Physical descriptors guided image translation for generalizable critical heat flux detection

Accurate detection of the critical heat flux (CHF) in boiling heat transfer is vital for ensuring the safety and reliability of thermal systems. Image-based, non-intrusive CHF detection models have emerged as powerful tools for improving the monitoring and design of heat exchangers. However, their generalizability across experimental setups remains limited due to domain shifts in imaging conditions and physical configurations. To address this challenge, this study introduces a novel physical descriptors guided unsupervised image-to-image translation framework designed to enhance cross-domain CHF classification by preserving physical property consistency during domain translation. The proposed model extracts bubble-level physical characteristics from boiling images and employs three newly introduced domain-guided loss functions, Blob Count Loss, Blob Mean Area Loss, and Blob Standard Deviation Area Loss, to incentivize the generator network to maintain physical property consistency between input and translated images. Experiments show that the model outperforms existing cross-domain CHF detection methods, achieving up to 26.2% higher Balanced Accuracy and 25.1% higher AUC across domains. An ablation study further confirms that jointly enforcing these physical constraints yields the best overall performance. Beyond CHF detection, the model offers a generalizable framework that may also be applicable to other cross-domain image translation tasks involving blob-based morphological structures extracted through image segmentation, such as biomedical imaging of kidney glomeruli or cell morphologies. The source code for this work can be found on the project’s official repository repository ¹ 1 https://github.com/Hindawi91/BubbleSync-GAN . .