Student Posters Repository

Subject motion in MRI remains an unsolved problem; motion during image acquisition may cause artefacts that severely degrade image quality. In the clinic, if an image with motion artefacts is acquired, it will often be reacquired. This provides a source from which a large number of motion-degraded images, along with their respective re-scans, could be collected. These pairs of images could be used to train a neural network to identify the mapping relationship between an image with motion artefacts and a high quality, artefact free image. Inspired by previous work demonstrating MR image reconstruction with machine learning,[1,2] our objective is to train a neural network to perform motion corrected image reconstruction on image data with simulated motion artefacts. We simulate motion in previously acquired brain images and use the image pairs (corrupted + original) to train a deep neural network (DNN). The DNN was developed and trained using the TensorFlow library [4]. We show that the  images predicted by the DNN, from motion-corrupted k-space, have improved image quality compared to the motion-corrupted images. 

References: [1] Zhu, Bo. et al., Image reconstruction by domain transform manifold learning, 2017 [2] Hammernik K, et al., Learning a Variational Network for Reconstruction of Accelerated MRI Data, 2017 [3] Forstmann BU, et al., Multi-modal ultra-high resolution structural 7-Tesla MRI data repository. 2014 [4] Abadi M. et al., TensorFlow, 2015.

Free surface problems are of great interest since, for example, one may be interested in how water waves will interact and affect ships and offshore structures so that they can be designed optimally. These problems are modeled by systems of time dependent partial differential equations. At each time step, the free surface changes according to certain nonlinear boundary conditions. Hence, the problem includes determining the time dependent domain, which is computationally expensive and mathematically challenging. In this poster, we present the solution to the linear free surface problem for irrotational flows which is modeled by Laplace’s equation. In order to discretize the problem, we apply Finite Element Methods (FEMs) in which a piecewise polynomial approximation to the solution is sought over a discretized domain. In particular, we apply a hybridizable discontinuous Galerkin method which, in general, results into a smaller linear system compared to other discontinuous FEMs. The implementation is done in the C++ library MFEM which allows parallel computing. We show two numerical results, one where the analytical solution is known, and another one where we simulate waves in a water tank.

The Galactic Centre (GC) is a unique place to study the extreme dynamical processes occurring near a massive black hole (MBH). In this work I study the role of supernova (SN) explosions occurring in massive binary stellar systems lying in a discy structure within the innermost pc. In this poster I show the results of a suite of regularized 3-body simulations of binaries orbiting the central MBH, while the primary star undergoes a SN explosion. The simulations are evolved by means of a fully regularized N-body code that implements the Mikkola’s algorithmic regularization. This code is designed for studying the dynamical evolution of few-body systems in which strong gravitational encounters are frequent and the mass ratio between the interacting objects is large. This scheme removes the singularity of the two-body gravitational potential and allows extremely high integration accuracy. In my study, SN kicks scatter the lighter stars in the pair on completely new orbits, with higher eccentricity and inclination. In contrast, stellar black holes (BHs) and massive stars retain memory of the orbit of their progenitor star. My results suggest that SN kicks are not sufficient to eject BHs from the GC: all BHs that form in situ in the central parsec of our Galaxy probably remain in the GC.

The development of a cloud-based parallel framework is proposed for analysis and interpretation of multi-media data to make intellectual interaction between socially-assistive robot and a human. In recent years, the effects of Big Data Analytics have received special attention due to their potential use in robotics. A dynamic and parallel planning of a cloud-based framework has been proposed for the collective learning by robots. Human-robot interaction requires more human-like robots that have cognitive capacities about environmental interpretation in association with objects and the surroundings. Collective robot learning is a new form of distributed and parallel computing through sharing of information and cooperation between robots. Hierarchical Deep Visual Learning Architecture is a scalable and distributed framework for visual learning using a hierarchical deep learning approach. It consists of different layers, where each layer will be a learning strategy to effectively improve the insights. The different layers are namely, Visual Feature and Signature Learning, Visual Object Inferencing, Contextual Expansion, and Communication. These layers are based on an extension of neural network models such as Convolution Neural Network, Long Short-Term Memory Network. The final layer is combinations of models from Neural Network and Semantic Web acting as the knowledge base for the robot.

Domain‐specific language (DSL) platforms for programs are promising approaches to hide the complexities to develop HPC Application from end‐users and to achieve the portability of program codes. However, these platforms do not have the portability of themselves. These platforms must have peculiar program optimizers and specific runtimes on each HPC system, which were ported by brute force to achieve the portability. In this study, the author focused on this problem and developed a DSL platform with a portable program optimizer, which can recursively optimize programs for system hierarchies. The platform has a framework to construct an individual program optimizer for an HPC system by assembling optimization function for each hierarchical system component. The platform has two primary functions, the runtime library components for optimizer and the compiler of kernel codes and kernel‐interface definitions. The author implemented the runtime library components as the aspects of Aspect‐Oriented Programming.

In this poster, time-domain mixed finite element simulations for Maxwells equations in bounded three-
dimensional domains are presented. The electric and magnetic fields are discretized with Nédeléc and Raviart
Thomas finite elements in space. Symplectic and Backward Euler methods are employed for temporal discretization. The obtained fields are also visualized on 3D meshes. The proposed methods are accurate both in space and
time up to order 4, and parallel in space. In case of symplectic time integration they are energy conserving.

Due to high computational costs, stellar evolution modeling is typically done in the 1D approximation where spherical symmetry is assumed. However, some hydrodynamical phenomena are intrinsically multidimensional, e.g. mixing of chemical elements due to convection. In common 1D evolution codes, these are parametrized as diffusion processes with certain diffusion coefficients. The exact value of these coefficients typically suffer from high uncertainties. To improve this, we conduct 2D and 3D hydro-simulations of shells inside a star by solving the compressible Euler-equations using a finite-Volume approach. From these simulations, we derive quantities like the amount of mixing or how far convective plumes penetrate into adjacent convectively stable regions. These results can help to improve the current treatment in 1D stellar evolution codes. Here, we show results for rotationally induced shear instabilities and convection in a He-burning shell obtained with our "Seven League Hydro (SLH)" code.

In order to study the dynamics of proteins and other biomolecules, we use Molecular Dynamics simulations, i.e. solving the equations of motion for systems in atomic detail with starting configurations obtained from experiments. In particular, we are interested in the gradients of free energies, as they are the universal driving forces of chemical and biophysical systems and related to fundamental processes such as binding or protein folding. Due to entropic contributions, full values of free energies can rarely be obtained, as this would require sampling the entire high dimensional conformational space. However, the difference in free energy can be determined more accurately if the phase spaces of the two states have sufficient overlap. To fulfill this requirement, we use a set of methods named Alchemical Transformations, in which simulations in intermediate states, consisting of an interpolation of the start and end states’ Hamiltonians, are conducted. We show that the choice of such pathways significantly influences the statistical accuracy of the estimates. The optimized pathways are implemented into the widely used “GROningen MAchine for Chemical Simulations” (GROMACS) open-source software package.