Student Posters Repository

Ultracold atoms provide a playground for physicists to observe quantum mechanical effects at a macroscopic level. Low temperatures facilitate the creation of Bose-Einstein condensates, in which large numbers of particles occupy the same quantum state and behave in concert. The mean-field theory predicts diverging probabilities when two Bose-Einstein condensates are suddenly coupled together. We here present the effects of extending to the many-body theory and adding in the effect of a continuous measurement acting on the system. Our results indicate that only in the infinite particle limit do we reproduce that classical result of diverging probabilities; for all finite N, quantum effects regularize the singularities. This marks a pioneering study of the quantum-to-classical transition in the presence of singularities.

Real time human face processing is a quite challenge in today's scientific study. How to save the time and let the system work much more rapid is the core target to this problem.

It is comprehensive work in different areas such as network, data transmission, cloud computing, objective detection and recognition, etc. My main study is to use the current stable and efficient algorithm and make it in parallel version to find a faster way to improve this progress. 

A new generation of radio telescopes are being commissioned to pursue audacious sciences goals at unprecedented scale. To illustrate, it is often boasted that the instantaneous data rates of these facilities will rival global internet traffic. This daunting requirement is pushing the forefront of high performance computing and big data analysis and management, by necessitating the development of novel solutions and techniques. Here we present the application of Bayesian optimal estimation techniques to the measurement problem in full-sky radio interferometry. These methods were previously thought to be computationally intractable in this setting, given the size of the matrices that require inverting. We will discuss their implementation for the Canadian Hydrogen Intensity Mapping Experiment (CHIME, chime.phas.ubc.ca) Pathfinder telescope, a 256-element interferometer currently surveying the northern hemisphere in 1024 frequency bands between 400-800 MHz. The Pathfinder is a 1/8 scale testbed for full CHIME whose primary science goal is to map the distribution of cosmic neutral hydrogen over about 10% of the universe’s observable volume, and use this data to constrain the time evolution of Dark Energy and over 4 billion years of its history. Time permitting I will also discuss efforts to produce Monte Carlo simulated full-sky maps of the cosmological neutral hydrogen distribution, which enter as the input to end-to-end simulations, critical for understanding the complicated systematics of a survey like CHIME. 

Turbulent jets are encountered in many applications including process technology, combustion, aerospace jet exhausts and the like. Hence, it becomes very important to accurately model these flows to be able to predict the behaviours as experimentation is quite difficult. Most of these flows are at very high Reynolds numbers which adds to the difficulty. When a scalar, passive or active is a part of the flow, discontinuities and shocks can be expected due to the hyperbolic and non-linear nature of the governing equations. Therefore methods are required to model these flows with high efficiency optimizing both computational power and memory. Compact finite difference methods are high order methods that can be very efficient for such flows. They also usually require solutions of tri-diagonal linear systems which make them a good candidate for such modelling purposes. With the usage of MPI libraries and partitioning, the governing equations have been solved using a combination of the compact finite difference methods and the WENO methods. Results show and underscore the need for robust methods which can efficiently perform on high performance clusters.

Aerodynamic fragmentation or aerobreakup, i.e. the breakup of a droplet of a high density fluid in a surrounding flow of lower density fluid, occurs in a wide range of technical applications. A common example is the breakup of a fuel drop in air flow in a liquid fuel combustion engine. To investigate the underlying physics of aerobreakup both experimentally and numerically, a standard test case is the interaction of a droplet with a shock. After the shock has passed the droplet, the droplet breaks up due to the velocity gradient at the interface. First two- and three-dimensional results of droplet breakup are presented. The focus is on investigating influences of different numerical schemes and analyzing different physical phenomena occurring during the breakup process. The current simulations are limited by our shared-memory approach, which restricts us to the use of maximum 300 million degrees of freedom. Our goal is to perform simulations with several billion degrees of freedom. Hence, distributed-memory feasibility of our algorithm is required and needs to be implemented to gain full advantage of HPC systems. An outlook will be given on simulations which are to be run once the hybrid-parallelization is available.

The use of non-­uniform rational B-­splines (NURBS) as  basis functions to discretise the solution variables as well  as the problem geometry in a boundary element formulation  has become known as the isogeometric boundary element  method (IGABEM). IGABEM has been shown to offer improved convergence properties over conventional boundary element method (BEM) algorithms  based on piecewise polynomial approximation spaces. Using the IGABEM formulation, the NURBS information from computer aided design (CAD) can be  used directly in boundary element analysis without  geometry clean­-up or mesh generation. However this analysis has been performed only for smooth acoustic solutions while in most real-­world engineering design and analysis acoustic problems, geometric corners and  discontinuities in boundary conditions raise the possibility of  weak discontinuities in the solution field. In the current work, we consider such problems. The motivation is to study acoustic problems at low frequency within the passenger compartment of an  automobile, and this problem is characterised by panels with piecewise continuous impedance boundaries. We develop a discontinuous IGABEM formulation based on NURBS and compare its performance against conventional BEM scheme. 

Parallel Krylov Subspace Methods are commonly used for solving large sparse linear systems. Facing the development of extreme scale platforms, the minimization of synchronous global communications become very critical to obtain good efficiency and scalability. In this poster, we highlight the recent development of a Distributed and Parallel solver for the large scale non-Hermitian linear systems. The unite and conquer GMRES/Least Square-ERAM (UCGLE) is a hybrid method which combines distributed and parallel algorithms to optimize the communications and accelerate the convergence. This method minimizes the residuals of the right-hand side, using the eigenvalues computed asynchronously in parallel. The approximated eigenvalues can be reused, and sometimes improved during the resolution of different systems, to improve the reusability and fault tolerance of method. Using a software engine, which distributes parallel implementations of Krylov methods into clusters of hundreds of computing nodes, we experiment our method with generated large-scale sparse matrices, evaluating the importance of some parameters. We obtained better convergence and performance than using some classical preconditioners. In conclusion, we propose future improvements and research perspectives.

One-sixth of the world’s population depends on snowmelt for their water supply. Current methods for estimating the amount of water contained in a snowpack, the snow water equivalent (SWE), rely on a small number of homogenous observations and an assumption that spatial patterns of snowpack are stationary over time. However, as climate change warms mountain regions, more winter precipitation will fall as rain instead of snow and snow will be confined to higher elevations. Therefore, future SWE distributions may not mirror previous patterns, necessitating a new approach to SWE estimation. Terrain and vegetation features, which are relatively stationary contributors to variation in SWE, have the potential to drive model-based estimates of SWE distributions. However, little is known of the effect of generalizability of these models across time and across model scale. In this study, we apply multiple regression tree approaches to a gridded snow depth timeseries from the Tuolomne River Basin from 2013-2016. The models are run at native 3m scale and at larger, resampled scales. Physical significance of temporal variance in parameter estimates is assessed via out-of-sample predictive skill and is discussed in the context of water resource forecasts.