Open source · Six connected systems
DeepLense builds the infrastructure around the research: simulation, survey-data preparation, machine learning, and auditable scientific agents.
The software stack
Simulation and survey observations become model-ready evidence. The core repository supports the experiments; typed agents make the workflow programmable and auditable.
A typical path through the stack—not a requirement. Every repository can be used and extended independently.
Repository index
Each project has a bounded role. Start with the scientific question, then enter the stack where your data or experiment begins.
The central research repository for datasets, trained approaches, benchmarks, and contributor projects studying dark-matter structure through strong gravitational lensing.
Researchers and contributors reproducing results or extending the model zoo.
The foundational simulator behind the early DeepLense datasets, generating controlled strong-lensing images with lenstronomy, pyHalo, and Colossus.
Controlled synthetic benchmarks and reproduction of the group’s foundational experiments.
A next-generation simulation pipeline that combines real astronomical source images from GalaxiesML with physically sampled galaxy and halo parameters.
More realistic simulation-to-survey studies where source morphology and parameter priors matter.
The Rubin Image Preparation and Processing Lensing Engine connects Vera C. Rubin Observatory LSST data to DeepLense-ready inputs and model predictions.
Retrieving, normalizing, and analyzing strong-lens candidates in survey-scale observations.
A provenance-first agentic pipeline that separates Rubin data retrieval, deterministic preprocessing, model qualification, and scientific decisions behind typed interfaces.
Auditable automation across survey preparation, model integration, and bounded simulation or training workflows.
The companion framework for DLens: typed Pydantic AI agents coordinate simulation, model search, evaluation, and experiment planning in auditable loops.
Closed-loop scientific machine-learning experiments that remain inspectable and reproducible.
Read the spotlight paper →Code, methods, and research lineage stay visible to the people building on them.
The software exposes scientific structure instead of hiding it behind generic abstractions.
Tools are judged by reproducibility, downstream utility, and honest limits.
Build with us