Student Posters Repository

Several emerging microscopy technologies, including those used in BRAIN Initiative projects, are capable of generating petabytes of data in a single brain image. Handling this data is an ongoing challenge because the data is too large to store or process on a single desktop computer. To address this challenge, I have led a group of undergraduate students to build nTracer2, a cloud-based platform to enable PB-scale brain image visualization and neuron tracing in the cloud. nTracer2 utilizes Google Neuroglancer’s web browser-based application for remote client data access. I developed HDF5*, a novel schema of the broadly used HDF5 data format which has been optimized for maximum efficiency when accessed through archival storage systems, such as is found on many supercomputers. This innovation enables browsing whole-brain image data even through a smartphone. nTracer2 consists of three independent components: a data server that synthesizes the proper views from the raw dataset and streams them to the user end, a database backend that supports parallel data analysis and result curation in the cloud, and a neuron tracing interface that is added to the user’s browser. This modular structure allows other developers to integrate third-party functions into nTracer2 as plugins or to reuse different nTracer2 modules in their own platforms. The nTracer2 modules are designed to be scalable to large compute servers to support centralized data access. We are actively collaborating with the Brain Image Library (BIL) to deploy nTracer2 for the visualization of fMOST and other images. With the ability to sustain the maximum data transfer rate from the BIL storage cluster, in our user testing, nTracer2 enabled 5 users to smoothly visualize the same whole-brain image dataset at the same time from different geological locations across the US, with many more users likely achievable. I expect that these innovations will meet the urgent needs of various BRAIN Initiative projects to make the large image datasets accessible to the public. As a general cloud-based image visualization, annotation, and data management platform, nTracer2 will provide a viable solution to meet the increasing microscopy demands of the scientific community.

We have been developing a code for high-resolution galaxy simulations using N-body / smoothed-particle hydrodynamics (SPH) method. We aim to use more than 10 billion particles to reproduce individual stars in the galaxy simulations. These codes are implemented using OpenMP / MPI hybrid parallel computing. In the simulations, small integration timesteps for a small fraction of short-timescale regions become a bottleneck, especially when using massively parallel computing, because they worsen the scalability. For future higher-resolution galaxy simulations, we need to overcome this problem. The regions affected by a supernova (SN) often have the smallest timestep in galaxy simulations. We, therefore, adopt a Hamiltonian splitting method, in which SN regions are integrated with small timesteps using a lower number of CPUs, but the entire galaxy is integrated with a shared timestep (global timestep) using massively parallel CPUs. We are implementing such a Hamiltonian splitting method in our SPH code. This approach needs to pick up particles in regions affected by SN explosions (the target particles) by the subsequent global step in advance.

In this work, we present a deep learning model to predict the region where the shell due to an SN explosion expands during one global step. We develop the deep learning model based on Memory In Memory (Wang et al. 2018), which predicts subsequent videos from previous frames. We expand the dimension of the model’s tensors and I/O mechanisms from two to three dimensions (3D) to learn and forecast the time series of distributions of physical quantities in 3D simulations. In addition, we develop an algorithm to identify the target particles using image processing of the density distribution predicted by the deep learning model. The algorithm can identify the target particles better than the method based on an analytic solution. This work demonstrates the possibility of achieving high parallelization efficiency on state-of-the-art supercomputers such as Frontier and Fugaku.

The coronavirus (CoV) causing the COVID-19 pandemic, SARS-CoV-2, uses −1 programmed ribosomal frameshifting (−1 PRF) to control the relative expression of key viral proteins. The pseudoknot region on the mRNA that stimulates the -1 PRF can serve as a strategic target for therapeutics against SARS-CoV-2. My research aims to computationally identify potential drug-like candidates that can bind to the pseudoknot region and suppress -1 PRF in SARS-CoV-2. To achieve this, we first model the 3D structure of the pseudoknot using extensive GPU based molecular dynamics simulations. Next, we identify small-molecule ligands that can bind to the pseudoknot by computationally screening large chemical libraries using CPU based docking simulations. Compounds selected from this screening will then be subjected to experimental tests.  Also, given the increasing human contact with major CoV reservoirs like bats, new CoV diseases will likely continue to emerge in the near future, generating new public-health challenges. It is thus essential to find anti-viral drugs that are effective against multiple CoVs. We aim to extend our work to computationally search for small-molecule drugs targeting -1 PRF in multiple CoVs to identify compounds with potential pan-CoV therapeutic activity.

