PID3Net: a deep learning approach for single-shot coherent X-ray diffraction imaging of dynamic phenomena

Abstract

This paper introduces a deep learning (DL)-based method for phase retrieval tailored to single-shot, multiple-frame coherent X-ray diffraction imaging (CXDI), designed specifically for visualizing local nanostructural dynamics within a larger sample. Current phase retrieval methods often struggle with achieving high spatiotemporal resolutions, handling dynamic imaging, and managing computational costs, which limits their applicability in observing nanostructural dynamics. This study addresses these gaps by developing a novel method that leverages a feedforward architecture with a physics-informed strategy utilizing measurement settings, enabling the reconstruction of dynamic “movies" from time-evolving diffraction images of the illuminated area. The method incorporates key enhancements, such as temporal convolution blocks to capture spatiotemporal correlations and a unified TV regularization applied to the reconstructed object, resulting in improved noise reduction and spatial smoothness. An expanded evaluation framework, including multiple metrics and systematic sensitivity analysis, is employed to comprehensively assess the method’s performance and robustness. Proof-of-concept experiments, including numerical simulations and imaging experiments of a moving Ta test chart and colloidal gold particles (dispersed in aqueous polyvinyl alcohol solutions) with synchrotron hard X-rays, validate the high imaging performance of this method. Experimental results demonstrate that structures in the sample have been successfully reconstructed at short exposure times, significantly outperforming both traditional methods and current DL-based methods. The proposed method provides efficient and reliable reconstruction of dynamic images with low computational costs, making it suitable for exploring fast-evolving phenomena in synchrotron- or free-electron laser-based applications requiring high spatiotemporal resolutions.