SequenceSync-GAN: sequence-aware domain translation for generalizable boiling crisis detection

Boiling crisis, or critical heat flux (CHF), is a key challenge in thermal engineering. It occurs when heat transfer from a heated surface to a boiling liquid collapses, causing a rapid surface temperature rise and potentially leading to system failure. Accurate, non-intrusive CHF detection using boiling images holds promise for improving heat exchanger design and operation. While machine learning models have been proposed for image-based CHF detection, their generalizability across different experimental setups (i.e., cross-domain scenarios) remains limited. Unsupervised image-to-image (UI2I) translation has attempted to address this, but often fails to maintain class consistency across domains. In this study, we present Sequence-Synchronized Generative Adversarial Network (SequenceSync-GAN), a UI2I framework designed to enhance cross-domain generalization in CHF detection by leveraging temporal sequence information. SequenceSync-GAN adapts boiling image datasets to a pre-trained CHF detection model through temporally consistent domain translation. Key contributions include: (1) a temporal data loading strategy, (2) a sequential annotation scheme, (3) a temporal discriminator architecture, and (4) loss functions that enforce sequential consistency during training. Tested on two boiling image datasets, SequenceSync-GAN demonstrated an 8 -15 % improvement gains over state-of-the-art UI2I baselines. This work serves as a proof-of-concept, demonstrating the potential of temporally-aware UI2I models for generalizable CHF detection and offering a foundation for future applications such as defect detection and time-dependent image classification in engineering systems. The source code for this work can be found on the project’s official repository.