UC Irvine

Efficient thermal system design is pivotal in cooling applications. The complex nature of boiling phenomena combined with the emergence of diverse flow patterns or regimes pose challenges. Despite advances in classification methods, machine learning (ML) models struggle to discern flow regimes from optical data obtained via high-speed cameras. This thesis presents a compelling case for learning from neuromorphic event representations of flow boiling, which offers insights into hidden properties of flow patterns. We delve into the application of diverse ML algorithms on event data for the task of flow regime classification. Through an analysis of convolutional neural networks, long short-term memory models, and event-based spiking neural networks, we highlight the superior accuracy and performance achieved by event data classification, as well as the unique insights provided by a novel Fourier-based classification approach. We evaluate the different ML algorithms on two event datasets — a video-to-event emulated dataset covering five flow boiling regimes, as well as a dataset combining the emulated data with a neuromorphic event camera dataset encompassing two regimes, for a total of seven flow boiling regimes. This research not only advances our understanding of flow boiling, but also showcases the potential of leveraging new data representations.