Student Posters Repository

The Discontinuous Galerkin Spectral Element Method (DGSEM) is a combination of Finite Element Method (FEM) and Finite Volume method. The current implementation is a high-order (spectral) method. The code achieves adaptivity by estimating the error of each element and decide whether to use the solution refinement or grid-coarsen. The code is implemented in a parallel format using both OpenMP and OpenMPI. The adaptivity would cause load imbalance during the computation procedure. The research involves implements a multilevel domain partition tool to evenly spread the elements among the processors and applies load balancing algorithms to speed up the make-span.  

Classical learning theory states that when the number of parameters of the model is too large compared to the data, the model will overfit and the generalization performance deteriorates. However, it has been empirically shown that deep neural networks (DNN) can achieve high generalization capability by training with an extremely large amount of data and model parameters, which exceeds the predictions of classical learning theory. 

One drawback of this is that training of DNN requires enormous calculation time. Therefore, it is necessary to reduce the training time through large scale parallelization. 

Straightforward data-parallelization of DNN degrades convergence and generalization. In the present work, we investigate the possibility of using second-order methods to solve this generalization gap in large-batch training. 

This is motivated by our observation that each mini-batch becomes more statistically stable, and thus the effect of considering the curvature plays a more important role in large-batch training. 

The aim of the project is the development of a numerical framework to support the design of a hypersonic UAV. A Variational Multiscale framework was widely and successfully used for the investigation of incompressible flows and it was validated with challenging cases in many engineering fields (marine, energy and aerospace). The ultimate goal is to extend the framework for the study of compressible flows. This will allow us to predict the fluid dynamic behaviour around complex geometries for a very wide range of Mach and Reynolds numbers. An MPI parallelized numerical framework for the solution of high-speed compressible flows is presented. The stabilized finite element discretization method is adopted to solve the Navier-Stokes equations for compressible flows using the pressured-based primitive variables. The Streamline Upwind Petrov-Galerkin (SUPG) stabilization method is used with a discontinuity capturing operator, which provides additional stability near the discontinuities, such as shock waves.

The formulation described above is applied to several benchmark cases. In this work we present some preliminary results obtained from 1D and 2D simulations of high-speed flows, with the aim of illustrating how well the framework handle the discontinuities (shock waves in this case)

We introduced a pipeline for screening functional peptides i.e antimicrobial peptide by coupling artificial intelligence with molecular modeling and simulation methods. We massively generated the peptide sequences and checked the activity via fast but effective simulations. The results suggest good sequence candidates for experimentalists to synthesize and evaluate.

Electromagnetic turbulent flows determine the thermal insulation of magnetically confined plasmas. Using nuclear fusion as an energy source is possible on earth, but the efficiency and profitability are limited by the thermal insulation of these plasmas – the better the insulation, the smaller and cheaper the device can be. Reliable extrapolation, like in a wind tunnel, from small experiments (0.2-1 m radius) to large reactors (>7 m) is not yet possible – making computer models crucial.

Turbulence emerges from non-linearity: the flow carries itself. Plasma flows are carried by electromagnetic fields and contain very different space and time scales, making the complete physical model computationally intractable - hierarchical sets of reduced models with decoupled scales are used instead. This requires a modular and flexible computational framework.

Instabilities on scales of ~1 mm must be resolved as well as flows on reactor scale - efficient use of HPC resources is therefore mandatory. Our code is specifically designed to consider complex anisotropic magnetic geometries of real reactors and therefore utilizes a special MPI + OpenMP parallelization scheme. For realistic predictive simulations the performance must still increase, though, by farther domain decomposition and utilization of GPUs.

Resilience is one of the major topics in modern high-performance computing (HPC) research. With millions of processors, the “mean-times between failure” become a relevant issue for simulation scientist around the world. Incorporating countermeasures on the hardware side is expensive, difficult, time-consuming or all at once. Thus, much attention has therefore been paid to “algorithm-based fault tolerance” strategies which exploit specific features of numerical methods to continue working even after a processor crashes or a bit flips. This research aims at using novel methods from the field of parallel-in-time integration techniques for detecting and correcting these faults. ”Parallel-across-the-steps” methods like Parareal or PFASST share features that make them natural candidates for algorithmic-based fault tolerance: they hold copies of the (approximate) solution at different times on different processes and they are iterative as well as hierarchical. First proofs-of-concept show that these properties allow to derive recovery strategies to continue integrating forward in time even when nodes fail or bits flip.

In the 21st century, people are witnessing a rapid rise of urbanization throughout the world. Cities that are becoming densly populated units with complex functions, and it is a significant challenge for cities to prepare against disasters effectively. Safeguarding urban resilience after natural disasters rely on the accurate understanding of the system and the interactions between multiple critical infrastructures. The objective of this research is to model the interactions between the traffic and pipeline networks in the San Francisco Bay Area under earthquake hazards. The population of the study area is over 7 million and simulating their individual behavior needs a powerful computing system. We conducted traffic simulations on an undamaged road network and compared it against the network when the San Francisco-Oakland Bay Bridge is closed. The results indicate the effect of the bridge closure on agents’ route choice and their travel time. In addition, we conduct traffic simulations assuming pipe damage and flooding as a direct result of ground movements to understand the performance and interactions of the traffic and pipeline networks under earthquake scenarios.

Turbulent multiphase flows are common in a wide number of industrial applications and nature;  dispersions or emulsions of two immiscible phases are frequently found, for example, in the food and chemical processing industry, the extraction and transport of crude oil and oil spills in seas and oceans. In many cases an additional element is often present: surfactants are present as impurities or are added to enhance the process efficiency. Surfactants, also known as surface active agents, are amphiphilic molecules, characterised by polar heads and non-polar tails, which accumulate over the interface between the two phases and locally reduce surface tension of this interface. 

These molecules, even at extremely low concentrations, have a strong effect on the dynamics of the entire multiphase flow: the lower surface tension at the interface enhances breakage and deformation of the interface, thus generating more and smaller droplets. This change in the morphology of the dispersed phase strongly affects the overall efficiency of the process: smaller droplets are more dispersed and the total interfacial area is larger, thus favouring mixing and transfer of mass, energy and species through the interface. Numerical simulations are used to shed light on the physics underlying these extremely complex phenomena.