Student Posters Repository

It has been suggested that mergers of stellar mass black hole binaries embedded in accretion disks of active galactic nuclei (AGN) may be a significant contributor to the observed LIGO merger rate, e.g., Stone et al. (2017). Additionally, nuclear star cluster (NSC) stars can be captured by AGN disks and massive stars may form in the gravitationally unstable outer regions of AGN disks, as has been suggested to be the case in the parsec scale disk of young stars in our own Galaxy, which have the potential to evolve into black holes. The presence of these compact objects within AGN disks could
significantly affect disk evolution as well as contribute to our understanding of black hole distributions in the Universe. As such, the goal of this research is to investigate the nature of black hole dynamics and interactions within AGN accretion disks via three dimensional global simulations using the GIZMO code. Running simulations of AGN accretion disks with embedded black holes, we will examine occurrences of black hole binaries, possible black hole merger rates, black hole migration torques & migration rates, and the effects of these dynamic interactions on AGN disk structure and evolution.
This research will inform future work regarding black hole merger rates, black hole distributions, and AGN accretion disk structure and evolution.

Glioblastoma (GBM) is an aggressive central nervous system malignancy arising from glial cells with poor prognosis and limited treatment options. Extensive intra-tumoral phenotypic heterogeneity and cellular plasticity resulting from transcriptional cell state diversity within GBM have posed challenges to effective therapy. While genetic mechanisms have been extensively studied, non-genetic processes that contribute to phenotypic diversity have not been adequately explored. Our central hypothesis is that the tumor microenvironment (TME) modulates the epigenetic remodeling of both malignant and non-malignant cells, leading to intra-tumoral phenotypic diversity and plasticity. To address this hypothesis, we conducted multimodal single-cell sequencing (transcriptome, methylome, and chromatin accessibility) on four distinct anatomical regions of GBM tumors from patient MRI images to examine cell-to-cell changes. We plan to computationally identify the regional differences across regions, define common and unique cell-cell interactions, and explore the spatial growth patterns. 

The widespread adoption of electric vehicles depends on an effective and equitable charging infrastructure network. My research explores the simulation of the complex system of vehicles, charging stations, and policy to evaluate existing networks and predict future infrastructure needs. Presently, I am leveraging the distributed for-loops provided by Julia to simulate thousands of vehicles independently of one another. This embarrassingly parallel easily scales to multiple compute nodes. However, without interactions between vehicles, the model neglects significant effects, such as charger congestion. Modeling these interactions would require communicating vehicle information between nodes and efficiently sharing the compute load. Switching to an MPI-based approach will enable me to scale to the millions of vehicles required for city-level simulations. Additionally, I have been exploring using Scientific Machine Learning to accelerate battery models, such as the Doyle-Fuller-Newman model. My current approach using recurrent neural networks has been effective, providing a 400x speedup for drive cycles. However, the current approach is limited to modeling a single cell. Expanding the model to perform pack-level simulations requires a multi-GPU paradigm due to the model’s size. By attending the IHPCSS, I hope to develop the skills to implement efficient MPI algorithms and multi-GPU codes, enabling me to perform larger and more complex simulations of charging infrastructure.

Regulating the process that allows cancer cells to indiscriminately proliferate at a pace that is significantly higher than their rate of death, in contrast to healthy cells, is the only way to understand how to cure cancer. Although a number of apoptotic regulators and cancer biomarkers have been identified, the issue is that each organism has a different quantity of proteins, ligands, and receptors. The dynamic fluctuation in responses to cancer therapy is brought on by the variety of these traits. We developed a mathematical model to investigate the effects of caspase mediated apoptotic signaling brought on by extracellular activation in order to address this. Three potential thresholds of caspase 3 necessary for the activation of apoptosis  were utilized to simulate the efficacy over a variety of patient profiles and treatment methods. 

This work focuses on melt flow through solid rocks in the lower and middle crust. Melt forms in local sites and migrates in ways that depend on local and global conditions. These phenomena leave their signature in rocks that eventually solidify and are now found on the Earth’s surface. The aim of the project is to assess the conditions in which they formed by analyzing structures that are predicted by numerical experiments. 

Our numerical code uses an innovative two-grid, hybrid approach to simulate the two phases (solid and fluid) and their interaction. On the first grid, a continuum model that solves porous flow simulates the pervasive flow of the molten phase. A discrete-element model is solved on the second grid, which consists of a network of springs (elastic behaviour of the host rock) and allows for fracture formation.

In the example reported in the poster, fracture networks are analyzed using topological features to assess how their geometry depends on melt viscosity and melt rate. We observed that low melt rates and low viscosity promote longer fractures and therefore a higher permeability. On the other hand, isolated cracks are indicators of higher viscosity values and fast melt rates.

1- Enhancing speed using a GPU: I am interested in optimizing the code to leverage the computational power of a GPU, which can potentially accelerate the simulation process.

Public transportation with high ridership rates is not only a sustainable alternative to automobile travel but also a crucial transportation mode to support zero vehicle households, vulnerable populations, and marginalized communities.  The reliability of the public transportation system is important in encouraging choice users to take transit.  Extreme weather is one of the most common factors affecting the operation of public transportation.  While departments of transportation and municipal governments are becoming increasingly aware of public transportation system vulnerability to extreme weather conditions, it is difficult to prioritize transit investment in reliability without a clear understanding of the costs.  This study presents a framework to assess transit network robustness to extreme weather events at the city level, using open source data.  The analysis first constructs transit networks using open source GTFS data, followed by a robustness assessment with scenarios informed by FEMA flood information.  Network analysis is then conducted to assess the overall characteristics of the system.  A large-scale analysis of 19 of the most populated cities in the United States identifies network features relevant to the transit network's robustness.  Our analysis shows that 100-year and 500-year flood significantly influence the transit network, cutting off peripheral transit stations and causing the entire transit network to "shrink." The evaluation and benchmarking approach proposed in this study is proven useful in assessing the flood risks of the transit network.  Based on the analysis, most transit networks studied do not have the capacity to withstand 100-year floods.

It has been shown that the solution to a dense linear system can be accelerated by using mixed precision iterative refinement relying on approximate LU-factorization. While most recent work has focused on obtaining such a factorization at a reduced precision, we investigate an alternative via low-rank approximations. Using the hierarchical matrix format, we are able to benefit from the reduced complexity of the LU-factorization, while being able to compensate for the accuracy lost in the approximation via iterative refinement. The resulting method is able to produce results accurate to a double precision solver at a lower complexity of O (n^2) for certain matrices. We evaluate our approach for matrices arising from BEM for 2-dimensional problems. First, we analyse the convergence behaviour of the method, assuring that we are able to adhere to the same error bounds as mixed precision iterative refinement. Afterwards, we evaluate the performance in terms of the execution time, comparing it to a general dense solver from LAPACK and preconditioned GMRES. On some matrices we are able to achieve a speedup of more than 7 times when compared to those reference algorithms.

Immunotherapy drugs restore the ability of the body's immune system to recognise and respond to cancer. While immunotherapy has an impressive response in many cancers, including melanoma, unfortunately it is not effective in all patients, and can also be associated with substantial adverse side effects. With the gut microbiome having emerged as an important factor influencing both response to immunotherapy treatment and the risk of developing side effects, there is need for a greater understanding of how the microbiome exerts this influence. Using HPC, I hope to identify the bacterial species that are associated with this effect as well as the underlying mechanism: in other words, "who is there" and "what are they doing". Answering these questions will improve treatment outcomes and create possibilities for personalised medicine.