Student Posters Repository

This poster presents an overview of the data processing pipeline employed for the dark matter search data collected by SuperCDMS (Cryogenic Dark Matter Search) detectors. SuperCDMS is a cutting-edge experiment designed to detect the elusive particles constituting dark matter. The poster focuses on key steps in the processing pipeline including triggering, event reconstruction, and post-processing.

My research attempts to address the cost of the I/O bottleneck on HPC applications. Initially, I developed a c based library to test the I/O bandwidth on HPC machines using different libraries such as MPIIO, HDF5 and ADIOS2. To improve the “effective” I/O bandwidth, I tested our hypothesis that the hyperthreads of physical cores in HPC machines could act as dedicated I/O servers and reduce the wall time of a typical HPC application. 

To enable asynchronous I/O and computational processing, I developed an I/O framework, iocomp, that supports both synchronous and asynchronous I/O approaches, while abstracting away complexities associated with multiple I/O back-ends such as MPIIO, HDF5, and ADIOS2. iocomp can be used as an I/O server for different computational kernels requiring minimal changes to the user code. The STREAM and HPCG benchmarks are adapted as examples of client processes to perform computation, adding a step to write their data through the iocomp library.

The results demonstrated that adding a separate set of nodes to run a dedicated I/O server improved the performance for the STREAM and the HPCG benchmarks for all I/O back-ends tested. However, hyperthreads resulted in lower or at best comparable performance, depending on which I/O back-end was used. To address this issue, I am currently developing a shared memory approach to enable the hyperthreads and their physical cores to benefit from their shared memory using MPI one sided communication. 

An effective tsunami warning system can save lives and valuable infrastructure, but requires real-time tsunami prediction in order to accurately respond to the tsunami risk. Numerical modeling is often used in tsunami prediction, but can be hindered by the computational cost, and thus may not facilitate real-time modeling, or may have to sacrifice accuracy. In order to preserve accuracy of tsunami hazard predictions while achieving real-time speed, we developed a deep-learning model to predict the maximum tsunami height at the shoreline trained on synthetic ocean bottom pressure data at four observation points. We generated input conditions for 1300 tsunami scenarios in the Cascadia Subduction Zone to train the model, which were sourced from the earthquake database Melgar et al. (2016).  We then utilized these input conditions to run the nonlinear shallow-water model GeoClaw developed by LeVeque et al. (2011) for 1300 corresponding simulations. The deep-learning model was trained by utilizing ocean bottom pressure from the observation points to predict tsunami wave height at the shoreline. The model was validated with records from the U.S. West Coast and the Crescent City shoreline, which were withheld from the training dataset. To apply this model in practice, the offshore information from the DART buoys system is necessary. 

Advances in molecular dynamics (MD) software alongside enhanced computational hardware have allowed for simulations to significantly expand our knowledge of biomolecular structure and dynamics. It has allowed for the extension of conformational sampling times from nanoseconds to the microsecond level and beyond. Accuracy and reproducibility of force fields used in biomolecular MD simulations are critical in producing biologically relevant data. The Amber nucleic acid force fields have been used widely since the mid-1980s, and improvement of these force fields has been a community effort with artifacts revealed, corrected, and reevaluated by various research groups. Here, Amber force fields for use with double-stranded DNA are highlighted and we present the assessment of two recently developed force field parameter sets (OL21 and Tumuc1). We observe the improvement of OL21 and Tumuc1 compared to previous generations of the Amber DNA force fields. We did not detect any significant improvement in the performance of Tumuc1 compared to OL21 despite the reparameterization of bonded force field terms in the former; however, we did note the deterioration of Tumuc1 performance when modeling Z-DNA.

In environmental microbiology, protists (microbial eukaryotes) are notoriously recalcitrant. The complex genomes of these organisms are peppered with largely unknown repetitive sequences. Increasingly complicated pangenomes have been identified within species of eukaryotes, whereby individual genetic subvariants within a metapopulation may harbor substantial genetic diversity with phenotypic consequences. I will discuss how some of my new high-quality transcriptome references can be used in conjunction with in situ metatranscriptomic and metagenomic data in the oceans and computational tools I have developed to process these data. Further, I will discuss a new algorithm I have developed for improving the accurate identification of protistan diversity in situ using meta-omic datasets. I will discuss the computational challenges of manipulating these massive multi-omic datasets on high-performance computing systems, and the benefit that algorithm development can have on making accurate assumptions. These computational advances are critical for making inference about the important ecosystem roles of ecologically-essential and globally-ubiquitous protists on relevant scales.

Relativistic magnetic reconnection has been proposed as an important nonthermal particle acceleration (NTPA) mechanism capable of producing power-law spectra and high-energy emissions. Power-law spectra are in general characterized by three parameters: The power-law index, the high-energy cutoff, and the low-energy cutoff (i.e., the injection energy). In relativistic magnetic reconnection, particle injection has received considerably less attention than acceleration to high energies, despite injection being a critical step in the NTPA chain. One unanswered question that has significant implications for both physical understanding and astronomical observations is how the injection energy depends on the upstream magnetization σ. Furthermore, we still do not understand how injection by the reconnection electric field, Fermi kicks, and pickup acceleration scales with σ. We uncover these relationships by measuring the injection energy over several σ in 3D reconnection. We also investigate how the flux rope kink instability, a uniquely 3D phenomenon, impacts particle injection. Lastly, we will discuss the observational implications of our results.

We are engaged in research on electrocardiograms(ECGs). ECGs are recordings of the electrical activity of the heart, expressed as a waveform. It is mainly used in hospitals to evaluate cardiac diseases such as arrhythmia. In recent years, it is also said that autonomic nervous system activity can be evaluated by analyzing fluctuations in heart rate variability from ECGs. In this context, we are conducting the following two studies.

The first is the study of individual identification using ECGs. It is said that the shape, size, and position of the human heart differs from person to person. In addition, ECGs can be used in situations where security is required, such as biometric identification, because they do not remain, as fingerprints do, and cannot be easily obtained, as is the case with facial images. We have employed a machine learning approach to build an ECG-based personal identification system. Currently, the identification error rate is about 7% for public datasets. We plan to compare our system with other machine learning methods to improve its accuracy.

The second is a study to synchronously measure facial expressions and ECGs to clarify the correspondence between subjects' emotions and each physiological marker (RMSSD, LF/HF, etc.) obtained from the ECGs. In our system, facial expressions are captured at 5~10 fps during ECG measurement and measured with a high resolution setting, aiming for more detailed analysis. In the future, we plan to conduct measurement experiments while showing subjects videos that elicit emotions.

We present a custom Particle-In-Cell code for plasma simulation written in C++/CUDA. This algorithm allows to describe plasma properties such as density, electron temperature and potential on two-dimensional domains. To do so, the code is based on an iterative method comprising the following operations: plasma particles are generated, spatial distribution of plasma density is obtained and exploited to solve the Poisson equation (GMRES method is implemented), electromagnetic field is calculated (only static magnetic fields are considered) and used to move the particles. Our code features a comprehensive collision cycle, including 50+ different processes for describing hydrogen plasmas, which allows to achieve rate-driven equilibria. Our scientific purpose is to employ this code for investigating the phenomena underlying the plasma behaviour in large negative ion sources for fusion applications, ultimately aiming at providing useful information for improving source design and operation.

Our group "Computational Photonics" is specialized on the inverse design of nanophotonic devices.

We implemented a parallel "topology optimization" routine in our in-house electromagnetic solver,

enabling the design of arbitrary optical materials and large-scales structures. Due to the large storage 

requirements as well as a necessary acceleration of the optimization run time, an efficient HPC approach is essential.