High Performance Finite Element Methods with Application to Simulation of Vertical Axis Wind Turbines and Diffusion MRI
Time: Wed 2019-12-04 10.15
Subject area: Applied and Computational Mathematics Biological Physics Computer Science Engineering Mechanics Mathematics Biological and Biomedical Physics
Doctoral student: Van Dang Nguyen , Beräkningsvetenskap och beräkningsteknik (CST)
Opponent: Professor Mats Larson, Umeå universitet, Umeå, Sverige
Supervisor: Professor Johan Hoffman, Numerisk analys, NA; Associate Professor Johan Jansson, Beräkningsvetenskap och beräkningsteknik (CST)
Finite element methods have been developed over decades, and together with the growth of computer power, they become more and more important in dealing with large-scale simulations in science and industry.The objective of this thesis is to develop high-performance finite element methods, with two concrete applications: computational fluid dynamics (CFD) with simulation of turbulent flow past a vertical axis wind turbine (VAWT), and computational diffusion magnetic resonance imaging (CDMRI). The thesis presents contributions in the form of both new numerical methods for high-performance computing frameworks and efficient, tested software, published open source as part of the FEniCS/FEniCS-HPC platform. More specifically, we have four main contributions through the thesis work.
First, we develop a DFS-ALE method which combines the Direct finite element simulation method (DFS) with the Arbitrary Lagrangian-Eulerian method (ALE) to solve the Navier-Stokes equations for a rotating turbine. This method is enhanced with dual-based a posteriori error control and automated mesh adaptation. Turbulent boundary layers are modeled by a slip boundary condition to avoid a full resolution which is impossible even with the most powerful computers available today. The method is validated against experimental data with a good agreement.
Second, we propose a partition of unity finite element method to tackle interface problems. In CFD, it allows for imposing slip velocity boundary conditions on conforming internal interfaces for a fluid-structure interaction model. In CDMRI, it helps to overcome the difficulties that the standard approaches have when imposing the microscopic heterogeneity of the biological tissues and allows for efficient solutions of the Bloch-Torrey equation in heterogeneous domains. The method facilitates a straightforward implementation on the FEniCS/ FEniCS-HPC platform. The method is validated against reference solutions, and the implementation shows a strong parallel scalability.
Third, we propose a finite element discretization on manifolds in order to efficiently simulate the diffusion MRI signal in domains that have a thin layer or a thin tube geometrical structure. The method helps to significantly reduce the required simulation time, computer memory, and difficulties associated with mesh generation, while maintaining the accuracy. Thus, it opens the possibility to simulate complicated structures at a low cost, for a better understanding of diffusion MRI in the brain.
Finally, we propose an efficient portable simulation framework that integrates recent advanced techniques in both mathematics and computer science to enable the users to perform simulations with the Cloud computing technology. The simulation framework consists of Python, IPython and C++ solvers working either on a web browser with Google Colaboratory notebooks or on the Google Cloud Platform with MPI parallelization.