Student Posters Repository

We, the HPC community, are a special branch of computing where any technique is acceptable if it comes with more performance. But as a result, HPC often requires to take care of very low-level details such as spotting false sharing, while working on hybrid DM/SM software applied to large distributed heterogeneous architectures. The software developed requires time, effort and HPC expertise to be understood, maintained or improved, due to its great complexity. Paired with that fact that Exascale and its multi-million node scalability are in foreseeable future, writing HPC codes the way we do today will not be sustainable for long.
To that end, my PhD research investigates the use of behavioural types; similar to data types, but which express aspects of parallelism to be exploited by the compiler to infer additional parallelisation and HPC optimisation techniques to the code. For instance, one can think of the "pure elemental" keywords in High Performance Fortran, providing information about the behaviour of the function to which they are assigned. This additional information is later exploited by the compiler to apply automatic parallelisation over arrays.
From the above, it is easily understandable that this research is consubstantial with HPC.

The field of computational materials science relies heavily on electronic structure methods like Density Functional Theory (DFT) for material characterization and design. Although these methods have had great success, their algorithmic asymptotic scalings limit the system sizes that can be studied. Recently, deep learning has had great success applied to industrial problems like computer vision and speech recognition. In our work, we show how one can use a deep neural network to calculate total energy predictions that rival the accuracy of DFT but at a fraction of the computational cost. We then describe a new deep neural network topology we call extensive deep neural networks. These neural networks can be used to calculate extensive properties of large scale atomistic systems with O(N) scaling (N is the number of unit cells in the system). We describe the structure of these networks and their applicability to 2D hexagonal materials.

Inferring non-mediated cause-effect relations from time series data is a central problem in the study of large scale systems, with applications in climate science, seismology, neuroscience, statistical mechanics, biophysics and even economics. For this reason many data-driven methods have been proposed to identify the underlying causal structure of complex systems. Most of these methods are either based on heuristics or provide guarantees of a correct inference in restrictive domains of application. We develop a novel methodology to infer non-mediated cause-effect relations in large scale systems from observational data. Our method combines two prominent and complimentary techniques for detecting causality: the method of the PC-algorithm and the method of Granger Causality. Further, we provide theoretically proven bounds where the application of our methodology gives correct results. We also find that the domain of application of the methodology is extensive: the exact causal structure is correctly inferred in any large scale system for which every feedback loop contains at least one non-negligible delay.

In pressure atomizers a liquid is highly accelerated before it is discharged. The acceleration results in a local decrease of the static pressure even below vapor pressure which leads to partial evaporation (cavitation) of the fluid. Liquid-embedded vapor structures are subsequently advected into regions of higher pressure where they collapse and emit intense shock waves with post-shock pressures of more than 10,000 bar. These collapse events can cause severe damage of the injector which is called ‘cavitation induced erosion’. On the other hand, the collapse-induced turbulence can enhance the spray break up, which is a desired effect. In order to numerically investigate the effects of cavitation we employ a density based finite volume method that takes into account full compressibility of all phases to capture the shock wave after the collapses. For the numerical model an implicit large eddy approach is utilized, where the truncation error of the discretization scheme serves as a physically consistent subgrid-scale model for turbulence. Compressible Large Eddy Simulations of realistic geometries are challenging because of the required high spatial and temporal resolutions.

Numerical modelling of flood problems has become more and more complex. Consequently, the computational requirements, particularly for computing very large domains, have been increasing rapidly. To deal with these problems, parallel computing can be a choice. The topic of this research is “High Performance Computing of Coupling 2D and 3D Numerical Modelling of Flood Propagation and its High Performance Interface and Visualisation”. In this research, the 2D (non)-hydrostatic model will be coupled with the 3D non-hydrostatic model to simulate flood propagation problems within a framework of hybrid programming which combines the OpenMP and MPI parallelisation technique. For areas where wet-dry interfaces are dominant or vertical profiles are not of main interest, the 2D model is employed, whereas the 3D model is applied for areas where non-hydrostatic pressure condition applies, e.g. vertically circulating flows. Our code will be performed using the OpenMP for shared memory parallelisation inside of each node and communicated with other nodes using the MPI which is developed for distributed memory parallelisation. Using this hybrid model, some advantages, such as eliminating domain decomposition at node and lower memory latency and data movement within node, can be achieved.

Borrowing for college has been increasing and the number of graduates having difficulty repaying student loans remains high. Besides student loans, family transfers are also important sources of funding for college education. Evidence shows that parents not only pay for their kids' college expanses, but also provide potential support when youths have repayment problems. Though previous literature finds little effects of increasing borrowing limits during college on college enrollment, the high student loans default rate indicates that youths may be constrained after college and this may affect their educational choices, especially for those with little parental resources. This paper examines the effects of borrowing constraints and parental transfers, both during college and after college, on educational attainment and human capital investment. Parent's inter vivos transfers are modelled endogenously, along with youth's education, repayment, and labor market choices. The model can provide quantitative implications on the insurance role played by parents and the effects when additional credit or insurance such as income-contingent repayment plan is introduced. The model is estimated using National Longitudinal Survey of Youth 1997. The effects of counterfactual policy such as relaxing borrowing constraint, providing tuition subsidies, and adopting alternative repayment plans on college enrollment and human capital investment are examined.

The semi‐arid Salt and Verde River basins in Arizona are susceptible to Atmospheric River (AR)‐related flooding. To understand the precipitation‐related impacts of climate change on extreme ARs affecting Arizona, a Pseudo‐Global Warming (PGW) method was used. High‐resolution ‘control’ and ‘future’ simulations of five intense historical AR events that affected the Salt and Verde River basins in Central Arizona were carried out using the WRF regional climate model. The PGW approach for future simulations involved adding a temperature ‘delta’ at different vertical levels to the historical initial and lateral boundary conditions of the input data, while keeping constant relative humidity. The ‘deltas’ were calculated using projected changes towards end of the twenty‐first century from an ensemble of nine GCMs for the RCP8.5 scenario. Future simulations showed an overall increase in vertically integrated transport of vapor and upward moisture flux at cloud base over the region for all events. The changes in precipitation at both domain and basin level were highly spatially heterogeneous. Precipitation increased in all future simulations, but in general, this increase remained less than the increase in column‐integrated water vapor. It was found that in most cases, cloud ice content decreased while cloud water content increased, indicating the increased role of warm‐rain processes in producing precipitation in the future simulations. Freezing levels rose by more than 600 m, and this along with increased temperature and greater role of warm‐rain processes led to a decrease of more than 80% in the amount of frozen precipitation during the events.

Using CT and MRI guided computational fluid dynamics (CFD) to simulate blood flow in patient-specific geometries is a growing field that is generating results on the order of terabytes. To effectively use this data in clinical decision making, physicians must know what computed values from simulations are most important in determining the severity of a patient’s condition. Here, we study the risk of varying degrees of aortic coarctation (congenital narrowing of the aorta). One risk is atherosclerotic progression, which has been shown to correlate with low wall shear stress. We study how the severity of the coarctation, in terms of downstream wall shear stress (WSS), is affected by gross morphological characteristics, patient attributes, and other comorbidities. Hemodynamic simulations of the aortic coarctation geometries were performed using HARVEY, a massively parallel hemodynamics package developed by Randles et al. We show that by choosing and developing physiologically relevant features, such as the variance of WSS within the coarctation, we are able to perform unsupervised classification of our simulated patient data. Our tools allow us to probe which computed results are important in predicting WSS and other hemodynamic variables.