Can AI reconstruct dark matter from cosmic shadows?

Reconstructing dark matter's 3D structure from weak gravitational lensing—the bending of light by invisible matter—is nearly impossible with single-viewpoint observations and uncertain galaxy distances. Researchers built a diffusion-model prior trained on high-resolution cosmological simulations, then combined it with weak-lensing physics to infer 3D dark matter maps. On realistic survey data, the method substantially outperforms existing approaches and produces posterior samples whose statistics match the underlying simulations, even under modest cosmological shifts.