Wu Longlong, Juhas Pavol, Yoo Shinjae, Robinson Ian
Computational Science Initiative, Brookhaven National Laboratory, Upton, NY 11973, USA.
Condensed Matter Physics and Materials Science Department, Brookhaven National Laboratory, Upton, NY 11973, USA.
IUCrJ. 2021 Jan 1;8(Pt 1):12-21. doi: 10.1107/S2052252520013780.
The reconstruction of a single-particle image from the modulus of its Fourier transform, by phase-retrieval methods, has been extensively applied in X-ray structural science. Particularly for strong-phase objects, such as the phase domains found inside crystals by Bragg coherent diffraction imaging (BCDI), conventional iteration methods are time consuming and sensitive to their initial guess because of their iterative nature. Here, a deep-neural-network model is presented which gives a fast and accurate estimate of the complex single-particle image in the form of a universal approximator learned from synthetic data. A way to combine the deep-neural-network model with conventional iterative methods is then presented to refine the accuracy of the reconstructed results from the proposed deep-neural-network model. Improved convergence is also demonstrated with experimental BCDI data.
通过相位恢复方法从其傅里叶变换模量重建单粒子图像，已在X射线结构科学中得到广泛应用。特别是对于强相位物体，如通过布拉格相干衍射成像（BCDI）在晶体内部发现的相域，传统的迭代方法由于其迭代性质，既耗时又对初始猜测敏感。本文提出了一种深度神经网络模型，该模型以从合成数据中学习到的通用逼近器的形式，对复杂的单粒子图像进行快速准确的估计。然后提出了一种将深度神经网络模型与传统迭代方法相结合的方法，以提高所提出的深度神经网络模型重建结果的准确性。实验BCDI数据也证明了收敛性的提高。
