Designing Deep Learning Algorithm for Image Classification

Collaborating Professor: Rahul Gomes

Department: Computer Science

Dr. Gomes’ lab is focused on optimizing deep algorithms that enable faster image classification with less parameters. Deep learning algorithms are more efficient than traditional machine learning since they can automatically extract relevant features from images. Since these processes are computationally expensive, students in Dr. Gomes' lab are exploring new optimization techniques for image classification. These techniques can be applied to Internet of Things (IoT) devices, where power conservation is critical as well as imaging devices like CT scanners equipped with real-time automated image segmentation of patient CT scans. 

From Local to Global: Understanding Pattern and Process in Plant Evolution

Collaborating Professor: Nora Mitchell 

Department: Biology

Dr. Nora Mitchell is using the Blugold Center for High-Performance Computing to study evolution and hybridization in plants at population and macroevolutionary scales. This work includes using population genomics software that analyzes DNA data to understand population structure in sunflowers in the Midwest and phylogenetic software to understand larger patterns in plant evolution.

High-Performance Geospatial Computing and Deep Learning

Collaborating Professor: Papia Rozario

Department:Geography and Anthropology

"Research in my GIS-Remote Sensing group at the Dept. of Geography and Anthropology is utilizing HPC nodes to develop deep learning algorithms which are capable of multiclass semantic segmentation of high-resolution multispectral imagery collected from Unmanned Aerial Vehicles (UAV). Multispectral satellite imagery in remote sensing has been used for vegetation monitoring when compared with conventional Red-Green-Blue (RGB) imagery. However, processing these images requires significant computing power. Students in my research group are processing this imagery collected from MicaSense Altum, mounted on a UAV. The cluster is providing significant resources for batch image processing and monitoring vegetation."

Opinion Mining in Software Engineering 

Collaborating Professor: Rakib Islam

Department: Computer Science

Opinions that are often embedded with sentiments (i.e., positive, or negative) play a crucial role in shaping the perceptions of other software engineers towards the entities (e.g., tools, libraries, and APIs) that are important to develop high quality and secured software. Given the plethora of information posted in unstructured formats in technical social forums, it is a challenging task for a software engineer to mine those opinions manually and make informed decisions about those entities. Students working in Professor Islam's lab develop an automated software system using the Blugold Supercomputing Cluster that is capable of mining technical discussions from the unstructured content of social forums and generate opinions on various aspects (e.g., bug, security, usability) from the entities.

Studying Accretion Through Computational Modeling of Stellar Astrophysical Systems

Collaborating Professor: William Wolf

Department: Physics and Astronomy

Students in Dr. Wolf's research group will construct computational models of stellar systems to better understand how white dwarf stars evolve through successive accretion-driven thermonuclear runaways known as novae. They will be using the open-source scientific instrument MESA on HPC nodes on the BGSC. The precise topics for the “assumptions” to be studied may include any of the following: mixing mechanisms active during accretion and outburst, the steadiness of accretion throughout the quiescent phase, irradiation of donor star by active nova, and composition of accreted material. Students will learn the basics of stellar structure and evolution through direct instruction and literature review.

Macromolecular Crowding using Multi-scale Simulations

Collaborating Professor: Sanchita Hati

Department: Chemistry and Biochemistry

Many proteins are drug targets, and accurate knowledge of proteins’ structure-dynamics-function relationships are critical to design and develop potent drug molecules with minimal side effects. However, there is a major obstacle in identifying the right drug candidate; most laboratory studies are conducted in dilute conditions and therefore fail to obtain accurate structural and thermodynamic information about the target protein in the crowded intracellular environment. Through a systematic study with spectroscopy, kinetics, and multi-scale simulations, we ask if molecular crowding affects enzyme catalysis. The enzyme under study is multi-domain Escherichia coli prolyl-tRNA synthetase - a member of the aminoacyl-tRNA synthetase family. The structure-dynamics-function relationship under crowded conditions may hold clues to new-generational antibiotics.

