Student Posters Repository

QM/MM simulations have become a powerful tool to model chemical processes in solution and
biomolecular systems, at a reasonable computational time with the necessary accuracy. It has
been applied in drug design [1], enzyme/protein engineering [2] and complex systems like ionic
liquids[3]. One of the main applications of QM/MM simulations is to investigate the kinetics
of chemical processes. We have used our QM/MM software package [4], LICHEM (Layered
Interacting CHEmical Models), to performed geometry optimizations and minimum-energy path
(MEP) transition state (TS) search using the nudged elastic band (NEB) method and the quadratic
string (QSM) method to study the limiting step dealkylation reaction catalyzed by ALKBH3.

To most efficiently utilize modern parallel architectures, the memory access patterns of algorithms must make heavy use of the cache architecture: successively accessed data must be close in memory (spatial locality) and one piece of data must be reused as many times as possible (temporal locality).

Unstructured mesh algorithms are notoriously difficult in this sense, especially due to computations that indirectly modify data, leading to race conditions. In this work we address this problem through a number of optimisations on GPUs, specifically the use of the shared memory and a two-layered colouring strategy to cache the data. We also look at different block layouts to analyse the trade-off between data reuse and the amount of synchronisation.

We developed a standalone library that can transparently reorder the operations done and data accessed by a kernel, without modifications to the algorithm by the user. Using this, we performed measurements on relevant scientific kernels from different applications, such as Airfoil, Volna, Bookleaf, Lulesh and miniAero; using Nvidia Pascal and Volta GPUs. We observed significant speedups (1.2 -- 2.4x) compared to the original codes.

A frequency domain method for predicting phonon frequencies and lifetime values using phonon spectral energy density is being investigated. This method involves computing phonon mode eigenvectors through lattice dynamics calculations and obtaining trajectories of atoms through molecular dynamics simulations, which are used to calculate the phonon spectral energy density (SED) at each wavevector and polarization. Phonon lifetimes are calculated by fitting the phonon SED data to Lorentzian functions which are used as an input for solving the three-dimensional phonon Boltzmann transport equation to manifest the atomic scale nature of thermal transport.

The International Barcode of Life project (iBOL) continuously records and catalogs specimen information and stores data on the Barcode of Life Database system (BOLD). Advances in DNA analysis have lead to a rapid increase in the volume of data available for researchers involved with the iBOL project to study. This has resulted in modern statistical techniques playing an increasingly important role in the analysis of such large volumes of data. 

BOLD.R was developed to allow users to access data from BOLD directly into R via current APIs maintained by the BOLD system. Users can access their own private data by logging in to BOLD through BOLD.R or they can access public data without the need to login. Data accessed using BOLD.R is stored in R with a consistent internal structure thereby allowing the user to employ the suite of functions provided by BOLD.R in conjunction with existing R packages. Therefore BOLD.R provides the flexibility to use data retrieved from BOLD over a wide range of applications. A future implementation of BOLD will harness the power of cloud computing by allowing users to run scripts on BOLD hardware or on the hardware of an HPC partner.

Our research groups have developed the software system, HDDM_EMPS, to conduct fluid simulations based on a particle method called the Explicit Moving Particle Simulation (EMPS). HDDM_EMPS formerly adopted a wall-particles model to satisfy boundary conditions. Since wall-particles models allocate particles to express boundaries and require more memory space with larger-scale simulations, the research groups replaced it with a polygon-walls model in the fluid-analysis system that sets up polygons instead of particles for boundaries. The latter model has a potential to reduce computational costs by saving memory space, but the system with the model was not well load-balanced. In the former system, ParMETIS was utilized to stabilize load balancing based on the number of particles. However, the number of particles in a polygon-walls model becomes distinct from the one in a wall-particles model. Furthermore, some particles in the polygon-wall model hold a different weighting value due to polygon-walls algorithms, which search the closest particle located at every node of polygon walls. Hence, I have designed load balancing algorithms and demonstrated the efficiency of a polygon-wall boundary model compared to a wall-particles one in the HDDM_EMPS system.

Making modern high-performance architectures accessible to specific data analytics needs has become a field of research on its own. This naturally entails tackling performance engineering issues, where hardware characteristics can help the tuning of algorithms and data structures to fulfill the needs of real-life applications. Machine learning provides a vast amount of scenarios that deal with large amounts of data that need efficient algorithms and data structures, making it perfectly suited for a high-performance treatment. Additionally, input data and problem parameters are often uncertain, needing a complex treatment. Adaptive sparse grids, which are increasingly being used in numerous fields of study, provide a very powerful tool in reducing complexity by tackling the issues of high-dimensional spaces. In the context of data mining, sparse grids have already been successful as a dimensional-aware alternative to more traditional approaches and are especially well suited to modern multi- and many-core systems.

In that sense, the focus of this work is on developing sparse grid algorithms optimized for the latest high-performance clusters. These codes are applied to solving several computationally-demanding problems in machine learning and other fields that lend themselves to a sparse grid treatment, in both deterministic and stochastic scenarios.

The best explanation we have for the behavior of the world around us is the Standard Model of particle physics. However, for all its success, the Standard Model itself raises many important theoretical questions.  Further, the Standard Model is known to be incomplete: for example, it does not explain dark matter, dark energy, or the observed excess of matter over antimatter in the Universe. The field of Beyond the Standard Model (BSM) physics works towards finding a more complete description of nature.

One idea in BSM physics supposes that the Higgs boson is not a fundamental particle, but rather a tightly-bound composite particle, glued together by some new strongly-coupled force like the strong nuclear force. To test this idea, we must make predictions to compare with experiment. However, strongly-coupled quantum field theories are mathematically complicated. The only way to test these theories is by direct simulation with lattice gauge theory, a set of Monte Carlo techniques originally invented to study the strong nuclear force. We apply these techniques to investigate a particular composite Higgs model.