Student Posters Repository

Modeling of collision-coalescence is one of the main differences between various Lagrangian particle-based cloud microphysics models. Most of these models use the all-or-nothing (AON) algorithm of the super-droplet method. This algorithm gives the correct mean number of collisions, but too large variance in the number of collisions. Variance decreases with the number of super-droplets (SDs) used. It is not well understood how the increased variance affects precipitation. The goal of our study is to understand convergence of AON, with respect to the number of SDs and to the time step, in cloud simulations.

We perform box simulations of collision-coalescence and 2D and 3D simulations of a cumulus congestus (CC) cloud. Box simulations show that mean droplet size distribution (DSD) converges for a 0.1s time step and 1000 SDs. Variance of the DSD is not sensitive to the time step and is inversely proportional to the number of SDs. Simulations of CC are done dynamically, i.e., with a resolved flow field, and kinematically, i.e., for a predefined flow-field. Mean precipitation in CC varies with the number of SDs per cell in a non-trivial way. It does not converge even for 100000 per cell. This suggests that the increased variance in AON may affect mean precipitation. Variance in the amount of rain in CC decreases with the number of SDs, but only in kinematic simulations. In dynamic simulations the variance in rain is more strongly affected by differences in realization of the flow field than by differences in realization of the AON algorithm.

I present a study of the potential for Convolutional Neural Networks (CNNs) to enable separation of astrophysical transients from image artifacts, a task known as “real-bogus” classification without requiring a template subtracted (or difference) image which requires a computationally expensive process to generate, involving image matching on small spatial scales in large volumes of data. We compare the efficiency of two CNNs with similar architectures, one trained with the templates, search, and difference images and one that takes as input the template and search only. There is a decrease in efficiency associated with the loss of information in input, finding that the testing accuracy is reduced from ∼ 96% to ∼ 91.1%. Improvement of the performance would require more data and training time. Present and future research is focused on understand how the latter model learns the required information from the template and search by finding patterns in the multidimensional latent space with the implementation of visualization techniques and unsupervised learning. Improvement of CNNs performance and efficiency of unsupervised models would require more data and training time.

Core-collapse supernovae (CCSNe) are explosions of massive stars which have run out of nuclear fuel at their core. Modern observations suggest a high degree of asymmetry in the CCSNe ejecta, resulting from aspherical explosions. These asymmetries may depend upon the detailed stellar profile, the hydrodynamical instabilities that arise at the composition interfaces of the stars, and the initial distribution of energy and ejecta introduced by the neutrino-heated, convectively driven central engine. To make a connection between the observed asymmetries and the stellar properties, long-time simulations of the CCSNe, a few seconds after the collapse until the free expansion of the ejecta in the circumstellar medium are necessary.  We use hydrodynamic code to perform these simulations, understand the origin of the ejecta asymmetry, and to constrain the stellar properties of the observed remnants.

The Centre for High Performance Computing (CHPC) supports and facilitates the use of their compute resources for researchers in southern Africa, however there is a lack of core skills and knowledge of HPC among researchers. Additionally, limited collaboration and resource constraints hinder the development of HPC infrastructure. Language barriers further exacerbate the situation due to the presence of 11 official languages in South Africa. To address these challenges, the research proposes a set of training events. The Coding Summer School and HPC Winter School offer comprehensive training in essential skills. The HPC Eco-Systems Project focuses on OpenHPC Training, aiming to enhance the ecosystem of HPC skills and resources. The research also emphasizes the importance of establishing a hybrid end-to-end training infrastructure and promoting collaboration among Southern African research institutes. Furthermore, it highlights the need to address language challenges and seize opportunities for growth in African languages.

The 13th Coding Summer School, held in February this year, serves as an example of the proposed training events. The curriculum covers Bash, Python and scientific computing training for HPC. The collaboration of research institutes play a vital role in ensuring the success of the summer school. The event offers both in-person and online learning options, enabling continued engagement and participation. Moreover, efforts are made to overcome language barriers through African language translation. Adequate infrastructure support, including internet access, laptops, power, and suitable venues, is provided to facilitate the learning experience.

As evidence of the initiative's impact, the Coding Summer School 2023 attracted 535 active students, and 350 students successfully obtained certifications. These numbers reflect the positive response and the effectiveness of the training program. By addressing the challenges associated with training HPC skills in resource-constrained environments, this research contributes to enhancing the capacity of African researchers and fostering the development of HPC expertise in the region.

Hubbard model helps us understand the interactions between particles on a lattice. But this model cannot be solved analytically beyond one dimension. Hence, various numerical methods have been developed for simulating a lattice system. One such method is the Hamiltonian Monte Carlo (HMC) which is suitable even for higher dimensions. To implement HMC, fermionic action needs to be discretized, one has the freedom to formulate lattice action in many ways as long as the proper continuum limits are recovered. These differences in the discretizations lead to the differences in structure of the fermion matrix. Implementation of the source code for fermion matrix, testing the code and lastly, benchmarking were done on CPU as well as GPU as a part of my research.

The code discussed in this thesis was written for the Nanosystem Simulation Library (NSL). The library is in development phase and is being developed to carry out Lattice simulations on nanoscales for structures such as graphene, carbon nanotubes, etc. 

