Student Posters Repository

Strongly-coupled Brillouin backscattering in plasma is a promising route towards the amplification of high-intensity laser pulses in compact systems. However, to ensure the production of high-quality beams, Brillouin amplification schemes require the control of parasitic effects like pre-heat of the plasma by the pump beam, filamentation of the pump and probe beams, as well as other competing instabilities. Advances in the production and modeling of laser pulses with complex spatio-temporal structure, such as beams with a flying focus, offer new strategies to control these issues.

In this work, we perform multidimensional simulations of Brillouin amplification with a flying focus pump beam using the particle-in-cell code OSIRIS. We show that the use of a flying focus can effectively mitigate the development of premature instabilities triggered by the pump before it interacts with the probe beam. Further, we explore deviations from established nonlinear theory of Brillouin amplification in this new regime, and report on its impact on efficiency and beam quality. We perform parameter scans with different flying focus pump beam properties and plasma conditions and identify the optimal range of parameters for efficient Brillouin amplification with flying focus beams.

Solving large-scale linear systems is essential for numerically solving partial differential equations from real-world models. My research focuses on optimizing linear solvers for high-order finite element discretization on GPUs. To enhance the solver, I consider three key aspects: selecting a good initial guess, designing an effective preconditioner, and choosing an optimal stopping criterion. I demonstrate a numerical example using the proposed stopping criterion and compare its performance to the commonly-used relative residual norm criterion.

Sensitivity analysis with chemical transport models (CTM) enhances the understanding of atmospheric processes. A common application is understanding the relationship between input variables (e.g., the emissions of select species) and output variables (e.g., species concentrations or concentration-dependent metrics). Traditional forward methods like the finite difference method (FDM) and the complex step method for calculating sensitivities are prone to errors such as subtractive cancellation and truncation, especially for second-order sensitivities. Currently, the adjoint of GEOS-Chem enables the calculation of the first-order sensitivities of a selected concentration-based metric with respect to spatially heterogeneous emissions of all species. Probing the influence of a selected emissions source or species on all concentrations relies on the finite difference approach, which is computationally expensive and unstable for second-order sensitivities. Here, we are implementing a novel method for evaluating both first- and second-order sensitivities with machine precision, as previously demonstrated for CMAQ (Liu et al., GMD, 2024). The capability of computing numerically exact first-order and second-order sensitivities with the hyperdual step method is shown. We demonstrate this novel application by determining the effects of precursor NO_x and biogenic isoprene emissions on ground-level tropospheric O₃ concentrations. The hyperdual-step method is validated by comparing with FDM sensitivities for the first order and with a hybrid hyperdual-finite difference method (HYD-FDM) for the second order. This study represents the first application of the hyperdual-step method to a global atmospheric CTM, specifically GEOS-Chem.

Developing the next generation of nuclear fission and fusion reactors is a critical part of the transition toward low-carbon energy sources. To ensure that these reactors are safe, long-lasting, and economically viable, it is important to consider the degradation of key reactor components, such as corrosion of structural materials. In this study, we investigate the impact of radiation on nuclear reactors that use molten fluoride salts, such as FLiNaK and FLiBe, two commonly considered fuels and breeding blankets. We find that radiation has the potential to accelerate and decelerate corrosion of the structural material. However, little computational research has been done to show how radiation influences corrosion. Using high performance computing, we will conduct atomic simulations of radiation damage in molten fluoride salts. The results of this study should indicate that the chemical structure of molten salt plays a key role in the rate at which corrosion takes place. Improving our understanding of radiation-induced corrosion in molten fluoride salts will allow us to develop nuclear reactors that are safer, operate longer, and cost less.

Materials that can withstand extremely high temperature environments are desirable to identify for the next generation of energy technologies. Such environments are are often difficult to study experimentally, therefore computational techniques find a relevant niche here - especially for exploring microstructural mechanisms and dynamics such as melting, solidification and growth. Here, we are pursuing a novel computational approach: aiming to couple crystal plasticity (CP) simulations and machine learning (ML) for the purpose of exploring unique multi-principle metal alloy compositions capable of high temperature strength. The constitutive equations of the crystal plasticity formulation allow us to leverage control over the physics (i.e. we can choose the physics to incorporate into our models) at the mesoscale level, while the machine learning enables us to better explore the feature space of multi-principle refractory alloys.

