Student Posters Repository

Each year, landslides cause more than 5000 deaths. The frequency and severity of flooding continue to escalate due to the effects of drought and global warming, making landslides an ongoing concern. My research has a goal to better understand and make predictions which include granular media, water, and air. This process involve the study of small particles, rocks, and boulders falling into the water.

For simulation, we are using Computational Fluid Dynamics (CFD) methods combined with the Discrete Element Method (DEM), allows the simulation of solid bodies with arbitrary shapes. Such problems are typically referred to as Fluid-Structure Interaction problems. To improve the simulation model of this complex system, we have developed a force model based on the open-source code CFDEMcoupling which is a combination of the OpenFOAM and LIGGGHTS codes. OpenFOAM uses domain decomposition to run in parallel and LIGGGHTS codes based on LAMMPS software, which uses the MPI. The important goal of the project as well is being reproducible, for this goal we are using containers and avoiding GUI.

Ligand-gated ion channels (LGICs) are central receptors of electrochemical signaling in cells across evolution. These channels open ion-selective pores through the cell membrane in response to neurotransmitter release and transit to a desensitized/closed state in the presence of neurotransmitters, before resetting to their initial, structurally distinct resting/closed state. A diverse set of allosteric modulators, including neurosteroids, anesthetics, and lipids modulate their functions in myriad different ways, suggesting a complex conformational landscape of these protein functions. Here, we used a molecular dynamics simulations based goal-oriented adaptive sampling method, named fluctuation amplification of specific traits (FAST) in combination with Markov state modeling to characterize the energetics of the opening of the pore. Though functional evidence of these states is present in the literature, structural characterization with experimental techniques such as cryo-EM was not feasible before, possibly due to the transient nature of these states. Given the high pharmaceutical relevance of LGICs in diseases such as depression and epilepsy, exploring uncharted territories in the conformational landscape of these proteins by computational methods can provide an unprecedented ability to regulate biomolecular function by designing state-dependent drugs.

Cosmic rays (CRs) are one of the primary constituents of the ISM and are the main driver of ionization in star-forming molecular clouds. They are accelerated by high-energy shocks from sources external to the clouds such as supernova remnants and by the shocks at the collapsing protostars inside the cloud.  Despite their impact on gas temperature, kinematics, and chemistry, they have not been included in previous simulations of star formation. We run simulations of a collapsing turbulent molecular cloud with the GIZMO code that includes explicitly modeling CR transport from galactic CRs. We assess the impact of CRs on star formation and post-process the simulations to compute their effect on observational diagnostics. We find that in environments where the CR background is comparable to that of the Milky Way, the effect of including CR transport on the star formation in the cloud is minimal. However, in high CR environments such as near CR acceleration sources, CRs have a significant impact on the gas dynamics and increase the star formation in the cloud.

BRAF-mutant melanoma cells under prolonged BRAF inhibition enter a state of balanced division and death, termed “idling.” The idling state may act as a haven into which cancer cells can escape to survive drug treatments, and eventually acquire genetic resistance mutations, driving tumor recurrence. Here, we present a Boolean model of the crosstalk between the MAPK signaling pathway and store-operated calcium entry that regulates Ca2+ within the ER. The model simulations are in qualitative agreement with experimental calcium flux assays for 24-hr and 8-day BRAF inhibition treatments. Our model explains the observations in terms of a time delay in reactivation of calcium channels after ERK inactivation, assuming activation of a secondary gene expression program after addition of BRAF inhibitor. Model simulations show an increase in the cytosolic calcium level after addition of external calcium only in cases of long-duration BRAF inhibition. Therefore, our Boolean model sheds light on the crosstalk between BRAF inhibition and Ca2+ transport in the idling phenotype, contributing to the development of effective anti-cancer therapies that specifically target drug-resistant cells.

The Kuramoto model is one of the most representative models of coupled phase oscillators, commonly used in the description of synchronization phenomena in complex systems. The aim of this project is to study, from a mathematical and computational point of view, three linear conversions of the Kuramoto model in the case of identical oscillators with global coupling. A comparison with the original Kuramoto dynamics is also conducted.

Magnetic confined fusion plasmas play a key role in solving the global energy crisis. However, experiments can be costly and challenging to construct, while limiting their representation to a narrow parameter space. Different starting points for an optimization routine are physical descriptions like magnetohydrodynamics or gyrokinetics characterizing various scales of the system. Since most turbulence arises from microinstabilities that grow in scale through the inverse energy cascade, our focus is on kinetic theory, specifically gyrokinetics, which offers the best description of these microinstabilities. To optimize devices like a stellarator, simulations provide valuable insights. Several gyrokinetic codes exist, but in this study, we concentrate on "stella". This code is parallelized with MPI and runs on various different machines e.g Marconi(Bologna, Italy) or Cobra(Garching, Germany). As part of my work, I attempted to enhance the communication pattern used in the code.

Few low luminosity AGN - including the one in our galaxy - have been observed to produce flares. The flares represent sudden brightening of the source and can span a wide wavelength range, from radio to gamma rays. The physical mechanisms behind these flares are not well understood but magnetic reconnection is a promising candidate. Recent advances in HPC have enabled the use of high resolution general-relativistic magnetohydrodynamic (GRMHD) simulations that can accurately model the large scale dynamics of  these systems. We produce a flare lightcurve from a state of the art GRMHD simulation. To do so, we identify regions of magnetic reconnection, provide a model for photon emission and ray trace photons to produce a lightcurve. The final lightcurve compares well with the observations of M87 flares in terms of variability.

Modern ML algorithms rely on HPC clusters to train complex learning models that comprise of millions of parameters. Despite the use of high throughput networking technologies such as InfiniBand, communication has become one of the major processing bottlenecks in training large scale ML models. I focus on issues arising due to the "limited network bandwidth" in HPC clusters when training large scale ML models.

In this poster I discuss a several novel algorithms that reduce the network bandwidth usage in ML training. I implemented these algorithms using the Horovod package (github.com/horovod/horovod), which is a platform that facilitates ML model training on HPC clusters. Horovod virtually arranges the compute nodes in a ring topology (ring-all-reduce algorithm), and it internally employs MPI to facilitate communication. The next steps of this project requires extending Horovod to support Multi-terminal Source Coding, which is a compression technique that exploits message similarity across different HPC nodes. While existing ML software packages that use MPI currently support applying compression prior to communication, they do not exploit the similarity of messages across compute notes. Such exploitations could reduce the communication cost and up the efficiency of hardware usage.

The CRISPR-Cas9 system has become a leading tool for gene editing. However, the design of the guide RNAs used to target specific regions is not trivial. Design tools need to identify target sequences that will maximise the likelihood of obtaining the desired cut, and minimise the risk of off-target modifications. Achieving this across entire genomes is also computationally challenging. There is a clear need for a tool that can meet both objectives while remaining practical to use on large genomes. Here, I present Crackling, a method for whole-genome identification of suitable CRISPR targets. The method maximises the efficiency of the guides by combining the results of multiple scoring approaches. On experimental data, the set of guides it selects are better than those produced by existing tools. The method also incorporates a new approach for faster off-target scoring, based on Inverted Signature Slice Lists (ISSL). This approach provides a gain of an order of magnitude in speed, while preserving the same level of accuracy. This makes Crackling a faster and better method to design guide RNAs at scale. Crackling is available at https://github.com/bmds-lab/Crackling