Practicum Abstracts

This project focused on using the Cloud Resolving model radar SIMulator (CR-SIM) in order to (1) evaluate the differing radar signatures produced by 4 different Weather Research and Forecasting (WRF) model microphysics schemes and (2) compare outputs to observations to determine which microphysics schemes produced the most realistic radar signatures.

Over this summer, we have focused on building both a LSTM autoencoder as well as an LSTM to add transformers to that will learn and characterize chromatin statuses based on the sequence structure. For the LSTM autoencoder, we are able to directly use one hundred cells which have all been profiled for a set of 5 histone markers with CHiP-seq. We take the peak calling data for each cell for each of these 5 histone markers and use it to generate a training dataset for the LSTM autoencoder. We then can perform a similar analysis of the embedding as is done with ChromHMM’s states (ChromHMM is a software developed by the Kellis lab at MIT). For the LSTM that we plan to add transformers to, we are creating a training dataset using activation only histone marker CHiP-seq data where we use the sequences from the sample’s bam file as the input. In particular, we label this data’s sequences using FastSK, so that we can then train the LSTM with this labeled dataset.

The goal of this project was to describe and understand the interactions between periodic arrays of metal nanoparticles and 2-dimensional semiconductor layers. The former are known to support lattice resonances arising from the coherent coupling between all elements in the array, while the latter support excitons, which consist of the bound interactions between an excited electron and the hole left behind by its excitation.

(Full/manuscript title might be something like: Machine Learning Assisted Prediction of Local Geometries from X-ray Absorption Near Edge Spectroscopy) Using a data-driven approach, we utilize ab initio (FEFF9) x-ray absorption near edge spectroscopy (XANES) to predict coordination environments of absorbing atoms in a variety of different materials. Phase 1 of the project involved writing high-throughput algorithms to pull structural data from the Materials Project database, process the crystal structures, assign symmetry information to each absorbing atom, generate XANES and construct a usable training database. In phase 2, we train a machine learning classifier to predict the coordination environment (specifically, if an atom is octahedral, square pyramidal or tetrahedral coordinated) from the XANES. Hence, this is an inverse project in which we are using the output of electronic structure calculations to predict the input. Such a tool if generalized to experimental spectra could be a useful in situ tool for monitoring chemical reactions in a variety of compounds. Ultimately, we report macro (class-averaged) F1 scores in the mid-80%'s and accuracies as high as the mid-90%'s. This is exceptional compared to human level accuracy, which ranges from 30-50%.

For my practicum at Brookhaven National Laboratory, I developed faster computational methods for the Multiresolution Adaptive Numerical Environment for Scientific Simulation (MADNESS). MADNESS is a framework for scientific simulation that uses adaptive resolution methods in a multiwavelet basis. In short, it is a petascale programming environment that maintains backward compatibility with tools like MPI and Global Arrays. It also provides a suite of numerical capabilities to solve problems using a high level of composition. Built upon these numerical tools are a number of applications with a focus on chemistry and materials science. My original proposal was to implement a density functional theory (DFT) analysis of periodic systems to study electronic structure. As we analyzed the shortcomings of previous attempts at a fully periodic DFT code, we uncovered some inefficiencies in the way that MADNESS computed inner products and generalized algebraic problems (gaxpy's). My advisor, Robert Harrison, encouraged me to get "under-the-hood" in MADNESS and implement a faster inner produce and gaxpy operation. Much of my practicum was devoted to this implementation.

In order to study the role of sub-nm Au clusters in photocatalytic water splitting, we are investigating the electronic structure of the interface between sub-nm Au and CdS surfaces using first-principles electronic structure calculations. We are exploring the structural, electronic, and optical properties of Au clusters of different particle composition, size, charge state, and surface termination. We study the interactions between the nanoparticles and the semiconductor substrate, with a focus on (1) the adsorption geometry and interaction strength between nanoparticles and the substrate, (2) charge transfer and energy level alignment at the interface, and (3) how these properties impact photocatalytic activity of the combined system. We rely heavily on parallel computation with codes such as VASP and Quantum ESPRESSO and utilize the high-performance computing resources at BNL. Ultimately, we aim to provide a fundamental understanding of the physics that underlies the observed dramatic enhancement of catalytic activity, and enable rational design of materials for water splitting.

We found the structural and energetic differences between the two naturally occurring forms of crystalline cellulose using VASP, a quantum mechanical Density Functional Theory (DFT) code. In nature, these two types of cellulose compose the majority of biomass and are the primary targets of large-scale renewable ethanol production. In addition, preliminary analysis of the solvation and crystalline decomposition in water were performed via classical molecular dynamics simulations.

The conversion of cellulose to ethanol is a key process in reducing the dependence on fossil fuels. Cellulose is a polymer of glucose which resists breakdown into its monomeric subunits making it difficult to ferment into ethanol and other biofuels. There is currently a worldwide effort to find and/or develop new enzymes to hydrolyze cellulose. Improving the performance of known hydrolytic enzymes is also a major goal. Cellulose binds specifically to certain regions of proteins which are complimentary to the carbohydrate functional groups; therefore, protein structure is vital for recognition by cellulose substrates. From experiments, it is hypothesized that hydrophobic aromatic protein residues arranged parallel to the binding grove may stabilize substrate binding between a protein and cellulose. To better understand the effects of aromatic groups arranged parallel to the aliphatic surfaces of cellulose, a simplified model system was considered. One benzene molecule which is chemically similar to the aromatic rings in the protein residues Tryptophan and Phenylalanine is positioned in a stacked configuration relative to a glucose molecule, the subunit of cellulose. This system has been studied and simulated using molecular dynamics (MD) previously, but our interest was in increasingly the computational sampling time. In the previous study, dynamics were calculated for only 150 ps. Today, typical MD simulations collect data for 5-10 ns of simulated time. Calculated thermodynamic properties of a system are influenced by the sampling time. Because 150 ps is a very short time period, the goal for this project was to repeat the simulations for longer time periods in order to determine if the original sampling period was sufficient for the properties to be stable with time. From the simulations, the potential of mean force (PMF) for the approach of glucose and benzene were calculated as a function of the separation distance between those molecules. The PMF is an effective free energy potential which depends on state variables of the system, and it captures the effects of solute rearrangement in the medium. Obtaining the PMF vs. distance plot was a main objective because the comparison between the plots for the two different simulation lengths can be used to evaluate the effects of longer computational sampling time.

The Lattice QCD density-density autocorrelation function was computed, using QCDOC at Brookhaven National Lab, using three different discretizations of staggered fermion both above and below the temperature of the confinement-deconfinement crossover. For all three discretizations, it was found that the breaking of quark flavor symmetry by the staggered prescription resulted in numerical artifacts which are not yet under control. Once these artifacts are managed, the structure function will be forthcoming.

During the practicum I had the opportunity to work with Michael Creutz while getting an introduction to high energy physics and in particular lattice gauge theory. Over the summer I explored several directions including the lattice formulation of QCD, Witten's non-abelian bosonization in 1+1 dimension, and domain wall fermions.