Search

We develop a neural network model capable of predicting the margin to the boiling crisis (i.e., the departure from nucleate boiling ratio, DNBR) from high-resolution infrared measurements of the bubble dynamics on surfaces with different morphologies and wettability (or wickability). We use a feature ranking algorithm, i.e., minimum redundancy maximum relevance, to elucidate the importance of fundamental boiling parameters, i.e., nucleation site density, bubble departure frequency, growth time, and footprint radius, in predicting the boiling crisis. We conclude that these parameters are all necessary and equally important. This result has profound implications, as it undermines the general validity of many observations and mechanistic models that attempt to predict the critical heat flux (CHF) by describing how a single boiling parameter changes with the heat flux or from one surface to another. Notably, the neural network model can predict the DNBR on CHF-enhancing surfaces of different wickability without using any input information related to the surface properties. This result suggests that, at least on the considered surfaces, surface wickability enhances the CHF by modifying the bubble dynamics, i.e., the aforesaid boiling parameters, rather than acting as an additional heat removal mechanism. [ABSTRACT FROM AUTHOR]

• We develop a machine learning methodology for online detection of dry areas. • It uses IR thermometry and U-net based CNN. • It is as accurate as manual processing but much faster. • It can be used online and quasi-real-time. • Its capabilities are independent of the operating pressure and surface morphology. Capturing the dynamics of dry area patches on a boiling surface is crucial to understanding and evaluating mechanistic heat flux partitioning and boiling crisis models. We present and evaluate a technique leveraging convolutional neural networks (CNN) to identify and track dry area patches online (i.e., as we run the experiments) and in quasi-real-time (i.e., as fast as we collect data) from high-resolution infrared (IR) thermometry measurements. First, the time-dependent IR radiation distribution emitted by the boiling surface is collected using an IR camera. Then, we leverage the segmentation capability of UNET-based CNNs to isolate the dry patches. Precisely, we train the UNETs to recognize dry patches with ground-truth dry area maps obtained through a high-resolution optical phase detection diagnostic and demonstrate that UNETs can map dry areas with at least 90 % accuracy for the heating surfaces we tested at pressures up to 2 bars. The results open the potential for automated detection and tracking of dry areas for IR heaters nontransparent to visible light, paving the way for autonomous boiling heat transfer experimentation. [ABSTRACT FROM AUTHOR]