AI and the World's Hard Problems · Astronomy & space · arXiv · 2026-10-26
Physics-Informed Neural Networks vs U-Net: Application to thermal dust emission recovery in the microwave
Separating Galactic thermal dust emission is essential for CMB experiments above 100 GHz. The authors build a physics-informed neural network (PINN) that embeds the modified black body spectral model as a hard, differentiable constraint, and compare it with a U-Net trained under the same pipeline on simulated Planck HFI patches. Both networks reach comparable reconstruction quality, with SSIM consistently above 0.96, and the PINN guarantees physically admissible outputs without reducing accuracy.