A Novel Framework for Model Reduction and Data-Driven Modeling of Fluid-Structure System: Application to Flapping Dynamics

点击翻译按钮获取中文摘要

Advances in high-performance computing (HPC) have empowered us to perform large-scale simulations for billions of variables in complex coupled multifield, multidomain and multiphase systems. While the multifield represents several interacting physical fields (e.g., fluid, solid, acoustic), the multidomain implies the solution of these fields over separate geometric domains. Over the past six years in my research group, the high-fidelity simulations using the first-principle physical laws (i.e., continuum equations) have been providing invaluable insight for the development of new design and devices in aerospace, offshore and marine engineering. Despite efficient algorithms and powerful supercomputers, the multifield (e.g., fluid-structure interaction) simulations are somewhat inefficient hence less attractive with regard to design optimization, parameter space exploration and the development of control and monitoring strategies for engineering systems. On the other hand, current state-of-the-art methods for parametric investigation and control of flow dynamics and fluid-structure interactions are primarily based on semi-empirical methods and nonlinear effects such as vortex shedding, turbulent wake dynamics and wake interference, large deformation due to fluid-structure coupling are typically discarded. The proposed research program will focus on addressing fundamental and applied challenges during the integration of our in-house HPC-based high-fidelity solver with the emerging field of data science and machine learning while promoting interdisciplinary research and education in UBC. I believe that the new framework based on the physical-model and data-driven computing will revolutionize engineering predictions and design of next-generation systems. For example, in our recent studies for the unsteady flow dynamical predictions of canonical bodies, we have achieved over 4-5 orders of magnitudes improvements in the performance gain for the prediction of unsteady forces via data-driven modeling using deep learning. Such improvements on academic problems are very promising for the pressing needs for optimization of large-scale dynamical systems. We will employ our novel framework for modeling of flapping foil dynamics for the extraction of marine hydrokinetic (MHK) energy in ocean currents. Using our high-fidelity solver, inverted flexible foils immersed in fluid flow are recently found to exhibit large-amplitude flapping, which can be converted to electricity using piezoelectric devices. We aim to explore our multifidelity framework for a broad range of physical parameters and configurations of MHK devices. There are numerous challenges with regard to transient chaotic and/or multi-scale phenomenon and offline-online decompositions of the nonlinear dynamics. Finally, the research program will provide efficient tools, physical insight, and practical guidance and will foster a training environment for graduate students and postdoctoral fellows.