Data in Brief

Dataset link


a b s t r a c t
Boiling is used for the thermal management of high-energydensity devices and systems.However, sudden thermal runaway at boiling crisis often results in catastrophic failures.Machine learning is a promising tool for in-situ monitoring of boiling-based systems for preemptive control of boiling crisis.A carefully acquired and well-labeled dataset is a primary requirement for utilizing any data-driven learning framework to extract valuable descriptors.Here, we present a comprehensive dataset of boiling acoustics presented in our recent work [1].We collect the audio files through meticulously controlled near-saturated pool boiling experiments under steady-state conditions.To this end, we connect a highsensitivity hydrophone to a pre-amplifier and a data acquisition unit for accurate and reliable acquisition of acoustic signals.We organize the audio files into four categories as per the respective boiling regimes: background or natural convection (BKG, 2 − 5 W / cm 2 ), nucleate boiling (NB, 8 − 140 W / cm 2 ), excluding those at higher heat flux values preceding the onset of boiling crisis or the critical heat flux (Pre-CHF, ≈ 145 W / cm 2 ), and transition boiling (TB, uncontrolled).Each audio file label provides explicit information about the heat flux value and the experimental conditions.This dataset, consisting of 2056 files for BKG, 13367 files for NB, 399 files for Pre-CHF, and 460 files for TB, serves as the foundation for training and evaluating a deep learning strategy to predict boiling regimes.The dataset also includes acoustic emission data from transient pool boiling experiments conducted with varying heating strategies, heater surface, and boiling fluid modifications, creating a valuable dataset for developing robust data-driven models to predict boiling regimes.We also provide the associated MATLAB® codes used to process and classify these audio files.
© 2023 The Author(s

Value of the Data
• The provided dataset [2] presents a unique collection of boiling sound, previously unavailable for scientific research purposes.We carefully recorded audio files during various pool boiling experiments conducted in a controlled and quiet environment.• The dataset encompasses acoustic emissions for a wide range of steady-state heat flux values, spanning from 2 W / cm 2 − 145 W / cm 2 , for water on a heated smooth copper surface.It showed the capability to train a convolutional neural network (CNN) and predict boiling regimes solely based on boiling acoustic emissions.• The dataset contains well-labeled acoustic signatures of natural convection, nucleate boiling, departure from nucleate boiling, and transition boiling regimes, and includes the MATLAB® codes to process the dataset.In addition to allowing independent validation/verification of our results, such a freely available dataset may also help fellow researchers evaluate other classification algorithms to improve upon boiling regime prediction.
• The dataset also contains acoustic emissions from various transient boiling experiments, such as with pure water, aqueous ionic liquid solution, aqueous surfactant solution, and on a nanostructured surface.Such diverse datasets may help fellow researchers evaluate the generalizability and robustness of typical machine learning (ML) algorithms.• One can leverage the boiling acoustics dataset to develop novel data-driven models to extract and quantify boiling descriptors such as bubble departure frequency, number density and size, surface heat flux, and heat transfer coefficient, among others.• The MATLAB codes can also be employed for real-time prediction of boiling regimes to avoid impending boiling crisis on unseen datasets.

Background
This dataset has a broad range of applications in heat transfer, industrial safety and efficiency.It is crucial for predicting boiling regimes, covering all boiling stages from nucleate to transition, in systems such as power plants including nuclear reactors, chemical processing, and food industry among others.This dataset enables the development of a regime identification tool, which can help prevent catastrophic failure, which is typically referred to as the boiling crisis.
Acoustic signatures of boiling sound are valuable for monitoring quenching applications as well.The sound during quenching closely resembles boiling sound, aiding in identifying high heat transfer phases [ 3 ].During quenching, if the acoustic profiles resemble those of high heat flux boiling sounds in the dataset, it indicates the suitability of the coolant fluid.Additionally, correlating with heat flux enables exclusive heat flux quantification through acoustic analysis.
Furthermore, this dataset is invaluable for detecting cavitation, a pervasive industrial challenge.Cavitation and boiling both involve the formation and collapse of vapor bubbles within the liquid.Hence, by analyzing unique acoustic signatures in boiling sound recordings, it may allow for early detection and monitoring of cavitation events in hydraulic systems, pumps, and propellers, averting severe damage.During critical operations, correlating acoustic patterns between boiling and cavitation sounds may ensure precise and timely identification, empowering industries to take proactive measures to mitigate detrimental effects of cavitation.

