Student Posters Repository

Accreting black hole systems such as X-ray binaries (black holes with a close stellar companion that provide the black hole its fuel) and active galactic nuclei (supermassive, accreting black holes at the center of galaxies) exhibit variability in their luminosity on many timescales ranging from milliseconds to tens of days, and even hundreds of days. The mechanism(s) driving this variability is poorly understood. We seek to employ numerical techniques to study accretion disks including more complicated physics traditionally ignored in the fluid equations describing accretion disks in order to more accurately understand their behavior over time, including variability, which cannot be done analytically. I present a proof-of-concept 3D global simulation using the grid-based, hydrodynamic code PLUTO, in which both serial and massively parallel computations are possible, of a simplified thin disk model about a central black hole. I also generate a simplistic synthetic light curve that displays the variability in luminosity of the simulation over time, a critical step in comparing to observational data. The ultimate goal will be to include radiation and evolve the simulation over a period of many days (a challenge for hydrodynamic codes) in order to study instabilities that arise due to a central, radiating source and to create a more sophisticated simulated light curve for comparison to observations.

In the low-redshift universe, the majority of neutral hydrogen (HI) is found inside galaxies and is thought of as a biased tracer of the underlying dark matter distribution of the Universe. Therefore, measurement of the large-scale fluctuations of the 21cm-wavelength radiation emitted by the ground state spin flip transition of HI is a tool with which to trace galactic evolution and to do precision cosmology. We present a detection of cross-correlation between HI measured by the Parkes 21cm Intensity Mapping Survey and 2dF Galaxy Redshift Survey galaxies. The cross-power spectrum is detected at a significance of 11.9σ and exhibits lower power on small scales relative to the expected dark matter power spectrum. This decrement hints at lack of HI clustering or at a small correlation coefficient between HI and optical galaxies, on small scales.

PyGBe is a Python library that applies the boundary integral method for biomolecular electrostatics and simple nanoparticle plasmonics. It uses a continuum model for biomolecular electrostatics and handles localized surface plasmon effects quasi-statically.

PyGBe is efficient in time-to-solution and offers new problem-solving options to scientists in molecular biology, biochemistry and applied physics. It handles solvent-filled cavities and Stern layers; computes protein-surface electrostatic interactions; probes protein orientation near charged nanosurfaces; and obtains localized surface plasmon resonance with an eye towards nano-scale biosensor calculations.

PyGBe uses a boundary element method (BEM) that formulates the matrix-vector product as an N-body problem. This formulation has direct O(N^2) complexity, but PyGBe applies a Barnes-Hut algorithm to accelerate each iteration of a GMRES solver to O(N logN), for N unknowns. It exploits NVIDIA GPU hardware on the most computationally intensive parts of the code using CUDA kernels in the treecode, interfacing with PyCUDA. Some parts of the code are written in C++, wrapped using SWIG.

A new synchronization mechanism called Spatial Lock for parallel
geometric algorithms is presented. We demonstrate that Spatial 
Locks can ensure thread synchronization in geometric applications 
that perform parallel operations over objects in 2D or 3D space. 
A parallel algorithm for mesh simplification was implemented using 
Spatial Locks to show its usefulness when parallelizing geometric 
application with ease. Experimental results illustrate the advantage 
of using this synchronization mechanism.

Recent super computers are capable of numerical simulations that generates vast amounts of data (sizes in petabytes). Extracting relevant information from that data requires sophisticated analysis and visualisation methods. Current workflow for data analysis is often done after the simulation is performed; during the simulation raw data is written to main storage, and later time, this data is read into dedicated systems for analysis. This traditional workflow has some disadvantages, such as the strong rely on the file system, the inability to observe the evolution for a large simulation, and the modification of parameters on the fly. With GPUs powering super computers a natural way to think handling visualisations will be using these resources to perform graphics. However, the locality of data, the nature of the window system to be render, such as X server, and the API to be used for rendering e.g. OpenGL are some of many challenges to overcome. Our research is focus on exploiting GPUs recent hardware advantages (pascal architecture) such as the built-in encoder/decoder for graphics compression streaming. Our main objective is to run real time simulations and visualisations, such as Molecular Dynamics (MD) using interactive and low power devices, such as Tablets and Virtual Reality headset.