We present a machine learning approach for predicting the light curves of kilonovae, important transients associated with the merger of compact objects such as neutron stars and black holes. Our software pipeline uses Gaussian processes and principal component analysis (PCA) to predict new light curves at arbitrary points in the parameter space, and then evaluates Markov chain Monte Carlo (MCMC) chains to estimate the properties of a kilonova given a set of observed light curves. We train our model on radiative transfer simulations and demonstrate its performance on a sample simulation, showing good agreement between the predicted and ground truth light curves. Our approach has the potential to aid in the detection and follow-up observations of kilonovae, and can also have implications for gravitational wave observations and the study of the universe's expansion.

Cosmological hydrodynamical simulations are crucial to understanding the formation and evolution of galaxies. They incorporate models for the main physical processes that give rise to the formation of these structures. Namely, the gravitational interaction between dark matter and baryons, the hydrodynamics of the gaseous component, the formation of stars and black holes and how they feed back energy into the gas. Thus, these simulations are very expensive and require an extensive dynamic range: the large scale of the Universe should be included whilst resolving individual galaxies.
I have extended PKDGRAV3, a massively parallel GPU-accelerated code previously used for gravity-only simulations, to handle hydrodynamics and galaxy formation. This extension has been designed and implemented modularly, focusing on continuous testing and performance.

During the last year, I have been helping other researchers to improve the efficiency of their codes, typically by parallelizing the most demanding computation and providing support.

The most elementary electrochemical process is the transfer of an electron from an electrode to a redox species in the electrolyte.  As defined by Marcus Theory, the reorganization energy quantifies the solvent-dependence of an electron transfer. As new electrolytes are being employed and as applications of electrochemical systems are being further explored, the solvent effects on the rate of the electron transfer need to be revisited.  Recent experiments have found that the reorganization energy is decreased near the electrode, and computational studies have suggested that the image charge of the electrodes may be responsible. The effects of the image charge and the electronic solvent polarization need to be quantified. We employ polarizable force fields and fixed voltage molecular dynamics simulations to study the consequences of instantaneous polarization on the reorganization energy. We explore the electrode-distance dependence of the reorganization energy through the redox reaction of a ferrocene molecule in an acetonitrile electrolyte with graphene electrodes. We find that instantaneous electronic polarization causes the reorganization energy of the oxidation of ferrocene to decrease to ~10 kJ/mol near the electrode, whereas a bulk system has a reorganization energy near ~100 kJ/mol. From these findings, we conclude that the image charge and electronic polarization of the solvent molecules greatly reduce the barrier to electron transfer. To further explore solvent effects, we will look at different electrolytes and apply a voltage to induce an electrochemical double layer—which modulates the solvent environment near the electrode.

The field of modern aviation has been intrigued by the potential of sustained hypersonic flight. However, several aerodynamic challenges impede the efficient operation of hypersonic vehicles, including the laminar to turbulent boundary layer transition. This transition increases heat transfer and aerodynamic drag, which are significant drawbacks of hypersonic boundary-layer flow. There are several approaches to stabilize the boundary layer and delay the transition. However, each approach comes with a disadvantage. To investigate the effects of these methods, we have developed a high-order CPU parallel compressible Navier-Stokes solver, OK-DNS, that utilizes a fifth-order weighted essentially non-oscillatory (WENO) scheme and third-order Total Variation Diminishing Runge-Kutta (TVD-RK) method. The parallelization in the code is carried out with MPI implementation. Although solving Navier-Stokes equations with high accuracy provides high-fidelity solutions, it is an expensive process. To that end, we developed a high-order linearized compressible Navier-Stokes solver, OK-LST, to solve the eigenvalue problem obtained from the linear stability theory (LST) that focuses on the linear stage of the transition process. The OK-LST solver is a hybrid GPU-CPU parallel solver that utilizes a fourth-order compact difference scheme for local eigenvalue problems and a second-order finite difference method for global eigenvalue problems. Compared to OK-DNS, OK-LST is significantly fast. Thus, it is a great solver for the preliminary studies of novel transition control ideas. Both solvers are extensively validated and utilized in investigating control methodologies for hypersonic boundary layer transition.

Sodium montmorillonite (Na-MMT) is one of the most commonly found swelling clay minerals that have diverse technological and engineering applications. The nanomechanical properties of this mineral have been extensively explored computationally utilizing molecular dynamics (MD) simulations to depict the molecular-level changes at different environmental conditions. As the environmentally found Na-MMT clays are generally sized within hundreds of nanometers, all-atomistic MD simulations of clays within this size range are challenging due to computational inefficiency. Atomistically informed modeling, a coarse-grained (CG) modeling technique can be employed to overcome the spatiotemporal limitation. The current study introduces a CG modeling strategy to develop a computationally efficient model of Na-MMT clay with a typical size over ~100 nm by shrinking the atomistic platelet thickness and reducing the number of center-layer atoms. Utilizing the “strain-energy conservation” approach, the developed CG model can well preserve in-plane tension, shear, and bending behaviors of atomistic counterparts. Remarkably, the CG tactoid model of Na-MMT, a hierarchical multilayer structure, can recreate the interlayer shear and adhesion, as well as d-spacing among the clay layers as of atomistic one to a good approximation, while gaining significantly improved computational speed (i.e., several thousand times faster than the atomistic one). Our study establishes the efficacy of the CG modeling framework, paving the way for bottom-up multiscale prediction of mechanical behaviors of clay minerals.