Data Description
The dataset includes two main folders: "Audio Files and MATLAB Code used for Advance Prediction of CHF", and "Transient Experiments Data."The first folder contains four subfolders: BKG, NB, Pre-CHF, and TB, each comprising raw audio files.These acronyms represent different boiling regimes: BKG stands for the background, indicating natural convection without bubbles; NB means nucleate boiling; excluding those at higher heat flux values preceding the onset of boiling crisis or the critical heat flux, subsequently termed as Pre-CHF; and TB represents the transition boiling regime.The audio files within these subfolders are one second each and are recorded at various heat flux values.Each filename includes the corresponding heat flux information.For example, a file in the NB subfolder labeled as "106Wcm2(100).wav,"indicates that it was acquired at a heat flux of 106 W / cm 2 .The numerical value within the parentheses, such as 100, denotes the specific audio file corresponding to the 100th second after the steady-state condition was attained and data acquisition was started.Additionally, this folder contains MATLAB codes for training a convolutional neural network (CNN), and a "readme" file is provided for further guidance.
The embedded CNN code, a fundamental component of our research methodology, functions as follows: Initially, the audio data files which have been thoughtfully pre-categorized into groups, corresponding to distinct boiling regimes, undergo a crucial transformation into spectrograms.These spectrograms serve as visual representations of the frequency components within the audio data at each time step.Each spectrogram is then meticulously labeled to correspond with its specific boiling regime category.Notably, these spectrograms are treated as image data during processing, thus enabling the neural network to leverage its deep learning capabilities effectively.
Following this transformation, the CNN code employs a suite of convolutional-based filter algorithms to systematically extract pertinent features from the spectrogram data.Subsequently, the code autonomously trains a sophisticated deep learning-based neural network, utilizing the distinctive features extracted from the spectrogram data of the training set.This trained neural network, the ultimate output of the CNN code, possesses the ability to proficiently identify and classify boiling regimes.It is essential to emphasize that the input in this process is represented by the spectrogram of the audio file, while the output is the classification of the specific boiling regime.
We have included an illustrative figure ( Fig. 1 ) which elucidates the intricate training and testing steps of the CNN code.This visual representation provides a clear and concise breakdown of the process, aiding readers in grasping the methodology with ease.The second folder, "Transient Experiments Data," contains five .matfiles.For illustrative purposes, let us consider one sample mat file named "TIME_ACOUSTICS_HEAT_FLUX_AND_ PERCENTAGE_CHF_DATA_FOR_CUF_AND_IONIC_LIQUID_SOLUTION_CASE.mat."This file consists of four columns: the first column represents the time in seconds, the second column contains temporal pressure data in Pascal for the acoustic emissions (AE), the third column provides heat flux in W / m 2 .The fourth column represents the heat flux value in terms of the percentage of the heat flux at boiling crisis/CHF.Furthermore, the .matfile includes information about the specific experiment, such as the test surface (e.g., CUF representing a flat copper surface) and the fluid used (e.g., an aqueous ionic liquid solution).The first four .matfiles in the second folder align with the content above.However, the fifth mat file requires further explanation, as it corresponds to a particular case.This file is generated from an experiment where the heating power is suddenly increased to a significantly higher value, causing the heated surface to traverse through all boiling regimes in a short timeframe.

