On problems and computational imaging solutions for ptychography

Ptychography is a modern microscopy technique still in development within the broader family of Coherent Diffraction Imaging (CDI) based methods. They have the potential to revolutionize the study of multiple scientific fields by providing higher resolutions of larger areas at faster speeds including chemical speciation. Its principle is that a beam of coherent light is shined onto a sample in a scanning fashion and the acquired diffraction patterns are computationally merged to recreate the whole specimen's complex image. This research is multidisciplinary; with pivotal point in electronic/computer engineering it solves problems in this modern computational imaging technique, which is used in applied physics, with applications in fields such as electrochemistry, biology, nanomaterials and other disciplines that benefit from microscopy. Based on recent advances in Artificial Intelligence, this research makes use of Machine Learning and Optimisation methods for solving certain problems. The results of this research provided solutions to various problems of ptychography, improving the technique. Moreover, the thesis delivers an implementation of the proposed solutions as a complete software framework based on modern engineering paradigms such as GPU computing. The development of this work is based on an actual laboratory that implements ptychography. This laboratory is the X-ray spectromicroscopy beamline TwinMic at the synchrotron radiation facility Elettra Sincrotrone Trieste, in collaboration with its Scientific Computing group. The thesis also contains the result of two beamtime experiments and discusses the proposals of an additional 2 that are already granted. The main contributions of this research range over the topic of spatial coherence, positions refinement, and parameters tuning. Partial coherence of the source deteriorates the images that can be reconstructed: we proposed a refined reconstruction algorithm (M-RPIE) where the large computational field of view of the method allows for a sparser scanning. Position errors impact negatively ptychography, thus a part of the research was devoted to providing a solution. This led to: a metric based approach; an analysis of the dynamics of the error signal; a method for the automatic control of the position feedback gain; a method to include position refinement coefficients within an optimisation process Ptychography requires a multitude of parameters which currently are manually tuned. Using advanced Deep Learning techniques, we proposed a reconstruction algorithm which automatically regresses the propagation distance and the position correction coefficients within an optimisation-based process. Fourier ptychography was also explored and implemented in the software framework. We proposed a CNN model for the generation of a prior which can be effectively used to seed the reconstruction. Since the research is multidisciplinary, certain results were utilised in other fields such as X-ray Fluorescence, Computed Tomography , super-resolution for forensics, CNNs for face recognition and depth estimation.

Ptychography is a modern microscopy technique still in development within the broader family of Coherent Diffraction Imaging (CDI) based methods. They have the potential to revolutionize the study of multiple scientific fields by providing higher resolutions of larger areas at faster speeds including chemical speciation. Its principle is that a beam of coherent light is shined onto a sample in a scanning fashion and the acquired diffraction patterns are computationally merged to recreate the whole specimen's complex image. This research is multidisciplinary; with pivotal point in electronic/computer engineering it solves problems in this modern computational imaging technique, which is used in applied physics, with applications in fields such as electrochemistry, biology, nanomaterials and other disciplines that benefit from microscopy. Based on recent advances in Artificial Intelligence, this research makes use of Machine Learning and Optimisation methods for solving certain problems. The results of this research provided solutions to various problems of ptychography, improving the technique. Moreover, the thesis delivers an implementation of the proposed solutions as a complete software framework based on modern engineering paradigms such as GPU computing. The development of this work is based on an actual laboratory that implements ptychography. This laboratory is the X-ray spectromicroscopy beamline TwinMic at the synchrotron radiation facility Elettra Sincrotrone Trieste, in collaboration with its Scientific Computing group. The thesis also contains the result of two beamtime experiments and discusses the proposals of an additional 2 that are already granted. The main contributions of this research range over the topic of spatial coherence, positions refinement, and parameters tuning. Partial coherence of the source deteriorates the images that can be reconstructed: we proposed a refined reconstruction algorithm (M-RPIE) where the large computational field of view of the method allows for a sparser scanning. Position errors impact negatively ptychography, thus a part of the research was devoted to providing a solution. This led to: a metric based approach; an analysis of the dynamics of the error signal; a method for the automatic control of the position feedback gain; a method to include position refinement coefficients within an optimisation process Ptychography requires a multitude of parameters which currently are manually tuned. Using advanced Deep Learning techniques, we proposed a reconstruction algorithm which automatically regresses the propagation distance and the position correction coefficients within an optimisation-based process. Fourier ptychography was also explored and implemented in the software framework. We proposed a CNN model for the generation of a prior which can be effectively used to seed the reconstruction. Since the research is multidisciplinary, certain results where utilised in other fields such as X-ray Fluorescence, Computed Tomography , super-resolution for forensics, CNNs for face recognition and depth estimation.