Computer Learns to See Invisible Dark Matter ✛ ✦

Original: "Joint inference of weak lensing convergence map and cosmology with diffusion models"

· Benjamin Remy, Chihway Chang, Rebecca Willett

arXiv:2606.31988 · 2026-06-30 · CC BY 4.0 · 1 min

The JADE program uses distorted galaxies to map dark matter in seconds and determine the properties of the Universe.

Galaxies in telescopes look slightly lopsided — not due to optical flaws, but because their light has been bent by the pull of invisible dark matter. This gravitational ripple is like patterns on water betraying a hidden boulder at the bottom. The JADE program (machine learning) trains on millions of computer simulations of galaxies and in a fraction of a second turns a real image into a map of hidden mass. At the same time, it calculates the fraction of mysterious dark energy that accelerates the expansion of the cosmos, and works even for the most distant objects with noticeable reddening of light.

Where classical analysis takes days, JADE handles it in 0.2 seconds — without losing accuracy.

Light bending by mass was predicted by Einstein, Zwicky first suggested hidden matter, and Rubin proved that galaxies would fly apart without it. Now we don't just believe in the invisible: we see it in real time.

Whole graph →

JADE was trained on just 100,000 simulations: that's like a handful of grains of sand compared to the millions of calculations needed by old methods. It creates each map in 0.2 seconds.

In Peter Watts' novel 'Blindsight,' invisible alien ships give themselves away only by distorting starlight — exactly like dark matter.

\kappa(\boldsymbol{\vartheta}) = \frac{3 H_0^2 \Omega_m}{2c^2} \int_0^{\chi_{\lim}} d\chi \, \frac{q(\chi)}{a(\chi)} f_K(\chi) \delta(f_K(\chi) \boldsymbol{\vartheta}, \chi)

Equation (1) from the paper: convergence as a weighted integral of the density contrast along the line of sight.

\tilde{\kappa}_E + i\tilde{\kappa}_B = \left( \frac{k_1^2 - k_2^2}{k^2} + i \frac{2k_1 k_2}{k^2} \right) (\tilde{\gamma}_1 + i\tilde{\gamma}_2)

Equation (3): relation between shear and convergence in Fourier space under the flat-sky approximation.

Scientists

Fritz Zwicky Vera Rubin Roberto D. Peccei Helen Quinn R. Brent Tully Bohdan Paczyński

Concepts

weak lensing posterior distribution deep learning in medical imaging Weakly Supervised Learning gravitational lensing modeling medical image to genomic mapping least squares single pixel imaging confusion matrix Diffusion Models integrated spectrometer symbolic regression neural network microlensing reconstruction convergence composite analysis nested sampling diffusion mri jackknife resampling rejection sampling autoregressive model

I am the author

Computer Learns to See Invisible Dark Matter ✛ ✦