UC Irvine

Liquid injection systems and subsequent atomization behaviors are vital in many power generation and propulsion systems. These systems are inherently complex, owing to coupling of nonlinear processes in turbulent, multiphase flows. As a result, understanding and predicting the dynamical behaviors is inhibited through traditional system analysis which greatly impacts the ability to predict future states or control the nonlinear flow.

Koopman analysis has emerged as a data-driven approach in extracting dynamical features and physical understanding of nonlinear flows through a modal decomposition. This dissertation investigates such Koopman analysis techniques to improve the understanding of atomization systems, with an emphasis on generating compact and interpretable representations. Commonplace techniques of proper orthogonal decomposition (POD) and dynamic mode decomposition (DMD) are applied and evaluated on canonical atomization systems, highlighting their respective strengths and disadvantages as a means to perform a modal decomposition. Although POD and DMD have seen widespread use, they are inherently limited in capturing underlying dynamical processes of nonlinear data, which is demonstrated and verified on a hierarchy of system complexity.

To overcome the limitations of POD and DMD, a deep learning-based extension of Koopman analysis, in the form of a deep convolutional Koopman network (CKN), is proposed for extracting dynamical features of spatio-temporal atomization data. The CKN is end-to-end trainable with an architecture which can successfully be applied to a wide range of fluid flows. The CKN admits a more compact and, importantly, a more interpretable modal decomposition for improving the physical understanding of these systems. Highly accurate long-term future state predictions are achieved across multiple systems using an identical architecture. Indeed, these findings extend to spatio-temporal data in general which exhibit periodic, dynamical behaviors.