Exploring the use of Machine Learning for Bragg Coherent Diffraction Imaging: Gap Inpainting and Phase Retrieval

Abstract:

This doctoral thesis explores the application of machine learning algorithms to the analysis of Bragg coherent diffraction imaging (BCDI) patterns, focusing on two recurring challenges in this field. The first concerns the presence of missing intensity regions in experimental three-dimensional images, caused by gaps between detectors. These gaps introduce artefacts in the reconstructed real-space objects. The second challenge lies in the limited convergence and high computational cost of classical phase-retrieval algorithms, particularly for highly strained crystals. In such cases, iterative methods require long computation times and manual parameter tuning by experts, limiting the efficiency and reproducibility of the analyses.

Within this context, two convolutional neural network (CNN) architectures were designed, trained, and evaluated to address these problems. The first CNN tackles the issue of detector gaps by performing three-dimensional inpainting reconstruction on sub-volumes of the diffraction patterns. This sub-volume approach enables faster and more stable training while making the model more adaptable to different experimental conditions. The second CNN is designed to directly predict the complex phase lost during measurement, thus providing an estimate of the real-space object without relying on long iterative reconstructions.

Both models were validated on simulated and experimental datasets. The inpainting model significantly reduces artefacts caused by detector gaps, while the phase-predictive CNN enables faster and more reliable convergence for strongly strained particles. Furthermore, an additional section of the thesis presents an alternative phase-retrieval framework combining GPU-accelerated gradient descent with physics-based constraints. This hybrid approach outperforms classical iterative algorithms as well as deep-learning-assisted reconstructions, offering more accurate and efficient results, particularly for datasets that are difficult to invert.