Light from a distant galaxy travels for billions of years, and on its way its path is slightly bent: this is how the invisible mass called dark matter reveals itself. This bending slightly stretches or compresses the galaxy's image, like a reflection in wavy glass. Catching such tiny shape changes is no easy task, since galaxies themselves can be elongated, spiral, or completely irregular, and the telescope adds blur and noise.
To separate the true signal from random distortions, scientists built a neural network that knows an important rule from the start: if the image is rotated by 90 degrees or mirrored, the output should not change. Such symmetry helps the network avoid inventing extra features. The network was then carefully calibrated using its own computations to ensure accurate answers. In tests with single galaxies similar to those the future LSST survey will see, the new approach produced 10% less noise at low signal, equivalent to increasing the number of observed galaxies by 20%. Meanwhile, the systematic error was below 0.2%—a strict threshold set for future observations.
This approach is a step toward making dark matter distribution maps more reliable. After all, it is through weak distortions of galaxy shapes that astronomers map the invisible mass of the universe. The more precise the measurements, the clearer how this mass gathers into filaments and clumps, holding galaxies together. And so, we will better understand what our world is actually made of.
Weak gravitational lensing distorts a galaxy's shape by only 1% on average, as if a circle became an oval with a difference of one hundredth of its diameter.
In the novel "Darkness Between the Stars," dark matter maps help heroes chart a course through invisible gravitational currents — precise measurements of galaxy shapes are vital there.