Deep-learning-powered desmearing for small-angle scattering

Smearing effects in small-angle scattering (SAS) measurements significantly compromise data analysis, arising from the convolution of theoretical scattering curves with the point spread function of the measurement system. This paper presents a deep-learning-based desmearing network (DSNet) designed to effectively mitigate smearing effects in SAS data. By integrating the processes underlying scattering data smearing, DSNet necessitates only a limited simulation dataset for pre-training. Both simulation and experimental results have demonstrated that DSNet exhibits robust noise resilience and exceptional generalization performance across diverse sample types, and achieves superior desmearing capabilities compared with the classical Lake method and Wiener filter.