Deep learning the sound of boiling for advance prediction of boiling crisis

Advance prediction of boiling crisis is critical to the safety and economy of many thermal systems. Here, we perform steady-state near-saturated boiling experiments on a plain copper surface and acquire the acoustic emissions (AEs) in natural convection, nucleate, and transition boiling regimes. We use the corresponding AE spectrograms to train a convolutional neural network, which shows a validation accuracy of 99.92% against the ground truth. We next evaluate the trained network on experiments with water and aqueous solutions of ionic liquid and surfactant on plain and nanostructured copper surfaces with time-varying heat inputs. Despite the variations in boiling surfaces, working fluids, and the heating strategy between the training and the evaluation datasets, the network accurately predicts the respective boiling regimes. Finally, we use the insights to perform advance prediction of boiling crisis for mitigating thermal runaway-induced accidents in boiling-based systems. [Display omitted] Spectrogram of bubble acoustics serves as the fingerprint of boiling regimesCNN-based deep learning algorithm uses spectrograms for boiling feature extractionCNN enables identification of various boiling regimes in diverse experimentsAdvance prediction of boiling crisis mitigates thermal runaway-induced failure The absence of prognostic tools constrains near-critical heat flux operations during boiling due to the threat of the looming temperature excursion. Sinha et al. report an acoustic emission-based deep learning framework, which enables advance prediction of boiling crisis to mitigate thermal runaway-induced failures in boiling-based systems.