Traditionally, HPC systems were dominated by large, long-running scientific simulations with predictable resource requirements. Nevertheless, the rapid development of science and technology continuously brings changes to those requirements. Those changes cause more challenges for optimizing the system performance to better fit the user needs. While system throughput remains the primary target of optimization, it is also crucial to look into the power consumption of jobs and maximize efficiency by minimizing power consumption and improving the number of jobs executed. 

In this work, we deeply observe the workloads and ways to optimize the largest supercomputer in the UK: ARCHER2. We look into power usage, job sizes, and other metrics to define the workload and visualize it. This is the intermediate step to predicting future workloads and system requirements and suggesting what changes could be implemented to improve throughput and decrease power consumption.

The seasonal snowpack plays a critical role in our ecosystem. Current ground observations and remote sensing techniques are insufficient to understand the spatiotemporal dynamics of snow. To address this issue, data assimilation (DA) techniques combined with models with remotely sensed snow depth observations can improve daily estimates of snow depth and snow water equivalent (SWE). This study addresses this gap by developing an open-source, cloud-based system for integrating Sentinel-1 (S1) remotely sensed data with a Particle Batch Smoother (PBS) DA scheme for estimating SD at 500m. The system uses the Multiscale Snow Data Assimilation System (MuSA) to generate ensembles with ERA5 meteorological forcing data. As a test case, we apply the system to assimilate 500m SD from S1 using PBS in the East River Colorado Basin in 2019 and 2020. PBS SD results were spatially compared to Airborne Snow Observatory (ASO) LiDAR, Moderate Resolution Imaging Spectroradiometer (MODIS) snow disappearance, and snow frequency data. Temporal validation included comparing the PBS SD data with the Snow Telemetry Station (SNOTEL) and ground data that we collected on Snodgrass Mountain. The analysis showed a bias of +0.8m between the assimilated and observed SD. The system tends to overestimate SD in shallow snowpack and underestimate it in deep snowpack. Despite these limitations, the cloud-based system has the potential to provide estimates of daily snow depth and other snow characteristics at a 500m resolution for snow and watershed studies. Future enhancements aim to reduce bias by incorporating diverse input data streams to improve performance over larger areas and different snow environments.

Since the inception of density functional theory (DFT), practitioners of DFT have sought more accurate, reliable, and efficient density functional approximations (DFAs) in order to predict the physical and chemical properties of materials. Despite current DFAs’ remarkable success in predicting mechanical properties [1] and crystallographic structures [2], they still exhibit significant failures in the prediction of molecular bond dissociation [3–5], band gaps in solids [6-10], and polymorph energy differences in transition metal oxides [11–13]. Many of these failures can be attributed to the breaking of certain exact physical constraints, namely the tilted plane condition [14-16], resulting in piecewise linearity errors with respect to electron count N and magnetization M. Here we present a new DFT+U type functional named BLOR after the authors [17], which is specifically designed to correct for the local analogues of these piecewise linearity errors. The BLOR functional has been shown to significantly improve the total energies of dissociated homo-nuclear molecular systems compared to both bare DFT (PBE) as well as all other DFT+U-type functionals tested. 

Gravitational waves are ripples in spacetime in general relativity, and like light waves, they come in many different frequencies. One of these frequencies is in the nanohertz regime, whose origin is often attributed to Supermassive Blackholes. Pulsar timing arrays (PTA) are sensitive to these frequencies and specifically sensitive to the stochastic gravitational background, which is the sum of all the overlapping gravitational waves in the universe. 

Last summer, multiple PTAs worldwide came forward with significant evidence for the stochastic gravitational wave background, but none could claim detection. However, these PTAs are also all a part of the consortium of consortiums, the International Pulsar Timing Array (IPTA). The IPTA combined the individual PTAs' data to create a more sensitive dataset. However, combining PTA datasets takes a significant amount of time and effort. Therefore, we have developed a version of the optimal statistic that does not require a direct combination of the datasets. Instead, we combine the Bayesian posteriors from individual PTA analyses. This method will serve as an early indicator of what we will expect in future International PTA datasets. Here, we demonstrate a simple version of our early indicator optimal statistical analysis on the Parkes Pulsar Timing Array, European Pulsar Timing Array, and the North American Nanohertz Observatory for Gravitational Waves(NANOGrav).