Physics is a useful prior.
Known symmetries and equations can reduce the search space.
Research record · 2019—2026
Machine learning,
under physical law.
We study how scientific structure—symmetry, geometry, equations, and data lineage—can make AI more efficient, interpretable, and useful for discovery.
Research thesis
Dark matter leaves structure in light.
A foreground galaxy bends the light of a more distant source into arcs and rings. Small structures inside the lensing galaxy’s dark-matter halo can perturb those images.
Our program develops methods to simulate, discover, enhance, and interpret these weak signals—while measuring the distance between synthetic benchmarks and observational evidence.
Simulation is not observation.
Domain shift must be treated as a scientific problem.
Image quality is not evidence.
Reconstruction must be tested on downstream physical tasks.
Efficiency changes the question.
Compact models and faster simulation unlock new experiments.
Publications
The complete collection.
20 distinct contributions. Local PDFs are provided from the group archive.
2026
2025
2024
2023
2020—2021
No publications match that search.
Read the evidence correctly
Benchmarks are waypoints,
not discoveries.
Most quantitative results in this collection use simulated or survey-like data. Near-perfect AUC on a controlled benchmark is not near-perfect dark-matter identification in the sky.
- “Euclid-like” and “HST-like” refer to simulation configurations.
- Super-resolution can generate plausible detail that was never recorded.
- Domain adaptation depends on a representative target set.
- The hybrid quantum result does not claim quantum advantage.