Learning new physical descriptors from reduced-order analysis of bubble dynamics in boiling heat transfer

Understanding bubble dynamics during boiling is challenging due to the drastic changes in system parameters, such as nucleation, bubble morphology, temperature, and pressure. In this study, principal component analysis (PCA), an unsupervised dimensionality reduction algorithm, is used to extract new physical descriptors of boiling heat transfer from pool boiling experimental images without labeling and training. The dominant frequency and amplitude of the time-series principal components (PCs) are analyzed, where the first few dominant PCs are used to approximate the instantaneous bubble morphologies, drastically reducing the data dimensions. The results show that the dominant frequency and amplitude can be used as new physical descriptors to distinguish different boiling regimes. The dominant frequency of the first PC is found to increase with heat flux in the discrete bubble regime until it reaches a peak and then decreases with heat flux in the bubble interference and coalescence regime, where the former is believed to be associated with the increase in bubble nucleation sites and the latter is associated with the bubble coalescence during pool boiling. The dominant frequency and amplitude extracted from the present unsupervised learning are qualitatively compared to the bubble count and size extracted from a supervised deep-learning algorithm, and the approach appears highly robust over multiple datasets and heater surfaces. To predict future boiling states for mitigating boiling crises, bidirectional long short-term memory (BiLSTM) neural network is used to estimate the future variations of PCs and hence the bubble dynamics, from time-series PCs. The PCA-BiLSTM models predict reduced-order bubble images well and show significantly higher prediction accuracy compared to the Convolutional-LSTM.