An array of radio telescopes is like a camera missing several lenses. Ever since Karl Jansky picked up the radio hiss of the Milky Way, astronomers have been trying to piece together a whole picture from fragmentary signals. But the antennas are spaced apart, and the mathematically “developed” image — the dirty map — ends up smeared. Small star clusters blend together. Faint details drown in noise.
PhySR takes a different approach. It’s like an optician who knows exactly which lenses are missing and reconstructs the picture from a blurry cast. It’s a deep learning neural network with a distortion model built inside: how the antenna beam responds to a signal and how the point spread function smears it. This physics “hint” lets the network learn to remove blur and boost resolution — no perfect training image needed.
On simulated data from the upcoming SKA-Mid telescope, the method produced excellent results. With a fourfold sharpness boost, the quality metrics PSNR (peak signal-to-noise ratio) and SSIM (structural similarity index) improved by roughly 13 dB and 0,38 compared to traditional methods. The spread of errors shrank. Crucially, PhySR is physically consistent: the restored image agrees with the original measurements. Now astronomers can discern brightness variations in the centers of galaxies or delicate gas filaments without waiting for new antennas to be built. The old “camera” with incomplete lenses gains perfect focus — thanks to knowing the physics.
To snapshot a faraway radio source, telescopes sync their signals to within billionths of a second — the tiniest timing slip would wreck the whole image.