Dolphin

In strong-lensing systems, the background lensed galaxies are often star-forming spirals with complex morphologies. Reconstructing that light during lens modelling is numerically hard: current procedures impose smoothness regularization on the reconstructed light, but residual artefacts remain, and the results can look unrealistic. A more informative prior on the light distribution would help — but a spiral galaxy’s intricate structure cannot be captured by a conventional analytic function whose parameters can be constrained.

Dolphin solves this by developing a generative neural network — for example a normalizing flow or a diffusion model — that maps a few tens of input parameters to a realistic spiral-galaxy light distribution. This learned model then acts as a function with a built-in empirical prior during lens modelling. It is trained on galaxy image libraries from the Hubble Space Telescope and JWST.

As a deliverable, the trained model is integrated into the widely used lens-modelling package lenstronomy, making the method readily usable.