| Summary: | Modern electron microscopes are fitted with ever larger charge-coupled device (CCD) cameras capable of faster acquisition rates which in turn drives a concomitant increase in the bandwidth of data that is being collected and the amount of information in our datasets. At the same time, current increases in computational performance are largely being delivered through the addition of parallel execution units rather than explicit increases in the speed of single processors, this means techniques that cannot exploit their inherent parallelism are seeing little performance benefit from the generational improvements in computer processors. Many techniques used in electron microscopy to process these large datasets have not been adapted to utilise the modern methods available for parallel data processing which can lead to lengthy offline data processing techniques which could otherwise be performed in near real-time. Reimagining these methods to suit highly parallel computational architectures such as graphics processing units (GPUs) can offer improved performance orders of magnitude higher than their central processing unit (CPU) counterparts. In this thesis I have looked specifically at the case of transmission electron microscopy (TEM) image simulation via the multislice procedure, and exit wave reconstruction (EWR), which can both potentially see huge benefits by adapting these algorithms to exploit their parallelism. Software has been developed for performing multislice simulations using GPU computation where the increase in computational power also allows for modifications to be made which can increase the accuracy of the simulations at the expense of simulation time. The multislice software developed here has no minimum slice thickness limitations and the slice thickness no longer has to be coupled to the structure being simulated to ensure accuracy. The CCD detector characteristics and electron dose have also been incorporated within the simulation process. The use of GPUs has allowed these simulations to be performed in vastly less time than CPUs based equivalent simulations. Software has also been developed for performing EWR on either multicore CPUs or GPUs which lowers the time required to perform EWR sufficiently that real-time reconstruction at typical CCD frame-rates is a distinct possibility. This EWR software additionally features mutual information (MI) based image alignment which can handle accurate image alignment in cases where other methods are prone to failure. These software are used to aid in the investigation of fluorinated graphene conformation via multislice simulation and EWR, and in the study of self-assembled block co-polymer assemblies also by EWR.
|
|---|