Characterization of Photochemically Excited Molecules 

Collaborating Professor: Stephen Drucker

Department: Chemistry and Biochemistry

Drucker's work is supported by active NSF-RUI and ACS-PRF individual investigator awards, sustaining both laboratory and computational investigations. The computational component of Drucker's research is expanding as new methods become available. The ongoing external funding provides critical salary support for a growing number undergraduate collaborators interested in computational science. However, the group's productivity in this area is threatened by the scarcity of adequate computational infrastructure on the UWEC campus. Excited-state calculations are typically very resource-intensive. For example, to calculate a set of ab initio harmonic frequencies for a single excited state often requires 100 GB of main memory along with 300 GB of scratch space, with 20-30 processors running in parallel. In recent months, Drucker's group has relied on resources available from an XSEDE allocation to run such jobs, but these are capped at relatively short run times. This limits the size of molecule or basis set that can be used. Funding via the Campus Cyberinfrastructure program would ameliorate this situation need by providing a reservoir of high-memory/high-storage nodes available locally for longer-running jobs.

The ultimate goal of the Dr. Drucker's research is to improve our understanding of the chemistry that occurs after molecules absorb light. During these photochemical events, energy from the sun or other light sources is used to weaken targeted chemical bonds in a molecule and thereby promote reactions that could not occur at a reasonable rate under normal lab conditions. The key event in photochemical reactions is a process known as electronic excitation, in which certain chemical bonds are altered (typically weakened). To predict the course of photochemical reactions, it is desirable to learn how extensively the chemical bonds within an electronic excited molecule have been compromised. Drucker's group is achieving this goal by using the experimental technique of jet-cooled laser spectroscopy. The technique allows us to determine structural parameters of the excited-state species.

This experimental information is immediately useful to chemists who wish to predict photoreactivity via computational modeling. The field of computational chemistry has grown considerably in the past several decades as computers have become faster. An ongoing goal for computational chemists is to ascertain the possible products in a given photochemical reaction environment and to predict quantitatively the relative yields of the various products under the specified conditions. This information has great value for helping to design photochemical reactions with the greatest product yields and minimal waste. An important caveat is that as computational chemists develop increasingly sophisticated modeling techniques, they constantly need to check their predictions against experimental data on electronic excited molecules. This process is necessary to validate a particular computational method or establish its limits of applicability so the method can be confidently applied to new, untested reactions.

The experimental information gathered in Drucker's laboratory is meeting this critical demand. In the past nine years, Drucker's group has used cavity ringdown or laser-induced fluorescence spectroscopy to characterize the low-lying excited states of ?,?-unsaturated carbonyl molecules—for example, the prototype acrolein molecule (propenal, CH₂=CH–CH=O) as well as cyclic enones such as 2-cyclopenten-1-one, 4-cyclopenten-1,3-dione, 4H-pyran-4-one, and 2-cyclohexen-1-one. In the most recent work, Drucker's group has used spectroscopic results to test predictions available from high-end ab initio techniques, going beyond "black-box" methods such as density functional theory.

Computational Study of Advanced Energy Materials

Collaborating Professor: Ying Ma

Department: Material Science and Biomedical Engineering

Energy storage has become unprecedentedly important in the utilization of intermittent energies such as solar and wind.⁷ It is also a critical component that enables advanced portable electronics, laptops, and electric vehicles. Lithium ion batteries (LIBs) are the most widely used energy storage devices in consumer electronics, however, limited energy density is still a major challenge. Even after decades of development, current LIB technology can only deliver a maximal specific capacity of about 300 mAh/g, which hinders its applications in a few key markets such as extended-range electric vehicles. Compared to the anode that generally has a much larger capacity, the cathode is the limiting factor where lithium ions are intercalated into the host structure of the cathode materials. Such an intercalation chemistry is essential for the stability and cyclability that are required for successful commercialization, however it is also the underlying reason that is responsible for the limited capacity. A significant increase in the capacity is unlikely, unless new chemistry and new materials are explored. In this regard, novel cathode materials and cell configurations have been actively sought and their electrochemical behaviors extensively studied.¹² Among many different materials systems, sulfur (S₈) cathode is attractive due to its exceptionally high theoretical specific capacity of 1672 mAh/g, highest among all known solid cathode materials ¹⁰.Unlike intercalation compounds where the cathode structure remains stable, a series of complex electrochemical conversion reactions take place in sulfur and various lithium polysulfides are formed during charge and discharge, as shown in Fig. 3. Substantial structural and morphological changes are observed as a result, and the structural stability requirement in the case of intercalation materials is relieved, thus leading to much improved capacity. Unfortunately, at the same time, such structure and morphological changes also lead to many serious problems such that a successful commercialization of Li-S batteries remains elusive.