Experimental Design, Materials and Methods
The pool boiling test rig includes the heater assembly, as depicted in Fig. 2 [ 4 ].The boiling chamber consists of a transparent cylindrical Borosilicate glass container with a capacity of 5 liters.The glass container is covered with a transparent acrylic lid with holes to maintain atmospheric pressure during experiments.A silicone rubber heater (Make: Marathon, 220 V, 500 W) is wrapped around the boiling chamber to achieve the desired subcooling.The heater is also connected to a proportional integral derivative (PID) controller (Make: Omega) and a thermocouple.During the boiling experiments, an ultrasensitive miniature hydrophone (Make: B&K, Type: Piezoelectric transducer, Model: 8103) is used to capture acoustic emissions.It is connected to a pre-amplifier (Make: B&K, Model: Nexus 2692) and a data acquisition unit (Make: B&K, Model: 3160-A-042).Next, a detailed discussion on the components and equipment used for acquiring the acoustic emissions is presented.
To capture the acoustic emissions during boiling, a small-sized hydrophone of the piezoelectric transducer type is utilized.Specifically, the hydrophone model is 8103, manufactured by B&K.This hydrophone is designed to be highly sensitive and enables accurate measurement of sound in the frequency range of 0 . 1 Hz to 180 kHz .It possesses a receiving sensitivity of −211 dB re 1 V / μPa , which allows for precise detection of even low-intensity acoustic signals.One advantage of this hydrophone is its omnidirectional nature, meaning it can capture acoustic signals from all directions.Therefore, the orientation of the hydrophone does not affect the quality or accuracy of the recorded acoustic data.Table 1 presents a summary of the operational conditions and additional specifications of the hydrophone.It is worth noting that the hydrophone is specifically designed for usage in underwater environments, as stated in the datasheet.A charge amplifier is an electronic device that functions as an integrator of electrical current.It generates a voltage output proportional to the accumulated charge of the input signal.This type of amplifier is particularly well-suited for piezoelectric sensors like hydrophones.One key feature of a charge amplifier is its ability to establish a virtual ground for the input signals.Essentially, if any charge accumulates on the plates of the sensor or the input parasitic capacitance of the amplifier, it will create a voltage across the input of the amplifier.The charge amplifier immediately compensates for this voltage and prevents any distortion by drawing or supplying an equivalent amount of charge current through the negative feedback network.For our experiments, we utilized the NEXUS 2692-C charge amplifier specifically designed for use with the hydrophone model 8103.The specifications and details of this charge amplifier can be found in Table 2 .
Acoustic signals of the boiling are stored using a data acquisition card (Model: 3160-A-042).A complete stand-alone analyzer test system is created by combining inputs and generator outputs.The module is excellent for system stimulation applications, such as audio signals.The frequency range of all input and output channels is 0 to 51.2 kHz.Specifications of the DAQ card are provided in Table 3 .
Steady-state experiment were performed under near-saturated conditions, with the pool temperature maintained at approximately 96 ± 1 • C .The power level was ramped up in a stepped manner to obtain a range of heat flux values, spanning from approximately from 2 W / cm 2 − 145 W / cm 2 .The selection of these heat flux values aimed to acquire AE data in various boiling regimes, including the natural convection regime (where two-phase heat transfer is absent),

Fig. 1 .
Fig. 1.Details of the CNN training and testing.
). Published by Elsevier Inc.This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/by-nc-nd/4.0/ ) Phase-change heat transfer, Boiling, Acoustic emissions, Deep learning, Convolutional neural network, Boiling regime prediction, Critical heat flux Data format Waveform audio files, AV, MAT files, Data files are raw Type of data Audio files (.wav), Audio files (.mat), Matlab codes (.mat) Data collection An ultrasensitive miniature hydrophone (Make: B&K, Type: Piezoelectric

Table 1
Summary of specifications of the hydrophone.

Table 2
Summary of specifications of the charge amplifier.