Experimental Bubble Shape Classification Dataset for Gas–Liquid Two-Phase Flow

This dataset contains experimental single-bubble images for bubble shape classification in gas–liquid two-phase flow. The bubbly-flow images were acquired using a Phantom T2410 high-speed camera with LED backlighting. Background artifacts were suppressed using Lambert–Beer-based image processing, and individual bubbles were extracted from the flow-field images through manual annotation and YOLO-based detection. The dataset contains 2,480 grayscale single-bubble images with a spatial size of 128 × 128 pixels. The images are classified into three morphological categories: 1,211 spherical bubbles, 774 ellipsoidal bubbles, and 495 other bubbles. Bubble morphology is characterized using aspect ratio, circularity, and solidity. For reproducibility of the RDMoF study, the dataset additionally provides the fixed train/validation/test split used in the experiments (1,984/248/248 samples). The original class labels are retained as the ground-truth labels. The 20% symmetric training-label-noise configuration used in the RDMoF experiments is provided separately and does not alter the original dataset labels. The dataset is intended to support research on bubble morphology analysis, gas–liquid two-phase flow, machine-learning-based image classification, and model generalization under noisy labels.