How the group worksResearch
One program, two frontiers.
We use the constraints of physical science to build better AI—and use advances in AI to ask better scientific questions.
Core AI
Building systems that reason.
- Agents
- Foundation Models
- Generative Models
- Representation Learning
- Interpretability & Evaluation
The program
From photons to foundations.
Each layer addresses a different bottleneck between simulated universes and scientific evidence.
Simulate
Generate controlled lensing systems across candidate dark-matter models.
Find
Discover rare lenses in noisy, survey-scale observations.
Learn
Build representations shaped by symmetry, geometry, and physical law.
Infer
Connect weak image morphology to testable physical hypotheses.
Selected work
Recent research signals.
FlowLensing
A controllable flow-matching model for strong-lensing simulation, designed for speed and scientific conditioning.
200×+reported speedup over intensive traditional simulators
Lens-LeJEPA
Task-aware self-supervision that encodes exact lens symmetries—and documents when those invariances become harmful.
+11.6accuracy points with 100 labels per class
DLens
Closed-loop agents orchestrate simulation, architecture search, and evaluation through typed, auditable interfaces.
8.4×fewer parameters than the comparison ResNet-34
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.
“The right representation depends on the scientific question.”
Symmetry can make classification more efficient, yet erase detail needed for reconstruction. Sharp images can look convincing while adding no useful physical information. Our work treats these tensions as research results—not footnotes.
News & field notes
Latest from the group.
Community
Seven DeepLense projects join Google Summer of Code 2026.
Work spans scientific agents, foundation models, quantum learning, fast simulation, super-resolution, and real-lens finding.
Meet the cohort ↗ ResearchSelective memory for long-horizon agents.
An ICML workshop study asks when learned retention beats simple memory policies under noisy experience.
Read the paper ↗ EventDeepLearn Hackathon
Open scientific machine-learning challenges become hands-on, collaborative experiments.
Open the challenge repo ↗Our ecosystem
Open science works in community.
DeepLense is a project of Machine Learning for Science (ML4SCI), the open-source umbrella organization connecting researchers and contributors across scientific disciplines.
Build in the open