One-way wave propagation algorithms, an active area of research, are at the heart of many modern seismic imaging and inversion processes, such as the Full Wavefield Migration (FWM) and Joint Migration Inversion (JMI) processes. These processes are used in large seismic surveys as mechanisms for the construction of representations (images) of the subsurface structures and properties. However, this comes at a high computational cost, due to the large amount of data that needs to be processed. To a large extent, the computation time of FWM and JMI is determined by forward and backward propagation of seismic waves using one-way methods. An effective approach to one-way wave propagation is the so called phase-shift plus interpolation (PSPI) algorithm. One aspect of my work is to reveal the performance bottlenecks of this algorithm, propose, and develop optimization strategies that enable high computational efficiency across CPU and GPU architectures. At the same time, I am working on providing the algorithms through architecture-agnostic routines that work as a portability layer, which invokes on the back-end platform-specific kernels, that are implemented using CUDA/HIP and OpenMP. In addition, I aim to develop an MPI layer that will enable effective scaling of these processes across multi-node supercomputers.

The development of an elastic turbulence regime is one of the remarkable characteristics of the highly viscous polymer solutions that have been observed in experiments. The flow of polymer solution in this regime displays irregularities typical of turbulent flows even at low velocity and high viscosity (i.e., for vanishing Reynolds number). As a consequence of turbulent motion at small scales, elastic turbulence can reveal an efficient technique for mixing in very low Reynolds flows (e.g., in microchannels). Despite its great technological interest, elastic turbulence is still only partially understood from a fundamental point of view. For understanding the physics of elastic turbulence for the viscoelastic non-Newtonian fluids, the Taylor-Couette flow, and serpentine channel flow geometries are considered. The elastic turbulence is analyzed as a function of the Weissenberg number within the framework of the Oldroyd-B model using the numerical simulations performed with the OpenFOAM®. The analysis focuses on the statistical behavior of energy and velocity fluctuations.

In our research, we are aiming at developing innovative designs and engineering models in renewable energy applications mainly for power output augmentation. Various projects are presented including airborne wind energy, diffuser augmented wind turbines, wind solar tower, and wake modeling applications. In our projects, various designs and techniques are introduced to achieve optimized results. On one hand, to check the reliability of our designs, we have conducted various experimental tests that showed promising increase in power generation. On the other hand, numerical simulation of the new prototypes is crucial in our work. It provides a fast, cheap, reliable, and full-scale analysis that is hard to be performed experimentally on large modern wind turbine systems. Computational Fluid Dynamics (CFD) are used to numerically solve the fluid flow problem alongside various numerical models to capture the physics of the problem as well as the flow turbulence. Such kind of problems requires constructing a huge domain with a number of elements in the order of millions to capture the flow physics with reasonable accuracy. To proceed with numerical analysis, we used our inhouse code which is developed using an efficient OpenMP-MPI hybrid parallelization. Besides, various open-sources libraries were used alongside some commercial tools using openMP-MPI parallelization. The simulations were performed on Kyushu University Supercomputer System ITO.

The aim of this project it to investigate, apply and evaluate a broad range of optimisation techniques to improve the performance of the meshing aspect of a Formula 1 race car’s modelling workflow. A F1 car’s aerodynamic capabilities are a key factor in determining it's race pace. These aerodynamics are computationally simulated using Computational Fluid Dynamics (CFD) solvers. Geometries of the simulated components are represented using complex meshes made up of millions of cells which are generated and refined through "mesh generation" or "meshing" prior to the simulation. For this project, the meshing routine must be performed before simulating even the smallest changes in geometry and so any latencies resulting from the process add to the simulation time and consequently the overall development time. This work will endeavour to improve the performance while maintaining the accuracy of the meshing routine. The specific meshing routine that this project focuses on is parallelised for multicore CPU execution using MPI. Benchmarking and analysis of the CPU-based implementation will inform its optimisations such as incorporating alternative HPC platforms or using non-traditional methods like machine learning.

How are the Black Holes born? What is the origin of the second most compact object in the Universe- Neutron Stars. Black holes and Neutron Stars are born when the Massive Stars die with a core-collapse supernova explosion (CCSN). The explosion is so energetic and bright that it lights up the entire galaxy. Core-collapse supernovae are the main source of heavy elements including the oxygen we breathe. These elements are synthesized during the stellar evolution of Massive Stars and released into the Universe when they explode. CCNS mechanism is very complex; Therefore, we use multi-dimensional simulations to study the CCSN problem. The main questions we trying to answer are: How do Massive Stars explode? Which Stars explode and which collapse forming Black holes? To investigate these questions we need to run hundreds of multi-dimensional simulations that are computationally very expensive. Therefore, our strategy is to 1. develop an analytic understanding which will help us to simplify simulations, and 2. optimize the code to make it faster.