5.8 KiB
| title | chunk | source | category | tags | date_saved | instance |
|---|---|---|---|---|---|---|
| Coherent diffraction imaging | 2/4 | https://en.wikipedia.org/wiki/Coherent_diffraction_imaging | reference | science, encyclopedia | 2026-05-05T10:03:59.662514+00:00 | kb-cron |
== Reconstruction == In a typical reconstruction the first step is to generate random phases and combine them with the amplitude information from the reciprocal space pattern. Then a Fourier transform is applied back and forth to move between real space and reciprocal space with the modulus squared of the diffracted wave field set equal to the measured diffraction intensities in each cycle. By applying various constraints in real and reciprocal space the pattern evolves into an image after enough iterations of the HIO process. To ensure reproducibility the process is typically repeated with new sets of random phases with each run having typically hundreds to thousands of cycles. The constraints imposed in real and reciprocal space typically depend on the experimental setup and the sample to be imaged. The real space constraint is to restrict the imaged object to a confined region called the "support". For example, the object to be imaged can be initially assumed to reside in a region no larger than roughly the beam size. In some cases this constraint may be more restrictive, such as in a periodic support region for a uniformly spaced array of quantum dots. Other researchers have investigated imaging extended objects, that is, objects that are larger than the beam size, by applying other constraints. In most cases the support constraint imposed is a priori in that it is modified by the researcher based on the evolving image. In theory this is not necessarily required and algorithms have been developed which impose an evolving support based on the image alone using an auto-correlation function. This eliminates the need for a secondary image (support) thus making the reconstruction autonomic. The diffraction pattern of a perfect crystal is symmetric so the inverse Fourier transform of that pattern is entirely real valued. The introduction of defects in the crystal leads to an asymmetric diffraction pattern with a complex valued inverse Fourier transform. It has been shown that the crystal density can be represented as a complex function where its magnitude is electron density and its phase is the "projection of the local deformations of the crystal lattice onto the reciprocal lattice vector Q of the Bragg peak about which the diffraction is measured". Therefore, it is possible to image the strain fields associated with crystal defects in 3D using CDI and it has been reported in one case. Unfortunately, the imaging of complex-valued functions (which for brevity represents the strained field in crystals) is accompanied by complementary problems namely, the uniqueness of the solutions, stagnation of the algorithm etc. However, recent developments that overcame these problems (particularly for patterned structures) were addressed. On the other hand, if the diffraction geometry is insensitive to strain, such as in GISAXS, the electron density will be real valued and positive. This provides another constraint for the HIO process, thus increasing the efficiency of the algorithm and the amount of information that can be extracted from the diffraction pattern.
== Algorithms == One of the most important aspects of coherent diffraction imaging is the algorithm that recovers the phase from Fourier magnitudes and reconstructs the image. Several algorithms exist for this purpose, though they each follow a similar format of iterating between the real and reciprocal space of the object (Pham 2020). Furthermore, a support region is frequently defined to separate the object from its surrounding zero-density region (Pham 2020). As mentioned earlier, Fienup developed the initial algorithms of Error Reduction (ER) and Hybrid Input-Output (HIO) which both utilized a support constraint for real space and Fourier magnitudes as a constraint in reciprocal space (Fienup 1978). The ER algorithm sets both the zero-density region and the negative densities inside the support to zero for each iteration (Fienup 1978). The HIO algorithm relaxes the conditions of ER by gradually reducing the negative densities of the support to zero with each iteration (Fienup 1978). While HIO allowed for the reconstruction of an image from a noise-free diffraction pattern, it struggled to recover the phase in actual experiments where the Fourier magnitudes were corrupted by noise. This led to further development of algorithms that could better handle noise in image reconstruction. In 2010, a new algorithm called oversampling smoothness (OSS) was created to use a smoothness constraint on the imaged object. OSS would utilize Gaussian filters to apply a smoothness constraint to the zero-density region which was found to increase robustness to noise and reduce oscillations in reconstruction (Rodriguez 2013).
== Generalized Proximal Imaging (GPS) == Building upon the success of OSS, a new algorithm called generalized proximal smoothness (GPS) has been developed. GPS addresses noise in the real and reciprocal space by incorporating principles of Moreau-Yosida regularization, which is a method of turning a convex function into a smooth convex function (Moreau 1965) (Yosida 1964). The magnitude constraint is relaxed into a least-fidelity squares term as a means of lessening the noise in the reciprocal space (Pham 2020). Overall, GPS was found to perform better than OSS and HIO in consistency, convergence speed, and robustness to noise. Using R-factor (relative error) as a measurement for effectiveness, GPS was found to have a lower R-factor in both real and reciprocal spaces (Pham 2020). Moreover, it took fewer iterations for GPS to converge towards a lower R-factor when compared to OSS and HIO in both spaces (Pham 2020).
== Coherence ==