IPE 12-2022 Master Thesis or Internship: Leveraging parallel hardware for advanced tomography and laminography

Computed Tomography and Laminography are indispensable tools for studying inner structures in materials and biological spices. New instruments relaying on different types of energy-matter interaction principles, technological advances in instrumentation and detection technologies open a way for increasing number of new applications. Many of these applications, however, pose a challenge for a high quality reconstruction due to low signal-to-noise ratio, limited or non-standard acquisition trajectory, etc. Traditional reconstruction algorithms are ill suited for the low-dose, limited angle, or spectral tomography. Instead regularized iterative reconstruction techniques are recommended to improve reconstruction quality through incorporation of available prior information. While a proof-of-concept has been already established in mathematical community, there are significant challenges in adapting them for a wider use in labs and research facilities. One of the main aspects is extremely high requirements in terms of memory consumption and computing power.

There is an ongoing international effort to build a modular reconstruction framework providing an extensive library of regularization techniques and also practically viable for cutting-edge scientific and industrial applications where conventional algorithms fail. You will join the project and will contribute your numerical and computing skills to optimize the performance of reconstruction and regularization pipelines. The project is performed as a collaboration between the University of Manchester (UK), STFC (UK), DTU (Denmark) and KIT.

We can offer multiple tasks widely varying in terms of complexity and focus. The available options range from simple integration and benchmarking tasks (offered as internship only) to a more complex projects like careful optimization of selected regularization algorithms for recent hardware architectures. The tomographic reconstruction is a data-intensive task and the performance is heavily dependent on efficient usage of processor caches. We are interested in both fine-tuning of traditional multi-threaded CPU codes and also in leveraging the massively parallel architectures like GPUs and FPGAs. To reduce memory consumption, mixed-precision computations are envisaged. The actual task within project will be offered during the interview depending on applicant skills and interests.