This paper proposes a 3D machine learning (ML) model, combining supervised and transfer learning, for phase retrieval in coherent X-ray diffraction imaging (CDI). The model significantly improves the accuracy of 3D morphological information recovery from experimental Bragg CDI data compared to traditional iterative methods. The model can provide rapid, accurate results beneficial for real-time experiments and can be used without prior training, leveraging a self-defined loss function for robust phase retrieval.

Longlong Wu, Shinjae Yoo, Ana F. Suzana, Tadesse A. Assefa, Jiecheng Diao, Ross J. Harder, Wonsuk Cha, Ian K. Robinson