In Ma's lab, computational methods are employed to reveal the microscopic mechanisms associated with the complex energy conversion reactions in the sulfur cathode. Using first principles calculations based on the density function theory (DFT), we studied the structure and energetics of various surface of S₈ and the cathode/electrolyte interfaces. Lithium diffusion barriers into those surfaces/interfaces have also been systematic investigated. Significant dependence of the diffusion barriers on the surface orientations are observed, indicating the possibility to control the reaction pattern through controlling the surface orientations. Furthermore, ab initio molecular dynamics (AIMD) simulations have been performed to study the initial lithiation and delithiation, and the elementary processes that lead to the formation of polysulfides have been revealed. It is found that polysulfides form through a three-step process, i.e., the breakdown of the S₈ rings in the cathode, the formation of small Li_xS_y molecules that are subsequently dissolved into the electrolyte (Fig. 4), and the recombination of Li_xS_y molecules into various polysulfides within the electrolyte. Such simulations provide important insights in designing efficient strategies to trap polysulfides and prevent capacity loss.

All these important results were obtained through faculty-student collaborative research where undergraduate students were actively involved. Our research in Li-S batteries has led to two presentations by undergraduate students at the 229^th electrochemical society meeting and an invited talk for the upcoming 146th Annual Meeting and Exhibition of the Minerals, Metals &Materials Society. As with other DFT calculations, our studies are computationally expensive. Most of our jobs involve parallel computing with hundreds of CPU cores and require high speed internode communication, which is beyond the capabilities of our on campus computational cluster: the BGSC. Currently, our research is supported by the computational resources provided by the University of Wisconsin – Madison and through XSEDE. Unfortunately, the availability of both resources are limited. Acquiring of additional on campus resources will significantly expedite our research progress, thus enabling further scientific discoveries in the field of advanced energy materials through research involving undergraduate students

Condensed-Phase Effects on the Structural Properties of Molecular Complexes

Collaborating professor: Jim Phillips

Department: Chemistry and Biochemistry

Phillips group is involved in studying the effect of condensed phase on the structures of primarily electron donor-acceptor complexes (e.g., CH₃CN–BF₃) but more recently hydrogen-bonded/proton-transfer systems as well.¹⁹ We employ a combination of experiment and theory. Using low-temperature IR spectroscopy, we assess the extent of structural change^{18b, 20} across various media via measured frequency shifts. Computations are vital, not only for characterizing gas-phase systems;they also provide direct insight into the mechanism for medium-induced structural change and enable us to identify promising targets and focus the experimental effort.

Much of our initial work dealt with nitrile-BF₃ systems, which exhibit large gas-solid structure differences,²¹ and even inert matrix media (e.g., solid argon) compress their B-N bonds. Currently, we are exploring the nitrile and imine complexes of the Group IV halides (MX₄: M = Si, Ge;X=F, Cl), and their monoalkyl analogs (MX₃R). The substituents in these systems have linear arrangements that would facilitate the incorporation of medium-sensitive structural motifs into larger assemblies. As such, a stimulus (e.g., electric field) could produce a force along the length of the structure (Fig. 5), which could be useful for nanotechnology applications. We initially predicted substantial condensed-phase sensitivity for CH₃CN–SiF₄, but experiments indicated no such response. We are now exploring complexes with stronger donors;nitriles with larger substituents (e.g., (CH₃)₃CCN), and imines (e.g., pyridine and its fluorinated analogs). A survey of N-Si potentials will enable us to identify promising experimental targets. For N–MX₃R complexes, the first issue is the most stable overall geometry;there are four possible geometric isomers to consider, each with four possible (symmetric) conformers. As such, we will survey a series of CH₃CN, C₅H₅N, and NH₃ complexes of the MX₃R acids (M=Ge, Si;X=F, Cl;R=CH₃-), identify the most stable isomers, and in turn, scan the potentials to identify the systems most prone to condensed-phase effects.

