Generalizable physical descriptors of pool boiling heat transfer from unsupervised learning of images

Boiling processes are notoriously difficult to analyze visually due to the complex interactions between vapor bubbles and the surface. To aid in the quantitative analysis of these phenomena, this repository provides the high-speed videos, manually annotated pool boiling images, and MATLAB analysis toolkit associated with the study "Generalizable physical descriptors of pool boiling heat transfer from unsupervised learning of images" (International Journal of Heat and Mass Transfer, 255 (2026) 127894). The dataset comprises experiments conducted with different working fluids (water and HFE-7100) and heater surfaces (plain and microstructured copper and silicon) to investigate the effect on bubble morphology. Conventional physical descriptors, such as bubble size, bubble count, and vapor area fraction, as well as the descriptors derived from Principal Component Analysis (PCA), were extracted from the abovementioned dataset. The results demonstrate strong positive correlations between the PCA-derived descriptors and the conventional parameters, confirming that dominant amplitude correlates with bubble size and vapor area fraction, while dominant frequency correlates with bubble count. The dataset and accompanying tools therefore provide a basis for applying and validating an unsupervised learning approach that can act as a robust surrogate for traditional, time-consuming manual labeling techniques.

Methods
This repository provides a comprehensive dataset and a MATLAB-based analysis toolkit designed for quantitative analysis of pool boiling phenomena.

Dataset Information

The root directory contains separate folders for each experimental dataset (PCu-H2O, FCu-H2O, PSi-HFE, and SSi-HFE). Each of these folders contains the experimental data as downsized high-speed videos (.mp4) captured at various heat loads. For quantitative analysis, a manually annotated subset of randomly selected high-speed image frames (.jpg) is provided in an annotatedBubbles subfolder, along with a structured JSON (.json) file containing the corresponding bubble contour coordinates. This JSON data is organized as a structure where each image entry contains a FileName and a Bubbles object, which in turn holds the x and y coordinates for each individually labeled contour (e.g., b0001, b0002, etc.).

Image Analysis Toolkit

The DataProcessTools folder, also located under the root directory, contains the MATLAB analysis toolkit, which is divided into two main parts:

Conventional Descriptors: Scripts are provided for manual annotation of bubble contours (bubbles_annotator_v6.m) and for the statistical calculation of traditional physical parameters (bubbleStat_calculator_v4.m). These parameters include bubble count (Nb​), average bubble area (Ab​​), average bubble radius (Rb​​), and vapor area fraction (VAF).
Generalizable Descriptors: This is a two-part workflow for unsupervised feature extraction. The first script, PC_calculator_v4.m, processes image sequences to perform dimensionality reduction and calculate the time series of the principal components (PCs). The second script, dominantDescriptor_calculator_v4.m, then applies a Fast Fourier Transform (FFT) to the time series of the principal component to determine the Dominant Frequency (Df​) and Dominant Amplitude (Da​) of the boiling process.