Self-supervised restoration of weak signal from noisy scattering and diffraction data in reciprocal space

Reciprocal-space imaging and measurement underpin a wide range of modern characterization techniques in scientific research. Nevertheless, radiation damage and limited photon budgets constrain the high-quality and efficient acquisition of noise-vulnerable weak signals in high-dynamic-range (HDR) reciprocal-space. To overcome the inherent unavailability of noise-free ground truth in most scientific measurement scenarios, here we introduce a self-supervised denoising network, SeDeNet. By leveraging a self-supervised denoising framework incorporating nonlinear correction, SeDeNet enables high-fidelity restoration of weak signal features in HDR scenarios using only noisy data. Extensive experiments on X-ray scattering and diffraction data from diverse samples validate the superior denoising performance and generalization of SeDeNet. Moreover, we demonstrate that SeDeNet achieves approximately 20-fold reduction in acquisition time while preserving ≥ 95% accuracy in one-dimensional grating sample’s parameter extraction. These results highlight a promising approach for the efficient acquisition of low-noise HDR scattering data, while maintaining physical consistency in structural parameter analysis.