2022 HPC Parallel Programming Workshop

Accurate, high-resolution modeling of multiphase fluid-fluid and fluid-solid flows is extremely computationally intensive. We will focus on the example of densely packed spherical particles suspended in a fluid in a planar channel. Neutrally buoyant, non-Brownian particles in a low Reynolds number pressure-driven flow display an irreversible net particle migration towards the center of the channel. Simulations of densely  packed particles present several numerical challenges, including accurate representation of the forces between particles and resolving the fluid flow between particles. This talk will cover how high performance computing techniques are used to make the computations of 10,000+ particles feasible using the Force Coupling Method. We will then discuss how the data generated can be used in physics-informed machine learning to gain additional insight of the system. By using the continuum-scale suspension balance model we can accurately learn additional information about the particles, such as the particle stresses, that we cannot measure directly from experiments.

Glass is tempered to improve strength, however, after fracture initiation it fragments into many small pieces. The number of fragments depend on the material, thickness, and the amount of stored residual energy from the tempering process. Historically, empirical studies have been performed to understand this phenomenon. This work will discuss the use of peridynamic theory to directly simulate the fragmentation process of various glass plate thicknesses and residual energies. The resulting trend suggest a parameter that characterizes fragmentation susceptibility and a fracture surface roughness and stored energy relationship. These can then be used to predict the number of fragments created upon failure after characterizing parameters with experimental data.

We present the construction and implementation of a recently proposed embedded-boundary/arbitrary-Lagrangian-Eulerian (or EBM-ALE) method for solving many-physics problems in a parallel computing environment. The EBM-ALE features stabilized finite element methods for both fluids and solids on unstructured grids, a projected level set method to capture the material interfaces between fluids, a ghost-type strategy to enforce fluid-fluid transmission conditions, and a partitioned coupling scheme between fluids and solids. Although the problem is complicated, we shall discuss several strategies in data localization and communication cost reduction while implementing the algorithm, with the help of the data infrastructure provided by the Trilinos scientific software libraries.