Research

Physics-Informed Deep Learning for the Information-Limited Universe

Next-generation cosmological surveys that will elucidate dark energy and dark matter—Rubin/LSST, Euclid, and Roman—are fundamentally information-limited experiments. Their statistical power will not be constrained by photon counts alone, but by our ability to extract cosmological information from complex, blended, noisy pixel data while controlling systematic bias at unprecedented accuracy and precision.

We develop a new paradigm for scientific inference in this regime: physics-informed, symmetry-constrained deep learning architectures that operate directly at the pixel level and produce calibrated probabilistic posteriors suitable for precision cosmology.

Rather than applying machine learning as a performance tool, we embed physical structure directly into neural architectures. We enforce discrete and continuous symmetries, incorporate analytic response operators, and produce full posterior distributions through likelihood-aware neural models. This transforms deep learning from a “black-box” predictor into a structured inference engine consistent with physical law.

Our approach reframes the role of AI in cosmology:

  • From catalog-level regression to pixel-level inference

  • From simulation-calibrated correction to analytic response modeling

  • From point estimates to information-preserving posterior distributions

  • From heuristic architectures to symmetry-aware representation learning

Our DeepDISC framework unifies detection, deblending, classification, and photometric redshift inference within a single pixel-based probabilistic architecture. We extend this philosophy to weak gravitational lensing cosmic shear inference, targeting sub-permil level accuracy through symmetry-enforced networks with embedded analytic calibration.

The long-term vision is an end-to-end, survey-scale Astro-AI stack that preserves cosmological information while satisfying next-generation systematics requirements. Our work establishes a new intellectual foundation for scientific machine learning in the era of petascale astronomical surveys—one in which symmetry, operator theory, and probabilistic inference are not add-ons, but architectural principles.