Vision transformer for multi-domain phase retrieval in coherent diffraction imaging

Area of Science:

• Materials Science
• Crystallography
• Computational Imaging

Background:

• Bragg coherent diffraction imaging (BCDI) phase retrieval is challenging for crystals with significant distortions (strong-phase regime).
• The phase domain problem, involving block displacements and sharp domain walls, leads to split Bragg peaks and complex fringe patterns.
• Classical iterative solvers often struggle with convergence and solution stability in these complex scenarios.

Purpose of the Study:

• To develop a novel, unsupervised method for solving the multi-domain phase retrieval problem in the strong-phase regime of BCDI.
• To address limitations of traditional solvers in handling complex crystal structures and distortions.

Main Methods:

• Introduction of an unsupervised Vision Transformer (ViT) model incorporating Fourier attention.
• Utilizing token mixing and multiscale Fourier attention to couple global reciprocal-space information.
• Directly solving phase retrieval from measured 2D Bragg diffraction intensities.

Main Results:

• Validated on synthetic Voronoi multi-domain crystals with strong-phase contrast and noise.
• Tested on experimental data from SrTiO3, Fe3O4 (during phase transition), and La0.5Ca0.5MnO3 nanocrystals.
• Achieved reciprocal-space mismatch (χ²) comparable to or better than baseline methods.
• Demonstrated preservation of domain-resolved phase reconstructions with increasing domain numbers on synthetic data.

Conclusions:

• The Fourier ViT method offers a robust solution for strong-phase BCDI phase retrieval, particularly for multi-domain materials.
• The approach shows competitive performance against iterative methods and CNNs, especially under challenging conditions.
• This AI-driven technique advances the capability to image complex crystalline structures with high fidelity.