High Performance Computing For Microfluidics On Graphics Processing Units

Abstract

Computational science has recently emerged as a third essential scientific tool along with theoretical study and experiment. Despite its substantial success and increasing popularity, computational science offers a multi- tude of challenges. A crucial one is the problem of efficient hardware uti- lization for elaborate physical simulations. As the microprocessors are get- ting more specialized and computers – more complex and diverse, the gap between offered and attained performance widens, reaching orders of magnitude. Our first goal within this work is to close this gap. Our second goal is to explore unique features of microfluidic flows with possibly millions of suspended cells or objects, within complex geometries and for extended periods of time. Spanning different spatial and temporal scales, these flows not only raise many scientific questions, but also pose significant computational challenges. In this thesis we develop a feature-packed particle based code for meso- scale flows, reaching unprecedented time-to-solution and scaling. It allows to simulate bigger problems, on larger clusters, at faster rates than state- of-the art software, enabling previously impossible exploration of systems and phenomena. With the help of our software, we study inertial particle migration and report several previously unexplored features and configurations, that may guide future development of microfluidic devices for focusing and separat- ing living cells. Finally, we evaluate parametrization of our numerical mod- els in Bayesian manner, which can be used for more accurate and robust future simulations.