Addressing the Domain Shift in Computer Vision for Boiling Crisis Detection

Boiling crisis detection, also known as critical heat flux (CHF), represents a formidable challenge in thermal engineering and is of paramount importance to various industrial processes, particularly in nuclear reactors and power generation systems. This phenomenon occurs when the heat transfer rate from a heated surface to a boiling liquid abruptly deteriorates, leading to a sudden increase in surface temperatures and potentially catastrophic consequences. Accurate detection of the boiling crisis enables the safe and efficient operation of technological applications involving boiling heat transfer, thereby preventing disasters such as reactor meltdowns and equipment damage, making it a vital and urgent research frontier. This dissertation addresses the critical challenge of detecting CHF from boiling images of out-of-distribution domains in heat transfer applications. Leveraging machine learning and computer vision, the proposed framework utilizes state-of-the-art unsupervised image-to-image translation models, incorporating temporal factors and physical properties to enhance cross-domain CHF detection. Contributions include a novel cross-domain CHF detection framework, a new evaluation metric (DIPS), and advancements in unsupervised image-to-image translation models (SequenceSync-GAN) and (BubbleSync-GAN). The models showcase promising results and potential real-world applications. Future work aims to enhance the translation model by integrating multimodal learning using sensory data to capture more comprehensive information for improved cross-domain classification. In summary, this dissertation contributes innovative solutions to the domain shift problem in boiling heat transfer, advancing unsupervised CHF detection through machine learning and computer vision.