Deep learning strategies for critical heat flux detection in pool boiling

Image-based deep learning (DL) models are employed to enable the detection of critical heat flux (CHF) based on pool boiling experimental images. Most machine learning approaches for pool boiling to date focus on a single dataset under a certain heater surface, working fluid, and operating conditions. For new datasets collected under different conditions, a significant effort in re-training the model or developing a new model is required under the assumption the new dataset has a sufficient amount of data. This research is to explore strategies of DL adapting to new datasets with limited data available. The insights gained could help improve the practicality and reliability of DL for boiling regime studies. Specifically, convolutional neural networks (CNN) and transfer learning (TL) are studied. Using a base model trained and tested for one public dataset (DS1), CNN and TL models are trained with a small portion of a new public dataset (DS2) and tested for the rest of DS2. Results show that TL outperforms CNN by having much higher accuracy and a much lower false negative rate for scarce data (less than5% DS2). When 1% DS2 is used for re-training in CNN versus fine-tuning in TL, the TL model can detect the CHF with an accuracy of 94.79% and a false negative rate of 0.0997, compared with the CNN model with an accuracy of 85.10% and a false negative rate of 0.3237. To further demonstrate the advantage of TL over CNN, an in-house dataset (DS3) is acquired. With less than 0.05% DS3 being used, the TL model can detect the CHF with an accuracy of 95.31% and a false negative rate of 0.0016, compared with the CNN model with an accuracy of 85.91% and a false negative rate of 0.1263. It is observed that TL has much higher robustness than CNN by having more consistent results and smaller standard deviations over multiple trials using stratified random sampling from both DS2 and DS3. Besides, the training time for TL is significantly lower than CNN when limited data used in the re-training and fine-tuning for both DS2 and DS3. These results demonstrate the ability of TL for handling data scarcity in pool boiling applications with potentials for real-time implementations.