Domain-Domain Cross-talk in Aminoacyl-tRNA Synthetases

Collaborating Professors: Sudeep Bhattacharyay and Sanchita Hati

Department: Chemistry and Biochemistry

Aminoacyl-tRNA synthetases catalyze a key reaction in protein biosynthesis - the aminoacylation of tRNA i.e. the ester bond formation between tRNA and amino acid. These enzymes consist of multiple functional domains for the correct selection of substrates (an amino acid and the corresponding tRNA) and catalyzing the ester bond formation. The cross-talk between these domains is prerequisite for the proper functioning (substrate recognition and catalysis) of these enzymes. We are interested in knowing the underlying mechanism of domain-domain communications in these multi-domain enzymes. Using computations and experiments, we are trying to explore how the global domain dynamics and local side chain fluctuations modulate inter-domain communications and thereby effecting substrate recognition and catalysis in these enzymes. A detailed characterization of the interplay of dynamics and how they influence substrate recognition/catalysis could aid in developing potent and species–specific inhibitors against pathogenic enzymes? 

Students working on these projects will have opportunity to get hands-on experience with cutting-edge tools of computational quantum chemistry including electronic structure calculations, hybrid quantum mechanical/molecular mechanical calculations; theoretical formalisms including free energy perturbation, thermodynamic integration, charge-neutralization calculations, and intermolecular energy decomposition analysis. They will also be using various bioinformatics tools and statistical coupling analysis to identify the evolutionarily coupled residues in a protein. The experimental part of the project will provide exposure to several biochemical and molecular biology techniques that include PCR, site-directed mutagenesis, plasmid purification, cloning and sequencing, in vitro transcription, protein purification, gel electrophoresis, enzyme kinetics, and binding assays using radioisotopes. Spectroscopic studies will involve Circular dichroism and florescence measurements to monitor changes in secondary and tertiary structures of proteins.

High Precision Calculations on Few-Electron Systems

Collaborating Professor: Fred King

The focus of King's research effort is on the calculation of the properties of few-electron systems to high accuracy, supported by detailed studies of computational convergence rates. Most of my recent activity has been directed at the four-electron members of the beryllium isoelectronic series. Recently published work on some of these systems has employed two different computational strategies. The first approach employed explicitly correlated Gaussian functions as the basis set for carrying out quantum mechanically solutions of the variational problem. This approach, in the hands of a few groups, has produced the most accurate results available. This is despite the fact that Gaussians basis sets suffer from two well-known deficiencies: they have the wrong asymptotic behavior, that is, they decay too quickly away from the nucleus. Secondly, they do not allow cusp conditions to be satisfied, and these conditions are known to be important for an accurate description of electron correlation. These two limitations can be overcome by working with basis sets of enormous size, which require enormous computational resources (usually ownership of a major-sized cluster). The Gaussian basis sets have one compelling advantage, the underlying mathematical issues required to evaluate the necessary formulas is relatively straight-forward –the driving force for the use of Gaussian basis sets. The second approach employed is called the HY-CI (Hylleraas–Configuration Interaction). The most recent results from this approach have been impressive, but the basis set sizes have been exceeding large ~80,000 terms, employed by authors having access to huge blocks of computer resources.

The approach employed by King's research group is a Hylleraas technique, which avoids the use of Gaussian basis functions and replaces them with exponentially decaying functions multiplied by functions that depend on the inter-electronic separation distance explicitly. This approach has been shown to be the most accurate for two- and three-electron atomic and small molecular systems. The long-term goal is to employ this approach for four-electron atomic and molecular systems. The major stumbling block with this approach is the extreme mathematical complexity of the underlying integrals that must be evacuated. A numerical approach has been published recently to evaluation of the required integrals,²⁸ and most recently, analytical formulas for a number of the more complex integrals have worked out.²⁹ With two research students, reasonably accurate results were published for selected members of the beryllium isoelectronic series using compact wave functions.³⁰ Most recently, working with a research student, it was possible to obtain a numerical evaluation scheme for some of the most recalcitrant integrals that arise, exploiting gpu resources. The GPU approach shows a major gain in computational speed over standard OpenMP calculations.

The current approach we employ is to do the optimization phase in double precision arithmetic, and then carry the calculations of various properties such as the energy of selected states, isotope shifts, kinetic and potential energy contributions, and others, in quadruple precision. Because of resource limitations on the current cluster environment, the latter calculations are carried out in a highly fragmented manner.