Due to their long coherence length, topologically protected edge states in the quantum Hall regime represent promising candidates for semiconductor qubits in electron quantum optics devices. The realization of such architectures requires suitable potential profiles for the dynamical control of the flying qubits, which is the target of our research.

We model a multichannel Mach-Zender interferometer where the first two parallel spin-degenerate edge states interfere. The scattering problem for non-interacting delocalized edge channels is initially solved, allowing us to extract the energy-dependent transmission coefficients and to expand the core system including leads, temperature or bias effects, at a low computational cost. We then study the dynamics of single charge carriers as strongly localized gaussian wavepackets of edge states. Specifically, the time evolution of the electron state is computed with a parallel implementation of the split-step Fourier method.

We introduce a second electron in the model in order to address the effect of electron-electron interaction on the mixing of the two edge states. We aim to assess the quantum entanglement created between the two carriers exploiting two-particle interference in Hong-ou-Mandel and Handbury-Brown-Twiss experiments.

Plasma simulations offer the means to understand the plasma physics within an ion thruster and can aid the design of new thruster concepts. A widely applied method is the Particle in Cell scheme, simulating the trajectories of superparticles consisting of many real particles.  The simulation of an ion thruster is difficult as spatially the electron Debye length of around 1mm has to be resolved on a domain that covers the dimensions of the thruster and the plume that can extend to a distance from the thruster exit of some m. The electron plasma frequency has to be resolved, resulting in timesteps of about 10 ns, for a simulation which should cover a time in the range of up to one second.  Even with modern hardware, state-of-the-art features such as similarity scaling, along with efficient computational methods and code parallelization have to be used to make simulation of an ion thruster conceivable for only short times.

This poster is part of my master thesis on the "Development of a lightweight spectral element parallel framework". After developing a stable and extensible framework for 1D spectral element computations, we proceed with a study on how parallel programming can be utilized in solvers for systems of equations that come as a result from spectral domain discretization. Spectral element methods, provide a lot of benefits in the solution of differential equations compared to more conventional domain discretization methods such as finite differences and finite elements. By substituting the linear basis functions with higher order piecewise polynomials, we can achieve high accuracy and exponential convergence rates without the need of a finer mesh. The case examined here is that of a spectral element solver for nonlinear wave equations and is validated against the exact solution for a standing wave. The system assembly is performed by a Matlab code ,and PETSc is used to parallelize the solver. Several methods and preconditioners are tested, both for serial algorithmic scalability and weak parallel scalability. The problem is memory bound and parallel scalability is assessed up to 40 cores on the DTU cluster. Results show that parallel scalability can be obtained and further study is required to extend scalability to more cores and embed the parallel functionality in the spectral element framework.

Cell-linked method is commonly used for classic molecular dynamics codes to find the nearest particle neighbors of a particle faster. This method consists of a domain decomposition into cells cutoff radius sized in a grid. Then the neighbor search just needs to test particles in the 26 neighbor cells. The Cell-linked method can be parallelized by iterating on cells with TBB (Intel Thread Building Bloc) and combine with domain decomposition (MPI). For heterogeneous simulations, cell densities are constantly modified and cells can be empty. Simulation performances decrease due to a cell memory overhead for empty cells and scheduling cell with different weights. To take into account these disadvantages, we propose an adaptive mesh refinement (AMR) algorithm based on k-tree. Then, we will be able to optimize cache coherence with Z-curves, to balance particles between cells and to decrease cell memory cost.

In 2017s, an estimated 250,000 new cases of breast cancer are expected to be diagnosed. A significant challenge in the diagnosis and treatment of this disease is tumor localization. In many cases (20-40%) cancerous cells are left behind after the lumpectomy surgery. Thus, a second surgery is required to remove the remaining cancer, which could be devastating. This research aims at removing the entire tumor in the first surgery while the patient is in the operating room. We will use the new technology of pulsed terahertz to image the margins of excised breast tumors. Current results show that terahertz imaging is capable of distinguishing between cancerous and healthy tissues.

In the proposed project, a three dimensional computational electromagnetic model will be developed to simulate the interactions between the terahertz signals and breast tumor tissues. The computational resources required to achieve this goal could be prohibitive. We propose to implement a parallelized version of the finite difference time domain method using the graphics processing units (GPUs) clusters. The visualization of the transmitted and reflected terahertz signals from the excised breast tumor will help us understand the way this wave interacts with the tissues, leading to significant advancement in tumor margin assessment.