Date of Award
Fall 2021
Document Type
Open Access Dissertation
Department
Mechanical Engineering
First Advisor
Yi Wang
Abstract
The microfluidic concentration gradient generator (μCGG) is an important device to generate and maintain concentration gradients (CGs) of biomolecules for understanding and controlling biological processes. However, determining the optimal operating parameters of μCGG is still a significant challenge, especially for complex CGs in cascaded networks. To tackle such a challenge, this study presents multi-fidelity surrogate and reduced-order model-based optimization methodologies for accurate and computationally efficient design of μCGGs.
The surrogate-based optimization (SBO) method is first proposed for the design optimization of μCGGs based on an efficient physics-based component model (PBCM). Various combinations of regression and correlation functions in Kriging and different adaptive sampling (infill) techniques are examined to establish the design process with refined model structures. In order to combine the simulation data from different sources with varying fidelities and computational costs for improved design efficiency and accuracy, a novel multi-fidelity surrogate-based optimization (MFSBO) method is presented. For the first time, a new computation-aware adaptive sampling strategy based on expected improvement reduction (EIR) is proposed to accelerate the convergence of MFSBO. EIR-based infill determines the data source and infill location by hypothetically interrogating the effect of samples and simulation fidelities on the reduction of the expected improvement. It also enables low-fidelity batch infills within a dynamically varying trust region to improve exploration on the fly. Subsequently, a new data sparsification technique based on the reduced design space and data filtering (RDS&DF) is investigated to eliminate redundant data and reduce the modeling time for improved optimization efficiency, hence addressing the long-standing “big data” issues associated with MFSBO. RDS&DF is also combined with EIR-based infill technique, enabling both parsimony and computational awareness for MFSBO. Finally, a multi-fidelity reduced-order modeling (MFROM) method is developed to enable model reusability and completely replace the CFD simulation when different μCGGs need to be designed. The key innovation of MFROM is using the proper orthogonal decomposition to obtain the low-dimensional representation of the high-fidelity CFD data and the low-fidelity PBCM data and a kriging model to bridge the fidelity gap between them in the modal subspace, yielding compact MFROM applicable within the broad trade space. As a result, MFROM is highly compatible with GPU-enabled optimization by utilizing its massively parallelized computing threads. The excellent agreement between the designed CGs and the prescribed CGs demonstrates the unprecedented accuracy and efficiency of the proposed multi-fidelity modeling and optimization methodologies.
In conclusion, given their non-intrusive, data-driven natures, both (MF)SBO and MFROM are versatile and can serve as a new paradigm for μCGG design.
Rights
© 2021, Haizhou Yang
Recommended Citation
Yang, H.(2021). Multi-Fidelity Surrogate and Reduced-Order Model-Based Microfluidic Concentration Gradient Generator Design. (Doctoral dissertation). Retrieved from https://scholarcommons.sc.edu/etd/6867