The principal expected outcomes include the evaluation of accurate values for the ground states and selected excited state properties of a number of members of the beryllium isoelectronic series. Each of the principal systems under investigation is expected to take around 60 –100 cMs (core-mega seconds) to complete, to yield results that can serve as bench marks for other computational approaches. At the current rate of progress, where jobs are run typically 180+ cores 24/7, this would take at least two years to complete on the current BGSC setup, for just the first few members of the Be series. The fragmented nature of how the calculations are being carried out, is far from ideal. It is expected that some of the GPU approaches we will employ in our integral code might have wider application in other computational chemistry and physics codes.

Ping-pong Kinetics and Substrate-Shuttling in Quinone Reductases

Collaborating Professors:  Sanchita Hati and Sudeep Bhattacharyay

Department: Chemistry and Biochemistry

We are studying the redox chemistry of an enzyme family, quinone reductases, using theory and molecular simulations. Quinone reductases are detoxifying flavoenzymes and are responsible for in vivo conversion of prodrug to an anti-cancer drug. The electron transfer occurs through a classic one-site ping-pong mechanism. In this mechanism, the cofactor flavin undergoes cyclic redox reactions and the active site shuttles two substrates - a hydride donor and a hydride acceptor. Using theory and simulations, we are trying to understand the molecular-level details of kinetics. We hope to use the simulated data to learn the interplay of protein dynamics and substrate-shuttling. A complete understanding of the processes could pave the way to design mechanism-based activators and inhibitors of these enzyme systems.

Protein dynamics and its Role in Enzymatic Function

Collaborating Professors: Sanchita Hati and Sudeep Bhattacharyay

Department: Chemistry and Biochemistry

Proteins play central roles in biological processes. They are dynamic molecules and it has been proposed that a protein sequence has evolved to preserve those structural dynamics that are essential for functions like substrate binding, allosteric communications, and catalysis. The relationship between protein dynamics and their key functions is an evolving perspective in enzymology. The more complex the architecture of a protein is, the fuzzier is the picture at the molecular-level. Our research is focused on molecular-level understanding of how protein dynamics influence enzymatic processes such as electron transfer reaction and chemical bond formation. We have chosen two important families of enzymes for our study- Quinone Reductases and Aminoacyl-tRNA Synthetases.

Simulation of Planetary Bodies in Outer Solar Systems

Collaborating Professor: Paul Thomas

Research in the Thomas group is focused on N-body simulations of perturbations of the Kuiper Belt from mass distributions in the outer solar system. Brown and Batygin²² have shown that the statistical clustering of the arguments of perihelion of the population of the scattered disk component of the Kuiper Belt can be explained by the gravitational perturbation effects of a single object "Planet 9" with an approximate mass of 10 times that of the Earth at a heliocentric distance of 700 AU and an orbital eccentricity of 0.6. However, this model has been criticized by other investigators. Numerical modeling indicates that the observed regularities in the Kuiper Belt population can also be explained as the perturbations produced by a heliocentric disk of similar total mass and location to the proposed Planet 9. In addition, the stability of the orbit of a single distant object may leave it subject to disruptive perturbations from nearby stars over the age of the solar system.

The detailed behavior of both scenarios (single "Planet 9" vs. distributed mass disk) is proposed to be simulated using long-term integrations of planetary motion with the Mercury6 code.²⁵ We plan to include the major planets of the solar system and follow the evolution of a set of candidate Kuiper Belt objects. Our goal will be to examine the statistical alignments of the arguments of perihelion for both scenarios to determine which explains the current behavior of known scattered disk Kuiper Belt objects best. By starting with a small scale representation we can quickly analyze an interaction between a handful of objects and make necessary adjustments. From there, we can add levels of complexity, adding Trans Neptunian Objects and other minor planets. This should considerably increase the computing complexity of the program and require greater resources to execute. Eventually, this program will lend well to a multi gigayear execution of the solar system with minor bodies, such as Kuiper Belt Objects to determine the effect Planet Nine would have on a large time scale. Although this would yield a vast amount of information about how bodies of the solar system would interact, it would also require a large amount of computing power.

This program recognizes when a close encounter occurs and in order to preserve accuracy, executes via different protocols. This is what makes Mercury6 a hybrid integrator. The benefit of separating processes like this is the increased accuracy, however it uses considerably more computing power to execute in this manner. In a system where many close encounters may occur, a personal computer is quickly overwhelmed. Being able to allocate separate encounters to separate processors is the most efficient method of handling a situation with many close encounters.